CN110222728B - Training method and system of article identification model and article identification method and equipment - Google Patents
Training method and system of article identification model and article identification method and equipment Download PDFInfo
- Publication number
- CN110222728B CN110222728B CN201910404189.3A CN201910404189A CN110222728B CN 110222728 B CN110222728 B CN 110222728B CN 201910404189 A CN201910404189 A CN 201910404189A CN 110222728 B CN110222728 B CN 110222728B
- Authority
- CN
- China
- Prior art keywords
- image
- article
- false
- true
- authentication
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/018—Certifying business or products
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Business, Economics & Management (AREA)
- Life Sciences & Earth Sciences (AREA)
- Entrepreneurship & Innovation (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- Image Analysis (AREA)
Abstract
The application discloses a training method and a system of an article identification model, and an article identification method and equipment, wherein the method comprises the following steps: acquiring a plurality of training sample images of an article to be identified, wherein the training sample images comprise a true article image and a false article image; fake image is obtained by using a confrontation generation network to fake the true object image; and using the countermeasure generating network to make the false article image true to obtain a true image; inputting the fake image into a confusion discrimination network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the false result of the false image. According to the training method of the article identification model, the confrontation generation network and the confusion discrimination network are combined, and meanwhile, the mode that the confrontation generation network and the confusion discrimination network are alternately and iteratively trained is adopted, so that the confrontation generation network and the confusion discrimination network can be trained mutually and jointly learned, and the identification capability and the robustness of the article identification model are improved.
Description
Technical Field
The present application relates to the field of image recognition, and in particular, to a training method and system for an article identification model, and an article identification method and apparatus.
Background
The deep learning technology is a technology which enables a machine to process and understand information such as images and voice through an artificial neural network. Deep learning is a popular field of academic research, and has been successfully applied in various fields such as face learning and natural language processing. Deep learning can accurately locate and learn to output different kinds of articles and classify according to their subtleties. In the aspect of image generation, the confrontation generation network can synthesize images of different articles which are hard to distinguish by human eyes according to requirements.
Luxury leather labels are important authentication points for authenticating the genuineness of luxury goods. The font and the engraving process of the leather label are difficult to completely imitate when a fake product is made. At present, the way of authenticating luxury goods at home and abroad is mainly manual authentication, and a plurality of problems exist. Firstly, at present, no unified industrial standard and industrial standard exist in China, professional qualification authentication is lacked, and the identification capability and quality of a manual identification worker are difficult to guarantee; secondly, the manual identification is high in cost and time-consuming, and the high cost and time overhead directly limit the application range of luxury goods identification and directly limit the convenient and efficient identification service which can be obtained by the general public in daily shopping; thirdly, the fashion and process of luxury genuine goods are constantly changing, the way of imitation of counterfeit goods is also constantly changing, and luxury goods are of various styles, and the authentication accuracy of manual authenticators is difficult to guarantee.
Therefore, how to automatically and accurately identify the truth of luxury leather labels based on the artificial intelligence technology of deep learning is a problem to be solved by practitioners in the field.
Disclosure of Invention
In view of the above-described drawbacks of the related art, an object of the present application is to provide a training method, a system, and an item authentication method, apparatus, client, and computer storage medium for an item authentication model.
To achieve the above and other related objects, a first aspect of the present application provides a method for training an item authentication model, comprising the steps of: acquiring a plurality of training sample images of an article to be identified, wherein the training sample images comprise a true article image and a false article image; utilizing a countermeasure generating network to counterfeit the true article image to obtain a counterfeit image; and using the countermeasure generating network to simulate the fake article image to obtain a simulated image; inputting the fake image into a confusion discrimination network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
In certain embodiments of the first aspect, the method further comprises the step of locating the identified point locations in the images of the genuine article and the images of the counterfeit article to extract the characteristic information.
In certain embodiments of the first aspect, the step of locating the positions of the identified points in the images of the real and false objects to extract the characteristic information comprises locating the positions of the identified points in the images of the real and false objects using a fast recursive convolutional neural network and extracting the characteristic information using a depth separable convolutional network.
In certain implementations of the first aspect, the countermeasure generation network includes a spurious generator and a spurious generator, wherein: the step of obtaining the fake image is to utilize the fake generator to fake the true article image to obtain a fake image; the step of obtaining the false image is to utilize the false generator to make the false article image true to obtain the false image.
In certain embodiments of the first aspect, the false or true generator comprises a depth residual network.
In certain embodiments of the first aspect, further comprising the step of training the challenge generating network: truing the genuine article image with the truing generator to obtain an authentication image that is genuine; truing the false image with the truing generator to obtain a reset image that is true; calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image to maintain a loss function value of the countermeasure generation network within a desired range.
In certain embodiments of the first aspect, further comprising the step of training the challenge generating network: counterfeiting the fake item image with the counterfeiting generator to obtain an authentication image that is fake; faking the authenticity image with the faking generator to obtain a reset image that is false; calculating an image similarity of the fake item image to the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
In certain embodiments of the first aspect, the confusion discrimination network comprises a fake item discriminator and a true item discriminator, wherein the fake image is input to the fake item discriminator to learn to output a result that the fake image is true; and inputting the false image to the genuine article discriminator to learn to output a result that the false image is false.
In certain embodiments of the first aspect, the method further comprises the step of training the confusion discrimination network: inputting the counterfeit image to the counterfeit article discriminator to learn to output a result that the counterfeit image is true; inputting the true article image to the fake article discriminator to learn so as to output the true article image as a result of predicting true; inputting the false article image to the false article discriminator to learn to output a result that the false article image is predicted false.
In certain embodiments of the first aspect, the method further comprises the step of training the confusion discrimination network: inputting the genuine article image to the genuine article discriminator to learn to output the genuine article image as a result of predicting genuine; inputting the false article image to the true article discriminator to learn to output the false article image as a result of predicting false; inputting the counterfeit image to the genuine article discriminator to learn to output a result that the counterfeit image is false.
In certain embodiments of the first aspect, the method further comprises the step of training the confusion discrimination network: inputting the true article image into the confusion discrimination network learning to output the true article image as a result of predicting true; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false; inputting an unknown true and false image into the confusion discrimination network learning to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result or output the unknown true and false image and the true object image or the false object image as different attribute prediction results; and obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction results and the true prediction result or the false prediction result.
In certain embodiments of the first aspect, the confusion discrimination network comprises a siense network.
In certain embodiments of the first aspect, the article comprises a luxury bag with a leather slip or an article with handwriting.
A second aspect of the present application provides a training system for an item authentication model, comprising: the system comprises a sample input module, a verification module and a verification module, wherein the sample input module is used for acquiring a plurality of training sample images of an article to be identified, and the training sample images comprise a real article image and a false article image; the countermeasure generation module is used for counterfeiting the true article image by utilizing a countermeasure generation network to obtain a counterfeit image; and using the countermeasure generating network to simulate the fake article image to obtain a simulated image; the confusion judging module is used for inputting the fake image into a confusion judging network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
In certain embodiments of the second aspect, the system further comprises a feature extraction unit for locating the identified point locations in the images of the genuine articles and the images of the counterfeit articles to extract the characteristic information.
In certain embodiments of the second aspect, the feature extraction unit locates the authentication point locations in the true item image and the false item image using a fast recursive convolutional neural network and extracts the characteristic information using a depth separable convolutional network.
In certain embodiments of the second aspect, the challenge generation module comprises: the fake generator is used for carrying out fake making on the real article image to obtain a fake image; and the fake generator is used for truing the fake article image to obtain a trueness image.
In certain embodiments of the second aspect, the false or true generator comprises a depth residual network.
In certain embodiments of the second aspect, the authenticity generator is further for authenticating the genuine article image to obtain an authentication image that is genuine; truing the false image to obtain a true reset image; and calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image so as to keep the loss function value of the countermeasure generation network within a desired range.
In certain embodiments of the second aspect, the counterfeiting generator is further configured to counterfeit the counterfeit item image to obtain a; truing the false image to obtain a false reset image; and calculating an image similarity of the fake article image and the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
In certain embodiments of the second aspect, the confusion discrimination module comprises: a fake article discriminator for learning an input fake image to output a result that the fake image is true; and the genuine article discriminator is used for learning the input genuine image so as to output a result that the genuine image is false.
In certain embodiments of the second aspect, the false item discriminator is further configured to learn the input true item image to output the true item image as a result of predicting true; and learning the input false article image to output the false article image as a result of predicting false.
In certain embodiments of the second aspect, the genuine article discriminator is further configured to learn the input genuine article image to output the genuine article image as a result of predicting genuineness; and learning the input false article image to output the false article image as a result of predicting false.
In certain embodiments of the second aspect, the confusion discrimination module is further configured to input the genuine article image to the confusion discrimination network learning to output a result that the genuine article image is predicted genuine; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false; inputting an unknown true and false image into the confusion discrimination network learning to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result or output the unknown true and false image and the true object image or the false object image as different attribute prediction results; and obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction result and the true prediction result or the false prediction result.
In certain embodiments of the second aspect, the confusion discrimination module comprises a siense network.
In some embodiments of the second aspect, the article comprises a luxury bag with a leather slip or an article with handwriting.
A third aspect of the present application provides an article authentication method comprising the steps of: acquiring a shot image of an article to be authenticated; the image comprises at least one target authentication point; identifying the image containing the target identification point by using an article identification model obtained by the training method of the article identification model according to the first aspect; and outputting the result that the object to be authenticated is true or the result that the object to be authenticated is false.
In certain embodiments of the third aspect, the article comprises a luxury bag with a leather tag or an article with handwriting, and the target identification point is a location of the leather tag or handwriting.
A fourth aspect of the present application provides an article authentication apparatus comprising: the shooting device is used for acquiring a shot image of the article to be authenticated; the image comprises at least one target authentication point; a memory for storing program code; one or more processors configured to invoke program code stored in the memory to perform the method of item authentication according to the third aspect.
In certain embodiments of the fourth aspect, the article comprises a luxury bag with a leather tag or an article with handwriting, and the target identification point is a location of the leather tag or handwriting.
A fifth aspect of the present application provides an article authentication client, which is loaded in a smart device, and includes: the intelligent equipment comprises an input module, a storage module and a display module, wherein the input module is used for calling a shooting device of the intelligent equipment to acquire a shot image of an article to be authenticated when an authentication instruction input by a user is received; the image comprises at least one target authentication point; a processing module calling a program code stored in the smart device to perform the item authentication method according to the third aspect; and the display module is used for displaying and outputting the result that the article to be authenticated is true or the result that the article to be authenticated is false.
In some embodiments of the fifth aspect, the input module further includes a selection unit, configured to, when receiving an authentication instruction input by a user, invoke an image library of the smart device for the user to select an image of an article to be authenticated; the image includes at least one target authentication point.
In certain embodiments of the fifth aspect, the article comprises a luxury bag with a leather tag or an article with handwriting, and the target identification point is a location of the leather tag or handwriting.
A sixth aspect of the present application provides a computer-readable storage medium storing a computer program for item authentication, which when executed implements the method of training an item authentication model according to the first aspect.
A seventh aspect of the present application provides a computer-readable storage medium storing a computer program for article authentication, which when executed, implements the article authentication method according to the third aspect.
In summary, the training method of the article authentication model, the training system of the article authentication model, the article authentication method, the article authentication device, the article authentication client, and the computer-readable storage medium provided by the present application have the following advantages: the countermeasure generating network and the confusion judging network are combined, the true making generator and the false making generator of the countermeasure generating network are utilized to make true or false of the training sample image, and the image after true or false making is input into the confusion judging network, so that the true article discriminator and the false article discriminator of the confusion judging network are trained, and the identification capability of the confusion judging network can be improved. Meanwhile, the mode of alternately and iteratively training the countermeasure generating network and the confusion judging network is adopted, so that the countermeasure generating network and the confusion judging network can be trained and learned together, the robustness and the accuracy of the countermeasure generating network and the confusion judging network are improved, and the identification capability and the robustness of the article identification model are further effectively improved. The article identification model obtained by training through the article identification model training method provided by the application is used for identifying the article, so that the accuracy rate is high, and the stability is strong.
Drawings
Fig. 1 is a flow chart illustrating an embodiment of a training method for an item authentication model according to the present application.
Fig. 2 is a schematic diagram of a training method of an article authentication model according to another embodiment of the present application.
Fig. 3 is a schematic diagram of a training process for generating a network against the countermeasure in the present application in one embodiment.
Fig. 4 is a schematic diagram of a training process against a generation network in another embodiment of the present application.
Fig. 5 is a schematic diagram of a training method of an article authentication model according to the present application in a further embodiment.
Fig. 6 is a schematic diagram illustrating a training process of the confusion decision network according to an embodiment of the present invention.
Fig. 7 is a schematic diagram of a training process of the confusion discrimination network in another embodiment of the present application.
Fig. 8 is a schematic diagram of a training process of the confusion discrimination network in another embodiment of the present application.
Fig. 9 is a schematic diagram of a training process of the confusion discrimination network in another embodiment of the present application.
Fig. 10 is a schematic structural diagram of a training system for an article authentication model according to an embodiment of the present application.
Fig. 11 is a flow chart illustrating an embodiment of the method for authenticating an article according to the present application.
Fig. 12 is a schematic view showing the structure of an article authentication apparatus according to an embodiment of the present application.
Fig. 13 is a schematic structural diagram of an article authentication client according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application is provided for illustrative purposes, and other advantages and capabilities of the present application will become apparent to those skilled in the art from the present disclosure.
Although the terms first, second, etc. may be used herein to describe various elements or parameters in some instances, these elements or parameters should not be limited by these terms. These terms are only used to distinguish one element or parameter from another element or parameter. For example, the first euclidean distance may be referred to as the second euclidean distance, and similarly, the second euclidean distance may be referred to as the first euclidean distance, without departing from the scope of the various described embodiments. The first euclidean distance and the second euclidean distance are both values describing calculated euclidean distances for both images, but they are not the same euclidean distance and do not have the same value unless the context clearly indicates otherwise. Similar situations also include the first distance and the second distance device, or the first discriminator and the second discriminator.
Also, as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context indicates otherwise. It will be further understood that the terms "comprises," "comprising," "includes" and/or "including," when used in this specification, specify the presence of stated features, steps, operations, elements, components, items, species, and/or groups, but do not preclude the presence, or addition of one or more other features, steps, operations, elements, components, species, and/or groups thereof. The terms "or" and/or "as used herein are to be construed as inclusive or meaning any one or any combination. Thus, "A, B or C" or "A, B and/or C" means "any of the following: a; b; c; a and B; a and C; b and C; A. b and C ". An exception to this definition will occur only when a combination of elements, functions, steps or operations are inherently mutually exclusive in some way.
The increasing development of electronic transaction platforms not only enriches the sales channels of articles, but also provides a convenient sales platform for counterfeiting or imitating articles, and users have to worry about the authenticity of the purchased articles. However, the article authentication capability is not easily possessed by ordinary consumers, and similarly, as an article sales platform, the authenticity of the article is not easily audited by an electronic transaction platform. As some fashionable luxury goods such as bags, watches and the like, article identifiers need to store a large amount of knowledge about related articles and to be familiar with the materials, working procedures, theme characteristics and the like used by the articles, and therefore, the requirements of users who purchase high-end articles on the accuracy, timeliness and the like of article identification cannot be met by adopting a manual identification mode.
Deep learning is an algorithm for performing characterization learning on data in machine learning. The motivation for the research originated from artificial neural networks is to build neural networks that simulate the human brain for analytical learning, interpreting image, sound and text data by a mechanism that mimics the human brain. Deep learning forms a more abstract class or feature of high-level representation properties by combining low-level features to discover a distributed feature representation of the data. Deep learning has become a representative machine learning technique and is widely used in the fields of images, sounds, texts, etc., however, no precedent for identifying luxury leather labels using deep learning techniques exists in the existing research patents and documents. The method creatively uses an artificial intelligence technology based on deep learning, combines a deep convolution network and a countermeasure generation network to train an article identification model, utilizes the trained article identification model to identify an article image, and can automatically identify the authenticity of articles such as luxury goods labels.
Referring to fig. 1, a schematic flow chart of an embodiment of the method for training the item authentication model of the present application is shown, and as shown in the drawing, the method for training the item authentication model includes the following steps:
step S11, obtaining a plurality of training sample images of the article to be identified, wherein the training sample images comprise a true article image and a false article image.
The article to be authenticated includes a luxury bag with a leather label or an article with handwriting. For example, identifying a bag with a label, such as a backpack, rucksack, handbag, messenger bag, travel bag, etc. of a certain brand; such as bracelets, rings, watches, pocket watches, contracts, etc. that have writing. Among other luxury items, in some embodiments, such as Amazos (Herme) bag, Louis Vuitton (Louis Vuitton) bag, Amani (Giorgio Armani), Baogeli (BVLGARRI), Fendi (FenDi), Paris Shimega (BALENCIAGA), Chanel (CHANEL) bag, Prada (PRADA) bag, BottegaVeneta (Bottegenaeta) bag, Gucci (Gucci) bag, Carz (COACH) bag, and Dior (Dior) bag, among others. In the examples provided in this application, the description is made for the time being taking the article to be authenticated as a luxury bag as an example.
It is easy to understand that the training sample image is composed of a plurality of acquired images, wherein the plurality of images include a true article image and a false article image, the true article image is an image of the article to be identified as a true article in advance and is shot for the true article, and the false article image is an image of the article to be identified as a fake article in advance and is shot for the fake article. In an embodiment, the prior identification may be by manual identification or technical identification.
The real article image and the false article image can be directly obtained by shooting through a camera device of the intelligent terminal, and can also be obtained through network transmission. In some embodiments, the means for acquiring the images of the real item and the false item include, but are not limited to: shooting an article to be identified by using camera equipment of an intelligent terminal such as a mobile phone, a tablet personal computer, a notebook computer and the like, and directly taking a shot image as a true article image or a false article image; in some embodiments, the true or false item image is obtained by downloading an item image on a web page or on a server from which true or false has been identified; in some embodiments, the image received by wired or wireless communication or the like is taken as a true item image or a false item image. The image format is a format of a computer storage image, such as a storage format of bmp, jpg, png, tiff, gif, pcx, tga, exif, fpx, svg, psd, cdr, pcd, dxf, ufo, eps, ai, raw, WMF, and the like.
In practical applications, on the one hand, the images may vary depending on the shooting conditions. For example, an image of an object to be authenticated taken in the sun is different from an image of the object to be authenticated taken in indoor lighting, and moreover, the color of the lighting affects the image. On the other hand, the images taken differ depending on the length of use or the degree of use of the article to be authenticated. For example, items that are used frequently for extended periods of time may wear more than the items that were just purchased, or may have experienced weathering that darkens the color, or may become stained during use, etc.
Thus, in some embodiments, data enhancement methods are used on the acquired images to expand the number of images to increase the robustness of the training. The data enhancement includes, for example, mirror symmetry, random cropping, rotation, stretching, shrinking, distortion, local bending, color conversion, random patch coverage, adding noise, and the like. For example, different distortion values are added to R, G and B channels in one image respectively to obtain a color-converted image, and both the original image and the color-converted image are used as training sample images. The channel refers to an RGB channel or a CMYK channel in an image, for example, for an RGB 64 × 64 image, a 64 × 64 × 3 vector may be used for representation; wherein, 3 is used for representing channels, namely Red (Red), Green (Green) and Blue (Blue) channels. For another example, according to different noise types, a noise random number is added to each of the plurality of images, and the original image and the plurality of images with noise added are both used as training sample images.
In practical applications, there are cases where the size and resolution of the acquired few images are very large or very small, and if they are not screened or preprocessed and directly input into the neural network for training, the accuracy of the item identification model may be affected. Therefore, in some embodiments, the method further comprises a step of screening the acquired images, and rejecting images which are not matched with the input images of the article identification model so as to ensure the quality of the training sample images and increase the stability of the article identification model training.
Since each acquired image is not necessarily identical in parameters such as definition, image size, resolution, etc., in some embodiments, in order to facilitate the subsequent steps to identify all the acquired images or the images retained after screening, the training method of the article identification model further includes a step of preprocessing the plurality of training sample images. The preprocessing step is used for adapting the acquired image and a neural network model input image to be called to obtain a training sample image. The step of preprocessing comprises one or more of size modification, scaling, noising, inversion, rotation, translation, scaling transformation, cropping, contrast transformation, random channel offset, filtering of the image. For example, the pre-processing includes rejecting images whose image size is smaller than a preset size threshold. As another example, the preprocessing includes performing sharpness analysis on the acquired image and selecting an image with sharpness meeting a preset sharpness condition as a training sample image. As another example, the pre-processing includes rotating the received 550 × 300 pixel image to 300 × 550. As another example, the pre-processing includes reducing or cropping the received 1000 × 800 pixel image and then rotating it to 300 × 550. As another example, the preprocessing includes sharpening, adding noise, and the like to the acquired beautified image to restore the real image. For another example, the preprocessing includes identifying the outline of the article in the image, and rotating, translating and cropping the image according to the position of the article image in the whole image, so as to minimize the background in the article image and obtain an image conforming to the size received by the subsequent confrontation generating network. As another example, the preprocessing includes grayscale flipping of the acquired images to facilitate subsequent suppression of the background or to highlight item features.
In a practical application scenario, the overall appearances of the luxury genuine bag and the luxury fake bag are very similar, and the difficulty in identifying the luxury genuine bag and the luxury fake bag is high from the overall appearance, so that identification points such as the fonts of leather labels, the engraving process and the like are key points for identifying the genuine bags and the fake bags. Thus, in some embodiments, a training sample image containing an item to be authenticated is obtained, and the method further comprises the step of locating the position of the authenticated point in the images of the real and false items to extract characteristic information, including font, stroke, texture, relative position and size, and the like. It will be appreciated that the location of the authentication point of the article to be authenticated may vary from article to article, for example in ladies' bags for luxury goods, the authentication point typically comprising a leather label, a zipper and a label comprising the brand of the article; for another example, in the case of a wristwatch, the identified point is usually located on a watchband, a dial, or the like; for example, in the case of printed matter such as a contract, the position of the authentication point is usually written on a writing trace or a stamp.
In practical implementations, the way to locate the position of the evaluation point includes, but is not limited to: 1) and (5) manually calibrating. For example, an image containing a luxury bag (corresponding to an article to be authenticated) is photographed using a mobile phone, and a range of a certain size, which contains a leather label (corresponding to an authentication point) of the luxury bag, is drawn on the image as an authentication point position. For another example, after acquiring a plurality of training sample images, all the training sample images are manually identified and calibrated one by one. 2) The identified point is located by a recursive convolutional neural network. For example, using a fast recursive convolutional neural network, a candidate region where the position of the identified point may exist is first found in each of the true article image and the false article image to obtain the multidimensional characteristics of the images, and then the accurate position of the identified point is obtained by using sub-networks such as target detection discrimination, search box position regression, and the like.
After the location of the identification point, the manner of extracting the characteristic information includes extracting the characteristic information by using a deep convolutional network. The deep convolutional network includes Xception (deep separable convolutional network), ResNet (deep residual error network), MobileNet, and the like. For example, for an image of a luxury bag that identifies a point location at a picotag location, a depth separable convolutional network is used to extract characteristic information such as font, stroke, texture, relative position and size from the picotag portion of the image. The depth separable convolution network respectively convolves each channel of the image and decouples the spatial position of the characteristic information in the image and the convolution among the channels, so that the capability of extracting the characteristic information is improved, and the network learning and optimization are facilitated.
Referring to fig. 2, which is a schematic diagram of an embodiment of the training method of the article authentication model of the present application, as shown, for example, a training sample image containing a luxury bag is input into a recurrent convolutional neural network that finds candidate regions in the image where the positions of the authentication points may exist (the candidate regions are shown as rectangular boxed regions in fig. 2). After a plurality of candidate areas are determined, the recursive convolutional neural network inputs the generated image containing the candidate areas into a deep convolutional network, and the deep convolutional network analyzes the candidate areas in the image by utilizing sub-networks such as target detection discrimination, search box position regression and the like, so that the accurate position of the identification point is obtained.
Step S12, fake the true object image by using a confrontation generating network to obtain a fake image; and utilizing the countermeasure generating network to simulate the fake article image to obtain a simulated image.
In order to further improve the authentication capability of the article authentication model and enable the neural network to really learn the nuances between the leather labels of luxury goods, the embodiment of the application trains the training sample image by using a pair of confrontation generation networks. In the training process, the true article image and the false article image are input into the countermeasure generating network, the countermeasure generating network performs counterfeiting on the true article image to obtain a fake image, and performs counterfeiting on the false article image to obtain a fake image.
Specifically, the counterfeiting refers to a process of performing image processing on the image of the real object to obtain a counterfeiting image. In some embodiments, the counterfeiting or forgery includes image processing of the false or true item image, such as mirror symmetry, rotation, distortion, local bending, color conversion, random patch coverage, adding noise, and the like. For example, any processing on an image (e.g., stretching, cropping, color transformation, etc.) made on an image containing a real luxury bag cannot affect the authenticity of the luxury bag real object. Therefore, no matter how the true article image is faked, the article corresponding to the obtained faking image is still true, that is, the faking image is still a true article image in nature. If the false image is identified, the identification result is still true. Similarly, the false positive refers to a process of image processing the false article image to obtain a false positive image. Similarly, no matter how the false article image is subjected to counterfeiting, the article corresponding to the obtained counterfeiting image is still a counterfeit article, and the counterfeiting image is still a false article image in nature; if the false image is identified, the identification result is still false.
In some embodiments, the countermeasure generation network includes a genuine generator and a fake generator that can generate a genuine image or a fake image different from a genuine article image or a fake article image from an input genuine article image or fake article image. The step of obtaining the fake image is to utilize the fake generator to fake the true article image to obtain a fake image; the step of obtaining the false image is to utilize the false generator to make the false article image true to obtain the false image.
Generally speaking, the more layers of the network in deep learning is more effective, i.e. the deeper the network is, the higher the training and testing accuracy is. However, in practical application, because of the problem of gradient disappearance, when the number of layers of the network is increased to a certain number, the accuracy of training and testing is reduced. To address the problem of gradient vanishing while increasing the depth of the network, thereby further increasing the accuracy of the item authentication model, in some embodiments the fraud generator or the fraud generator comprises a depth residual network. In some embodiments, a deep residual network of a plurality of residual blocks, which may be 8, 10, 18, 34, etc., connected together, may be used as a true generator and a false generator. The number of residual blocks described in the embodiments of the present application is merely an example, and does not limit the scope thereof.
For neural network models, complex models tend to be over-fit. Overfitting means that the model works much better for the training sample set as input than for the test sample set. Therefore, the loss function of the model needs to be balanced, so that the model has better generalization capability and the robustness of the model is improved.
Thus, in some embodiments, the method further comprises the step of training the challenge generating network. Referring to fig. 3, a schematic diagram of a training process for a challenge generation network according to an embodiment of the present application is shown, as shown in fig. 3, including: truing the genuine article image with the truing generator to obtain a genuine authentication image, truing the counterfeit image with the truing generator to obtain a genuine reset image; calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image to maintain a loss function value of the countermeasure generation network within a desired range.
And inputting the real article image in the training sample image to the counterfeiting generator for counterfeiting, and generating a counterfeiting image by the counterfeiting generator. Inputting the counterfeit image to the authenticity generator, the authenticity generator generating a true reset image. The fake image is true and the reset image is a true reset image because the image generated by the true object image after being subjected to the fake is still true. And inputting the true article image in the training sample image into the false positive generator for false positive, wherein the false positive generator generates the authentication image. Similarly, the image generated after the genuine article image is genuine, and therefore the generated authentication image is a genuine authentication image.
The calculation of the image similarity includes but is not limited to: calculating the euclidean distance, absolute value distance, histogram intersection, mahalanobis distance, manhattan distance, chebyshev distance, minkowski distance, normalized euclidean distance, included angle cosine, center moment, etc. of the true article image and the true authentication image and/or the true reset image. In the embodiments provided in the present application, the euclidean distance is calculated as an example for the first time. When the similarity of the images is calculated by utilizing the Euclidean distance, the Euclidean distance refers to the actual distance between two points in the multi-dimensional space, and the smaller the Euclidean distance is, the greater the similarity of the two images is.
There are a number of ways to calculate the euclidean distance of the genuine article image from the genuine authentication image and/or the genuine reset image, including but not limited to: for example, a first euclidean distance between the true article image and a true authentication image and a second euclidean distance between the true article image and a true reset image are respectively calculated, an average value or a root mean square of the first euclidean distance and the second euclidean distance is taken, so as to obtain a target euclidean distance, and whether a loss function value of the countermeasure generation network is kept within an expected range is judged according to the target euclidean distance; for another example, a first euclidean distance between the true article image and the true authentication image and a second euclidean distance between the true article image and the true reset image are respectively calculated, a larger value of the first euclidean distance and the second euclidean distance is taken, the euclidean distance corresponding to the larger value is taken as a target euclidean distance, and whether the loss function value of the countermeasure generation network is kept within an expected range is determined according to the target euclidean distance.
Similarly, in some embodiments, please refer to fig. 4, which is a schematic diagram of a training process of the countermeasure generation network in another embodiment of the present application, and as shown in fig. 4, the step of training the countermeasure generation network further includes: counterfeiting the counterfeit item image with the counterfeiting generator to obtain a false authentication image, counterfeiting the authenticity image with the counterfeiting generator to obtain a false reset image; calculating an image similarity of the fake item image to the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
And inputting the false article image in the training sample image into the false reality generator for false reality, wherein the false reality generator generates a false reality image. Inputting the authenticity image to the fraud generator, the fraud generator generating a false reset image. Since the image generated by the false article image after the false article image is false, the false image is false, and the reset image is a false reset image. And inputting the fake article image in the training sample image to the fake generator for fake making, wherein the fake generator generates an authentication image. Similarly, the image generated by the fake article image after being counterfeited is still false, so that the generated authentication image is a false authentication image.
The steps of training the confrontation generating network in this embodiment are similar to those described in the above embodiment, and the principle and flow of the training are referred to the above embodiment, which is not described herein again.
It should be appreciated that if the training of the true and false generators is not constrained, the true and false generators randomly do either true or false, and that there is a high probability that either over-true or over-false will occur. For example, if the genuine article image contains a luxury wallet, the corresponding authentication point location includes a label, a brand mark, etc., and if over-counterfeiting or over-counterfeiting occurs, it is likely that the label in the generated counterfeit image will be image-processed into another article that is not related to the label, or even to the wallet. In this case, it is meaningless to train the item authentication model with an image that is generated by counterfeiting, regardless of the item to be authenticated, and that has been missing the necessary attributes or features in the training sample image. Thus, during the training process, the image similarity of the fake item image to the fake authentication image and/or the fake reset image is calculated to keep the loss function value of the countermeasure generating network within a desired range. The loss function is used for estimating the degree of inconsistency (or deviation degree) between the predicted value and the true value of the model, and the smaller the value of the loss function is, the better the robustness of the model is. The loss function values include L1 norm loss functions, cross entropy loss functions, log logarithmic loss functions, squared loss functions, exponential loss functions, Hinge loss functions, absolute value loss functions, and the like. After the target Euclidean distance is obtained through calculation, the loss function value of the countermeasure generation network is obtained according to the target Euclidean distance, and the calculated loss function value is compared with a preset value or a preset range to judge whether the loss function value of the countermeasure generation network is kept in an expected range or not, so that the false generator and the false generator are prevented from being over-true or false.
The cross entropy loss function is used to determine how close the actual value is to the desired value. The cross entropy refers to the distance between the actual value and the expected value, and the smaller the cross entropy, the closer the probability distribution of the actual value and the expected value. In some embodiments, the discrimination capability and accuracy of the item discrimination model can be further enhanced by using the cycle loss and the self loss in the L1 norm loss function and training the challenge generating network using the challenge loss in the cross entropy loss function.
Step S13, inputting the false image into a confusion judgment network to output the true result of the false image; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
And inputting a fake image obtained after the fake object image is fake and a fake image obtained after the fake object image is fake into a confusion judging network, judging and learning the fake image and the fake image by the confusion judging network, and outputting a result that the fake image is true and a result that the fake image is false.
In some embodiments, the confusion discrimination network comprises training and optimizing the confusion discrimination network using the structure of the siemese network. For example, two true and false images containing the skin tag are used as input, and besides the identification results of the two images, the same-attribute prediction result or different-attribute prediction result is output to indicate whether the two images belong to the same class (both true or both false) or whether the two images belong to one true or one false. Thereby, the discrimination capability of the article discrimination model can be further enhanced.
As mentioned above, no matter how the true article image is faked, the article corresponding to the obtained faked image is still a true article. Therefore, the result of judging and outputting the fake image by using the confusion judging network is still that the fake image is true. Similarly, no matter how the false article image is subjected to the counterfeiting, the article corresponding to the obtained counterfeiting image is still a counterfeit article. Therefore, the result of the discrimination output of the false image by the confusion discrimination network is still that the false image is false.
In some embodiments, the confusion discrimination network comprises a fake article discriminator and a true article discriminator, wherein the fake image is input to the fake article discriminator to learn to output a result that the fake image is true; and inputting the false image to the genuine article discriminator to learn to output a result that the false image is false.
Referring to fig. 5, a schematic diagram of a training method for an item identification model according to another embodiment of the present application is shown, in which a solid line with an arrow indicates a training process when a real item image is input as a training sample image in one embodiment, and a dotted line with an arrow indicates a training process when a false item image is input as a training sample image in one embodiment. As shown, in one aspect, the true article image is input to a fraud generator in a countermeasure generation network, which generates fraud images, which are then input to a fraud discriminator in a confusion discrimination network, which discriminates the fraud images. As can be seen from the foregoing, the image obtained by counterfeiting the true article image is still true, and therefore, the false article discriminator outputs the result that the counterfeit image is true. On the other hand, the fake image generated by the fake generator is input into a fake generator in the countermeasure generation network, and the fake generator generates a true reset image; inputting the genuine article image into the authenticity generator, the authenticity generator generating a genuine authentication image. Whether the loss function value of the countermeasure generation network is kept within an expected range in the training process is judged by calculating the image similarity between the real article image and the real reset image and/or the real authentication image, and if the loss function value is beyond the expected range, the parameters of the countermeasure generation network need to be adjusted, so that the truth generator and the fake generator are prevented from being over-true or fake, and the accuracy of the article authentication model is further influenced.
Similarly, the false article image is input to a false image generator in a countermeasure generation network that generates a false image, which is then input to a true article discriminator in a confusion discrimination network that discriminates the false image. As can be seen from the foregoing, the image obtained by truing the false article image is still false, and therefore, the true article discriminator outputs the result that the truing image is false. On the other hand, the false image generated by the false generator is input into a false generator in the countermeasure generation network, and the false generator generates a false reset image; inputting the counterfeit item image into the counterfeiting generator, the counterfeiting generator generating a counterfeit authentication image. And calculating image similarity between the false article image and the false reset image and/or the false authentication image to judge whether the loss function value of the countermeasure generation network in the training process is kept within an expected range.
In some embodiments, the method further comprises the step of training the confusion discrimination network: inputting the counterfeit image to the counterfeit article discriminator to learn to output a result that the counterfeit image is true; inputting the true article image to the fake article discriminator to learn so as to output the true article image as a result of predicting true; inputting the false article image to the false article discriminator to learn to output a result that the false article image is predicted false. Referring to fig. 6 and 7, fig. 6 is a schematic diagram illustrating a training method of the confusion discrimination network according to an embodiment of the present invention, and fig. 7 is a schematic diagram illustrating a training method of the confusion discrimination network according to another embodiment of the present invention. As shown in fig. 6, in the training process, the counterfeit image marked as true is input into the counterfeit article discriminator, and the counterfeit article discriminator outputs a true result; inputting the true article image marked as true into the false article discriminator, wherein the false article discriminator outputs a result of predicting true; inputting the false article image marked as false into the false article discriminator, and outputting a result of predicting false by the false article discriminator. Obviously, the result of predicting true should indeed be true, and the result of predicting false should indeed be false.
Similar to training the fake generator and the fake generator, when training the fake article discriminator and the real article discriminator, the training process also needs to be constrained, so that the fake article discriminator and the real article discriminator improve the discrimination accuracy and keep the training convergence stable. Therefore, in some embodiments, the method further comprises comparing the result of the true prediction with the image of the true item, calculating a distance between the result of the true prediction and the image of the true item, and determining whether the distance is smaller than a preset value. Similarly, in some embodiments, the method further comprises comparing the result of predicting the false with a false article image, and determining whether the distance is smaller than a preset value by calculating the distance between the result of predicting the false and the false article image. In some embodiments, the method further comprises comparing the true result with a fake image, and determining whether the distance is less than a predetermined value by calculating the distance between the true result and the fake image.
In a specific embodiment, the true result, the false result, the result predicted to be true, and the result predicted to be false are set to a real number within a range of [0,1 ]. Wherein the "0" indicates a flag corresponding to false and the "1" indicates a flag corresponding to true. As mentioned earlier, before being input to the false article discriminator, the false, real and false article images have been marked by means of manual or technical marking, and obviously, the false and real article images are marked as "1" to represent that the image is true, and the false article image is marked as "0" to represent that the image is false. Under the above conditions, the output result of the fake article discriminator or the real article discriminator is also a real number in the range of [0,1 ]. The preset value is set, for example, to 0.5. Obviously, if the result output by the false article discriminator or the true article discriminator is less than 0.5, the input image is false, otherwise, the input image is true. When a fake image is used as input, the true result output by the fake article discriminator is compared with the mark value of the fake image, and a first distance between the two is calculated. Similarly, a second distance between the result of predicting true and the tag value of a true item image is calculated, and a third distance between the result of predicting false and the tag value of a false item image is calculated. And obtaining the distance sum of the first distance, the second distance and the third distance through calculation, and minimizing the distance sum through optimization algorithms such as Adam and SGD, so as to train the fake article discriminator.
Similarly, as shown in fig. 7, in some embodiments, the method further includes the step of training the confusion discrimination network: inputting the genuine article image to the genuine article discriminator to learn to output the genuine article image as a result of predicting genuine; inputting the false article image to the true article discriminator to learn to output the false article image as a result of predicting false; inputting the counterfeit image to the genuine article discriminator to learn to output a result that the counterfeit image is false.
Since the step of training the confusion decision network corresponding to fig. 7 is similar to that described above, the principle and process thereof refer to the above embodiments, and are not repeated here.
In some embodiments, the method further comprises the step of training the confusion discrimination network:
step S131, inputting the true article image into the confusion discrimination network learning to output the true article image as a result of true prediction; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false.
Step S132, inputting an unknown true and false image into the confusion discrimination network learning, so as to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result, or output the unknown true and false image and the true object image or the false object image as different attribute prediction results.
And step S133, obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction result and the true prediction result or the false prediction result.
Referring to fig. 8 and 9, fig. 8 is a schematic diagram illustrating a training process of the confusion judging network in an embodiment of the present application, and fig. 9 is a schematic diagram illustrating a training process of the confusion judging network in another embodiment of the present application. As shown in fig. 8, the true article image is input to the confusion discrimination network for discrimination learning, and since the result that the true article image is true is a priori result (that is, the true article image is a training sample image predetermined to be true), the result output by the confusion discrimination network is a result of true prediction, which indicates that the true article image is true. Then, an unknown true and false image is input to the confusion judging network for judging and learning. Because the unknown true and false image only contains true or false results, when the unknown true and false image is a true article image, the prediction result has the same attribute as that of the true article image, namely, the prediction result is true; when the unknown true and false image is a false article image, the attribute of the prediction result of the unknown true and false image is different from that of the true article image, namely true and false. Therefore, after the confusion discrimination network discriminates and learns the unknown true and false images, the unknown true and false images and the true object images are predicted results with the same attribute or predicted results with different attributes. Because the truth of the true article image is known, the unknown true and false image can be judged to be true or false according to the fact that the output result is the same attribute prediction result or different attribute prediction result.
Similarly, as shown in fig. 9, the false article image is input to the confusion judging network for judging and learning, and the output result of the confusion judging network is a result of predicting false, which indicates that the false article image is false. Then, inputting an unknown true and false image into the confusion judging network for judging and learning, and similarly, when the unknown true and false image is a false article image, the attribute of the prediction result of the unknown true and false image is the same as that of the prediction result of the false article image, namely, the unknown true and false image is false; when the unknown true and false image is a true article image, the prediction result of the unknown true and false image has different attributes from the prediction result of the false article image, namely true and false. Since the principle is similar to the above embodiment, please refer to the above embodiment for the specific processes and steps, which are not described herein again.
In the embodiment of the application, the countermeasure generation network and the confusion discrimination network are alternately trained, and the false generator in the countermeasure generation network learn how to convert the true article image and the false article image into each other, so as to generate the image which is false and true to confuse the confusion discrimination network. And the true article discriminator and the false article discriminator in the confusion discrimination network continuously learn whether the input image is true or false, so that network confusion generated by confrontation is avoided, and the discrimination capability is improved. Through the training mode of the cyclic iteration of the confrontation generation network and the confusion discrimination network, the neural network can be continuously perfected and optimized in the training process, so that the discrimination capability of the article discrimination model is improved.
In some embodiments, the neural network may be optimized by using an optimization algorithm such as an adam (adaptive motion estimation) optimization algorithm, an sgd (stored Gradient decision) optimization algorithm, or an RMSprop optimization algorithm, and weights of the neural network are iteratively updated based on training data, so that a cross entropy loss function of the neural network is minimized, convergence is more stable, and the neural network is prevented from falling into local optimization, thereby enhancing robustness of the item identification model.
The article identification model training method comprises the steps of constructing the article identification model based on the structure of a cyclic countermeasure generation network, combining the countermeasure generation network and a confusion discrimination network, utilizing a true making generator and a false making generator of the countermeasure generation network to make a true or false of a training sample image, and inputting the image after the true or false making into the confusion discrimination network, so that a true article discriminator and a false article discriminator of the confusion discrimination network are trained, and the discrimination capability of the confusion discrimination network can be improved. Meanwhile, the mode of alternately and iteratively training the countermeasure generating network and the confusion judging network is adopted, so that the countermeasure generating network and the confusion judging network can be trained and learned together, the robustness and the accuracy of the countermeasure generating network and the confusion judging network are improved, and the identification capability and the robustness of the article identification model are further effectively improved.
The training method of the item identification model is executed by a training system of the item identification model. Referring to fig. 10, which is a schematic structural diagram of a training system of an item identification model according to an embodiment of the present application, as shown in the figure, the training system 100 includes a sample input module 101, a countermeasure generation module 102, and a confusion determination module 103, wherein the training system of the item identification model is implemented by software and hardware in a computer device.
The computer device may be any computing device with mathematical and logical operations, data processing capabilities, including but not limited to: personal computer equipment, a single server, a server cluster, a distributed server, the cloud server and the like. The Cloud Service end comprises a Public Cloud (Public Cloud) Service end and a Private Cloud (Private Cloud) Service end, wherein the Public or Private Cloud Service end comprises Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS), Infrastructure-as-a-Service (IaaS), and the like. The private cloud service end is used for example for an Aliskian cloud computing service platform, an Amazon cloud computing service platform, a Baidu cloud computing platform, a Tencent cloud computing platform and the like.
The cloud server provides at least one remote image uploading service. The remote image upload service includes, but is not limited to, at least one of: goods shelving service, goods identification service, goods complaint service and the like. The service of goods listing, such as a merchant uploads an image of a goods to be sold and a related text description, the service of goods identification, such as a purchaser uploads an image of the goods to identify authenticity, and the service of goods complaint, such as a service of uploading an image of the goods for intervention and mediation of a third party (such as an electronic transaction platform) when the purchaser cannot agree with the merchant.
The computer device includes at least: memory, one or more processors, I/O interfaces, network interfaces, and input structures, among others. Wherein the memory is for storing a plurality of images of an item to be authenticated and at least one program. The memory may include high speed random access memory and may also include non-volatile memory, such as one or more magnetic disk storage devices, flash memory devices, or other non-volatile solid state storage devices.
The memory is for storing program code. The Memory may include Volatile Memory (Volatile Memory), such as Random Access Memory (RAM); the Memory may also include a Non-Volatile Memory (Non-Volatile Memory), such as a Read-Only Memory (ROM), a Flash Memory (Flash Memory), a Hard Disk (Hard Disk Drive, HDD), or a Solid-State Drive (SSD); the memory may also comprise a combination of memories of the kind described above. The memory may be configured to store a set of program codes, so that the processor may call the program codes stored in the memory to implement the functions of any one or more of the functional modules, such as the sample input module, the countermeasure generation module, and the confusion determination module, mentioned in the embodiments of the present application.
The processor may be comprised of one or more general-purpose processors, such as a Central Processing Unit (CPU). The processor may be configured to run a program of any one or more of the following functional blocks in the associated program code: the system comprises a sample input module, a confrontation generation module, a confusion discrimination module and the like. That is, the processor executing the program code may implement the functions of any one or more of the following functional modules: a sample input module, a confrontation generation module, a confusion discrimination module and the like. For the sample input module, the countermeasure generation module and the confusion determination module, reference may be made to the related explanations in the foregoing embodiments.
In certain embodiments, the memory may also include memory that is remote from the one or more processors, such as network-attached memory accessed via RF circuitry or external ports and a communication network, which may be the internet, one or more intranets, Local Area Networks (LANs), wide area networks (WLANs), Storage Area Networks (SANs), and the like, or suitable combinations thereof. The memory controller may control access to the memory by other components of the device, such as the CPU and peripheral interfaces. The memory optionally includes high-speed random access memory, and optionally also includes non-volatile memory, such as one or more magnetic disk storage devices, flash memory devices, or other non-volatile solid-state memory devices. Access to the memory is optionally controlled by a memory controller by other components of the device, such as a CPU and peripheral interfaces.
The one or more processors are operatively coupled with the network interface to communicatively couple the computing device to a network. For example, the network interface may connect the computing device to a local area network (e.g., a LAN), and/or a wide area network (e.g., a WAN). The processor is also operatively coupled to I/O ports that enable the computing device to interact with various other electronic devices, and input structures that enable a user to interact with the computing device. Thus, the input structures may include buttons, keyboards, mice, touch pads, and the like. In addition, the electronic display may include a touch component that facilitates user input by detecting the occurrence and/or location of an object touching its screen.
In a specific application scenario, the sample input module, the countermeasure generation module, and the confusion determination module may be software modules, and the software modules may be deployed on a server, or a virtual machine on the server, or a container on the server. In addition, the software modules may be deployed on the same server or different servers according to actual needs, which is not limited in this application.
In another case, the article authentication system can be realized by an application program (APP) loaded on an intelligent terminal, the intelligent terminal acquires a plurality of images of the article to be authenticated through shooting, uploads the images to a cloud server through a wireless network, and the cloud server performs authentication and then feeds back an identification result.
Such as portable or wearable electronic devices including, but not limited to, smart phones, tablets, smart watches, smart glasses, Personal Digital Assistants (PDAs), etc., it is to be understood that the portable electronic device described in the embodiments of the present application is but one example of an application and that the components of the device may have more or fewer components than shown, or a different configuration of components. The various components of the depicted figures may be implemented in hardware, software, or a combination of hardware and software, including one or more signal processing and/or application specific integrated circuits.
The smart terminal includes memory, a memory controller, one or more processors (CPUs), peripheral interfaces, RF circuitry, audio circuitry, speakers, microphones, input/output (I/O) subsystems, touch screens, other output or control devices, and external ports. These components communicate over one or more communication buses or signal lines.
The intelligent terminal supports various applications, such as one or more of the following: a mapping application, a rendering application, a word processing application, a website creation application, a disc editing application, a spreadsheet application, a gaming application, a telephone application, a video conferencing application, an email application, an instant messaging application, a fitness support application, a photo management application, a digital camera application, a digital video camera application, a web browsing application, a digital music player application, and/or a digital video player application.
As shown in fig. 10, the training system 100 includes a sample input module 101, a confrontation generation module 102, and a confusion discrimination module 103, wherein: the sample input module 101 is configured to obtain a plurality of training sample images of an object to be identified, where the training sample images include a true object image and a false object image. The countermeasure generation module 102 uses a countermeasure generation network to counterfeit the true article image to obtain a counterfeit image; and utilizing the countermeasure generating network to simulate the fake article image to obtain a simulated image. The confusion judging module 103 is used for inputting the fake image into a confusion judging network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
The article to be authenticated includes a luxury bag with a leather label or an article with handwriting. For example, identifying a bag with a label, such as a backpack, rucksack, handbag, messenger bag, travel bag, etc. of a certain brand; such as bracelets, rings, watches, pocket watches, contracts, etc. that have writing. In the examples provided in this application, the description is made for the time being taking the article to be authenticated as a luxury bag as an example.
It is easy to understand that the training sample image is composed of a plurality of acquired images, wherein the plurality of images include a true article image and a false article image, and the true article image is taken as a training sample, the true article image is an image of an article to be identified as a true article in advance and is shot for the true article, and the false article image is taken as an image of the article to be identified as a fake article in advance and is shot for the fake article. In an embodiment, the prior identification may be by manual identification or technical identification.
The real article image and the false article image can be directly obtained by shooting through a camera device of the intelligent terminal, and can also be obtained through network transmission. In some embodiments, the means for acquiring the images of the real item and the false item include, but are not limited to: shooting an article to be identified by using camera equipment of an intelligent terminal such as a mobile phone, a tablet personal computer, a notebook computer and the like, and directly taking a shot image as a true article image or a false article image; in some embodiments, the true or false item image is obtained by downloading an item image on a web page or on a server from which true or false has been identified; in some embodiments, the image received by wired or wireless communication or the like is taken as a true item image or a false item image. The image format is a format of a computer storage image, such as a storage format of bmp, jpg, png, tiff, gif, pcx, tga, exif, fpx, svg, psd, cdr, pcd, dxf, ufo, eps, ai, raw, WMF, and the like.
In practical applications, on the one hand, the images may vary depending on the shooting conditions. For example, an image of an object to be authenticated taken in the sun is different from an image of the object to be authenticated taken in indoor lighting, and moreover, the color of the lighting affects the image. On the other hand, the images taken differ depending on the length of use or the degree of use of the article to be authenticated. For example, items that are used frequently for extended periods of time may wear more than the items that were just purchased, or may have experienced weathering that darkens the color, or may become stained during use, etc.
Thus, in some embodiments, data enhancement methods are used on the acquired images to expand the number of images to increase the robustness of the training. The data enhancement includes, for example, mirror symmetry, random cropping, rotation, stretching, shrinking, distortion, local bending, color conversion, random patch coverage, adding noise, and the like. For example, different distortion values are added to R, G and B channels in one image respectively to obtain a color-converted image, and both the original image and the color-converted image are used as training sample images. The channel refers to an RGB channel or a CMYK channel in an image, for example, for an RGB 64 × 64 image, a 64 × 64 × 3 vector may be used for representation; wherein, 3 is used for representing channels, namely Red (Red), Green (Green) and Blue (Blue) channels. For another example, according to different noise types, a noise random number is added to each of the plurality of images, and the original image and the plurality of images with noise added are both used as training sample images.
In practical applications, there are cases where the size and resolution of the acquired few images are very large or very small, and if they are not screened or preprocessed and directly input into the neural network for training, the accuracy of the item identification model may be affected. Therefore, in some embodiments, the method further comprises a step of screening the acquired images, and rejecting images which are not matched with the input images of the article identification model so as to ensure the quality of the training sample images and increase the stability of the article identification model training.
Since each acquired image is not necessarily identical in parameters such as definition, image size, resolution, etc., in some embodiments, in order to facilitate the subsequent steps to identify all the acquired images or the images retained after screening, the training method of the article identification model further includes a step of preprocessing the plurality of training sample images. The preprocessing step is used for adapting the acquired image and a neural network model input image to be called to obtain a training sample image. The step of preprocessing comprises one or more of size modification, scaling, noising, inversion, rotation, translation, scaling transformation, cropping, contrast transformation, random channel offset, filtering of the image. For example, the pre-processing includes rejecting images whose image size is smaller than a preset size threshold. As another example, the preprocessing includes performing sharpness analysis on the acquired image and selecting an image with sharpness meeting a preset sharpness condition as a training sample image. As another example, the pre-processing includes rotating the received 550 × 300 pixel image to 300 × 550. As another example, the pre-processing includes reducing or cropping the received 1000 × 800 pixel image and then rotating it to 300 × 550. As another example, the preprocessing includes sharpening, adding noise, and the like to the acquired beautified image to restore the real image. For another example, the preprocessing includes identifying the outline of the article in the image, and rotating, translating and cropping the image according to the position of the article image in the whole image, so as to minimize the background in the article image and obtain an image conforming to the size received by the subsequent confrontation generating network. As another example, the preprocessing includes grayscale flipping of the acquired images to facilitate subsequent suppression of the background or to highlight item features.
In a practical application scenario, the overall appearances of the luxury genuine bag and the luxury fake bag are very similar, and the difficulty in identifying the luxury genuine bag and the luxury fake bag is high from the overall appearance, so that identification points such as the fonts of leather labels, the engraving process and the like are key points for identifying the genuine bags and the fake bags. Therefore, in some embodiments, the training system of the item authentication model further comprises a feature extraction unit for locating the position of the authentication point in the image of the true item and the image of the false item to extract the characteristic information. In some embodiments, the feature extraction unit locates the authentication point locations in the true item image and the false item image using a fast recursive convolutional neural network and extracts the characteristic information using a depth separable convolutional network.
It will be appreciated that the location of the authentication point of the article to be authenticated may vary from article to article, for example in ladies' bags for luxury goods, the authentication point typically comprising a leather label, a zipper and a label comprising the brand of the article; for another example, in the case of a wristwatch, the identified point is usually located on a watchband, a dial, or the like; for example, in the case of printed matter such as a contract, the position of the authentication point is usually written on a writing trace or a stamp.
In practical implementations, the way to locate the position of the evaluation point includes, but is not limited to: 1) and (5) manually calibrating. For example, an image containing a luxury bag (corresponding to an article to be authenticated) is photographed using a mobile phone, and a range of a certain size, which contains a leather label (corresponding to an authentication point) of the luxury bag, is drawn on the image as an authentication point position. For another example, after acquiring a plurality of training sample images, all the training sample images are manually identified and calibrated one by one. 2) The identified point is located by a recursive convolutional neural network. For example, using a fast recursive convolutional neural network, a candidate region where the position of the identified point may exist is first found in each of the true article image and the false article image to obtain the multidimensional characteristics of the images, and then the accurate position of the identified point is obtained by using sub-networks such as target detection discrimination, search box position regression, and the like.
After the location of the identification point, the manner of extracting the characteristic information includes extracting the characteristic information by using a deep convolutional network. The deep convolutional network includes Xception (deep separable convolutional network), ResNet (deep residual error network), MobileNet, and the like. For example, for an image of a luxury bag that identifies a point location at a picotag location, a depth separable convolutional network is used to extract characteristic information such as font, stroke, texture, relative position and size from the picotag portion of the image. The depth separable convolution network respectively convolves each channel of the image and decouples the spatial position of the characteristic information in the image and the convolution among the channels, so that the capability of extracting the characteristic information is improved, and the network learning and optimization are facilitated.
In order to further improve the authentication capability of the article authentication model and enable the neural network to really learn the nuances between the leather labels of luxury goods, the embodiment of the application trains the training sample image by using a pair of confrontation generation networks. In the training process, the true article image and the false article image are input into the countermeasure generating network, the countermeasure generating network performs counterfeiting on the true article image to obtain a fake image, and performs counterfeiting on the false article image to obtain a fake image.
Specifically, the counterfeiting refers to a process of performing image processing on the image of the real object to obtain a counterfeiting image. In some embodiments, the counterfeiting or forgery includes image processing of the false or true item image, such as mirror symmetry, rotation, distortion, local bending, color conversion, random patch coverage, adding noise, and the like. For example, any processing on an image (e.g., stretching, cropping, color transformation, etc.) made on an image containing a real luxury bag cannot affect the authenticity of the luxury bag real object. Therefore, no matter how the true article image is faked, the article corresponding to the obtained faking image is still true, that is, the faking image is still a true article image in nature. If the false image is identified, the identification result is still true. Similarly, the false positive refers to a process of image processing the false article image to obtain a false positive image. Similarly, no matter how the false article image is subjected to counterfeiting, the article corresponding to the obtained counterfeiting image is still a counterfeit article, and the counterfeiting image is still a false article image in nature; if the false image is identified, the identification result is still false.
In some embodiments, the countermeasure generation module includes a fraud generator for counterfeiting the genuine article image to obtain a fraud image, and a fraud generator for counterfeiting the counterfeit article image to obtain a fraud image. The step of obtaining the fake image is to utilize the fake generator to fake the true article image to obtain a fake image; the step of obtaining the false image is to utilize the false generator to make the false article image true to obtain the false image.
Generally speaking, the more layers of the network in deep learning is more effective, i.e. the deeper the network is, the higher the training and testing accuracy is. However, in practical application, because of the problem of gradient disappearance, when the number of layers of the network is increased to a certain number, the accuracy of training and testing is reduced. To address the problem of gradient vanishing while increasing the depth of the network, thereby further increasing the accuracy of the item authentication model, in some embodiments the fraud generator or the fraud generator comprises a depth residual network. In some embodiments, a deep residual network of a plurality of residual blocks, which may be 8, 10, 18, 34, etc., connected together, may be used as a true generator and a false generator. The number of residual blocks described in the embodiments of the present application is merely an example, and does not limit the scope thereof.
For neural network models, complex models tend to be over-fit. Overfitting means that the model works much better for the training sample set as input than for the test sample set. Therefore, the loss function of the model needs to be balanced, so that the model has better generalization capability and the robustness of the model is improved.
Thus, in some embodiments, the authenticity generator is further configured to authenticate the genuine article image to obtain an authentication image that is genuine; truing the truing image to obtain a true reset image; and calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image so as to keep the loss function value of the countermeasure generation network within a desired range.
Similarly, in some embodiments, the counterfeiting generator is further configured to counterfeit the counterfeit item image to obtain a false; faking the faking image to obtain a false reset image; and calculating an image similarity of the fake article image and the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
And inputting the real article image in the training sample image to the counterfeiting generator for counterfeiting, and generating a counterfeiting image by the counterfeiting generator. Inputting the counterfeit image to the authenticity generator, the authenticity generator generating a true reset image. The fake image is true and the reset image is a true reset image because the image generated by the true object image after being subjected to the fake is still true. And inputting the true article image in the training sample image into the false positive generator for false positive, wherein the false positive generator generates the authentication image. Similarly, the image generated after the genuine article image is genuine, and therefore the generated authentication image is a genuine authentication image.
The calculation of the image similarity includes but is not limited to: calculating the euclidean distance, absolute value distance, histogram intersection, mahalanobis distance, manhattan distance, chebyshev distance, minkowski distance, normalized euclidean distance, included angle cosine, center moment, etc. of the true article image and the true authentication image and/or the true reset image. In the embodiments provided in the present application, the euclidean distance is calculated as an example for the first time. When the similarity of the images is calculated by utilizing the Euclidean distance, the Euclidean distance refers to the actual distance between two points in the multi-dimensional space, and the smaller the Euclidean distance is, the greater the similarity of the two images is.
There are a number of ways to calculate the euclidean distance of the genuine article image from the genuine authentication image and/or the genuine reset image, including but not limited to: for example, a first euclidean distance between the true article image and a true authentication image and a second euclidean distance between the true article image and a true reset image are respectively calculated, an average value or a root mean square of the first euclidean distance and the second euclidean distance is taken, so as to obtain a target euclidean distance, and whether a loss function value of the countermeasure generation network is kept within an expected range is judged according to the target euclidean distance; for another example, a first euclidean distance between the true article image and the true authentication image and a second euclidean distance between the true article image and the true reset image are respectively calculated, a larger value of the first euclidean distance and the second euclidean distance is taken, the euclidean distance corresponding to the larger value is taken as a target euclidean distance, and whether the loss function value of the countermeasure generation network is kept within an expected range is determined according to the target euclidean distance.
It should be appreciated that if the training of the true and false generators is not constrained, the true and false generators randomly do either true or false, and that there is a high probability that either over-true or over-false will occur. For example, if the genuine article image contains a luxury wallet, the corresponding authentication point location includes a label, a brand mark, etc., and if over-counterfeiting or over-counterfeiting occurs, it is likely that the label in the generated counterfeit image will be image-processed into another article that is not related to the label, or even to the wallet. In this case, it is meaningless to train the item authentication model with an image that is generated by counterfeiting, regardless of the item to be authenticated, and that has been missing the necessary attributes or features in the training sample image. Thus, during the training process, the image similarity of the fake item image to the fake authentication image and/or the fake reset image is calculated to keep the loss function value of the countermeasure generating network within a desired range. The loss function is used for estimating the degree of inconsistency (or deviation degree) between the predicted value and the true value of the model, and the smaller the value of the loss function is, the better the robustness of the model is. The loss function values include L1 norm loss functions, cross entropy loss functions, log logarithmic loss functions, squared loss functions, exponential loss functions, Hinge loss functions, absolute value loss functions, and the like. After the target Euclidean distance is obtained through calculation, the loss function value of the countermeasure generation network is obtained according to the target Euclidean distance, and the calculated loss function value is compared with a preset value or a preset range to judge whether the loss function value of the countermeasure generation network is kept in an expected range or not, so that the false generator and the false generator are prevented from being over-true or false.
In some embodiments, the confusion discrimination module comprises a fake article discriminator and a real article discriminator, the fake article discriminator is configured to learn the input fake image to output a result that the fake image is true, and the real article discriminator is configured to learn the input fake image to output a result that the fake image is false.
And inputting a fake image obtained after the fake object image is fake and a fake image obtained after the fake object image is fake into a confusion judging network, judging and learning the fake image and the fake image by the confusion judging network, and outputting a result that the fake image is true and a result that the fake image is false.
In some embodiments, the confusion discrimination module includes training and optimizing the confusion discrimination module using the structure of the siemese network. For example, two true and false images containing the skin tag are used as input, and besides the identification results of the two images, the same-attribute prediction result or different-attribute prediction result is output to indicate whether the two images belong to the same class (both true or both false) or whether the two images belong to one true or one false. Thereby, the discrimination capability of the article discrimination model can be further enhanced.
As mentioned above, no matter how the true article image is faked, the article corresponding to the obtained faked image is still a true article. Therefore, the result of the discrimination and output of the aliasing discrimination module on the spurious images is still true. Similarly, no matter how the false article image is subjected to the counterfeiting, the article corresponding to the obtained counterfeiting image is still a counterfeit article. Therefore, the result of the discrimination and output of the false image by the confusion discrimination module is still that the false image is false.
In some embodiments, the false article discriminator is further configured to learn the input true article image to output the true article image as a result of predicting true; and learning the input false article image to output the false article image as a result of predicting false.
Inputting the fake images marked as true into the fake article discriminator in the training process, and outputting true results by the fake article discriminator; inputting the true article image marked as true into the false article discriminator, wherein the false article discriminator outputs a result of predicting true; inputting the false article image marked as false into the false article discriminator, and outputting a result of predicting false by the false article discriminator. Obviously, the result of predicting true should indeed be true, and the result of predicting false should indeed be false.
Similar to training the fake generator and the fake generator, when training the fake article discriminator and the real article discriminator, the training process also needs to be constrained, so that the fake article discriminator and the real article discriminator improve the discrimination accuracy and keep the training convergence stable. Therefore, in some embodiments, the method further comprises comparing the result of the true prediction with the image of the true item, calculating a distance between the result of the true prediction and the image of the true item, and determining whether the distance is smaller than a preset value. Similarly, in some embodiments, the method further comprises comparing the result of predicting the false with a false article image, and determining whether the distance is smaller than a preset value by calculating the distance between the result of predicting the false and the false article image. In some embodiments, the method further comprises comparing the true result with a fake image, and determining whether the distance is less than a predetermined value by calculating the distance between the true result and the fake image.
In a specific embodiment, the true result, the false result, the result predicted to be true, and the result predicted to be false are set to a real number within a range of [0,1 ]. Wherein the "0" indicates a flag corresponding to false and the "1" indicates a flag corresponding to true. As mentioned earlier, before being input to the false article discriminator, the false, real and false article images have been marked by means of manual or technical marking, and obviously, the false and real article images are marked as "1" to represent that the image is true, and the false article image is marked as "0" to represent that the image is false. Under the above conditions, the output result of the fake article discriminator or the real article discriminator is also a real number in the range of [0,1 ]. The preset value is set, for example, to 0.5. Obviously, if the result output by the false article discriminator or the true article discriminator is less than 0.5, the input image is false, otherwise, the input image is true. When a fake image is used as input, the true result output by the fake article discriminator is compared with the mark value of the fake image, and a first distance between the two is calculated. Similarly, a second distance between the result of predicting true and the tag value of a true item image is calculated, and a third distance between the result of predicting false and the tag value of a false item image is calculated. And obtaining the distance sum of the first distance, the second distance and the third distance through calculation, and minimizing the distance sum through optimization algorithms such as Adam and SGD, so as to train the fake article discriminator.
Similarly, in some embodiments, the genuine article discriminator is further configured to learn the input genuine article image to output the genuine article image as a result of predicting genuine; and learning the input false article image to output the false article image as a result of predicting false. The steps are similar to those described above, and the principle and flow refer to the above embodiments, which are not described herein again. The confusion judging module is also used for inputting the true article image into the confusion judging network to learn so as to output the true article image as a result of predicting true; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false; inputting an unknown true and false image into the confusion discrimination network learning to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result or output the unknown true and false image and the true object image or the false object image as different attribute prediction results; and obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction result and the true prediction result or the false prediction result.
For example, the true article image is input to the confusion judging module for judging and learning, and since the result that the true article image is true is a priori result (that is, the true article image is a training sample image determined to be true in advance), the result output by the confusion judging module is a result of predicting true, which indicates that the true article image is true. Then, an unknown true and false image is input to the confusion judging module for judging and learning. Because the unknown true and false image only contains true or false results, when the unknown true and false image is a true article image, the prediction result has the same attribute as that of the true article image, namely, the prediction result is true; when the unknown true and false image is a false article image, the attribute of the prediction result of the unknown true and false image is different from that of the true article image, namely true and false. Therefore, after the confusion judging module judges and learns the unknown true and false images, the unknown true and false images and the true object images are predicted results with the same attribute or predicted results with different attributes. Because the truth of the true article image is known, the unknown true and false image can be judged to be true or false according to the fact that the output result is the same attribute prediction result or different attribute prediction result.
Similarly, the false article image is input into the confusion judging module for judging and learning, and the result output by the confusion judging module is a result of predicting false, which indicates that the false article image is false. Then, inputting an unknown true and false image into the confusion judging module for judging and learning, and in the same way, when the unknown true and false image is a false article image, the attribute of the prediction result of the unknown true and false image is the same as that of the prediction result of the false article image, namely, the unknown true and false image is false; when the unknown true and false image is a true article image, the prediction result of the unknown true and false image has different attributes from the prediction result of the false article image, namely true and false. Since the principle is similar to the above embodiment, please refer to the above embodiment for the specific processes and steps, which are not described herein again.
In the embodiment of the application, the countermeasure generation module and the confusion discrimination module are alternately trained, and the false generator in the countermeasure generation module learn how to mutually convert the true article image and the false article image, so as to generate the image which is false and true to confuse the confusion discrimination module. And the true article discriminator and the false article discriminator in the confusion discrimination module continuously learn whether the input image is true or false, so as to avoid confusion of the confrontation generation module and improve the discrimination capability. Through the training mode of the cyclic iteration of the countermeasure generation module and the confusion discrimination module, the countermeasure generation module and the confusion discrimination module can be continuously perfected and optimized in the training process, so that the discrimination capability of the article discrimination model is improved.
In some embodiments, the training system of the item identification model may further include an optimization module, where the optimization module optimizes the neural network by using an optimization algorithm, such as an adam (adaptive motion estimation) optimization algorithm, an sgd (stored Gradient decision) optimization algorithm, or an RMSprop optimization algorithm, and iteratively updates weights of the neural network based on training data, so as to minimize a cross-entropy loss function of the neural network, make convergence more stable, and avoid the neural network from falling into local optimization, thereby enhancing robustness of the item identification model.
It should be noted that fig. 10 is only one possible implementation manner of the embodiment of the present application, and in practical applications, the article authentication apparatus may further include more or less components, which is not limited herein. For the content that is not shown or described in the embodiments of the present application, reference may be made to the relevant explanation in the foregoing embodiments, which are not described herein again.
The training system of the article identification model, provided by the application, is used for constructing the article identification model based on the structure of the cyclic countermeasure generation network, combining the countermeasure generation network and the confusion discrimination network, utilizing the true generator and the fake generator of the countermeasure generation network to true or fake the training sample image, and inputting the image after true or fake into the confusion discrimination network, so that the true article discriminator and the fake article discriminator of the confusion discrimination network are trained, and the discrimination capability of the confusion discrimination network can be improved. Meanwhile, the mode of alternately and iteratively training the countermeasure generating network and the confusion judging network is adopted, so that the countermeasure generating network and the confusion judging network can be trained and learned together, the robustness and the accuracy of the countermeasure generating network and the confusion judging network are improved, and the identification capability and the robustness of the article identification model are further effectively improved.
Referring to fig. 11, a schematic flow chart of an embodiment of the article authentication method of the present application is shown, wherein the article authentication method includes the following steps:
step S21, acquiring the shot image of the article to be identified; the image includes at least one target authentication point.
The article to be authenticated includes a luxury bag with a leather label or an article with handwriting. For example, identifying a bag with a label, such as a backpack, rucksack, handbag, messenger bag, travel bag, etc. of a certain brand; such as bracelets, rings, watches, pocket watches, contracts, etc. that have writing. In the examples provided in this application, the description is made for the time being taking the article to be authenticated as a luxury bag as an example. It should be understood that the location of the target authentication point of an item to be authenticated may vary from item to item, such as for ladies' bags for luxury items, where the target authentication point typically includes a leather label, a zipper, and a label that includes the trademark of the item; for another example, in the case of a wristwatch, the target identification point location is typically on a wristband, a dial, or the like.
In some embodiments, obtaining a captured image of an item to be authenticated includes, but is not limited to: the method comprises the steps of shooting an article to be identified by using a camera of an intelligent terminal such as a mobile phone, a tablet personal computer and a notebook computer, and taking a shot image as an image of the article to be identified. Ways to locate the target authentication point location include, but are not limited to: 1) and (5) manually calibrating. For example, an image containing a luxury bag (corresponding to an article to be authenticated) is taken using a mobile phone, and a range of a certain size, which contains a label (corresponding to a target authentication point) of the luxury bag, is drawn on the image as a target authentication point position. For another example, after acquiring a plurality of training sample images, all the images are manually identified and calibrated one by one. 2) The identified point is located by a recursive convolutional neural network. For example, using a fast recursive convolutional neural network, a candidate region where a position of a discrimination point may exist is first found in each image to obtain a multidimensional feature of the image, and then a sub-network such as target detection discrimination, search box position regression, etc. is used to obtain an accurate position of the target discrimination point.
In some embodiments, locating the target authentication point may further include extracting characteristic information using a deep convolutional network. The deep convolutional network includes Xception (deep separable convolutional network), ResNet (deep residual error network), MobileNet, and the like. For example, for an image of a luxury bag that identifies a point location at a picotag location, a depth separable convolutional network is used to extract characteristic information such as font, stroke, texture, relative position and size from the picotag portion of the image. The depth separable convolution network respectively convolves each channel of the image and decouples the spatial position of the characteristic information in the image and the convolution among the channels, so that the capability of extracting the characteristic information is improved, and the network learning and optimization are facilitated.
Step S23, the image including the target identification point is identified by the article identification model obtained by the training method of the article identification model.
The image including the target authentication point is input into an article authentication model, which authenticates the image. The article identification model is obtained by training the article identification model training method.
And step S23, outputting the result that the article to be authenticated is true or the result that the article to be authenticated is false.
The article identification model identifies an input image containing a target identification point, and if the image is identified to be true, the article identification model outputs a result that the article to be identified is true; and if the image is identified to be false, the article identification model outputs the result that the article to be identified is false.
According to the article identification method, the shot image containing the article to be identified is identified through the pre-trained article identification model, so that the difference of identification results caused by different identification abilities of different identifiers in a manual identification mode is avoided, and the identification stability and accuracy are high. Meanwhile, on one hand, by using the article identification method provided by the application, a user can receive the returned identification result only by taking a picture, so that the economic cost and the time cost spent on manual identification are greatly reduced, and the requirement of the user for identifying the authenticity of the article in real time in daily life is really met. On the other hand, the article identification method provided by the application can effectively stop circulation of counterfeit articles and improve shopping experience of users.
Referring to fig. 12, which is a schematic structural diagram of an article authentication apparatus in an embodiment of the present application, as shown in the figure, the article authentication apparatus 110 includes a camera 111, a memory 112, and one or more processors 113, and the camera 111, the memory 112, and the processors 113 may be connected by a bus or in other manners, and the embodiment of the present application takes the example of connection by a bus as an example. The shooting device 111 is used for acquiring a shot image of an article to be authenticated; the image includes at least one target authentication point.
The shooting device 111 comprises an intelligent terminal and other devices or equipment with cameras, and is used for acquiring the shot images of the articles to be authenticated; the image includes at least one target authentication point. Such as portable or wearable electronic devices including, but not limited to, smart phones, tablets, smart watches, smart glasses, Personal Digital Assistants (PDAs), etc., it is to be understood that the portable electronic device described in the embodiments of the present application is but one example of an application and that the components of the device may have more or fewer components than shown, or a different configuration of components. The various components of the depicted figures may be implemented in hardware, software, or a combination of hardware and software, including one or more signal processing and/or application specific integrated circuits.
The article to be authenticated includes a luxury bag with a leather label or an article with handwriting. For example, identifying a bag with a label, such as a backpack, rucksack, handbag, messenger bag, travel bag, etc. of a certain brand; such as bracelets, rings, watches, pocket watches, contracts, etc. that have writing. The location of the authentication point of the article to be authenticated may vary from article to article, for example in ladies' bags for luxury goods, the authentication point typically comprises a leather label, a zipper, and a label comprising the trademark of the article; for another example, in the case of a wristwatch, the identified point location is typically on the wristband, face, etc. In some embodiments, obtaining a captured image of an item to be authenticated includes, but is not limited to: the method comprises the steps of shooting an article to be identified by using a camera of an intelligent terminal such as a mobile phone, a tablet personal computer and a notebook computer, and taking a shot image as an image of the article to be identified. The image format is a format of a computer storage image, such as a storage format of bmp, jpg, png, tiff, gif, pcx, tga, exif, fpx, svg, psd, cdr, pcd, dxf, ufo, eps, ai, raw, WMF, and the like.
The Memory may include Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read Only Memory (PROM), Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), and the like. The memory is used for storing a program, and the processor executes the program after receiving the execution instruction.
The processor comprises an integrated circuit chip having signal processing capabilities; or a general-purpose processor, such as a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), discrete gate or transistor logic, discrete hardware components, may implement or perform the methods, steps, and logic blocks disclosed in the embodiments of the present application. The general purpose processor may be a microprocessor, or any conventional processor such as a Central Processing Unit (CPU). The processor 113 is configured to call the program code stored in the memory 112 to execute the aforementioned item authentication method: acquiring a plurality of training sample images of an article to be identified, wherein the training sample images comprise a true article image and a false article image; utilizing a countermeasure generating network to counterfeit the true article image to obtain a counterfeit image; and using the countermeasure generating network to simulate the fake article image to obtain a simulated image; inputting the fake image into a confusion discrimination network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
The article identification device of the present application is used for executing the article identification method, and the principle and the specific flow refer to the above embodiments, which are not described herein again. It should be noted that fig. 12 is only one possible implementation manner of the embodiment of the present application, and in practical applications, the article authentication apparatus may further include more or less components, which is not limited herein. For the content that is not shown or described in the embodiments of the present application, reference may be made to the relevant explanation in the foregoing embodiments, which are not described herein again.
According to the article identification device, the user can receive the returned identification result only by taking the picture, so that the economic cost and the time cost of manual identification are greatly reduced, the requirement of the user for identifying the authenticity of the article in daily life is really met, meanwhile, the circulation of counterfeit articles can be effectively stopped, and the shopping experience of the user is improved.
The article authentication method provided by the application can also be executed by an article authentication client. The article authentication client is loaded in an intelligent device and is realized through software and hardware in the intelligent device. The loading form of the article authentication client on the intelligent device comprises an APP application program or an applet program (such as an applet program on a WeChat).
Referring to fig. 13, which is a schematic structural diagram of an article authentication client according to an embodiment of the present application, as shown, the article authentication client 120 loaded in an intelligent device includes an input module 121, a processing module 122, and a display module 123. Wherein: the input module 121 is configured to call a shooting device of the smart device to obtain a shot image of an object to be authenticated when receiving an authentication instruction input by a user; the image comprises at least one target authentication point; the processing module 122 invokes program code stored in the smart device to perform the aforementioned item authentication method; the display module 123 displays and outputs a result that the article to be authenticated is true or a result that the article to be authenticated is false.
Such as portable or wearable electronic devices including, but not limited to, smartphones, tablets, smartwatches, smart glasses, Personal Digital Assistants (PDAs), etc., it is to be understood that the portable or wearable electronic devices described in the embodiments of the present application are but one example of an application and that the components of the device may have more or fewer components than shown, or a different configuration of components. The various components of the depicted figures may be implemented in hardware, software, or a combination of hardware and software, including one or more signal processing and/or application specific integrated circuits.
The article includes a luxury bag with a leather label or an article with handwriting, the target authentication point being a location of the leather label or handwriting. For example, identifying a bag with a label, such as a backpack, rucksack, handbag, messenger bag, travel bag, etc. of a certain brand; such as bracelets, rings, watches, pocket watches, contracts, etc. that have writing. The location of the target authentication point of an item may vary from item to item, for example in ladies' bags for luxury items, where the target authentication point typically includes a leather label, a zipper, and a label that includes the trademark of the item; for another example, for a watch, the target identification point is usually located on a watchband, a dial, or the like; for example, the position of the target authentication point is usually the position of writing handwriting or a seal, etc., for a printed matter such as a contract. In the examples provided in this application, the description is made for the time being taking the article to be authenticated as a luxury bag as an example.
The item authentication client 120 obtains an authentication instruction input by a user through an input module 121 of the item authentication client, where the authentication instruction is used to instruct the item authentication client to call a shooting device of the smart device to obtain a shot image of an item to be authenticated. The manner of obtaining the identification instruction includes, but is not limited to: and monitoring text data input, touch data input, voice data input and the like of the intelligent equipment to obtain the identification instruction.
For example, a user inputs a specific character or character string on a display interface of the intelligent device, the character or character string is associated with an authentication instruction in advance, and when the input module monitors that the character or character string is input, the input module acquires the character or character string, so that a corresponding authentication instruction is acquired according to a preset association relationship. For another example, when a user performs touch operations such as clicking, sliding, dragging a virtual slider, gesture operations and the like on a display interface of the intelligent device, the touch operations are associated with the identification instruction in advance; the input module monitors the touch operation, and accordingly obtains a corresponding identification instruction according to a preset incidence relation. For another example, the input module monitors voice data of a user, and acquires a corresponding authentication instruction according to a preset association relationship between the voice data and the authentication instruction. After the identification instruction is obtained, the input module calls the image of the object to be identified, which is shot by the user through the shooting device of the intelligent equipment, and sends the image to the processing module.
In some embodiments, the input module further includes a selection unit, configured to, when receiving an authentication instruction input by a user, invoke an image library of the smart device for the user to select an image of an article to be authenticated; the image includes at least one target authentication point. After the input module acquires the identification instruction, the selection unit calls the image library to display all or part of images in the image library to a user for the user to select. The input module takes the image selected by the user as an image of the article to be authenticated and sends the image to the processing module. The image library is stored in a storage medium, which may be a storage device of a service end, a storage device of the smart device, or other storage media such as an SD card, a flash, and the like.
The target authentication point may be determined in advance by a manual marking or a technical marking before the input module calls the image, or may be determined by a manual marking or a technical marking after the input module calls the image. For the principle of performing manual marking or technical marking in advance before the input module calls the image, please refer to the embodiment corresponding to fig. 1, which is not described herein again. In some embodiments, the input module may further prompt a user to define one or more candidate regions on the image on a display interface of the smart device after calling the image, where the candidate regions include the target authentication point, so as to obtain an image of the to-be-authenticated item including the location of the authentication point.
The processing module 122 is connected to the input module 121, and is configured to call a program code stored in the smart device to execute the aforementioned item authentication method after receiving the image of the item to be authenticated sent by the input module 121. The method comprises the following steps: acquiring a shot image of an article to be authenticated; the image comprises at least one target authentication point; identifying the image containing the target identification point by using the article identification model obtained by the article identification model training method; and outputting the result that the object to be authenticated is true or the result that the object to be authenticated is false.
In some embodiments, the user may also directly shoot the position of the authentication point of the to-be-authenticated object (for example, a leather label, a brand mark, hardware, and the like) by using the shooting device of the smart device, and respectively obtain images such as a leather label image, a trademark image, a hardware image, and the like for the processing device to execute the object authentication method.
The display module 123 is connected to the processing module 122, and is configured to receive a result that the to-be-authenticated item is true or the to-be-authenticated item is false, which is sent by the processing module, and display the result on the display interface of the smart device.
In some embodiments, the identification result of the to-be-identified item may further be an image similarity or an image matching degree of the to-be-identified item and a genuine item, which is displayed in percentage form by the display module. For example, the processing module identifies the image and then sends the identification result to the display module, and the display module displays that the image matching degree is 95% on the display interface of the intelligent device, which indicates that the object to be identified is very likely to be a genuine object; or, if the display module displays that the image matching degree is 23% on the display interface of the intelligent device, it indicates that the article to be authenticated is likely to be a counterfeit article. It should be understood that the specific values of 95% and 23% are for illustration only and do not limit the range of values for outputting the image similarity or image matching.
To sum up, the article identification client obtains an input operation performed by a user on the intelligent device through an input module, obtains an identification instruction corresponding to the input operation according to a preset incidence relation between the input operation and the identification instruction, then the input module sends the identification instruction to a processing module, and the processing module executes an article identification method through an article identification model trained in advance based on a training method of the article identification model, so as to identify an image of the article to be identified, identify the authenticity of the image, and further identify whether the article to be identified is authentic or false. In an actual application scenario, a user may open an APP on a mobile phone (corresponding to a smart device) or open a WeChat applet, thereby starting an article authentication client to authenticate an article.
The utility model provides an article authentication client, through loading in the article authentication client of a smart machine, the user only need shoot the photo with the help of the shooting device of a smart machine, and pass through the article authentication client is appraised, receives the authentication result that returns immediately, has greatly reduced the economic cost and the time cost of artifical authentication cost, really satisfies the demand that the user immediately distinguished article true and false in daily life, and user experience is good. Meanwhile, the article identification client identifies the shot image containing the article to be identified through a pre-trained article identification model, so that the difference of identification results caused by different identification abilities of different identifiers in a manual identification mode is avoided, and the identification stability and accuracy are high.
The present application also provides a computer-readable and writable storage medium storing a computer program of a training method of an item authentication model, which when executed implements the above-described embodiment with respect to the training method of an item authentication model described in fig. 1 to 9.
The present application also provides a computer-readable and writable storage medium storing a computer program of an article authentication method that, when executed, implements the article authentication method described above with respect to fig. 11 of the embodiments.
The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application or portions thereof that substantially contribute to the prior art may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present application.
In the embodiments provided herein, the computer-readable and writable storage medium may include read-only memory, random-access memory, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory, a USB flash drive, a removable hard disk, or any other medium that can be used to store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer-readable medium. For example, if the instructions are transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. It should be understood, however, that computer-readable-writable storage media and data storage media do not include connections, carrier waves, signals, or other transitory media, but are intended to be non-transitory, tangible storage media. Disk and disc, as used in this application, includes Compact Disc (CD), laser disc, optical disc, Digital Versatile Disc (DVD), floppy disk and blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers.
In one or more exemplary aspects, the functions described in the training method of the item authentication model and the computer program of the item authentication method described herein may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. The steps of a method or algorithm disclosed herein may be embodied in a processor-executable software module, which may be located on a tangible, non-transitory computer-readable and/or writable storage medium. Tangible, non-transitory computer readable and writable storage media may be any available media that can be accessed by a computer.
The flowcharts and block diagrams in the figures described above of the present application illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The training method of the article identification model, the training system of the article identification model, the article identification method, the article identification device, the article identification client and the computer readable storage medium provided by the application combine the countermeasure generation network and the confusion discrimination network, utilize the true generator and the false generator of the countermeasure generation network to make true or false for the training sample image, and input the image after making true or false to the confusion discrimination network, thereby training the true article discriminator and the false article discriminator of the confusion discrimination network, and being capable of improving the discrimination capability of the confusion discrimination network. Meanwhile, the mode of alternately and iteratively training the countermeasure generating network and the confusion judging network is adopted, so that the countermeasure generating network and the confusion judging network can be trained and learned together, the robustness and the accuracy of the countermeasure generating network and the confusion judging network are improved, and the identification capability and the robustness of the article identification model are further effectively improved. The article identification model obtained by training through the article identification model training method provided by the application is used for identifying the article, so that the accuracy rate is high, and the stability is strong.
The above embodiments are merely illustrative of the principles and utilities of the present application and are not intended to limit the application. Any person skilled in the art can modify or change the above-described embodiments without departing from the spirit and scope of the present application. Accordingly, it is intended that all equivalent modifications or changes which can be made by those skilled in the art without departing from the spirit and technical concepts disclosed in the present application shall be covered by the claims of the present application.
Claims (35)
1. A method for training an item authentication model, comprising the steps of:
acquiring a plurality of training sample images of an article to be identified, wherein the training sample images comprise a true article image and a false article image;
utilizing a countermeasure generating network to counterfeit the true article image to obtain a counterfeit image; and using the countermeasure generating network to simulate the fake article image to obtain a simulated image;
inputting the fake image into a confusion discrimination network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
2. A method for training an item authentication model according to claim 1, further comprising the step of locating the position of the authentication point in the image of the true item and the image of the false item to extract characteristic information.
3. A method for training an article authentication model according to claim 2, wherein the step of locating the position of the authentication point in the image of the true article and the image of the false article to extract the characteristic information comprises: and positioning the positions of the appraisal points in the real article image and the false article image by using a fast recursive convolutional neural network, and extracting characteristic information by using a depth separable convolutional network.
4. The method for training an item authentication model according to claim 1, wherein the countermeasure generation network includes a forgery generator and a forgery generator, wherein: the step of obtaining the fake image is to utilize the fake generator to fake the true article image to obtain a fake image; the step of obtaining the false image is to utilize the false generator to make the false article image true to obtain the false image.
5. The method of claim 4, wherein the counterfeit generator or the genuine generator comprises a deep residual network.
6. The method for training the item authentication model according to claim 4, further comprising the step of training the challenge generation network:
truing the genuine article image with the truing generator to obtain an authentication image that is genuine;
truing the false image with the truing generator to obtain a reset image that is true;
calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image to maintain a loss function value of the countermeasure generation network within a desired range.
7. The method for training the item authentication model according to claim 4, further comprising the step of training the challenge generation network:
counterfeiting the fake item image with the counterfeiting generator to obtain an authentication image that is fake;
faking the authenticity image with the faking generator to obtain a reset image that is false;
calculating an image similarity of the fake item image to the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
8. The training method of an item authentication model according to claim 1, wherein the confusion discrimination network includes a false item discriminator and a true item discriminator, wherein the false image is input to the false item discriminator to be learned to output a result that the false image is true; and inputting the false image to the genuine article discriminator to learn to output a result that the false image is false.
9. The method for training the item identification model according to claim 8, further comprising the step of training the confusion discrimination network:
inputting the counterfeit image to the counterfeit article discriminator to learn to output a result that the counterfeit image is true;
inputting the true article image to the fake article discriminator to learn so as to output the true article image as a result of predicting true;
inputting the false article image to the false article discriminator to learn to output a result that the false article image is predicted false.
10. The method for training the item identification model according to claim 8, further comprising the step of training the confusion discrimination network:
inputting the false image to the genuine article discriminator to learn so as to output a result that the false image is false;
inputting the genuine article image to the genuine article discriminator to learn to output the genuine article image as a result of predicting genuine;
inputting the false article image to the true article discriminator to learn to output the false article image as a result of predicting false.
11. The method for training the item identification model according to claim 8, further comprising the step of training the confusion discrimination network:
inputting the true article image into the confusion discrimination network learning to output the true article image as a result of predicting true; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false;
inputting an unknown true and false image into the confusion discrimination network learning to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result or output the unknown true and false image and the true object image or the false object image as different attribute prediction results;
and obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction results and the true prediction result or the false prediction result.
12. The method of claim 8, wherein the confusion discrimination network comprises a siense network.
13. A training method of an article authentication model according to claim 1, wherein the article includes a luxury bag with a leather label or an article with handwriting.
14. A training system for an item authentication model, comprising:
the system comprises a sample input module, a verification module and a verification module, wherein the sample input module is used for acquiring a plurality of training sample images of an article to be identified, and the training sample images comprise a real article image and a false article image;
the countermeasure generation module is used for counterfeiting the true article image by utilizing a countermeasure generation network to obtain a counterfeit image; and using the countermeasure generating network to simulate the fake article image to obtain a simulated image;
the confusion judging module is used for inputting the fake image into a confusion judging network to learn so as to output the result that the fake image is true; and inputting the false image into the confusion discrimination network learning to output the result that the false image is false.
15. A system for training an article authentication model according to claim 14, further comprising a feature extraction unit for locating the positions of the authentication points in the images of the true article and the false article to extract characteristic information.
16. The training system of the item authentication model according to claim 15, wherein the feature extraction unit locates the positions of the authentication points in the true item image and the false item image using a fast recursive convolutional neural network, and extracts the characteristic information using a depth separable convolutional network.
17. The training system of the item authentication model of claim 14, wherein the challenge generation module comprises:
the counterfeiting generator is used for counterfeiting the true article image to obtain a counterfeiting image;
and the false article image is false, so that a false image is obtained.
18. The training system for an item authentication model of claim 17, wherein the fraud generator or the fraud generator comprises a deep residual network.
19. A training system for an item authentication model according to claim 17, wherein the authenticity generator is further configured to authenticate the authentic item image to obtain an authentic authentication image; truing the truing image to obtain a true reset image; and calculating an image similarity of the genuine article image and the genuine authentication image and/or the genuine reset image so as to keep the loss function value of the countermeasure generation network within a desired range.
20. The training system of the item authentication model of claim 17, wherein the counterfeiting generator is further configured to counterfeit the counterfeit item image to obtain a true; faking the faking image to obtain a false reset image; and calculating an image similarity of the fake article image and the fake authentication image and/or the fake reset image to maintain the loss function value of the countermeasure generation network within an expected range.
21. The system for training an item authentication model according to claim 14, wherein the confusion discrimination module comprises:
a fake article discriminator for learning an input fake image to output a result that the fake image is true;
and the genuine article discriminator is used for learning the input genuine image so as to output a result that the genuine image is false.
22. The system for training an item identification model according to claim 21, wherein the false item discriminator is further configured to learn the input true item image to output the true item image as a result of predicting true; and learning the input false article image to output the false article image as a result of predicting false.
23. The system for training an article authentication model according to claim 21, wherein the genuine article discriminator is further configured to learn the input genuine article image to output the genuine article image as a result of predicting genuine; and learning the input false article image to output the false article image as a result of predicting false.
24. The system for training an item identification model according to claim 14, wherein the confusion discrimination module is further configured to input the true item image to the confusion discrimination network learning to output the true item image as a result of predicting true; or inputting the false article image into the confusion judgment network to learn so as to output the false article image as a result of predicting false; inputting an unknown true and false image into the confusion discrimination network learning to output the unknown true and false image and the true object image or the false object image as the same attribute prediction result or output the unknown true and false image and the true object image or the false object image as different attribute prediction results; and obtaining the identification result of the unknown true and false image according to the same attribute prediction result or different attribute prediction result and the true prediction result or the false prediction result.
25. The system for training an item authentication model of claim 24, wherein the confusion discrimination module comprises a siense network.
26. A training system for an article authentication model according to claim 14, wherein the article includes a luxury bag with a leather label or an article with handwriting.
27. A method of authenticating an article, comprising the steps of:
acquiring a shot image of an article to be authenticated; the image comprises at least one target authentication point;
authenticating an image containing the target authentication point using an article authentication model obtained by the method of training an article authentication model according to any one of claims 1 to 13;
and outputting the result that the object to be authenticated is true or the result that the object to be authenticated is false.
28. The item authentication method as recited in claim 27, wherein the item comprises a luxury bag with a leather tag or an item with handwriting, and the target authentication point is a position of the leather tag or handwriting.
29. An article authentication apparatus, comprising:
the shooting device is used for acquiring a shot image of the article to be authenticated; the image comprises at least one target authentication point;
a memory for storing program code;
one or more processors for invoking program code stored in the memory to perform the item authentication method of any one of claims 27-28.
30. The article authentication device according to claim 29, wherein said article comprises a luxury bag with a leather tag or an article with handwriting, said target authentication point being the position of said leather tag or handwriting.
31. An article authentication client, loaded in a smart device, comprising:
the intelligent equipment comprises an input module, a storage module and a display module, wherein the input module is used for calling a shooting device of the intelligent equipment to acquire a shot image of an article to be authenticated when an authentication instruction input by a user is received; the image comprises at least one target authentication point;
a processing module that invokes program code stored in the smart device to perform the item authentication method of any one of claims 27-28;
and the display module is used for displaying and outputting the result that the article to be authenticated is true or the result that the article to be authenticated is false.
32. The item authentication client according to claim 31, wherein the input module further comprises a selection unit, configured to, upon receiving an authentication instruction input by a user, invoke the image library of the smart device for the user to select an image of an item to be authenticated; the image includes at least one target authentication point.
33. The item authentication client as recited in claim 31, wherein the item comprises a luxury bag with a leather label or an item with handwriting, the target authentication point being a location of the leather label or handwriting.
34. A computer-readable storage medium storing a computer program for item authentication, the computer program, when executed, implementing a method of training an item authentication model according to any one of claims 1 to 13.
35. A computer-readable storage medium storing a computer program for article authentication, wherein the computer program, when executed, implements the article authentication method of any one of claims 27-28.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910404189.3A CN110222728B (en) | 2019-05-15 | 2019-05-15 | Training method and system of article identification model and article identification method and equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910404189.3A CN110222728B (en) | 2019-05-15 | 2019-05-15 | Training method and system of article identification model and article identification method and equipment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110222728A CN110222728A (en) | 2019-09-10 |
CN110222728B true CN110222728B (en) | 2021-03-12 |
Family
ID=67821298
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910404189.3A Active CN110222728B (en) | 2019-05-15 | 2019-05-15 | Training method and system of article identification model and article identification method and equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110222728B (en) |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112651410A (en) * | 2019-09-25 | 2021-04-13 | 图灵深视(南京)科技有限公司 | Training of models for authentication, authentication methods, systems, devices and media |
CN111460195B (en) * | 2020-03-26 | 2023-08-01 | Oppo广东移动通信有限公司 | Picture processing method and device, storage medium and electronic equipment |
CN111553277B (en) * | 2020-04-28 | 2022-04-26 | 电子科技大学 | Chinese signature identification method and terminal introducing consistency constraint |
CN111683364B (en) * | 2020-04-30 | 2023-06-16 | 深圳大学 | Physical layer authentication method and system based on symbol imaginary part in spatial modulation system |
CN112115960A (en) * | 2020-06-15 | 2020-12-22 | 曹辉 | Method and system for identifying collection |
CN111881338A (en) * | 2020-08-03 | 2020-11-03 | 深圳一块互动网络技术有限公司 | Printed matter content retrieval method based on social software light application applet |
CN113658036B (en) * | 2021-08-23 | 2024-05-31 | 平安科技(深圳)有限公司 | Data augmentation method, device, computer and medium based on countermeasure generation network |
CN113902650B (en) * | 2021-12-07 | 2022-04-12 | 南湖实验室 | Remote sensing image sharpening method based on parallel deep learning network architecture |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107506770A (en) * | 2017-08-17 | 2017-12-22 | 湖州师范学院 | Diabetic retinopathy eye-ground photography standard picture generation method |
CN108520285A (en) * | 2018-04-16 | 2018-09-11 | 清华大学 | Article discrimination method, system, equipment and storage medium |
CN109583474A (en) * | 2018-11-01 | 2019-04-05 | 华中科技大学 | A kind of training sample generation method for the processing of industrial big data |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10319076B2 (en) * | 2016-06-16 | 2019-06-11 | Facebook, Inc. | Producing higher-quality samples of natural images |
US10706534B2 (en) * | 2017-07-26 | 2020-07-07 | Scott Anderson Middlebrooks | Method and apparatus for classifying a data point in imaging data |
-
2019
- 2019-05-15 CN CN201910404189.3A patent/CN110222728B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107506770A (en) * | 2017-08-17 | 2017-12-22 | 湖州师范学院 | Diabetic retinopathy eye-ground photography standard picture generation method |
CN108520285A (en) * | 2018-04-16 | 2018-09-11 | 清华大学 | Article discrimination method, system, equipment and storage medium |
CN109583474A (en) * | 2018-11-01 | 2019-04-05 | 华中科技大学 | A kind of training sample generation method for the processing of industrial big data |
Also Published As
Publication number | Publication date |
---|---|
CN110222728A (en) | 2019-09-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110222728B (en) | Training method and system of article identification model and article identification method and equipment | |
CN108520285B (en) | Article authentication method, system, device and storage medium | |
Hu et al. | A novel image steganography method via deep convolutional generative adversarial networks | |
US10474929B2 (en) | Cyclic generative adversarial network for unsupervised cross-domain image generation | |
US20230021661A1 (en) | Forgery detection of face image | |
CN111275784B (en) | Method and device for generating image | |
US20200125836A1 (en) | Training Method for Descreening System, Descreening Method, Device, Apparatus and Medium | |
CN107578459A (en) | Expression is embedded in the method and device of candidates of input method | |
CN108304789A (en) | Recognition algorithms and device | |
CN110428399A (en) | Method, apparatus, equipment and storage medium for detection image | |
US20200218772A1 (en) | Method and apparatus for dynamically identifying a user of an account for posting images | |
CN110232326A (en) | A kind of D object recognition method, device and storage medium | |
CN112818767A (en) | Data set generation method, data set forgery detection device, electronic device, and storage medium | |
KR20120066462A (en) | Method and system for providing face recognition, feature vector extraction apparatus for face recognition | |
CN110489659A (en) | Data matching method and device | |
CN108765532A (en) | Children paint this method for establishing model, reading machine people and storage device | |
CN112651333A (en) | Silence living body detection method and device, terminal equipment and storage medium | |
CN112651410A (en) | Training of models for authentication, authentication methods, systems, devices and media | |
CN109145703A (en) | Intelligent identification method, device, equipment and medium | |
CN112329586A (en) | Client return visit method and device based on emotion recognition and computer equipment | |
US20140050404A1 (en) | Combining Multiple Image Detectors | |
CN110428027A (en) | Fake method, device, equipment and medium are tested in identification based on LCD antifalsification label | |
CN115880347B (en) | Image processing method, electronic device, storage medium, and program product | |
CN115578768A (en) | Training method of image detection network, image detection method and system | |
CN114841340A (en) | Deep forgery algorithm identification method and device, electronic equipment and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |