WO2022052459A1 - 一种图像类别检测方法、***、电子设备及存储介质 - Google Patents

一种图像类别检测方法、***、电子设备及存储介质 Download PDF

Info

Publication number
WO2022052459A1
WO2022052459A1 PCT/CN2021/085892 CN2021085892W WO2022052459A1 WO 2022052459 A1 WO2022052459 A1 WO 2022052459A1 CN 2021085892 W CN2021085892 W CN 2021085892W WO 2022052459 A1 WO2022052459 A1 WO 2022052459A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
training
classification model
samples
model
Prior art date
Application number
PCT/CN2021/085892
Other languages
English (en)
French (fr)
Inventor
张莉
张正齐
王邦军
屈蕴倩
Original Assignee
苏州大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 苏州大学 filed Critical 苏州大学
Publication of WO2022052459A1 publication Critical patent/WO2022052459A1/zh

Links

Images

Classifications

    • 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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07D—HANDLING OF COINS OR VALUABLE PAPERS, e.g. TESTING, SORTING BY DENOMINATIONS, COUNTING, DISPENSING, CHANGING OR DEPOSITING
    • G07D7/00—Testing specially adapted to determine the identity or genuineness of valuable papers or for segregating those which are unacceptable, e.g. banknotes that are alien to a currency
    • G07D7/20—Testing patterns thereon
    • G07D7/2016—Testing patterns thereon using feature extraction, e.g. segmentation, edge detection or Hough-transformation

Definitions

  • the present application relates to the technical field of image processing, and in particular, to an image category detection method, a system, an electronic device, and a storage medium.
  • Image processing technology is a technology that uses a computer to process image information.
  • Image processing techniques may include image digitization, image enhancement and restoration, image data encoding, image segmentation, and image recognition, among others.
  • the purpose of this application is to provide an image category detection method, system, electronic device and storage medium, which can improve the accuracy of image classification.
  • the present application provides an image category detection method, the image category detection includes:
  • the classification model is trained by using the training samples; wherein, the classification model is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is a Fisher support vector machine based on L1 norm;
  • the training images in the training image set include positive example images and negative example images;
  • constructing a training sample according to the image feature of each training image in the training image set includes:
  • the negative example samples in the training samples are constructed according to the image features of the negative example images in the training image set.
  • obtaining the training image set includes:
  • the real banknote image is used as a positive image in the training image set, and the counterfeit banknote image is used as a negative image in the training image set.
  • construct training samples according to the image features of each training image in the training image set including:
  • the training samples are constructed according to image features and image categories of each of the training images.
  • constructing the training samples according to the image features and image categories of each of the training images includes:
  • the training samples are constructed according to the image features mapped to the target interval and the corresponding image categories.
  • the image features include variance, skewness, kurtosis and entropy of the image after wavelet transformation.
  • the method before using the training samples to train the classification model, the method further includes:
  • the preset model is is the classification model after training, ⁇ i is the coefficient of the classification model, k is the kernel function, b is the bias of the classification model, l is the total number of the training samples, x i is the ith training sample, x is the training sample substituted into the classification model;
  • the objective optimization problem is
  • the preset constraints are D y [K( ⁇ + - ⁇ - )+(b + -b - )1] ⁇ 1- ⁇ and ⁇ 0;
  • ⁇ K and ⁇ F are non-negative regularization parameters
  • K is the kernel matrix
  • N IG
  • I is the identity matrix
  • G is the preset matrix
  • Dy is the diagonal matrix
  • ⁇ is the slack variable
  • ⁇ ⁇ + - ⁇ -
  • b b + -b -
  • ⁇ + and ⁇ - is the auxiliary coefficient
  • b + and b - are the auxiliary biases.
  • the application also provides an image category detection system, the system includes:
  • a sample building module for obtaining a training image set, and constructing a training sample according to the image feature of each training image in the training image set;
  • a model training module used to train a classification model by using the training samples; wherein, the classification model is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is an L1 norm-based Fisher support vector machine;
  • the classification module is used to input the image to be detected into the trained classification model to obtain the image type of the image to be detected.
  • the present application also provides a storage medium on which a computer program is stored, and when the computer program is executed, the steps performed by the above-mentioned image category detection method are implemented.
  • the present application also provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and when the processor invokes the computer program in the memory, the steps performed by the above image category detection method are implemented.
  • the present application provides an image category detection method, including: acquiring a training image set, and constructing a training sample according to the image features of each training image in the training image set; using the training sample to train a classification model; wherein, the The classification model is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is a Fisher support vector machine based on the L1 norm; the image to be detected is input into the trained classification model to obtain the image to be detected. image type.
  • a training sample is constructed according to the image features of each training image in the training image set, and the training sample is used to train the classification model, so that the trained classification model can classify the images.
  • the classification model is a model constructed by the sparse Fisher support vector machine classification algorithm, and because the sparse Fisher support vector machine introduces the L1 norm to replace the original two norm, the classification model is sparse, and the classification model in this embodiment does not pay attention to training. Unimportant samples or noise samples in the samples have good generalization performance, so this application can improve the accuracy of image classification.
  • the present application also provides an image category detection system, an electronic device, and a storage medium, which have the above-mentioned beneficial effects, and will not be repeated here.
  • FIG. 1 is a flowchart of an image category detection method provided by an embodiment of the present application
  • FIG. 2 is a schematic structural diagram of an image category detection system provided by an embodiment of the present application.
  • FIG. 1 is a flowchart of an image category detection method provided by an embodiment of the present application.
  • S101 Obtain a training image set, and construct a training sample according to the image feature of each training image in the training image set;
  • a training image set can be obtained by taking images of the target object by a camera, and the training image set can include various types of training images.
  • the training images in the training image set include positive image and negative image.
  • real banknote images can be obtained by photographing real banknotes at a fixed position
  • counterfeit banknote images can be obtained by photographing counterfeit banknotes at the same fixed position.
  • Counterfeit banknote images are used as negative images in the training image set.
  • a training image set can also be obtained from a database, for example, images of cats of various breeds and non-cat images can be obtained from the database, and then the cat images can be used as positive images and the non-cat images can be used as negative images. This embodiment does not limit the types of training images in the obtained training image set, and can be flexibly set according to actual application scenarios.
  • the positive sample in the training sample can be constructed according to the image features of the positive image in the training image set; and the image feature of the negative image in the training image set can also be constructed.
  • a negative example sample in the training sample is constructed, and a training sample including a positive example sample and a negative example sample is obtained.
  • the classification model mentioned in this embodiment is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is an L1 norm-based Fisher support vector machine.
  • the traditional Fisher regularized support vector machine is not sparse.
  • a sparse Fisher support vector machine is obtained by introducing the L1 norm instead of the original two norm.
  • the training samples are used to train the classification model, and the trained classification model has the ability to identify image categories.
  • S103 Input the image to be detected into the trained classification model to obtain the image type of the image to be detected.
  • the image to be detected is an image of an unknown type, and by inputting the image to be detected into the trained classification model, the image type of the image to be detected can be determined according to the output result of the classification model.
  • a training sample is constructed according to the image features of each training image in the training image set, and the classification model is trained by using the training sample, so that the trained classification model can classify the images.
  • the classification model is a model constructed by the sparse Fisher support vector machine classification algorithm, and because the sparse Fisher support vector machine introduces the L1 norm to replace the original two norm, the classification model is sparse, and the classification model in this embodiment does not pay attention to training. Unimportant samples or noise samples in the samples have good generalization performance, so this embodiment can improve the accuracy of image classification.
  • a training sample may be constructed in the following manner: extracting image features of each of the training images in the training image set by means of wavelet transform extraction; The image features and image categories of the training images construct the training samples.
  • the above image features include the variance, skewness, kurtosis and entropy of the image after wavelet transformation.
  • the embodiment corresponding to FIG. 1 can also construct training samples in the following ways:
  • Step 1 extract the image features of each of the training images in the training image set by means of wavelet transform extraction
  • Step 2 map the image features of each of the training images to the target interval
  • Step 3 Construct the training samples according to the image features mapped to the target interval and the corresponding image categories.
  • this embodiment implements data standardization processing by mapping the image features to a unified target interval, so as to solve the comparability between the data. After the original data is standardized, each index is in the same order of magnitude, which is suitable for comprehensive comparative evaluation.
  • the model coefficients and offsets of the classification model can also be determined by solving the target optimization problem under preset constraints;
  • the preset model is is the classification model after training, ⁇ i is the coefficient of the classification model, b is the bias of the classification model, and l is the total number of training samples; x i is the ith training sample, x is the training sample that is substituted into the classification model, k is the kernel function, and i is the serial number of the training sample.
  • the preset constraint conditions are stD y [K( ⁇ + - ⁇ - )+(b + -b - )1] ⁇ 1- ⁇ and ⁇ 0;
  • min is the minimum value to be solved
  • ⁇ K and ⁇ F are non-negative regularization parameters
  • K is the kernel matrix
  • I is the identity matrix and G is the preset matrix.
  • D y is a diagonal matrix
  • ⁇ is the slack variable
  • L hinge ( ) is the hinge loss function
  • x i is the ith sample
  • ⁇ + and ⁇ - are the coefficients used to assist in the calculation of the classification model coefficients
  • b + and b - are the coefficients used to assist in the calculation of the classification model bias
  • T represents the transposed vector.
  • st is the abbreviation of subject to, and its meaning is to satisfy the following formula.
  • the hinge loss function is defined as follows:
  • the model coefficient ⁇ and bias b of the classification model can be obtained by the following formulas:
  • This embodiment also provides a system for discriminating genuine and counterfeit banknotes based on the sparse Fisher support vector machine.
  • the system can include a data preprocessing module, a data training module and a data prediction module.
  • the training samples need to be normalized; in the data training module, the classification model is trained with the training samples; in the data prediction module, the trained classification model is used to make predictions on the test set.
  • the number of features of each sample is d
  • l is the total number of samples in the training set. Normalize the data in the training set S, and map the sample points to the interval [0,1].
  • inequality constraints there are two sets of constraints to solve the above optimization problem, one is inequality constraints, the other is bound constraints; the number of inequality constraints and the number of bound constraints are as many as the number of samples, which is l.
  • the constraints are specifically stD y [K( ⁇ + - ⁇ - )+(b + -b - )1] ⁇ 1- ⁇ and ⁇ 0.
  • ⁇ K and ⁇ F are non-negative regularization parameters
  • G is the preset matrix
  • the feature vector of the banknotes to be predicted is obtained Then perform the data normalization operation described by the data preprocessing module, banknotes, otherwise the banknotes to be predicted are counterfeit banknotes.
  • sign is a sign function, and returns -1 for negative values and 1 for positive values.
  • the banknote authenticity detection scheme provided in this embodiment can speed up the classification efficiency and reduce the influence of error samples on the results, that is, improve the banknote authenticity detection scheme. Efficiency and accuracy of data classification. At present, with the upgrading of industrial cameras, more and more detailed data of banknotes are obtained. Under this situation, the accuracy of the classification of banknote data is also more prominent.
  • This embodiment overcomes the deficiencies of noise resistance and sparsity in the related art, and proposes a system for distinguishing genuine and counterfeit banknotes based on a sparse Fisher support vector machine, which adopts a sparse Fisher support vector machine classification algorithm. Due to the use of the L1 norm, the sparsity of the model is enhanced, the noise resistance is improved to a certain extent, and the generalization of the algorithm is further improved. Solving the optimization problem directly in the original space avoids the ill-conditioned situation when solving the dual space, maintains the sparseness of the model, and improves the classification accuracy.
  • the above-mentioned authentic and fake banknote discrimination system based on the sparse Fisher support vector machine will be described below through an embodiment in practical application.
  • This system has tested the Banknote (banknote) data set.
  • the data set uses an industrial camera to take pictures of banknotes at a fixed position to obtain a 400 ⁇ 400 training image, and then uses wavelet transform extraction to extract features from the image.
  • the latter data are used as training samples and test samples.
  • the dataset of training images contains a total of 1372 samples, 2/3 are randomly selected as training samples, and the remaining 1/3 are test samples. Each training sample and test sample contains 4 features.
  • the first column of features is the variance (continuous value) of the image after wavelet transformation
  • the second column of features is the skewness of the image after wavelet transformation (skewness) ( Continuous value)
  • the third column feature is the kurtosis (continuous value) of the image after wavelet transformation
  • the fourth column feature is the entropy (continuous value) of the image.
  • the numbers of genuine banknote samples and counterfeit banknote samples are 610 and 762 respectively.
  • the label of the genuine banknote sample is marked as +1
  • the label of the counterfeit banknote sample is marked as -1.
  • noise interference comparison is added, a certain proportion of samples are randomly selected for the entire data set, and the labels of these samples are set to opposite numbers.
  • ⁇ K and ⁇ F are non-negative regularization parameters
  • G is the preset matrix
  • the slack variable L hinge ( ) is the hinge loss function, which is defined as follows:
  • the feature vector is mapped to the interval [0,1], and then substituted into the discriminant function:
  • the authenticity of the banknote to be predicted can be authenticated according to the following rules: if it is 1, the banknotes to be predicted are genuine banknotes, otherwise the banknotes to be predicted are fake banknotes.
  • Table 1 lists the comparison between this embodiment and Fisher regularized support vector machine (FisherSVM) under certain parameter settings. It can be seen that under different parameter settings, this embodiment has better results than FisherSVM.
  • FisherSVM Fisher regularized support vector machine
  • FIG. 2 is a schematic structural diagram of an image category detection system provided by an embodiment of the present application.
  • the system can include:
  • a sample construction module 100 configured to obtain a training image set, and construct a training sample according to the image feature of each training image in the training image set;
  • the model training module 200 is used to train a classification model by using the training samples; wherein, the classification model is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is a Fisher support vector machine based on L1 norm Support Vector Machines;
  • the classification module 300 is configured to input the image to be detected into the trained classification model to obtain the image type of the image to be detected.
  • a training sample is constructed according to the image features of each training image in the training image set, and the classification model is trained by using the training sample, so that the trained classification model can classify the images.
  • the classification model is a model constructed by the sparse Fisher support vector machine classification algorithm, and because the sparse Fisher support vector machine introduces the L1 norm to replace the original two norm, the classification model is sparse, and the classification model in this embodiment does not pay attention to training. Unimportant samples or noise samples in the samples have good generalization performance, so this embodiment can improve the accuracy of image classification.
  • training images in the training image set include positive example images and negative example images
  • the sample building module 100 includes:
  • a positive example construction unit configured to construct a positive example sample in the training sample according to the image feature of the positive example image in the training image set
  • a negative example construction unit configured to construct a negative example sample in the training sample according to the image features of the negative example image in the training image set.
  • sample building module 100 includes:
  • an image capturing unit for capturing images of genuine banknotes and counterfeit banknotes at a fixed position
  • An image set construction unit configured to use the real banknote image as a positive example image in the training image set, and use the counterfeit banknote image as a negative example image in the training image set.
  • sample building module 100 includes:
  • a feature extraction unit configured to extract image features of each of the training images in the training image set by means of wavelet transform extraction
  • a training sample generating unit configured to construct the training sample according to the image feature and image category of each of the training images.
  • the training sample generation unit is configured to map the image features of each of the training images to a target interval; and to construct the training samples according to the image features mapped to the target interval and corresponding image categories.
  • the image features include variance, skewness, kurtosis and entropy of the image after wavelet transformation.
  • a parameter determination module configured to determine model coefficients and biases of the classification model by solving a target optimization problem under preset constraints before using the training samples to train the classification model;
  • the preset model is is the classification model after training
  • ⁇ i is the coefficient of the model
  • b is the bias of the model
  • l is the total number of samples in the training set
  • the objective optimization problem is The constraints of the objective optimization problem are stD y [K( ⁇ + - ⁇ - )+(b + -b - )1] ⁇ 1- ⁇ and ⁇ 0
  • ⁇ K and ⁇ F are non-negative regularization parameters
  • G is the preset matrix
  • the slack variable L hinge ( ) is the hinge loss function
  • x i is the ith sample
  • the present application also provides a storage medium on which a computer program is stored, and when the computer program is executed, the steps provided by the above embodiments can be implemented.
  • the storage medium may include: U disk, removable hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes.
  • the present application also provides an electronic device, which may include a memory and a processor, where a computer program is stored in the memory, and when the processor invokes the computer program in the memory, the steps provided in the above embodiments can be implemented.
  • the electronic device may also include various network interfaces, power supplies and other components.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种图像类别检测方法、***、电子设备及存储介质，该方法包括：获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型，上述方式能够提高图像分类准确性。

Description

一种图像类别检测方法、***、电子设备及存储介质
本申请要求于2020年09月14日提交中国专利局、申请号为202010961502.6、发明名称为“一种图像类别检测方法、***、电子设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及图像处理技术领域，特别涉及一种图像类别检测方法、***、一种电子设备及一种存储介质。
背景技术
图像处理技术是用计算机对图像信息进行处理的技术。图像处理技术可以包括图像数字化、图像增强和复原、图像数据编码、图像分割和图像识别等。
相关技术中，存在基于图像处理技术对图像进行分类的方案，例如存在利用Fisher正则化支持向量机对钞票的图像进行真伪分类的操作，以便实现真伪钞票鉴别。Fisher正则化支持向量机可以找到一个分类超平面并且同时最大化类间间隔和最小化类内散度，但是该Fisher正则化支持向量机构造的模型不具有稀疏性，也就是说该模型会受到训练样本中不重要的样本或噪声样本的影响，从而会导致分类器的泛化性能降低。
因此，如何提高图像分类准确性是本领域技术人员目前需要解决的技术问题。
发明内容
本申请的目的是提供一种图像类别检测方法、***、一种电子设备及一种存储介质，能够提高图像分类准确性。
为解决上述技术问题，本申请提供一种图像类别检测方法，该图像类别检测包括：
获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征 构建训练样本；
利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；
将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
可选的，所述训练图像集中的训练图像包括正例图像和负例图像；
相应的，根据所述训练图像集中每一训练图像的图像特征构建训练样本包括：
根据所述训练图像集中的正例图像的图像特征构建所述训练样本中的正例样本；
根据所述训练图像集中的负例图像的图像特征构建所述训练样本中的负例样本。
可选的，获取训练图像集包括：
在固定位置拍摄真钞图像和伪钞图像；
将所述真钞图像作为所述训练图像集中的正例图像，并将所述伪钞图像作为所述训练图像集中的负例图像。
可选的，根据所述训练图像集中每一训练图像的图像特征构建训练样本，包括：
通过子波变换提取的方式提取所述训练图像集中每一所述训练图像的图像特征；
根据每一所述训练图像的图像特征和图像类别构建所述训练样本。
可选的，所述根据每一所述训练图像的图像特征和图像类别构建所述训练样本，包括：
对每一所述训练图像的图像特征映射至目标区间；
根据映射至所述目标区间的图像特征和对应的图像类别构建所述训练样本。
可选的，所述图像特征包括图像经小波变换后的方差、偏态、峰度和图像的熵。
可选的，在利用所述训练样本对分类模型进行训练之前，还包括：
在预设约束条件下通过求解目标优化问题确定所述分类模型的模型系数和偏置；
其中，所述预设模型为
Figure PCTCN2021085892-appb-000001
为训练后的分类模型，α i为所述分类模型的系数，k为核函数，b为所述分类模型的偏置，l为所述训练样本的总数，x i为第i个训练样本，x为代入所述分类模型的训练样本；
所述目标优化问题为
Figure PCTCN2021085892-appb-000002
所述预设约束条件为D y[K(α +-α -)+(b +-b -)1]≥1-ξ和ξ≥0；γ K和γ F为非负正则化参数，K为核矩阵，N＝I-G，I为单位矩阵，G为预设矩阵，D y为对角矩阵，ξ为松弛变量，α＝α +-α -，b＝b +-b -，α +和α -为辅助系数，b +和b -为辅助偏置。
本申请还提供了一种图像类别检测***，该***包括：
样本构建模块，用于获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；
模型训练模块，用于利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；
分类模块，用于将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
本申请还提供了一种存储介质，其上存储有计算机程序，所述计算机程序执行时实现上述图像类别检测方法执行的步骤。
本申请还提供了一种电子设备，包括存储器和处理器，所述存储器中存储有计算机程序，所述处理器调用所述存储器中的计算机程序时实现上述图像类别检测方法执行的步骤。
本申请提供了一种图像类别检测方法，包括：获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
本申请在获取训练图像集后根据训练图像集中每一训练图像的图像特征构建训练样本，利用训练样本对分类模型进行训练，使训练后的分类模型能够对图像进行分类。分类模型为稀疏Fisher支持向量机分类算法构建的模型，且由于稀疏Fisher支持向量机引入L1范数代替原来的二范数，使分类模型具有稀疏性，本实施例中的分类模型不会关注训练样本中不重要的样本或噪声样本，具有良好的泛化性能，因此本申请能够提高图像分类准确性。本申请同时还提供了一种图像类别检测***、一种电子设备和一种存储介质，具有上述有益效果，在此不再赘述。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例所提供的一种图像类别检测方法的流程图；
图2为本申请实施例所提供的一种图像类别检测***的结构示意图。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描 述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
下面请参见图1，图1为本申请实施例所提供的一种图像类别检测方法的流程图。
具体步骤可以包括：
S101：获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；
其中，本步骤可以通过相机对目标物体拍摄图像得到训练图像集，训练图像集中可以包括多种类别的训练图像，例如训练图像集中的训练图像包括正例图像和负例图像。在具体应用中，可以在固定位置拍摄真钞得到真钞图像，在同样的固定位置拍摄伪钞得到伪钞图像，进而将所述真钞图像作为所述训练图像集中的正例图像，并将所述伪钞图像作为所述训练图像集中的负例图像。本实施例也可以从数据库中获取训练图像集，例如可以从数据库中获取各种品种的猫图像以及非猫图像，进而可以将猫图像作为正例图像，将非猫图像作为负例图像。本实施例不限定获取的训练图像集中训练图像的种类，可以根据实际应用场景灵活设置。
进一步的，在得到训练图像集后，可以根据所述训练图像集中的正例图像的图像特征构建所述训练样本中的正例样本；还可以根据所述训练图像集中的负例图像的图像特征构建所述训练样本中的负例样本，进而得到包括正例样本和负例样本的训练样本。
S102：利用所述训练样本对分类模型进行训练；
其中，本实施例中提到的分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机。传统的Fisher正则化支持向量机是不稀疏的，本实施例通过引入L1范数代替原来的二范数，得到了稀疏Fisher支持向量机。进一步的，本实施例利用训练样本对分类模型进行训练，训练后的分类模型具有识别图像类别的能力。
S103：将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
其中，待检测图像为未知类型的图像，通过将待检测图像输入训练后的分类模型，可以根据分类模型的输出结果确定待检测图像的图像类型。
本实施例在获取训练图像集后根据训练图像集中每一训练图像的图像特征构建训练样本，利用训练样本对分类模型进行训练，使训练后的分类模型能够对图像进行分类。分类模型为稀疏Fisher支持向量机分类算法构建的模型，且由于稀疏Fisher支持向量机引入L1范数代替原来的二范数，使分类模型具有稀疏性，本实施例中的分类模型不会关注训练样本中不重要的样本或噪声样本，具有良好的泛化性能，因此本实施例能够提高图像分类准确性。
作为对于图1对应实施例的进一步介绍，本实施例可以通过以下方式构建训练样本：通过子波变换提取的方式提取所述训练图像集中每一所述训练图像的图像特征；根据每一所述训练图像的图像特征和图像类别构建所述训练样本。上述图像特征包括图像经小波变换后的方差、偏态、峰度和图像的熵。
作为进一步的实施方式，图1对应的实施例还可以通过以下方式构建训练样本：
步骤1：通过子波变换提取的方式提取所述训练图像集中每一所述训练图像的图像特征；
步骤2：对每一所述训练图像的图像特征映射至目标区间；
步骤3：根据映射至所述目标区间的图像特征和对应的图像类别构建所述训练样本。
其中，由于不同的图像特征往往具有不同的量纲和量纲单位，这样的情况会影响到数据分析的结果。为了消除特征之间的量纲影响，本实施例通过将图像特征映射至统一的目标区间实现数据标准化处理，以解决数据之间的可比性。原始数据经过数据标准化处理后，各指标处于同一数量级，适合进行综合对比评价。
作为对于图1对应实施例的进一步介绍，在S102利用所述训练样本对分类模型进行训练之前，还可以在预设约束条件下通过求解目标优化问题确定所述分类模型的模型系数和偏置；
其中，所述预设模型为
Figure PCTCN2021085892-appb-000003
为训练后的分类模型，α i为分类模型的系数，b为分类模型的偏置，l为训练样本的总数；
Figure PCTCN2021085892-appb-000004
x i为第i个训练样本，x为代入分类模型的训练样本，k为核函数，i为训练样本的序号。
所述目标优化问题为
Figure PCTCN2021085892-appb-000005
所述预设约束条件为s.t.D y[K(α +-α -)+(b +-b -)1]≥1-ξ和ξ≥0；
其中，min为求解最小值，γ K和γ F为非负正则化参数，K为核矩阵，核矩阵K的第i行第j列为[K] ij＝k(x i,x j),i,j＝1,2,...,l。
Figure PCTCN2021085892-appb-000006
I是单位矩阵，G为预设矩阵。D y为对角矩阵，对角矩阵D y的第i行第i列为[D y] ii＝y i，ξ为松弛变量，第i个训练样本的松弛变量
Figure PCTCN2021085892-appb-000007
L hinge(·)为合页损失函数，x i为第i个样本，
Figure PCTCN2021085892-appb-000008
表示d维的实数空间，α +和α -为用来辅助计算分类模型系数的系数，b +和b -为用于辅助计算分类模型偏置的系数，T表示转置向量。s.t.为subject to的缩写，其含义为使得后面的公式满足。合页损失函数其定义如下：
Figure PCTCN2021085892-appb-000009
分类模型的模型系数α和偏置b可以通过以下公式得到：
α＝α +-α -；
b＝b +-b -。
本实施例还提供一种基于稀疏Fisher支持向量机的真伪钞票判别系 统。本***可以包括数据预处理模块、数据训练模块和数据预测模块组成。在数据预处理模块中，需要将训练样本归一化；在数据训练模块中，用训练样本对分类模型进行训练；在数据预测模块，用训练好的分类模型在测试集上进行预测。
一、数据预处理模块
首先以固定位置利用工业照相机对纸币进行拍照获得训练图像，然后利用特征提取器对训练图像进行图像特征提取，将提取的图像特征作为本***的训练集(即上文提到的训练样本)。训练集
Figure PCTCN2021085892-appb-000010
其中
Figure PCTCN2021085892-appb-000011
y i∈{±1}。标签为y i＝1的是真钞数据集合，标签为y i＝-1的是伪钞数据集合，每个样本的特征数为d，l为训练集样本总数。对训练集S中的数据进行归一化，将样本点映射到区间[0,1]中。
二、数据训练模块
本实施例确定的分类模型为：
Figure PCTCN2021085892-appb-000012
其中α i为模型的系数，b为模型的偏置。为获得模型的系数和偏置，需求解如下的优化问题：
Figure PCTCN2021085892-appb-000013
即解决上述优化问题的时有两组约束，一组是不等式约束，另一组是界约束；不等式约束的个数以及界约束的个数都和样本数一样多，为l个。约束具体为s.t.D y[K(α +-α -)+(b +-b -)1]≥1-ξ和ξ≥0。
γ K和γ F为非负正则化参数，核矩阵K的第i行第j列为[K] ij＝k(x i,x j),i,j＝1,2,...,l，N＝I-G,
Figure PCTCN2021085892-appb-000014
是单位矩阵，G为预设矩阵，对角矩阵D y的第i行第i列为[D y] ii＝y i，松弛变量
Figure PCTCN2021085892-appb-000015
L hinge(·)为合页损失函数其定义如下：
Figure PCTCN2021085892-appb-000016
求解上述优化问题之后，得到α＝α +-α -和b＝b +-b -，从而可以确定分类模型。
三、数据预测模块
对待预测钞票，按照数据预处理模块处理过程后，获得待预测钞票的特征向量
Figure PCTCN2021085892-appb-000017
然后执行数据预处理模块描述的数据归一化操作，将特
Figure PCTCN2021085892-appb-000018
钞，否则待预测的钞票为***。
Figure PCTCN2021085892-appb-000019
sign为取符号函数，负值返回-1，正值返回1。
由于图像样本中难免会存在由于人工失误、仪器测量误差等产生的噪音样本来干扰最终的实验结果，而那些不具有代表性样本的存在会造成数据冗余、降低训练效率。本实施例在训练钞票数据中有较多不具有代表性的样本的情况下，本实施例提供的钞票真伪检测方案可以加快分类效率、减小误差样本对结果的影响，即提高了对钞票数据分类的效率和准确率。目前随着工业照相机的升级，获取的钞票细节数据也越来越多，在这种形势下，本钞票数据分类的准确率也更为突出。
本实施例克服了相关技术中抗噪性和稀疏性的不足的问题，提出一种基于稀疏Fisher支持向量机的真伪钞票判别***，该***采用了稀疏Fisher支持向量机分类算法。由于L1范数的使用，增强了模型的稀疏性，一定程度上提高了抗噪性，进而提高了算法的泛化性。在原空间直接求解优化问题，避免了在对偶空间求解时存在的病态情况，也保持了模型的稀疏性，提高了分类准确度。
下面通过实际应用中的实施例说明上述基于稀疏Fisher支持向量机的真伪钞票判别***。本***对Banknote(钞票)数据集进行了测试，该数据集以固定位置利用工业照相机对纸币进行拍照，获得400×400的训练图像，然后利用子波变换提取对图像进行特征提取，对特征提取后的数据作为的训练样本和测试样本。训练图像的数据集一共包含1372个样本，随机选取2/3作为训练样本，剩余1/3为测试样本。每个训练样本和测试样本均包含4个特征，第一列特征是图像经小波变换后的方差(variance)(连续值)，第二列特征是图像经小波变换后的偏态(skewness)(连续值)，第三列特征是图像经小波变换后的峰度(curtosis)(连续值)，第四列特征是图像的熵(entropy)(连续值)。其中真钞样本和伪钞样本数量分别为610和762，这里 本实施例把真钞样本的标签记为+1，***样本的标签记为-1。
为了证明本实施例对噪声干扰的不敏感性，针对有标签噪声的样本，具有良好的分类性能和稳定性。本实施例增加噪声干扰对比，对整个数据集随机选取一定比例的样本，将这些样本的标签置为相反数。本实例对所有原始样本选取r＝{0％,5％,10％}的噪声样本(r＝0％则为原始样本)。对一个原始样本和两个标签噪声样本分别进行如下步骤。
具体实施过程如下：
一、数据预处理模块
输入特征提取后的样本集
Figure PCTCN2021085892-appb-000020
其中
Figure PCTCN2021085892-appb-000021
y i∈{±1}。标签为y i＝1的是真钞数据集合，标签为y i＝-1的是伪钞数据集合。对数据集S中的数据进行归一化，将样本点映射到区间[0,1]中。
二、数据训练模块
本实施例的分类模型为：
Figure PCTCN2021085892-appb-000022
其中α i为模型的系数，b为模型的偏置。为获得模型的系数和偏置，需求解如下的优化问题：
Figure PCTCN2021085892-appb-000023
其中，γ K和γ F为非负正则化参数，核矩阵K的第i行第j列为[K] ij＝k(x i,x j),i,j＝1,2,...,l，N＝I-G,
Figure PCTCN2021085892-appb-000024
是单位矩阵，G为预设矩阵，对角矩阵D y的第i行第i列为[D y] ii＝y i，松弛变量
Figure PCTCN2021085892-appb-000025
L hinge(·)为合页损失函数其定义如下：
Figure PCTCN2021085892-appb-000026
求解上述优化问题之后，得到α＝α +-α -和b＝b +-b -，从而可以确定判别函数。
三、数据预测模块
对458个剩余的钞票数据，将其特征向量映射到区间[0,1]中，接着代 入判别函数：
Figure PCTCN2021085892-appb-000027
Figure PCTCN2021085892-appb-000028
可以按照下述规则鉴定待预测钞票的真伪：若
Figure PCTCN2021085892-appb-000029
为1，则待预测的钞票为真钞，否则待预测的钞票为***。
表1列出了本实施例在某些参数设置下与Fisher正则化支持向量机(FisherSVM)的对比。可以看出在不同参数设置下，本实施例均比FisherSVM有较好的结果。
表1 Banknote数据集上噪声干扰的准确率对比
参数 本实施例准确度(％) FisherSVM准确度(％)
r＝0％ 88.06 86.05
r＝5％ 87.66 85.90
r＝10％ 86.94 85.70
请参见图2，图2为本申请实施例所提供的一种图像类别检测***的结构示意图；
该***可以包括：
样本构建模块100，用于获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；
模型训练模块200，用于利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；
分类模块300，用于将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
本实施例在获取训练图像集后根据训练图像集中每一训练图像的图像特征构建训练样本，利用训练样本对分类模型进行训练，使训练后的分类模型能够对图像进行分类。分类模型为稀疏Fisher支持向量机分类算法构 建的模型，且由于稀疏Fisher支持向量机引入L1范数代替原来的二范数，使分类模型具有稀疏性，本实施例中的分类模型不会关注训练样本中不重要的样本或噪声样本，具有良好的泛化性能，因此本实施例能够提高图像分类准确性。
进一步的，所述训练图像集中的训练图像包括正例图像和负例图像；
相应的，样本构建模块100包括：
正例构建单元，用于根据所述训练图像集中的正例图像的图像特征构建所述训练样本中的正例样本；
负例构建单元，用于根据所述训练图像集中的负例图像的图像特征构建所述训练样本中的负例样本。
进一步的，样本构建模块100包括：
图像拍摄单元，用于在固定位置拍摄真钞图像和伪钞图像；
图像集构建单元，用于将所述真钞图像作为所述训练图像集中的正例图像，并将所述伪钞图像作为所述训练图像集中的负例图像。
进一步的，样本构建模块100包括：
特征提取单元，用于通过子波变换提取的方式提取所述训练图像集中每一所述训练图像的图像特征；
训练样本生成单元，用于根据每一所述训练图像的图像特征和图像类别构建所述训练样本。
进一步的，训练样本生成单元用于对每一所述训练图像的图像特征映射至目标区间；用于根据映射至所述目标区间的图像特征和对应的图像类别构建所述训练样本。
进一步的，所述图像特征包括图像经小波变换后的方差、偏态、峰度和图像的熵。
进一步的，还包括：
参数确定模块，用于在利用所述训练样本对分类模型进行训练之前，在预设约束条件下通过求解目标优化问题确定所述分类模型的模型系数和偏置；
其中，所述预设模型为
Figure PCTCN2021085892-appb-000030
为训练后的分类模型，α i为模型的系数，b为模型的偏置，l为训练集样本总数；所述目标优化问题为
Figure PCTCN2021085892-appb-000031
所述目标优化问题的约束条件为s.t.D y[K(α +-α -)+(b +-b -)1]≥1-ξ和ξ≥0；γ K和γ F为非负正则化参数，核矩阵K的第i行第j列为[K] ij＝k(x i,x j),i,j＝1,2,...,l，N＝I-G,
Figure PCTCN2021085892-appb-000032
是单位矩阵，G为预设矩阵，对角矩阵D y的第i行第i列为[D y] ii＝y i，松弛变量
Figure PCTCN2021085892-appb-000033
L hinge(·)为合页损失函数，x i为第i个样本，
Figure PCTCN2021085892-appb-000034
表示d维的实数空间。
由于***部分的实施例与方法部分的实施例相互对应，因此***部分的实施例请参见方法部分的实施例的描述，这里暂不赘述。
本申请还提供了一种存储介质，其上存有计算机程序，该计算机程序被执行时可以实现上述实施例所提供的步骤。该存储介质可以包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
本申请还提供了一种电子设备，可以包括存储器和处理器，所述存储器中存有计算机程序，所述处理器调用所述存储器中的计算机程序时，可以实现上述实施例所提供的步骤。当然所述电子设备还可以包括各种网络接口，电源等组件。
说明书中各个实施例采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似部分互相参见即可。对于实施例公开的***而言，由于其与实施例公开的方法相对应，所以描述的比较简单，相关之处参见方法部分说明即可。应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以对本申请进行若干改进和修饰，这些改进和修饰也落入本申请权利要求的保护范 围内。
还需要说明的是，在本说明书中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的状况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。

Claims (10)

  1. 一种图像类别检测方法，其特征在于，包括：
    获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；
    利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；
    将待检测图像输入训练后的分类模型，得到所述待检测图像的图像类型。
  2. 根据权利要求1所述图像类别检测方法，其特征在于，所述训练图像集中的训练图像包括正例图像和负例图像；
    相应的，根据所述训练图像集中每一训练图像的图像特征构建训练样本包括：
    根据所述训练图像集中的正例图像的图像特征构建所述训练样本中的正例样本；
    根据所述训练图像集中的负例图像的图像特征构建所述训练样本中的负例样本。
  3. 根据权利要求2所述图像类别检测方法，其特征在于，所述获取训练图像集包括：
    在固定位置拍摄真钞图像和伪钞图像；
    将所述真钞图像作为所述训练图像集中的正例图像，并将所述伪钞图像作为所述训练图像集中的负例图像。
  4. 根据权利要求1所述图像类别检测方法，其特征在于，根据所述训练图像集中每一训练图像的图像特征构建训练样本，包括：
    通过子波变换提取的方式提取所述训练图像集中每一所述训练图像的图像特征；
    根据每一所述训练图像的图像特征和图像类别构建所述训练样本。
  5. 根据权利要求4所述图像类别检测方法，其特征在于，所述根据每一所述训练图像的图像特征和图像类别构建所述训练样本，包括：
    对每一所述训练图像的图像特征映射至目标区间；
    根据映射至所述目标区间的图像特征和对应的图像类别构建所述训练样本。
  6. 根据权利要求4所述图像类别检测方法，其特征在于，所述图像特征包括图像经小波变换后的方差、偏态、峰度和图像的熵。
  7. 根据权利要求1所述图像类别检测方法，其特征在于，在利用所述训练样本对分类模型进行训练之前，还包括：
    在预设约束条件下通过求解目标优化问题确定所述分类模型的模型系数和偏置；
    其中，所述预设模型为
    Figure PCTCN2021085892-appb-100001
    Figure PCTCN2021085892-appb-100002
    为训练后的分类模型，α i为所述分类模型的系数，k为核函数，b为所述分类模型的偏置，l为所述训练样本的总数，x i为第i个训练样本，x为代入所述分类模型的训练样本；
    所述目标优化问题为
    Figure PCTCN2021085892-appb-100003
    所述预设约束条件为D y[K(α +-α -)+(b +-b -)1]≥1-ξ和ξ≥0；γ K和γ F为非负正则化参数，K为核矩阵，N＝I-G，I为单位矩阵，G为预设矩阵，D y为对角矩阵，ξ为松弛变量，α＝α +-α -，b＝b +-b -，α +和α -为辅助系数，b +和b -为辅助偏置。
  8. 一种图像类别检测***，其特征在于，包括：
    样本构建模块，用于获取训练图像集，并根据所述训练图像集中每一训练图像的图像特征构建训练样本；
    模型训练模块，用于利用所述训练样本对分类模型进行训练；其中，所述分类模型为稀疏Fisher支持向量机分类算法构建的模型，所述稀疏Fisher支持向量机为基于L1范数的Fisher支持向量机；
    分类模块，用于将待检测图像输入训练后的分类模型，得到所述待检 测图像的图像类型。
  9. 一种电子设备，其特征在于，包括存储器和处理器，所述存储器中存储有计算机程序，所述处理器调用所述存储器中的计算机程序时实现如权利要求1至7任一项所述图像类别检测方法的步骤。
  10. 一种存储介质，其特征在于，所述存储介质中存储有计算机可执行指令，所述计算机可执行指令被处理器加载并执行时，实现如上权利要求1至7任一项所述图像类别检测方法的步骤。
PCT/CN2021/085892 2020-09-14 2021-04-08 一种图像类别检测方法、***、电子设备及存储介质 WO2022052459A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010961502.6A CN112085112A (zh) 2020-09-14 2020-09-14 一种图像类别检测方法、***、电子设备及存储介质
CN202010961502.6 2020-09-14

Publications (1)

Publication Number Publication Date
WO2022052459A1 true WO2022052459A1 (zh) 2022-03-17

Family

ID=73738030

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/085892 WO2022052459A1 (zh) 2020-09-14 2021-04-08 一种图像类别检测方法、***、电子设备及存储介质

Country Status (2)

Country Link
CN (1) CN112085112A (zh)
WO (1) WO2022052459A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112085112A (zh) * 2020-09-14 2020-12-15 苏州大学 一种图像类别检测方法、***、电子设备及存储介质
CN113627556B (zh) * 2021-08-18 2023-03-24 广东电网有限责任公司 一种图像分类的实现方法、装置、电子设备和存储介质
CN114140699B (zh) * 2021-11-25 2023-02-10 苏州大学 基于在线局部Fisher矢量机的SAR图像目标识别方法及***

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104915684A (zh) * 2015-06-30 2015-09-16 苏州大学 一种基于鲁棒多平面支持向量机的图像识别方法及装置
CN105975983A (zh) * 2016-04-28 2016-09-28 西安电子科技大学 基于稀疏Fisher矢量编码的数据分类方法
CN106066994A (zh) * 2016-05-24 2016-11-02 北京工业大学 一种基于Fisher判别的稀疏表示的人脸识别方法
CN108805054A (zh) * 2018-05-29 2018-11-13 苏州大学 一种人脸图像分类方法、***、设备及计算机存储介质
CN112085112A (zh) * 2020-09-14 2020-12-15 苏州大学 一种图像类别检测方法、***、电子设备及存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020009877A1 (en) * 2018-07-02 2020-01-09 Stowers Institute For Medical Research Facial image recognition using pseudo-images
CN109754001B (zh) * 2018-12-21 2023-06-27 中国科学院苏州纳米技术与纳米仿生研究所 图像分类方法、计算机存储介质及图像分类装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104915684A (zh) * 2015-06-30 2015-09-16 苏州大学 一种基于鲁棒多平面支持向量机的图像识别方法及装置
CN105975983A (zh) * 2016-04-28 2016-09-28 西安电子科技大学 基于稀疏Fisher矢量编码的数据分类方法
CN106066994A (zh) * 2016-05-24 2016-11-02 北京工业大学 一种基于Fisher判别的稀疏表示的人脸识别方法
CN108805054A (zh) * 2018-05-29 2018-11-13 苏州大学 一种人脸图像分类方法、***、设备及计算机存储介质
CN112085112A (zh) * 2020-09-14 2020-12-15 苏州大学 一种图像类别检测方法、***、电子设备及存储介质

Also Published As

Publication number Publication date
CN112085112A (zh) 2020-12-15

Similar Documents

Publication Publication Date Title
WO2022052459A1 (zh) 一种图像类别检测方法、***、电子设备及存储介质
WO2021121127A1 (zh) 样本类别识别方法、装置、计算机设备及存储介质
WO2020077895A1 (zh) 签约意向判断方法、装置、计算机设备和存储介质
CN111353491B (zh) 一种文字方向确定方法、装置、设备及存储介质
US20230030792A1 (en) Document authenticity identification method and apparatus, computer-readable medium, and electronic device
US20120328197A1 (en) Identifying matching images
Macarulla Rodriguez et al. Likelihood ratios for deep neural networks in face comparison
CN102722708A (zh) 一种薄片类介质分类方法及其装置
WO2022134580A1 (zh) 证件信息的获取方法及装置、存储介质、计算机设备
CN114971294A (zh) 数据采集方法、装置、设备及存储介质
WO2024051427A1 (zh) 硬币识别方法、***及存储介质
KR101912433B1 (ko) 영상 분석 기반 위조 지문 판별 장치 및 방법
CN113378609A (zh) 代理商代签名的识别方法及装置
WO2023070843A1 (zh) 多模态货币识别***、方法、存储介质和终端
CN114677333A (zh) 一种基于直方图的图像对比度增强检测方法
CN112464675B (zh) 一种言辞矛盾检测方法、装置、设备及存储介质
Sundravadivelu et al. Extensive Analysis of IoT Assisted Fake Currency Detection using Novel Learning Scheme
Li et al. Shadow determination and compensation for face recognition
Tong et al. A novel classification algorithm for new and used banknotes
Chennupati A structured approach to JPEG tampering detection using enhanced fusion algorithm
CN111160245A (zh) 一种动态签名识别方法及设备
CN117648409B (zh) 一种基于ocr的版式文件防伪识别方法
US20230316795A1 (en) Auto-Document Detection & Capture
CN112507137B (zh) 开放环境下基于粒度感知的少样本关系抽取方法及应用
Salih A Suggested System for Palmprint Recognition Using Curvelet Transform and Co-Occurrence Matrix

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: 21865522

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21865522

Country of ref document: EP

Kind code of ref document: A1