WO2018166114A1 - 图片识别的方法、***、电子装置及介质 - Google Patents
图片识别的方法、***、电子装置及介质 Download PDFInfo
- Publication number
- WO2018166114A1 WO2018166114A1 PCT/CN2017/091371 CN2017091371W WO2018166114A1 WO 2018166114 A1 WO2018166114 A1 WO 2018166114A1 CN 2017091371 W CN2017091371 W CN 2017091371W WO 2018166114 A1 WO2018166114 A1 WO 2018166114A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- picture
- image
- training
- recognition result
- recognition
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
Definitions
- the present invention relates to the field of image processing technologies, and in particular, to a method, system, electronic device, and computer readable storage medium for picture recognition.
- indecent pictures mainly include violent pictures and erotic images.
- the traditional methods are mainly used to determine whether a picture is an indecent picture. For example, for a violent picture, it is judged according to characteristics of a specific part of the human body and skin color. Whether the picture is a violent picture, and, for example, for a erotic picture, according to the characteristics of the sensitive part of the human body combined with the text to determine whether it is an erotic picture, etc., the method of identifying an indecent picture is generally not accurate and cannot be effective. The identification of some indecent pictures transmitted on the network has caused inconvenience to some network supervision departments or safety supervision departments.
- An object of the present invention is to provide a method, system, electronic device and medium for picture recognition, which aim to improve the accuracy of identifying indecent pictures.
- a first aspect of the present invention provides a method for picture recognition, where the method for picture recognition includes:
- a second aspect of the present application provides a system for picture recognition, where the system for picture recognition includes:
- An identification module configured to: after receiving the image to be identified, input the image to be identified into the recognition model generated by the pre-training, and output the recognition result;
- a determining module configured to determine, according to a relationship between the predetermined recognition result and the picture category, a picture category corresponding to the outputted recognition result, where the recognition result includes a first recognition result, a second recognition result, and a third recognition result,
- the picture category includes a normal picture, a violent picture, and an erotic picture, the first recognition result corresponds to a normal picture, the second recognition result corresponds to a violent picture, and the third recognition result corresponds to a erotic picture;
- a prompting module configured to prompt when the image category is violent or erotic, and mark and record the violent or erotic image to determine the source of the violent or erotic image.
- a third aspect of the present application provides an electronic device, including a processing device, a storage device, and a picture setting system.
- the picture recognition system is stored in the storage device, and includes at least one computer readable instruction, where the at least one computer readable instruction is Executed by the processing device to:
- a fourth aspect of the present application provides a computer readable storage medium having stored thereon at least one computer readable instruction executable by a processing device to:
- the invention has the beneficial effects that the invention recognizes a picture by using a recognition model obtained through a large amount of data training in the field of indecent picture recognition, and since the recognition model can cover various factors affecting indecent picture recognition, therefore, the traditional Compared with the identification method of the ya picture, the invention can effectively identify the indecent picture on the network, greatly improve the accuracy of the indecent picture recognition, and facilitate the execution of the work of some network supervision departments or security supervision departments.
- FIG. 1 is a schematic diagram of an application environment of a preferred embodiment of a method for recognizing a picture according to the present invention
- FIG. 2 is a schematic flow chart of a first embodiment of a method for recognizing a picture according to the present invention
- FIG. 3 is a schematic flow chart of a second embodiment of a method for recognizing a picture according to the present invention.
- step S02 is a schematic diagram of a refinement process of step S02 shown in FIG. 2;
- FIG. 5 is a schematic structural diagram of a first embodiment of a system for recognizing a picture according to the present invention
- FIG. 6 is a schematic structural diagram of a second embodiment of a picture recognition system according to the present invention.
- FIG. 7 is a schematic structural view of the preprocessing module shown in FIG. 6.
- FIG. 1 it is a schematic diagram of an application environment of a preferred embodiment of a method for implementing picture recognition according to the present invention.
- the application environment diagram includes an electronic device 1 and a terminal device 2.
- the electronic device 1 can perform data interaction with the terminal device 2 through a suitable technology such as a network or a near field communication technology.
- the terminal device 2 includes, but is not limited to, any electronic product that can interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device, for example, a personal computer, a tablet computer, a smart phone, or an individual.
- Digital Assistant (PDA) game console, Internet Protocol Television (IPTV), smart wearable device, etc.
- the electronic device 1 is an apparatus capable of automatically performing numerical calculation and/or information processing in accordance with an instruction set or stored in advance.
- the electronic device 1 may be a computer, a single network server, a server group composed of multiple network servers, or a cloud-based cloud composed of a large number of hosts or network servers, where cloud computing is a type of distributed computing, A super virtual computer consisting of a loosely coupled set of computers.
- the electronic device 1 includes, but is not limited to, a storage device 11, a processing device 12, and a network interface 13 that are communicably connected to each other through a system bus. It should be noted that FIG. 1 only shows the electronic device 1 having the components 11-13, but it should be understood that not all illustrated components are required to be implemented, and more or fewer components may be implemented instead.
- the storage device 11 includes a memory and at least one type of readable storage medium.
- the memory provides a cache for the operation of the electronic device 1;
- the readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, a card type memory, or the like.
- the readable storage medium may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1; in other embodiments, the non-volatile storage medium may also be external to the electronic device 1.
- a storage device such as a plug-in hard disk equipped with an electronic device 1, a smart memory card (SMC), a Secure Digital (SD) card, a flash card, or the like.
- SMC smart memory card
- SD Secure Digital
- the readable storage medium of the storage device 11 is generally used to store an operating system installed in the electronic device 1 and various types of application software, such as a program code of the system 10 for picture recognition in an embodiment of the present application. Further, the storage device 11 can also be used to temporarily store various types of data that have been output or are to be output.
- Processing device 12 may, in some embodiments, include one or more microprocessors, microcontrollers, digital processors, and the like.
- the processing device 12 is generally used to control the operation of the electronic device 1, for example, to perform control and processing related to data interaction or communication with the terminal device 2.
- the processing device 12 is configured to run program code or processing data stored in the storage device 11, such as the system 10 that runs the picture processing, and the like.
- the network interface 13 may comprise a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the electronic device 1 and other electronic devices.
- the network interface 13 is mainly used to connect the electronic device 1 with one or more terminal devices 2, and establish a data transmission channel and a communication connection between the electronic device 1 and one or more terminal devices 2.
- the picture recognition system 10 includes at least one computer readable instruction stored in the storage device 11, the at least one computer readable instruction being executable by the processing device 12 to implement the method of picture recognition of various embodiments of the present application. As described later, the at least one computer readable instruction can be classified into different logic modules depending on the functions implemented by its various parts.
- the processing device 12 when the system 10 for image recognition is executed by the processing device 12, the following operations are performed: first, after receiving the picture to be recognized sent by the terminal device 2, inputting the picture to be recognized into the recognition model generated by the pre-training Identifying and outputting the recognition result; then determining the picture category corresponding to the output recognition result according to the association relationship between the predetermined recognition result and the picture category; and finally sending the prompt information when identifying the picture type as a violent picture or a erotic picture
- the terminal device 2 prompts the end user and marks and records the violent picture or the erotic picture to determine the source of the violent picture or the erotic picture.
- the recognition result output by the system 10 for image recognition includes a first recognition result, a second recognition result, and a third recognition result.
- the picture category includes a normal picture, a violent picture, and an erotic picture
- the first recognition result corresponds to a normal picture
- the second identification corresponds to a violent picture
- the third recognition result corresponds to a erotic picture.
- the system 10 for picture recognition is stored in the storage device 11 and includes at least one computer readable instruction stored in the storage device 11, the at least one computer readable instruction being executable by the processing device 12 to implement the present A method of applying picture recognition of each embodiment is applied.
- the at least one computer readable instruction can be classified into different logic modules depending on the functions implemented by its various parts.
- FIG. 2 is a schematic flowchart of a preferred embodiment of a method for recognizing a picture according to the present invention.
- the method for recognizing a picture in this embodiment is not limited to the steps shown in the process, and in addition, in the steps shown in the flowchart, These steps can be omitted and the order between the steps can be changed.
- the method of image recognition includes the following steps:
- Step S1 After receiving the picture to be identified, input the picture to be identified into the recognition model generated by the pre-training for recognition, and output the recognition result.
- the picture may be identified by a system for picture recognition.
- the system for picture recognition may be implemented by software and/or hardware, and the system for picture recognition may be integrated into the server.
- the picture to be identified may be any picture on the network, such as pictures of different resolutions, different sizes or different picture contents.
- the present embodiment identifies the picture according to the content of the picture, and the picture to be identified is classified into an indecent picture and a normal picture, wherein, Ya pictures include violent pictures and erotic images.
- the recognition model generated by the pre-training may be a model for performing picture recognition using various machine learning algorithms, such as CNN (Convolutional Neural Network), RNN (Circular Neural Network), DNN (Deep Neural Network), etc., preferably,
- the recognition model is a deep convolutional neural network model.
- the recognition model is pre-applied with a large number of indecent images before being used to identify indecent images. Since a large number of indecent images can maximize or simulate scenes in indecent images, the trained recognition models can accurately identify Out of the picture.
- the recognition model After the picture to be recognized is input to the recognition model, the recognition model recognizes it and outputs the recognition result, and the recognition result can represent different results by numbers, for example, the number “0” indicates a recognition result, and the number “1” indicates Another way to identify the result.
- Step S2 Determine a picture category corresponding to the outputted recognition result according to the association relationship between the predetermined recognition result and the picture category.
- the recognition result includes a first recognition result, a second identification result, and a third identification result
- the identification result may be identified by a number, for example, the number “0” identifies the first recognition result, the number “1” identifies the second identification result, and the number "2" identifies the third recognition result.
- the image categories include normal pictures, violent pictures and erotic pictures.
- the recognition result is associated with the picture category in advance, and the association relationship is stored, the first recognition result corresponds to the normal picture, the second recognition result corresponds to the violent picture, and the third recognition result corresponds to the erotic picture.
- the picture category of the picture can be determined, for example, the output number “0” can be determined as a normal picture; the output number “1” can be determined as a violent picture; the output number “ 2”, can be determined as erotic images.
- Step S3 prompting when the image category is violent or erotic, and marking and recording the violent or erotic image to determine the source of the violent or erotic image.
- the prompt may be presented in the form of sound and/or light, for example, a "drip" sound is emitted when a violent picture or a erotic picture is recognized.
- a voice prompt for "violent pictures (or erotic images)" or an electronic device with a system for image recognition, which has a prompt light, and the light is blinking or blinking when a violent picture or a erotic picture is recognized. And keep it for a certain period of time (for example, 3 seconds) to prompt the relevant personnel to confirm the identified violent pictures or erotic images.
- the present embodiment uses a recognition model obtained through a large amount of data training to identify a picture in the field of indecent picture recognition. Since the recognition model can cover various factors affecting indecent picture recognition, therefore, with the traditional Compared with the identification method of the indecent picture, the embodiment can effectively identify the indecent picture on the network, greatly improve the accuracy of the indecent picture recognition, and facilitate the execution of the work of some network supervision departments or security supervision departments.
- the step S1 includes:
- S02 Perform image preprocessing on each sample image to obtain a training picture to be trained; performing image preprocessing on each sample image includes uniformly adjusting the size of the sample image, flipping the sample image, and the like to increase the data.
- the scale The scale.
- the training picture of the first preset ratio in the training picture is used as a training set, and the training picture of the second preset ratio in the training picture is used as a verification set; and the first preset ratio in the training picture may be used.
- the training picture is used as a training set, for example, 50% of the training pictures in the training picture are used as the training set, and the training picture of the second preset proportion in the training picture is used as the verification set, for example, 25% of the training pictures in the training picture are used as the verification set.
- the first preset ratio is greater than the second preset ratio.
- the accuracy rate threshold is a depth convolutional neural network model with an accuracy rate greater than or equal to a preset accuracy threshold as the recognition model in step S1 of the above embodiment.
- the step S02 includes:
- S021 Adjust each sample picture to a first picture of the same size (for example, the pixel is 384*384), and crop a second picture of a preset size (for example, the pixel is 256*256) on each first picture.
- S023 Calculate an average pixel image of the sample image based on the second image and the third image corresponding to each sample image, and obtain training corresponding to each sample image based on the second image, the third image, and the average pixel image corresponding to each sample image. image.
- Each pixel value of the pixel picture is an average value of pixel values of the corresponding second picture and the third picture corresponding pixel, for example, the pixel point X of the average pixel picture and the pixel point X1 of the second picture and the pixel point of the third picture, respectively Corresponding to X2, the pixel value of the pixel point X is an average value of the pixel values of all the pixel points X1 and the pixel point X2; and each pixel value in each of the second picture and the third picture corresponding to each sample picture is respectively subtracted from the corresponding average
- the pixel value of the corresponding pixel in the pixel image is used to obtain a training picture corresponding to each sample picture, thereby increasing the size of the training picture.
- the total number of layers of the deep convolutional neural network model is 22, including one input layer, 13 convolutional layers, 5 pooling layers, and 2 fully connected layers.
- 1 classification layer the detailed structure of the deep convolutional neural network model is shown in Table 1 below:
- the Layer Name column indicates the name of each layer
- Input indicates the input layer
- the input layer neurons are connected with the pixel of the training picture
- Conv indicates the convolution layer of the deep convolutional neural network model
- Conv1 indicates the deep convolutional neural network model.
- the first convolutional layer, Conv2 represents the second convolutional layer of the deep convolutional neural network model, and so on
- MaxPool represents the maximum pooled layer of the deep convolutional neural network model
- MaxPool1 represents the deep convolutional neural network.
- the first maximum pooling layer of the model Fc represents the fully connected layer in the deep convolutional neural network model
- Fc1 represents the first fully connected layer in the deep convolutional neural network model
- Softmax represents the Softmax classifier
- Batch Size table Shows the number of input images of the current layer
- Kernel Size represents the scale of the current layer convolution kernel (for example, Kernel Size can be equal to 3, indicating that the scale of the convolution kernel is 3x 3);
- Stride Size represents the moving step size of the convolution kernel, ie The distance moved to the next convolution position after a convolution;
- Pad Size indicates the size of the image fill in the current network layer.
- model function of the deep convolutional neural network model is:
- W is the weight matrix of the model function
- b is the bias term vector of the model function
- N is the number of training pictures in the training set
- x (i) is the training picture of the ith input
- y (i) is the ith input
- ⁇ is the weight attenuation term
- l is the serial number of the layer in the model function
- n l represents the total number of layers of the model function
- s l represents the number of neurons included in the first layer of the model function
- a weight value indicating the connection between the jth neuron of the model function and the i th neuron in the next layer.
- the statistic factor is set to 3*10 -4
- the probability of the connection weight of the fully connected layer being dropped (dropout) is set to 0.5
- the learning rate of the model training is initially set to 0.003 to ensure the efficiency and accuracy of the training.
- FIG. 5 is a functional block diagram of a preferred embodiment of the system 10 for picture recognition of the present invention.
- the system 10 for picture recognition may be divided into one or more modules, one or more modules being stored in the memory 11 and being processed by one or more processors (the processor 12 in this embodiment) It is performed to complete the present invention.
- the system 10 for picture recognition may be divided into an identification module 101, a determination module 102, and a prompting module 103.
- module refers to a series of computer program instruction segments capable of performing a specific function, which is more suitable than the program for describing the execution of the system 10 for picture recognition in the electronic device 1, wherein:
- the identification module 101 is configured to: after receiving the picture to be identified, input the picture to be identified into the recognition model generated by the pre-training for identification, and output the recognition result;
- system for picture recognition can be implemented by software and/or hardware, and the system for picture recognition can be integrated into the server.
- the picture to be identified may be any picture on the network, such as pictures of different resolutions, different sizes or different picture contents.
- the embodiment identifies the picture according to the content of the picture, and the picture to be recognized is identified into an indecent picture and a normal picture, wherein the indecent picture includes a violent picture and a erotic picture.
- the recognition model generated by the pre-training may be a model for performing image recognition using various machine learning algorithms, such as CNN (convolution neural network), RNN (circular neural network), DNN (deep) Degree neural network) and the like, preferably, the recognition model is a deep convolutional neural network model.
- the recognition model is pre-applied with a large number of indecent images before being used to identify indecent images. Since a large number of indecent images can maximize or simulate scenes in indecent images, the trained recognition models can accurately identify Out of the picture.
- the recognition model After the picture to be recognized is input to the recognition model, the recognition model recognizes it and outputs the recognition result, and the recognition result can represent different results by numbers, for example, the number “0” indicates a recognition result, and the number “1” indicates Another way to identify the result.
- the determining module 102 is configured to determine, according to a relationship between the predetermined recognition result and the picture category, a picture category corresponding to the outputted recognition result.
- the recognition result includes a first recognition result, a second identification result, and a third identification result
- the identification result may be identified by a number, for example, the number “0” identifies the first recognition result, the number “1” identifies the second identification result, and the number "2" identifies the third recognition result.
- the image categories include normal pictures, violent pictures and erotic pictures.
- the recognition result is associated with the picture category in advance, and the association relationship is stored, the first recognition result corresponds to the normal picture, the second recognition result corresponds to the violent picture, and the third recognition result corresponds to the erotic picture.
- the picture category of the picture can be determined, for example, the output number “0” can be determined as a normal picture; the output number “1” can be determined as a violent picture; the output number “ 2”, can be determined as erotic images.
- the prompting module 103 is configured to prompt when the image category is violent or erotic, and mark and record the violent or erotic image to determine the source of the violent or erotic image.
- the prompt may be presented in the form of sound and/or light, for example, a "drip" sound is emitted when a violent picture or a erotic picture is recognized.
- a voice prompt for "violent pictures (or erotic images)" or an electronic device with a system for image recognition, which has a prompt light, and the light is blinking or blinking when a violent picture or a erotic picture is recognized. And keep it for a certain period of time (for example, 3 seconds) to prompt the relevant personnel to confirm the identified violent pictures or erotic images.
- the recognition model is a deep convolutional neural network model
- the system for image recognition further includes:
- the preparation module 001 is configured to prepare a corresponding preset number of sample pictures for each picture category, and calibrate the picture categories corresponding to each sample picture; for each sample picture of the picture category, prepare a corresponding preset number of sample pictures, For example, the number of sample pictures of a normal picture is 100,000, the number of sample pictures of a violent picture is 100,000, and the number of sample pictures of a erotic picture is 100,000.
- the pre-processing module 002 is configured to perform image pre-processing on each sample picture to obtain a training picture to be trained; performing image pre-processing on each sample picture includes uniformly adjusting the size of the sample picture Integrity, flipping the sample image, etc., to increase the size of the data.
- the processing module 003 is configured to use, as a training set, a training picture of a first preset ratio in the training picture, and use a training picture of a second preset ratio in the training picture as a verification set;
- a preset proportion of the training picture is used as a training set, for example, 50% of the training pictures in the training picture are used as the training set, and the training picture of the second preset ratio in the training picture is used as the verification set, for example, 25% training in the training picture.
- the picture is used as a verification set, and the first preset ratio is greater than the second preset ratio.
- the training module 004 is configured to train a predetermined deep convolutional neural network model by using the training picture in the training set; wherein, the parameters of the deep convolutional neural network model in the initial training may adopt default parameters, and the parameters are continuously adjusted as the training progresses. .
- the verification module 005 is configured to verify the accuracy of the trained deep convolutional neural network model by using the training picture in the verification set. If the accuracy rate is greater than or equal to a preset accuracy threshold (for example, 0.98), the training ends, or If the accuracy is less than the preset accuracy threshold, increase the number of sample pictures corresponding to each picture category, and re-trigger the above preparation module 001, etc., to re-train until the accuracy of the deep convolutional neural network model is verified.
- the depth convolutional neural network model with the accuracy rate greater than or equal to the preset accuracy threshold is greater than or equal to the preset accuracy threshold, and is used as the recognition model in the identification module 101 of the above embodiment.
- the pre-processing module 002 includes:
- the adjusting unit 0021 is configured to adjust each sample picture to a first picture of the same size (for example, the pixel is 384*384), and trim the preset size (for example, the pixel is 256*256) on each of the first pictures. Two pictures.
- the processing unit 0022 is configured to perform a reversal of a preset direction (for example, a horizontal or vertical direction) on each of the second pictures, and perform a twisting operation according to the preset angle to obtain a third picture corresponding to each second picture, where
- the function of the flip and twist operation is to simulate various forms of pictures in the actual business scene.
- the flipping and twisting operations of the pictures can increase the size of the data set, thereby increasing the size of the training picture.
- the calculating unit 0023 is configured to calculate an average pixel image of the sample image based on the second image and the third image corresponding to each sample image, and obtain each sample based on the second image, the third image, and the average pixel image corresponding to each sample image.
- the training picture corresponding to the picture.
- each pixel value of the average pixel picture is an average value of pixel values of corresponding pixels of the second picture and the third picture, for example, an average pixel picture
- the pixel points X correspond to the pixel point X1 of the second picture and the pixel point X2 of the third picture, respectively, and the pixel value of the pixel point X is an average value of the pixel values of all the pixel points X1 and the pixel point X2; corresponding to each sample picture
- Each pixel value in each of the second picture and the third picture is respectively subtracted from the pixel value of the corresponding pixel in the corresponding average pixel picture to obtain a training picture corresponding to each sample picture, thereby increasing the size of the training picture.
- the total number of layers of the deep convolutional neural network model is 22, including one input layer, 13 convolutional layers, 5 pooling layers, and 2 fully connected layers, 1
- the classification structure and the detailed structure of the deep convolutional neural network model are shown in Table 1 above, and are not described here.
- model function of the deep convolutional neural network model is:
- W is the weight matrix of the model function
- b is the bias term vector of the model function
- N is the number of training pictures in the training set
- x (i) is the training picture of the ith input
- y (i) is the ith input
- ⁇ is the weight attenuation term
- l is the serial number of the layer in the model function
- n l represents the total number of layers of the model function
- s l represents the number of neurons included in the first layer of the model function
- a weight value indicating the connection between the jth neuron of the model function and the i th neuron in the next layer.
- the statistic factor is set to 3*10 -4
- the probability of the connection weight of the fully connected layer being dropped (dropout) is set to 0.5
- the learning rate of the model training is initially set to 0.003 to ensure the efficiency and accuracy of the training.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
一种图片识别的方法、***、电子装置及介质,包括:在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果(S1);根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别(S2),在识别出图片类别为暴力图片或***时进行提示,并对暴力图片或***进行标记及记录,以确定暴力图片或***的来源(S3)。由此,能够有效地识别出网络上的不雅图片,大大提高不雅图片识别的准确率,便于一些网络监管部门或者***部门的工作的执行。
Description
优先权申明
本申请基于巴黎公约申明享有2017年3月13日递交的申请号为CN201710147708.3、名称为“图片识别的方法及***”中国专利申请的优先权,该中国专利申请的整体内容以参考的方式结合在本申请中。
本发明涉及图像处理技术领域,尤其涉及一种图片识别的方法、***、电子装置及计算机可读存储介质。
不雅图片,简称NSFW(Not Suitable/Safe For Work,工作场所不合适/不安全浏览的图片),不雅图片主要包括暴力图片和***。目前,已经存在较多识别不雅图片的技术方案,在图片处理技术领域,目前主要使用传统的方法判断图片是否是不雅图片,例如,对于暴力图片,根据人体特定部位及皮肤颜色等特征判断图片是否是暴力图片,又如,对于***,根据人体敏感部位特征结合文字的方式来判断是否是***,等等,这种识别不雅图片的方法通常识别的准确率不高,不能有效地识别出网络上传播的一些不雅图片,给一些网络监管部门或者***部门的工作带来不便。
发明内容
本发明的目的在于提供一种图片识别的方法、***、电子装置及介质,旨在提高识别不雅图片的准确率。
为实现上述目的,本发明第一方面提供一种图片识别的方法,所述图片识别的方法包括:
S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;
S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;
S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
本申请第二方面提供一种图片识别的***,所述图片识别的***包括:
识别模块,用于在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;
确定模块,用于根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;
提示模块,用于在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
本申请第三方面提供一种电子装置,包括处理设备、存储设备及图片设别***,该图片识别***存储于该存储设备中,包括至少一个计算机可读指令,该至少一个计算机可读指令可被所述处理设备执行,以实现以下操作:
S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;
S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;
S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
本申请第四方面提供一种计算机可读存储介质,其上存储有至少一个可被处理设备执行以实现以下操作的计算机可读指令:
S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;
S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;
S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
本发明的有益效果是:本发明在不雅图片识别领域利用经过大量数据训练得到的识别模型对图片进行识别,由于识别模型能够涵盖各种影响不雅图片识别的因素,因此,与传统的不雅图片的识别方法相比,本发明能够有效地识别出网络上的不雅图片,大大提高不雅图片识别的准确率,便于一些网络监管部门或者***部门的工作的执行。
图1为本发明图片识别的方法的较佳实施例的应用环境示意图;
图2为本发明图片识别的方法第一实施例的流程示意图;
图3为本发明图片识别的方法第二实施例的流程示意图;
图4为图2所示步骤S02的细化流程示意图;
图5为本发明图片识别的***第一实施例的结构示意图;
图6为本发明图片识别的***第二实施例的结构示意图;
图7为图6所示预处理模块的结构示意图。
以下结合附图对本发明的原理和特征进行描述,所举实例只用于解释本发明,并非用于限定本发明的范围。
参阅图1所示,是本发明实现图片识别的方法的较佳实施例的应用环境示意图。应用环境示意图包括电子装置1及终端设备2。电子装置1可以通过网络、近场通信技术等适合的技术与终端设备2进行数据交互。
终端设备2包括,但不限于,任何一种可与用户通过键盘、鼠标、遥控器、触摸板或者声控设备等方式进行人机交互的电子产品,例如,个人计算机、平板电脑、智能手机、个人数字助理(Personal Digital Assistant,PDA),游戏机、交互式网络电视(Internet Protocol Television,IPTV)、智能式穿戴式设备等。
电子装置1是一种能够按照事先设定或者存储的指令,自动进行数值计算和/或信息处理的设备。电子装置1可以是计算机、也可以是单个网络服务器、多个网络服务器组成的服务器组或者基于云计算的由大量主机或者网络服务器构成的云,其中云计算是分布式计算的一种,由一群松散耦合的计算机集组成的一个超级虚拟计算机。
在本实施例中,电子装置1包括,但不仅限于,可通过***总线相互通信连接的存储设备11、处理设备12、及网络接口13。需要指出的是,图1仅示出了具有组件11-13的电子装置1,但是应理解的是,并不要求实施所有示出的组件,可以替代的实施更多或者更少的组件。
其中,存储设备11包括内存及至少一种类型的可读存储介质。内存为电子装置1的运行提供缓存;可读存储介质可为如闪存、硬盘、多媒体卡、卡型存储器等的非易失性存储介质。在一些实施例中,可读存储介质可以是电子装置1的内部存储单元,例如该电子装置1的硬盘;在另一些实施例中,该非易失性存储介质也可以是电子装置1的外部存储设备,例如电子装置1上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。本实施例中,存储设备11的可读存储介质通常用于存储安装于电子装置1的操作***和各类应用软件,例如本申请一实施例中的图片识别的***10的程序代码等。此外,存储设备11还可以用于暂时地存储已经输出或者将要输出的各类数据。
处理设备12在一些实施例中可以包括一个或者多个微处理器、微控制器、数字处理器等。该处理设备12通常用于控制电子装置1的运行,例如执行与终端设备2进行数据交互或者通信相关的控制和处理等。在本实施例中,处理设备12用于运行存储设备11中存储的程序代码或者处理数据,例如运行图片处理的***10等。
网络接口13可包括无线网络接口或有线网络接口,该网络接口13通常用于在电子装置1与其他电子设备之间建立通信连接。本实施例中,网络接口13主要用于将电子装置1与一个或多个终端设备2相连,在电子装置1与一个或多个终端设备2之间建立数据传输通道和通信连接。
图片识别的***10包括至少一个存储在存储设备11中的计算机可读指令,该至少一个计算机可读指令可被处理设备12执行,以实现本申请各实施例的图片识别的方法。如后续所述,该至少一个计算机可读指令依据其各部分所实现的功能不同,可被划为不同的逻辑模块。
在一实施例中,图片识别的***10被处理设备12执行时,实现以下操作:首先在接收到终端设备2发送的待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;然后根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别;最后在识别出图片类别为暴力图片或***时发送提示信息至终端设备2,对终端用户进行提示,并对暴力图片或***进行标记及记录,以确定暴力图片或***的来源。
其中,图片识别的***10输出的识别结果包括第一识别结果、第二识别结果及第三识别结果,图片类别包括正常图片、暴力图片及***,第一识别结果对应正常图片,第二识别结果对应暴力图片,第三识别结果对应***。在一实施例中,图片识别的***10存储在存储设备11中,包括至少一个存储在存储设备11中的计算机可读指令,该至少一个计算机可读指令可被处理设备12执行,以实现本申请各实施例的图片识别的方法。如后续所述,该至少一个计算机可读指令依据其各部分所实现的功能不同,可被划为不同的逻辑模块。
如图2所示,图2为本发明图片识别的方法较佳实施例的流程示意图,本实施例图片识别的方法并不限于流程中所示的步骤,此外流程图中所示步骤中,某些步骤可以省略、步骤之间的顺序可以改变。该图片识别的方法包括以下步骤:
步骤S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果。
本实施例中,可以由图片识别的***对图片进行识别,该图片识别的***可由软件和/或硬件实现,图片识别的***可以集成于服务器中。
待识别的图片可以是网络上的任意图片,例如不同分辨率、不同尺寸或者不同图片内容的图片等。本实施例在对图片进行识别的过程中,按照图片内容进行识别,待识别的图片被识别后分为不雅图片及正常图片,其中,不
雅图片包括暴力图片及***。
预先训练生成的识别模型可以是运用各种机器学习算法进行图片识别的模型,例如可以是CNN(卷积神经网络)、RNN(循环神经网络)、DNN(深度神经网络)等,优选地,该识别模型为深度卷积神经网络模型。识别模型在用于识别不雅图片之前,预先运用大量的不雅图片进行训练,由于大量的不雅图片能够最大程度地囊括或模拟不雅图片中的场景,因此训练好的识别模型能够准确识别出不雅图片。
在将待识别的图片输入至识别模型后,识别模型对其进行识别,并输出识别结果,识别结果可以用数字表示不同的结果,例如数字“0”表示一种识别结果,数字“1”表示另一种识别结果。
步骤S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别。
其中,识别结果包括第一识别结果、第二识别结果及第三识别结果,识别结果可以用数字进行标识,例如数字“0”标识第一识别结果、数字“1”标识第二识别结果及数字“2”标识第三识别结果。
其中,图片类别包括正常图片、暴力图片及***。预先将识别结果与图片类别进行关联对应,并将关联关系进行存储,第一识别结果对应正常图片,第二识别结果对应暴力图片,第三识别结果对应***。
在输出识别结果后,根据识别结果与图片类别的关联关系可以确定图片的图片类别,例如输出数字“0”,可确定为正常图片;输出数字“1”,可确定为暴力图片;输出数字“2”,可确定为***。
步骤S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
在识别模型识别出图片类别为暴力图片或***时进行提示,优选地,可以以声和/或光的形式进行提示,例如在识别出暴力图片或***时发出“滴滴”的提示音或者发出“暴力图片(或***)”的语音提示,或者安装有图片识别的***的电子装置上有提示灯,在识别出暴力图片或***时该提示灯处于亮的状态或者闪烁的状态,并保持一定的时间(例如3秒),以提示相关人员对该识别出的暴力图片或***进行确认。
然后对识别出来的暴力图片或***进行标记,例如暴力图片标记为“I”,***标记为“II”,以与其他正常图片进行区分,最后可以记录识别出来的暴力图片或***的相关图片信息,例如记录暴力图片或***的URL信息,以确定暴力图片或***的来源。
与现有技术相比,本实施例在不雅图片识别领域利用经过大量数据训练得到的识别模型对图片进行识别,由于识别模型能够涵盖各种影响不雅图片识别的因素,因此,与传统的不雅图片的识别方法相比,本实施例能够有效地识别出网络上的不雅图片,大大提高不雅图片识别的准确率,便于一些网络监管部门或者***部门的工作的执行。
在一优选的实施例中,如图3所示,在上述图2的实施例的基础上,所述步骤S1之前包括:
S01,为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;对于每一图片类别的样本图片,准备对应的预设数量的样本图片,例如正常图片的样本图片的数量为10万,暴力图片的样本图片的数量为10万,***的样本图片的数量为10万。
S02,对各样本图片进行图片预处理,以获取待训练的训练图片;对各样本图片进行图片预处理包括对样本图片进行大小的统一调整,对样本图片进行翻转扭曲等等,以增大数据的规模。
S03,以所述训练图片中第一预设比例的训练图片作为训练集,并以所述训练图片中的第二预设比例的训练图片作为验证集;可以以训练图片中第一预设比例的训练图片作为训练集,例如以训练图片中50%的训练图片作为训练集,以训练图片中第二预设比例的训练图片作为验证集,例如以训练图片中25%的训练图片作为验证集,第一预设比例大于第二预设比例。
S04,利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;其中,初次训练时深度卷积神经网络模型的参数可以采用默认参数,随着训练的进行参数不断调整。
S05,利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值(例如0.98),则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,重新执行上述的步骤S01至S05,以重新进行训练,直至验证深度卷积神经网络模型的准确率大于等于预设准确率阈值,以准确率大于等于预设准确率阈值的深度卷积神经网络模型作为上述实施例的步骤S1中的识别模型。
在一优选的实施例中,如图4所示,在上述图3的实施例的基础上,所述步骤S02包括:
S021,将各样本图片调整为相同大小(例如,像素为384*384)的第一图片,在各第一图片上裁剪出预设大小(例如,像素为256*256)的第二图片。
S022,对各个第二图片做预设方向(例如,水平或垂直方向)的翻转,以及按照预设的角度进行扭曲操作,以获得各个第二图片对应的第三图片,其中,翻转和扭曲操作的作用是模拟实际业务场景下各种形式的图片,通过图片的翻转和扭曲操作可以增大数据集的规模,从而增大训练图片的规模。
S023,基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
计算出各个样本图片对应的第二图片和第三图片的平均像素图片,平均
像素图片的各个像素值是对应的第二图片和第三图片对应像素的像素值的平均值,例如,平均像素图片的像素点X分别与第二图片的像素点X1和第三图片的像素点X2对应,像素点X的像素值是所有像素点X1和像素点X2的像素值的平均值;将各个样本图片对应的各个第二图片和第三图片中的各个像素值分别减去对应的平均像素图片中的对应像素的像素值值,以得到各个样本图片对应的训练图片,从而增大训练图片的规模。
优选地,在上述图2的实施例的基础上,深度卷积神经网络模型的总层数为22,包括1个输入层,13个卷积层,5个池化层,2个全连接层,1个分类层,深度卷积神经网络模型的详细结构如下表1所示:
表1
Layer Name | Batch Size | Kernel Size | Stride Size | Pad Size |
Input | 128 | N/A | N/A | N/A |
Conv1 | 128 | 3 | 1 | 1 |
Conv2 | 128 | 3 | 1 | 1 |
MaxPool1 | 128 | 2 | 2 | 0 |
Conv3 | 128 | 3 | 1 | 1 |
Conv4 | 128 | 3 | 1 | 1 |
MaxPool2 | 128 | 2 | 2 | 0 |
Conv5 | 128 | 3 | 1 | 1 |
Conv6 | 128 | 3 | 1 | 1 |
Conv7 | 128 | 3 | 1 | 1 |
MaxPool3 | 128 | 2 | 2 | 0 |
Conv8 | 128 | 3 | 1 | 1 |
Conv9 | 128 | 3 | 1 | 1 |
Conv10 | 128 | 3 | 1 | 1 |
MaxPool4 | 128 | 2 | 2 | 0 |
Conv11 | 128 | 3 | 1 | 1 |
Conv12 | 128 | 3 | 1 | 1 |
Conv13 | 128 | 3 | 1 | 1 |
MaxPool5 | 128 | 2 | 2 | 0 |
Fc1 | 4096 | 1 | 1 | 0 |
Fc2 | 2048 | 1 | 1 | 0 |
Softmax | 3 | N/A | N/A | N/A |
其中,Layer Name列表示每一层的名称,Input表示输入层,输入层的神经元与训练图片的像素连接;Conv表示深度卷积神经网络模型的卷积层,Conv1表示深度卷积神经网络模型的第1个卷积层,Conv2表示深度卷积神经网络模型的第2个卷积层,以此类推;MaxPool表示深度卷积神经网络模型的最大值池化层,MaxPool1表示深度卷积神经网络模型的第1个最大值池化层,Fc表示深度卷积神经网络模型中的全连接层,Fc1表示深度卷积神经网络模型中第1个全连接层,Softmax表示Softmax分类器;Batch Size表
示当前层的输入图像数目;Kernel Size表示当前层卷积核的尺度(例如,Kernel Size可以等于3,表示卷积核的尺度为3x 3);Stride Size表示卷积核的移动步长,即做完一次卷积之后移动到下一个卷积位置的距离;Pad Size表示对当前网络层之中的图像填充的大小。
优选地,深度卷积神经网络模型的模型函数为:
W为模型函数的权值矩阵,b为模型函数的偏置项向量,N为训练集中训练图片的数量,x(i)为第i次输入的训练图片,y(i)为第i次输入的训练图片对应的图片类别标识,τ为权值衰减项,l为模型函数中层的序号,nl表示模型函数的总层数,sl表示模型函数的第l层包含的神经元个数,表示模型函数第l层第j个神经元与下一层中的第i个神经元之间的连接的权重值。
优选地,规约化因子设为3*10-4,全连接层的连接权重被丢弃(Dropout)的概率设置为0.5,模型训练的学习率初始设置为0.003,以保证训练的效率及准确性。
请参阅图5,是本发明图片识别的***10较佳实施例的功能模块图。在本实施例中,图片识别的***10可以被分割成一个或多个模块,一个或者多个模块被存储于存储器11中,并由一个或多个处理器(本实施例为处理器12)所执行,以完成本发明。例如,在图5中,图片识别的***10可以被分割成识别模块101、确定模块102及提示模块103。本发明所称的模块是指能够完成特定功能的一系列计算机程序指令段,比程序更适合于描述图片识别的***10在电子装置1中的执行过程,其中:
识别模块101,用于在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;
本实施例中,图片识别的***可由软件和/或硬件实现,图片识别的***可以集成于服务器中。
待识别的图片可以是网络上的任意图片,例如不同分辨率、不同尺寸或者不同图片内容的图片等。本实施例在对图片进行识别的过程中,按照图片内容进行识别,待识别的图片被识别后分为不雅图片及正常图片,其中,不雅图片包括暴力图片及***。
预先训练生成的识别模型可以是运用各种机器学习算法进行图片识别的模型,例如可以是CNN(卷积神经网络)、RNN(循环神经网络)、DNN(深
度神经网络)等,优选地,该识别模型为深度卷积神经网络模型。识别模型在用于识别不雅图片之前,预先运用大量的不雅图片进行训练,由于大量的不雅图片能够最大程度地囊括或模拟不雅图片中的场景,因此训练好的识别模型能够准确识别出不雅图片。
在将待识别的图片输入至识别模型后,识别模型对其进行识别,并输出识别结果,识别结果可以用数字表示不同的结果,例如数字“0”表示一种识别结果,数字“1”表示另一种识别结果。
确定模块102,用于根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别。
其中,识别结果包括第一识别结果、第二识别结果及第三识别结果,识别结果可以用数字进行标识,例如数字“0”标识第一识别结果、数字“1”标识第二识别结果及数字“2”标识第三识别结果。
其中,图片类别包括正常图片、暴力图片及***。预先将识别结果与图片类别进行关联对应,并将关联关系进行存储,第一识别结果对应正常图片,第二识别结果对应暴力图片,第三识别结果对应***。
在输出识别结果后,根据识别结果与图片类别的关联关系可以确定图片的图片类别,例如输出数字“0”,可确定为正常图片;输出数字“1”,可确定为暴力图片;输出数字“2”,可确定为***。
提示模块103,用于在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
在识别模型识别出图片类别为暴力图片或***时进行提示,优选地,可以以声和/或光的形式进行提示,例如在识别出暴力图片或***时发出“滴滴”的提示音或者发出“暴力图片(或***)”的语音提示,或者安装有图片识别的***的电子装置上有提示灯,在识别出暴力图片或***时该提示灯处于亮的状态或者闪烁的状态,并保持一定的时间(例如3秒),以提示相关人员对该识别出的暴力图片或***进行确认。
然后对识别出来的暴力图片或***进行标记,例如暴力图片标记为“I”,***标记为“II”,以与其他正常图片进行区分,最后可以记录识别出来的暴力图片或***的相关信息,例如记录暴力图片或***的URL信息,以确定暴力图片或***的来源。
在一优选的实施例中,如图6所示,在上述图5的实施例的基础上,识别模型为深度卷积神经网络模型,所述图片识别的***还包括:
准备模块001,用于为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;对于每一图片类别的样本图片,准备对应的预设数量的样本图片,例如正常图片的样本图片的数量为10万,暴力图片的样本图片的数量为10万,***的样本图片的数量为10万。
预处理模块002,用于对各样本图片进行图片预处理,以获取待训练的训练图片;对各样本图片进行图片预处理包括对样本图片进行大小的统一调
整,对样本图片进行翻转扭曲等等,以增大数据的规模。
处理模块003,用于以所述训练图片中第一预设比例的训练图片作为训练集,并以所述训练图片中的第二预设比例的训练图片作为验证集;可以以训练图片中第一预设比例的训练图片作为训练集,例如以训练图片中50%的训练图片作为训练集,以训练图片中第二预设比例的训练图片作为验证集,例如以训练图片中25%的训练图片作为验证集,第一预设比例大于第二预设比例。
训练模块004,用于利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;其中,初次训练时深度卷积神经网络模型的参数可以采用默认参数,随着训练的进行参数不断调整。
验证模块005,用于利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值(例如0.98),则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,重新触发上述的准备模块001等,以重新进行训练,直至验证深度卷积神经网络模型的准确率大于等于预设准确率阈值,以准确率大于等于预设准确率阈值的深度卷积神经网络模型作为上述实施例的识别模块101中的识别模型。
在一优选的实施例中,如图7所示,在上述图6的实施例的基础上,所述预处理模块002包括:
调整单元0021,用于将各样本图片调整为相同大小(例如,像素为384*384)的第一图片,在各第一图片上裁剪出预设大小(例如,像素为256*256)的第二图片。
处理单元0022,用于对各个第二图片做预设方向(例如,水平或垂直方向)的翻转,以及按照预设的角度进行扭曲操作,以获得各个第二图片对应的第三图片,其中,翻转和扭曲操作的作用是模拟实际业务场景下各种形式的图片,通过图片的翻转和扭曲操作可以增大数据集的规模,从而增大训练图片的规模。
计算单元0023,用于基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
计算出各个样本图片对应的第二图片和第三图片的平均像素图片,平均像素图片的各个像素值是对应的第二图片和第三图片对应像素的像素值的平均值,例如,平均像素图片的像素点X分别与第二图片的像素点X1和第三图片的像素点X2对应,像素点X的像素值是所有像素点X1和像素点X2的像素值的平均值;将各个样本图片对应的各个第二图片和第三图片中的各个像素值分别减去对应的平均像素图片中的对应像素的像素值值,以得到各个样本图片对应的训练图片,从而增大训练图片的规模。
优选地,在上述的实施例的基础上,深度卷积神经网络模型的总层数为22,包括1个输入层,13个卷积层,5个池化层,2个全连接层,1个分类层,深度卷积神经网络模型的详细结构如上述表1所示,此处不再赘述。
优选地,深度卷积神经网络模型的模型函数为:
W为模型函数的权值矩阵,b为模型函数的偏置项向量,N为训练集中训练图片的数量,x(i)为第i次输入的训练图片,y(i)为第i次输入的训练图片对应的图片类别标识,τ为权值衰减项,l为模型函数中层的序号,nl表示模型函数的总层数,sl表示模型函数的第l层包含的神经元个数,表示模型函数第l层第j个神经元与下一层中的第i个神经元之间的连接的权重值。
优选地,规约化因子设为3*10-4,全连接层的连接权重被丢弃(Dropout)的概率设置为0.5,模型训练的学习率初始设置为0.003,以保证训练的效率及准确性。
以上所述仅为本发明的较佳实施例,并不用以限制本发明,凡在本发明的精神和原则之内,所作的任何修改、等同替换、改进等,均应包含在本发明的保护范围之内。
Claims (18)
- 一种图片识别的方法,其特征在于,所述图片识别的方法包括:S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
- 根据权利要求1所述的图片识别的方法,其特征在于,所述识别模型为深度卷积神经网络模型,所述步骤S1之前包括:S01,为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;S02,对各样本图片进行图片预处理,以获取待训练的训练图片;S03,以所述训练图片中第一预设比例的训练图片作为训练集,并以所述训练图片中的第二预设比例的训练图片作为验证集;S04,利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;S05,利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值,则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,以重新进行训练。
- 根据权利要求2所述的图片识别的方法,其特征在于,所述步骤S02包括:S021,将各样本图片调整为相同大小的第一图片,在各第一图片上裁剪出预设大小的第二图片;S022,对各第二图片进行翻转及扭曲操作,以获得各第二图片对应的第三图片;S023,基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
- 根据权利要求4所述的图片识别的方法,其特征在于,所述深度卷积神经网络模型的规约化因子为3*10-4,全连接层的连接权重被丢弃的概率为0.5,训练的学习率初始为0.003。
- 一种图片识别的***,其特征在于,所述图片识别的***包括:识别模块,用于在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;确定模块,用于根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;提示模块,用于在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
- 根据权利要求6所述的图片识别的***,其特征在于,所述识别模型为深度卷积神经网络模型,所述图片识别的***还包括:准备模块,用于为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;预处理模块,用于对各样本图片进行图片预处理,以获取待训练的训练图片;处理模块,用于以所述训练图片中第一预设比例的训练图片作为训练集,并以所述训练图片中的第二预设比例的训练图片作为验证集;训练模块,用于利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;验证模块,用于利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值,则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,以重新进行训练。
- 根据权利要求7所述的图片识别的***,其特征在于,所述预处理模块包括:调整单元,用于将各样本图片调整为相同大小的第一图片,在各第一图片上裁剪出预设大小的第二图片;处理单元,用于对各第二图片进行翻转及扭曲操作,以获得各第二图片对应的第三图片;计算单元,用于基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
- 根据权利要求9所述的图片识别的***,其特征在于,所述深度卷积神经网络模型的规约化因子为3*10-4,全连接层的连接权重被丢弃的概率为0.5,训练的学习率初始为0.003。
- 一种电子装置,包括处理设备、存储设备及图片识别***,该图片识别***存储于该存储设备中,包括至少一个计算机可读指令,该至少一个计算机可读指令可被所述处理设备执行,以实现以下操作:S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来源。
- 根据权利要求11所述的电子装置,其特征在于,所述识别模型为深度卷积神经网络模型,所述至少一个计算机可读指令还可被所述处理设备执行,在实现所述操作S1之前,以实现以下操作:S01,为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;S02,对各样本图片进行图片预处理,以获取待训练的训练图片;S03,以所述训练图片中第一预设比例的训练图片作为训练集,并以所 述训练图片中的第二预设比例的训练图片作为验证集;S04,利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;S05,利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值,则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,以重新进行训练。
- 根据权利要求12所述的电子装置,其特征在于,所述至少一个计算机可读指令被所述处理设备执行,实现所述操作S02包括:S021,将各样本图片调整为相同大小的第一图片,在各第一图片上裁剪出预设大小的第二图片;S022,对各第二图片进行翻转及扭曲操作,以获得各第二图片对应的第三图片;S023,基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
- 根据权利要求14所述的电子装置,其特征在于,所述深度卷积神经网络模型的规约化因子为3*10-4,全连接层的连接权重被丢弃的概率为0.5,训练的学习率初始为0.003。
- 一种计算机可读存储介质,其上存储有至少一个可被处理设备执行以实现以下操作的计算机可读指令:S1,在接收到待识别的图片后,将待识别的图片输入至预先训练生成的识别模型中进行识别,并输出识别结果;S2,根据预定的识别结果与图片类别的关联关系,确定所输出的识别结果对应的图片类别,其中,所述识别结果包括第一识别结果、第二识别结果及第三识别结果,所述图片类别包括正常图片、暴力图片及***,所述第一识别结果对应正常图片,所述第二识别结果对应暴力图片,所述第三识别结果对应***;S3,在识别出图片类别为暴力图片或***时进行提示,并对所述暴力图片或***进行标记及记录,以确定所述暴力图片或***的来 源。
- 根据权利要求16所述的存储介质,其特征在于,所述至少一个计算机指令在实现所述操作S1之前,还用于实现以下操作:S01,为各图片类别准备对应的预设数量的样本图片,并标定每一样本图片对应的图片类别;S02,对各样本图片进行图片预处理,以获取待训练的训练图片;S03,以所述训练图片中第一预设比例的训练图片作为训练集,并以所述训练图片中的第二预设比例的训练图片作为验证集;S04,利用所述训练集中的训练图片训练预定的深度卷积神经网络模型;S05,利用所述验证集中的训练图片验证训练后的深度卷积神经网络模型的准确率,若所述准确率大于等于预设准确率阈值,则训练结束,或者,若所述准确率小于预设准确率阈值,则增加各图片类别对应的样本图片的数量,以重新进行训练。
- 根据权利要求17所述的存储介质,其特征在于,所述至少一个计算机指令实现所述操作S02包括:S021,将各样本图片调整为相同大小的第一图片,在各第一图片上裁剪出预设大小的第二图片;S022,对各第二图片进行翻转及扭曲操作,以获得各第二图片对应的第三图片;S023,基于各样本图片对应的第二图片及第三图片计算得到该样本图片的平均像素图片,并基于各样本图片对应的第二图片、第三图片及平均像素图片获取各样本图片对应的训练图片。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710147708.3A CN106951848B (zh) | 2017-03-13 | 2017-03-13 | 图片识别的方法及*** |
CN201710147708.3 | 2017-03-13 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2018166114A1 true WO2018166114A1 (zh) | 2018-09-20 |
Family
ID=59466880
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2017/091371 WO2018166114A1 (zh) | 2017-03-13 | 2017-06-30 | 图片识别的方法、***、电子装置及介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN106951848B (zh) |
WO (1) | WO2018166114A1 (zh) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109657694A (zh) * | 2018-10-26 | 2019-04-19 | 平安科技(深圳)有限公司 | 图片自动分类方法、装置及计算机可读存储介质 |
CN109727238A (zh) * | 2018-12-27 | 2019-05-07 | 贵阳朗玛信息技术股份有限公司 | X光胸片的识别方法及装置 |
CN109947940A (zh) * | 2019-02-15 | 2019-06-28 | 平安科技(深圳)有限公司 | 文本分类方法、装置、终端及存储介质 |
CN110287991A (zh) * | 2019-05-22 | 2019-09-27 | 平安科技(深圳)有限公司 | 植物类生药真伪验证方法、装置、计算机设备及存储介质 |
CN111046859A (zh) * | 2018-10-11 | 2020-04-21 | 杭州海康威视数字技术股份有限公司 | 字符识别方法及装置 |
CN111402224A (zh) * | 2020-03-12 | 2020-07-10 | 广东电网有限责任公司广州供电局 | 一种用于电力设备的目标识别方法 |
CN111460247A (zh) * | 2019-01-21 | 2020-07-28 | 重庆邮电大学 | 网络图片敏感文字自动检测方法 |
CN112597828A (zh) * | 2020-12-11 | 2021-04-02 | 京东数字科技控股股份有限公司 | 网页识别模型的训练方法、装置、网页识别的方法 |
CN112907498A (zh) * | 2019-11-18 | 2021-06-04 | 中国商用飞机有限责任公司 | 一种孔隙识别方法、装置、设备及存储介质 |
CN114155417A (zh) * | 2021-12-13 | 2022-03-08 | 中国科学院空间应用工程与技术中心 | 图像目标的识别方法、装置、电子设备及计算机存储介质 |
Families Citing this family (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107480773B (zh) * | 2017-08-09 | 2020-11-13 | 北京小米移动软件有限公司 | 训练卷积神经网络模型的方法、装置及存储介质 |
CN107516112A (zh) * | 2017-08-24 | 2017-12-26 | 北京小米移动软件有限公司 | 对象类别识别方法、装置、设备及存储介质 |
CN107665333A (zh) * | 2017-08-28 | 2018-02-06 | 平安科技(深圳)有限公司 | 一种基于卷积神经网络的不雅图片识别方法、终端、设备及计算机可读存储介质 |
CN108052523A (zh) * | 2017-11-03 | 2018-05-18 | 中国互联网络信息中心 | 基于卷积神经网络的赌博网站识别方法和*** |
CN108415807A (zh) * | 2018-02-12 | 2018-08-17 | 成都睿码科技有限责任公司 | 一种爬取监控电子设备是否浏览不良信息的方法 |
CN108491866B (zh) * | 2018-03-06 | 2022-09-13 | 平安科技(深圳)有限公司 | ***鉴定方法、电子装置及可读存储介质 |
CN108509959A (zh) * | 2018-04-13 | 2018-09-07 | 广州优视网络科技有限公司 | 色情应用识别方法、装置、计算机可读存储介质和服务器 |
CN108734708B (zh) * | 2018-05-23 | 2021-02-19 | 平安科技(深圳)有限公司 | 胃癌识别方法、装置及存储介质 |
CN109145979B (zh) * | 2018-08-15 | 2022-06-21 | 上海嵩恒网络科技股份有限公司 | 敏感图像鉴定方法及终端*** |
CN109491970B (zh) * | 2018-10-11 | 2024-05-10 | 平安科技(深圳)有限公司 | 面向云存储的不良图片检测方法、装置及存储介质 |
CN109460485A (zh) * | 2018-10-12 | 2019-03-12 | 咪咕文化科技有限公司 | 一种图像库建立方法、装置和存储介质 |
CN109241951A (zh) * | 2018-10-26 | 2019-01-18 | 北京陌上花科技有限公司 | ***识别方法、识别模型构建方法及识别模型和计算机可读存储介质 |
CN109508746A (zh) * | 2018-11-16 | 2019-03-22 | 西安电子科技大学 | 基于卷积神经网络的脉冲星候选体识别方法 |
CN109886865B (zh) * | 2019-01-07 | 2024-01-23 | 平安科技(深圳)有限公司 | 自动屏蔽不良信息的方法、装置、计算机设备以及存储介质 |
CN112651413B (zh) * | 2019-10-10 | 2023-10-17 | 百度在线网络技术(北京)有限公司 | 低俗图的集成学习分类方法、装置、设备及存储介质 |
CN110807139B (zh) * | 2019-10-23 | 2023-09-01 | 腾讯科技(深圳)有限公司 | 图片识别方法、装置、计算机可读存储介质和计算机设备 |
CN111127327B (zh) * | 2019-11-14 | 2024-04-12 | 贝壳技术有限公司 | 一种图片倾斜检测方法及装置 |
CN113093967A (zh) * | 2020-01-08 | 2021-07-09 | 富泰华工业(深圳)有限公司 | 数据生成方法、装置、计算机装置及存储介质 |
CN112417847A (zh) * | 2020-11-19 | 2021-02-26 | 湖南红网新媒体集团有限公司 | 一种新闻内容安全监测方法、***、装置和存储介质 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101901346A (zh) * | 2010-05-06 | 2010-12-01 | 复旦大学 | 一种对彩色数字图像进行不良内容识别的方法 |
CN103164687A (zh) * | 2011-12-15 | 2013-06-19 | 阿里巴巴集团控股有限公司 | 一种色情图像侦测的方法和*** |
CN105808610A (zh) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | 一种互联网图片过滤方法及装置 |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1178148C (zh) * | 2001-07-16 | 2004-12-01 | 友立资讯股份有限公司 | ***检查*** |
CN101447020B (zh) * | 2008-12-12 | 2011-07-06 | 北京理工大学 | 基于直觉模糊的色情图像识别方法 |
CN104281833B (zh) * | 2013-07-08 | 2018-12-18 | 深圳市腾讯计算机***有限公司 | 色情图像识别方法和装置 |
US9953425B2 (en) * | 2014-07-30 | 2018-04-24 | Adobe Systems Incorporated | Learning image categorization using related attributes |
CN104182735A (zh) * | 2014-08-18 | 2014-12-03 | 厦门美图之家科技有限公司 | 训练优化的基于卷积神经网络的色情图像或视频检测方法 |
CN104992177A (zh) * | 2015-06-12 | 2015-10-21 | 安徽大学 | 基于深层卷积神经网络的网络色情图像检测方法 |
CN106127108B (zh) * | 2016-06-14 | 2019-07-16 | 中国科学院软件研究所 | 一种基于卷积神经网络的人手图像区域检测方法 |
-
2017
- 2017-03-13 CN CN201710147708.3A patent/CN106951848B/zh active Active
- 2017-06-30 WO PCT/CN2017/091371 patent/WO2018166114A1/zh active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101901346A (zh) * | 2010-05-06 | 2010-12-01 | 复旦大学 | 一种对彩色数字图像进行不良内容识别的方法 |
CN103164687A (zh) * | 2011-12-15 | 2013-06-19 | 阿里巴巴集团控股有限公司 | 一种色情图像侦测的方法和*** |
CN105808610A (zh) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | 一种互联网图片过滤方法及装置 |
Non-Patent Citations (1)
Title |
---|
REN, DONG ET AL.: "A Survey on special video content detection algorithms", vol. 9, 30 September 2016 (2016-09-30), pages 184 - 191 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111046859B (zh) * | 2018-10-11 | 2023-09-29 | 杭州海康威视数字技术股份有限公司 | 字符识别方法及装置 |
CN111046859A (zh) * | 2018-10-11 | 2020-04-21 | 杭州海康威视数字技术股份有限公司 | 字符识别方法及装置 |
CN109657694A (zh) * | 2018-10-26 | 2019-04-19 | 平安科技(深圳)有限公司 | 图片自动分类方法、装置及计算机可读存储介质 |
CN109727238A (zh) * | 2018-12-27 | 2019-05-07 | 贵阳朗玛信息技术股份有限公司 | X光胸片的识别方法及装置 |
CN111460247B (zh) * | 2019-01-21 | 2022-07-01 | 重庆邮电大学 | 网络图片敏感文字自动检测方法 |
CN111460247A (zh) * | 2019-01-21 | 2020-07-28 | 重庆邮电大学 | 网络图片敏感文字自动检测方法 |
CN109947940B (zh) * | 2019-02-15 | 2023-09-05 | 平安科技(深圳)有限公司 | 文本分类方法、装置、终端及存储介质 |
CN109947940A (zh) * | 2019-02-15 | 2019-06-28 | 平安科技(深圳)有限公司 | 文本分类方法、装置、终端及存储介质 |
CN110287991A (zh) * | 2019-05-22 | 2019-09-27 | 平安科技(深圳)有限公司 | 植物类生药真伪验证方法、装置、计算机设备及存储介质 |
CN110287991B (zh) * | 2019-05-22 | 2023-12-19 | 平安科技(深圳)有限公司 | 植物类生药真伪验证方法、装置、计算机设备及存储介质 |
CN112907498A (zh) * | 2019-11-18 | 2021-06-04 | 中国商用飞机有限责任公司 | 一种孔隙识别方法、装置、设备及存储介质 |
CN112907498B (zh) * | 2019-11-18 | 2024-04-30 | 中国商用飞机有限责任公司 | 一种孔隙识别方法、装置、设备及存储介质 |
CN111402224A (zh) * | 2020-03-12 | 2020-07-10 | 广东电网有限责任公司广州供电局 | 一种用于电力设备的目标识别方法 |
CN111402224B (zh) * | 2020-03-12 | 2023-12-05 | 广东电网有限责任公司广州供电局 | 一种用于电力设备的目标识别方法 |
CN112597828A (zh) * | 2020-12-11 | 2021-04-02 | 京东数字科技控股股份有限公司 | 网页识别模型的训练方法、装置、网页识别的方法 |
CN114155417A (zh) * | 2021-12-13 | 2022-03-08 | 中国科学院空间应用工程与技术中心 | 图像目标的识别方法、装置、电子设备及计算机存储介质 |
Also Published As
Publication number | Publication date |
---|---|
CN106951848A (zh) | 2017-07-14 |
CN106951848B (zh) | 2018-12-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2018166114A1 (zh) | 图片识别的方法、***、电子装置及介质 | |
WO2019120115A1 (zh) | 人脸识别的方法、装置及计算机装置 | |
JP6994588B2 (ja) | 顔特徴抽出モデル訓練方法、顔特徴抽出方法、装置、機器および記憶媒体 | |
WO2017220032A1 (zh) | 基于深度学习的车牌分类方法、***、电子装置及存储介质 | |
WO2019169688A1 (zh) | 车辆定损方法、装置、电子设备及存储介质 | |
WO2022105118A1 (zh) | 基于图像的健康状态识别方法、装置、设备及存储介质 | |
US10839238B2 (en) | Remote user identity validation with threshold-based matching | |
WO2019024808A1 (zh) | 语义分割模型的训练方法和装置、电子设备、存储介质 | |
CN110321845B (zh) | 一种从视频中提取表情包的方法、装置及电子设备 | |
KR20190138712A (ko) | 배치 정규화 레이어들 | |
CN110889446A (zh) | 人脸图像识别模型训练及人脸图像识别方法和装置 | |
CN111598038B (zh) | 脸部特征点检测方法、装置、设备及存储介质 | |
CN112085701B (zh) | 一种人脸模糊度检测方法、装置、终端设备及存储介质 | |
WO2019136960A1 (zh) | 一种爬取网站数据的方法、存储介质、服务器及装置 | |
US20190347472A1 (en) | Method and system for image identification | |
CN114746898A (zh) | 用于生成图像抠图的三分图的方法和*** | |
KR20220100812A (ko) | 안면 생체 검출 방법, 장치, 전자 기기 및 저장 매체 | |
CN111126268B (zh) | 关键点检测模型训练方法、装置、电子设备及存储介质 | |
CN110909784A (zh) | 一种图像识别模型的训练方法、装置及电子设备 | |
CN115050064A (zh) | 人脸活体检测方法、装置、设备及介质 | |
US11893773B2 (en) | Finger vein comparison method, computer equipment, and storage medium | |
CN112966685A (zh) | 用于场景文本识别的攻击网络训练方法、装置及相关设备 | |
WO2021042544A1 (zh) | 基于去网纹模型的人脸验证方法、装置、计算机设备及存储介质 | |
WO2022126917A1 (zh) | 基于深度学习的人脸图像评估方法、装置、设备及介质 | |
CN114926322B (zh) | 图像生成方法、装置、电子设备和存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 17900540 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 09/12/2019) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 17900540 Country of ref document: EP Kind code of ref document: A1 |