Fully supervised learning image classification method based on adaptive and margin repulsion loss

Abstract

In view of the limitations of the prior art, the present invention proposes a fully supervised learning image classification method based on adaptation and margin repulsion loss, which effectively expands the inter-class separation by adaptively controlling the misclassification loss; takes into account the most confusing classes so that the classifier can better distinguish different classes; the time complexity of this method is almost the same as that of the method based on intra-class margins; the model iteration process involved is more efficient, and the complementarity with the existing intra-class margin-based method and its superior performance are clearly demonstrated in the performance of image classification and face recognition applications.

Description

Technical Field

[0001] The present invention relates to the field of machine vision technology, in particular, to pattern recognition and image classification technology therein; more specifically, to a fully supervised learning image classification method based on adaptation and interval repulsion loss. Background Art

[0002] In image classification and recognition, extracting image features through deep convolutional neural networks (CNNs) is one of the most popular methods to improve classification accuracy. In order to enhance the generalization ability of CNNs and avoid overfitting, related works have proposed data augmentation and regularization as well as deeper neural networks, and have good performance in image recognition. With the increasing number of network layers, deep convolutional networks have continued to develop from early works to the current mainstream methods (such as ResNet, GoogleNet, and DenseNet). In ResNet, the scale of CNNs can reach more than 1,000 layers, and it performs well on public datasets, with strong generalization ability and powerful expression ability.

[0003] As the optimization of CNNs structure has achieved excellent representation capabilities, how to design an effective objective function loss has become a hot research topic. The Chinese invention application published on May 25, 2021: In an image classification method, a category center update strategy based on random batches is designed, and only the inter-class distance is calculated for the categories of samples in the random batches, and a dispersion loss is constructed accordingly, in order to allow the model to learn feature distributions with larger inter-class distances.

[0004] However, existing technologies focus directly on how to reduce the intra-class distance without specific constraints on inter-class separation. When the distance between a sample's feature vector and the weight vector of its class is small, the value of the loss function can be small even if the sample is similar to the wrong class. If there is no explicit loss to make the inter-class distance larger, the weight vectors of different classes may be close to each other, making it difficult for the learned classifier to accurately distinguish the class of the sample. Therefore, existing technologies still have certain limitations. Summary of the invention

[0005] In view of the limitations of the prior art, the present invention proposes a fully supervised learning image classification method based on adaptation and interval repulsion loss. The technical solution adopted by the present invention is:

[0006] A fully supervised learning image classification method based on adaptive and margin repulsion loss, comprising the following steps:

[0007] S1, obtain the image to be recognized;

[0008] S2, inputting the image to be identified into a preset image classification model to obtain a classification result of the image to be identified;

[0009] The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing category.

[0010] Compared with the prior art, the present invention provides a method for effectively expanding inter-class separation by adaptively controlling misclassification loss; the most confusing class is taken into account so that the classifier can better distinguish different classes; the time complexity of this method is almost the same as that of the intra-class interval-based method; the model iteration process involved is more efficient, and its complementarity with the existing intra-class interval-based method and its superior performance are clearly demonstrated in the performance of image classification and face recognition applications.

[0011] As a preferred solution, the objective function is expressed by the following formula:

[0012] L=L s +L f ;

[0013] Among them, L s represents the adaptive repulsion loss; L f Indicates interval repulsion loss.

[0014] Furthermore, the adaptive repulsion loss L s It is expressed as follows:

[0015]

[0016] Where N represents the number of samples sampled in one iteration during the training process; θ represents the angle between the feature vector and the weight vector; y i is the category label of the i-th sample; j represents the j-th classifier; m1, m2, and m3 are hyperparameters for controlling the intra-class interval size of the intra-class loss; Based on θ j and A function that adaptively increases the distance between classes; s is a hyperparameter used to control the radius of the hypersphere.

[0017] Furthermore, the j and Function that adaptively increases the distance between classes It is expressed as follows:

[0018]

[0019] Furthermore, the interval repulsion loss is expressed by the following formula:

[0020]

[0021] Where N is the number of samples sampled in one iteration during the training process; j is the jth classifier; W is the weight vector; W T x=||W||·||x||·cosθ; x represents the feature vector of the sample; i represents the i-th sample; y i is the category label of the i-th sample; θ represents the angle between the feature vector and the weight vector; m is a scalar used to determine whether the most confusing class needs to be separated.

[0022] Furthermore, the deep convolutional neural network is a deep convolutional neural network based on an intra-class method.

[0023] Furthermore, the in-class method is Sphereface or Cosface or Arcface.

[0024] The present invention also provides the following contents:

[0025] A fully supervised learning image classification system based on adaptive and interval repulsion loss includes an image acquisition module to be identified and a model identification module connected to each other; wherein:

[0026] The to-be-recognized image acquisition module is used to acquire the to-be-recognized image;

[0027] The model recognition module is used to input the image to be recognized into a preset image classification model to obtain a classification result of the image to be recognized;

[0028] The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing category.

[0029] A storage medium stores a computer program, which, when executed by a processor, implements the steps of the aforementioned fully supervised learning image classification method based on adaptation and interval repulsion loss.

[0030] A computer device comprises a storage medium, a processor and a computer program stored in the storage medium and executable by the processor, wherein when the computer program is executed by the processor, the steps of the aforementioned fully supervised learning image classification method based on adaptation and interval repulsion loss are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic flow chart of a fully supervised learning image classification method based on adaptation and interval repulsion loss provided by the present invention;

[0032] Figure 2 A schematic diagram of the training process of the deep convolutional neural network provided by the present invention;

[0033] Figure 3 A schematic diagram of the principle of RepulFace provided by the present invention;

[0034] Figure 4 is the most confusing class example of sample x;

[0035] Figure 5 Schematic diagram of the fully supervised learning image classification system based on adaptation and interval repulsion loss provided by the present invention. DETAILED DESCRIPTION

[0036] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0037] It should be clear that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0038] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0039] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0040] In addition, in the description of this application, unless otherwise specified, "multiple" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0041] In order to solve the limitations of the prior art, this embodiment provides a technical solution, and the technical solution of the present invention is further described below in conjunction with the drawings and embodiments.

[0042] Example 1

[0043] Please refer to Figure 1 , a fully supervised learning image classification method based on adaptive and margin repulsion loss, comprising the following steps:

[0044] S1, obtain the image to be recognized;

[0045] S2, inputting the image to be identified into a preset image classification model to obtain a classification result of the image to be identified;

[0046] The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing category.

[0047] Compared with the prior art, the present invention provides a method for effectively expanding inter-class separation by adaptively controlling misclassification loss; the most confusing class is taken into account so that the classifier can better distinguish different classes; the time complexity of this method is almost the same as that of the intra-class interval-based method; the model iteration process involved is more efficient, and its complementarity with the existing intra-class interval-based method and its superior performance are clearly demonstrated in the performance of image classification and face recognition applications.

[0048] Specifically, Figure 2 For example, the training process of the deep convolutional neural network combined with adaptive repulsion loss and interval repulsion loss is to extract the feature vector X from the picture in the training data set T, and then calculate the inner product with the weight vector W in the classifier to obtain the angle θ, use these angles θ to calculate the loss function L, and finally backpropagate the gradient of the loss function L to update the deep neural network N and the weight vector W of the classifier; repeat the above process until convergence to obtain the image classification model.

[0049] This embodiment aims to use the cosine similarity between samples and classes to adaptively adjust the separation between classes; since the cosine similarity does not need to be recalculated, the time complexity is reduced. The fully supervised learning image classification scheme provided in this embodiment can be called RepulFace, which mainly expands the distance between different classes by designing inter-class terms between samples, including an adaptive repulsion loss for expanding the separation between classes and an interval repulsion loss for the most confusing class.

[0050] RepulFace is implemented based on the inner product of the classifier weight vector and the sample feature vector, ensuring that the classifier that relies on the classification distance can classify the sample more accurately. The principle can be found in Figure 3 :For a given sample (labeled sample), the feature vector x of the sample is extracted through CNNs i ; This feature vector x i There is an angle between the weight vector of each class; j class represents the closest class other than the class with the same label as the sample, θ j Represents x i and y j The angle between the weight vectors of the classes; using θ j Calculate the interval repulsion to make the most confusing class as far away from this sample as possible. θ1,…,θ N Used to compute adaptive repulsion loss to enhance inter-class separation.

[0051] As a preferred embodiment, the objective function is expressed by the following formula:

[0052] L = l s +L f ; (1)

[0053] Among them, L s represents the adaptive repulsion loss; L f Indicates interval repulsion loss.

[0054] Specifically, the adaptive repulsion loss increases the distance between classes; while the most confusing class requires additional constraints, which can be solved by the interval repulsion loss. Since these two losses solve the problem of class separation, they can be effectively combined with existing methods. The following will first introduce the adaptive repulsion loss L proposed in this embodiment s , then the interval repulsion loss L f Provide explanation.

[0055] Adaptive repulsion loss L s

[0056] For ease of understanding, here is a brief description of the existing softmax cross entropy loss. Softmax cross entropy loss is widely used as an objective function in various deep convolutional neural networks. By optimizing the Softmax cross entropy loss, the network will be able to obtain good classification capabilities. The formula is as follows:

[0057]

[0058] Where C and N are the number of categories and the number of samples sampled in each iteration, respectively. is the yth in the last fully connected layer i The weight vector of each class. Represents the dimension of weights, usually set to 512. It is the feature vector of the i-th training sample in the current iteration sampling sample.

[0059] Based on Softmax cross entropy loss, first, the weight W and feature vector x are normalized: ||W||,||x||=1. Because W T x=||W||·||x||·cosθ (θ is the angle between the weight W and the feature vector x). The feature vectors of the training samples can be distributed on a unit hypersphere, and the hyperparameter s is set to represent the radius of the hypersphere:

[0060]

[0061] where θ j Represents the angle between the feature vector of the i-th sample and the weight vector of the j-th classifier.

[0062] Finally, when y i =j, cosθ can be j Replace it with the function ψ(θ) to let the network learn a smaller angle θ:

[0063]

[0064]

[0065] Among them, m1, m2, and m3 are hyperparameters that can control the intra-class interval size of the intra-class loss.

[0066] Obviously, in y i = j can reduce the intra-class distance and thus indirectly expand the inter-class interval. However, since there is no explicit inter-class regularization, this cannot adaptively control the inter-class distance. Ideally, when a sample is close to other classes than the class to which the sample belongs, the penalty for misclassification loss should be increased, and it is hoped that the inter-class distance will be constrained by both the sample and the class weight vector.

[0067] Therefore, this embodiment further modifies y i ≠j to increase the distance between classes; therefore, further, the adaptive repulsion loss L s It is expressed as follows:

[0068]

[0069] Where N represents the number of samples sampled in one iteration during the training process; θ represents the angle between the feature vector and the weight vector; y i is the category label of the i-th sample; j represents the j-th classifier; m1, m2, and m3 are hyperparameters for controlling the intra-class interval size of the intra-class loss; Based on θ j and A function that adaptively increases the distance between classes; s is a hyperparameter used to control the radius of the hypersphere.

[0070] Furthermore, the j and Function that adaptively increases the distance between classes It is expressed as follows:

[0071]

[0072] Specifically, When , it means that the i-th sample is a difficult sample; Determines the weight of the jth term in the loss function; the purpose of the above setting is to make the sample and the error class θ j When the angle is small, the enhancement of θ j At the same time, the sample and the correct class are also introduced Dynamically adjust the penalty from the perspective of is large, which means that the i-th sample is misclassified, and the penalty for the wrong category will be increased to expand the angle with the wrong category; this can make the weight vectors of each class as far away from each other as possible. When the distance between classes reaches a balanced state, these vectors can be evenly distributed on the hypersphere.

[0073] The most deceptive class of interval repulsion loss L f

[0074] Existing Softmax-based classifiers and their variants cannot guarantee that all classes have sufficient inter-class distances. For example, when When it is larger, some larger cosθ jThis will also make L3 (Formula (4)) get a smaller value. When the adaptive repulsion loss function is used to train the model, the classifier weight vector can be distributed on the hypersphere, but it may still have a problem: there may be a small θ between the sample and a certain class. j , this class is called the most confusing class of the sample. This problem needs to be corrected by additional regularization terms. Figure 4 As shown in the figure, the triangle represents the class center and the circle represents the sample; sample x belongs to Class 1; the distance from sample x to the center of Class 2 is smaller than the distance from sample x to the center of Class 3; therefore, Class 2 is the most confusing class for sample x. Obviously, sample x is too close to the center of Class 2, which will have a negative impact on the classifier.

[0075] In order to better solve the classification problem, the weight vectors of different classes should be as far away from each other as possible. In this embodiment, the maximum cosine similarity between the sample feature vector and the weight vector of other class classifiers is used to define the interval repulsion:

[0076]

[0077] Among them, m is a scalar used to determine whether it is necessary to separate the most confusing class. And (a,0) + It means if a>0, then (a,0) + The value of is a; otherwise, (a,0) + The value of is 0. The interval repulsion of the i-th sample is affected by the label not being y i , and the influence of the class weight vector closest to this sample. In this case, this interval repulsion can be regarded as the yth i The inter-class distance between the class and the most confusing class. At the same time, adjusting m can control a threshold to protect the class feature vectors that have been optimized. In order to better construct RepulFace, this embodiment takes the average value of the interval repulsion of all samples in one iteration; therefore, further, the interval repulsion loss is expressed by the following formula:

[0078]

[0079] Where N is the number of samples sampled in one iteration during the training process; j is the jth classifier; W is the weight vector; W T x=||W||·||x||·cosθ; x represents the feature vector of the sample; i represents the i-th sample; y i is the category label of the i-th sample; θ represents the angle between the feature vector and the weight vector; m is a scalar used to determine whether the most confusing class needs to be separated.

[0080] The adaptive repulsion loss in RepulFace is based on two different classes, so it can make different classes move away from each other to expand the separation between classes. In addition, the spacing repulsion loss in RepulFace is based on the inner product between different classes and samples, and also considers the relationship between classes and samples. Therefore, the optimization update of the weights of the network and classifier can be affected by both categories and samples. In addition, the method proposed in this embodiment is convex to the category weights, which allows the model of this embodiment to be optimized more effectively than other inter-class models. For example, the regularization of MHE and RegularFace is the inner product of two category weight vectors, which is non-convex in optimization for the weight vector.

[0081] Furthermore, the deep convolutional neural network is a deep convolutional neural network based on an intra-class method.

[0082] Furthermore, the in-class method is Sphereface or Cosface or Arcface.

[0083] In order to better explain the effect of the present invention, this embodiment conducts the following experiments on face and natural image datasets respectively:

[0084] Face dataset

[0085] LFW includes approximately 13,000 face images and more than 5,000 person identities. AgeDB-30 includes 16,488 face images and 568 person identities. The lighting, postures, and expressions in AgeDB-30 images are richer than those in LFW. 6,000 sets of image pairs were selected from LFW and AgeDB-30 to evaluate the RepulFace network trained in this embodiment. Each pair has a label to indicate whether the two images are the same face. According to the label, the number of matches with the label is calculated, and then divided by 6,000 to get the accuracy. As shown in Table 1, the RepulFace proposed in the present invention is superior to previous methods. Compared with ArcFace, the accuracy is improved by 0.51% in LFW and about 0.2% in AgeDB-30. In addition, RepulFace can also enhance the performance of SphereFace and CosFace. Practice has proved that the scheme of this embodiment can help such methods achieve higher accuracy without inter-class separation constraints. In addition, RepulFace successfully associates the weight vector of one class with the feature vectors different from that class.

[0086] Table 1. Experimental accuracy of LFW and AgeDB-30 (%)

[0087]

[0088]

[0089] MegaFace contains more than 1 million images with 690,000 identities. It is the first million-level dataset in the field of face recognition. The FaceScrub dataset consists of 100,000 images with 530 identities. The rules of MegaFace Challenge 1 are used in the experiment of this embodiment. The FaceScrub dataset is used as a gallery due to its high image quality, while the MegaFace dataset is used as an interference set. It selects a picture from the gallery as a probe, and then uses it to find another image with the same label from one million interference items randomly selected from the MegaFace dataset. L2 distance is used as a metric. This embodiment uses CASIA-Webface and Ms-celeb-1m datasets as small training sets and large training sets, respectively. In Table 2, RepulFace is compared with other methods. It shows that RepulFace can promote these in-class methods, such as RepulFace+ArcFace, which can improve the accuracy by about 1.22% compared with ArcFace in large training sets and 0.98% in small training sets. RepulFace enlarges the inter-class distance, which can complement the method based on reducing the intra-class distance; and it can correct the most confusing classes during the training phase so that these classes are evenly distributed on the hypersphere.

[0090] Table 2. Experimental accuracy of MegaFace (%)

[0091]

[0092]

[0093] Natural Image Datasets

[0094] CIFAR10 is a public dataset in image classification, which contains 50,000 training images and 10,000 test images, divided into 10 categories. MNIST is a handwritten digit dataset, which consists of 60,000 training images and 10,000 test images in 10 categories. Random cropping and horizontal flipping will be applied to CIFAR10. For MNIST, the experiment of this embodiment does not perform any preprocessing; this embodiment uses ResNet18 as the network structure and trains and tests all methods on these datasets. As shown in Table 3, the scheme provided in this embodiment is compared with CosFace and ArcFace on CIFAR10 and MNIST. Compared with the Softmax cross entropy loss function, both CosFac and ArcFace have been greatly improved. The scheme provided in this embodiment can further improve their performance. For example, RepulFace based on CosFace can achieve the highest accuracy on the two datasets, with an accuracy increase of approximately 0.35% and 0.09% respectively. RepulFace can improve the accuracy of ArcFace by 1.16% on CIFAR10. Such image classification results illustrate the superiority and practicality of RepulFace.

[0095] Table 3. Experimental accuracy of CIFAR10 and MINST (%)

[0096]

[0097]

[0098] Ablation Learning

[0099] RepulFace has two components: adaptive repulsion loss and interval repulsion loss. The following experiments combine these two components with ArcFace respectively to verify the effect of each component. In Table 4, ArcFace can improve the performance from 99.37% to 99.47% when combined with adaptive repulsion loss; this shows that adaptive repulsion loss can expand the inter-class separation, enabling the classifier to effectively identify different faces. Similarly, interval repulsion loss can improve performance by 0.06%, forcing the most confusing class to keep a distance from other samples. When the two components are combined with ArcFace, a maximum accuracy of 99.55% can be achieved, which proves that the two components can work together to improve performance.

[0100] Table 4. Experimental accuracy of ablation learning on LFW (%)

[0101]

[0102] Example 2

[0103] A fully supervised learning image classification system based on adaptive and margin repulsive loss, see Figure 5 , comprising an image acquisition module 1 to be identified and a model identification module 2 which are connected to each other; wherein:

[0104] The to-be-recognized image acquisition module 1 is used to acquire the to-be-recognized image;

[0105] The model recognition module 2 is used to input the image to be recognized into a preset image classification model to obtain a classification result of the image to be recognized;

[0106] The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing category.

[0107] Example 3

[0108] A storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the fully supervised learning image classification method based on adaptation and interval repulsion loss in embodiment 1 are implemented.

[0109] Example 4

[0110] A computer device comprises a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the fully supervised learning image classification method based on adaptation and interval repulsion loss in Example 1.

[0111] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A fully supervised learning image classification method based on adaptive and margin repulsion loss, characterized in that, The following steps are involved: S1, obtain the image to be recognized; S2, inputting the image to be identified into a preset image classification model to obtain a classification result of the image to be identified; The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing class; The objective function is expressed as follows: ； in, represents the adaptive repulsion loss; Indicates interval repulsion loss; The adaptive repulsion loss It is expressed as follows: ； Where N represents the number of samples sampled in one iteration during the training process; Represents the angle between the feature vector and the weight vector; is the category label of the i-th sample; Indicates classifiers; ; , , A hyperparameter for controlling the size of the intra-class margin for the intra-class loss; Based on and A function that adaptively increases the distance between classes; s is a hyperparameter used to control the radius of the hypersphere; The spacing repulsion loss is expressed by the following formula: ； Where N represents the number of samples sampled in one iteration during the training process; Indicates classifiers; represents the weight vector; ; Represents the feature vector of the sample; i represents the i-th sample; is the category label of the i-th sample; Represents the angle between the feature vector and the weight vector; A scalar that determines whether the most confusing class should be separated.

2. The fully supervised learning image classification method based on adaptive and margin repulsion loss according to claim 1, characterized in that: The basis and Function that adaptively increases the distance between classes It is expressed as follows: 。 3. The fully supervised learning image classification method based on adaptive and margin repulsion loss according to any one of claims 1 to 2, characterized in that: The deep convolutional neural network is a deep convolutional neural network based on the intra-class method.

4. The fully supervised learning image classification method based on adaptive and margin repulsion loss according to claim 3 is characterized in that: The method within the class is Sphereface, Cosface or Arcface.

5. A fully supervised learning image classification system based on adaptive and margin repulsion loss, characterized in that, A fully supervised learning image classification method based on adaptive and interval repulsion loss as claimed in any one of claims 1 to 4, comprising an image acquisition module (1) to be identified and a model identification module (2) connected to each other; wherein: The to-be-recognized image acquisition module (1) is used to acquire the to-be-recognized image; The model recognition module (2) is used to input the image to be recognized into a preset image classification model to obtain a classification result of the image to be recognized; The image classification model is obtained by training a deep convolutional neural network that combines an adaptive repulsion loss and an interval repulsion loss in an objective function; wherein the adaptive repulsion loss is used to expand inter-class separation, and the interval repulsion loss is used to process the most confusing category.

6. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the fully supervised learning image classification method based on adaptation and interval repulsion loss are implemented as described in any one of claims 1 to 4.

7. A computer device, characterized in that: It includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, and when the computer program is executed by the processor, the steps of the fully supervised learning image classification method based on adaptive and interval repulsion loss as described in any one of claims 1 to 4 are implemented.

Images