An image classification method, system, electronic device and storage medium

Abstract

The application discloses an image classification method and system, electronic equipment and storage medium, and relates to the field of computer vision. The method comprises the following steps: acquiring an image to be classified; and classifying the image to be classified by using an image classification network to determine the category of the image to be classified. The image classification network is obtained by training a residual network by using a training data set. The training data set is a training set in a CIFAR100 data set. The application fuses deep and shallow feature information on the basis of a backbone network of the residual network, and reconstructs a teacher network by using the fused feature information, so that the shallow layer can learn deeper semantic information, and the classification precision of the model is enhanced. In addition, by using the reconstructed teacher feature information and prediction result information for distillation, the hidden knowledge contained in the teacher can be more effectively used, and the accuracy of target image classification is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, and in particular to an image classification method and system, an electronic device and a storage medium. BACKGROUND

[0002] Although great progress has been made in deep learning, it is difficult to apply it to resource-constrained devices due to the huge amount of computation and the large number of parameters. In order to make the deep model more efficient, people explore the field of knowledge distillation. In 2006, Bucilua et al. first proposed the idea of transferring the knowledge of a large model to a small model. In 2015, Hinton formally proposed the widely known concept of knowledge distillation. The main idea of knowledge distillation is that the student model obtains the same accuracy as the teacher model by imitating the teacher model, and the key problem is how to transfer the knowledge of the teacher model to the student model.

[0003] Traditional knowledge distillation can be divided into response-based knowledge distillation and feature-based knowledge distillation. Response-based knowledge usually refers to the neural response of the last output layer of the teacher model. The main idea is to directly simulate the final prediction of the teacher model. Response-based knowledge distillation is a simple and effective model compression method and has been widely used in different tasks and applications.

[0004] Feature-based knowledge distillation comes from the intermediate layer and is a good extension of response-based knowledge. The feature map of the intermediate layer can be used as the knowledge to supervise the training of the student model. The most direct idea is to match the activation function values of the intermediate features. In particular, Zagoruyko and Komodakis (2017) proposed to use attention maps to represent knowledge; in order to match the semantic information between the teacher and the student, Chen et al. (2021) proposed cross-layer KD, which adaptively assigns layers in the teacher network to layers in each student network through attention positioning. However, the above two classical methods have two shortcomings, including: the first shortcoming is low knowledge transfer efficiency, which means that the student model hardly utilizes all the knowledge in the teacher model. It is still rare for a student model to outperform its teacher model; the second shortcoming is how to design and train a suitable teacher model. Existing distillation frameworks require a lot of effort and experiments to find the best teacher model architecture, which requires a relatively long time, for example, the traditional distillation method takes 14.67 hours to train the teacher network ResNet152 on CIFAR100, and 12.31 hours to train the student network ResNet50 in the second step.

[0005] In summary, the current distillation framework has the problems of long pre-training time of the teacher network, large size difference between the teacher network and the student network, and poor student accuracy. SUMMARY

[0006] The application aims to provide an image classification method, system, electronic device and storage medium to improve image classification accuracy.

[0007] To achieve the above-mentioned purpose, the application provides the following solutions.

[0008] An image classification method comprises:

[0009] Obtaining an image to be classified;

[0010] Classifying the image to be classified by using an image classification network to determine the class of the image to be classified; wherein the image classification network is obtained by training a residual network by using a training data set; and the training data set is a training set in a CIFAR100 data set.

[0011] Optionally, the residual network comprises a convolutional layer, four cascaded residual blocks, a pooling layer and a fully connected layer connected in sequence; wherein the residual block comprises two 3x3 convolutional layers and a 1x1 convolutional layer connected by a skip connection.

[0012] Optionally, the residual network is trained by using the training data set, and the training specifically comprises:

[0013] Obtaining a training data set;

[0014] Performing data enhancement processing on the images in the training data set to obtain a processed training data set;

[0015] Inputting the processed images in the processed training data set into a current residual network to obtain a prediction result;

[0016] Calculating a loss function value of the prediction result and a class label corresponding to the processed images in the processed training data set by using a loss function;

[0017] According to the loss function value, calculating the gradient of each parameter in the current residual network by using a back propagation algorithm;

[0018] According to the gradient and a learning rate, updating the parameters and weights of the current residual network by using an Adam optimization algorithm, and returning to the step of inputting the processed images in the processed training data set into the current residual network to obtain a prediction result until the processed training data set is traversed or the loss function value is less than a preset value, and taking the current residual network as an image classification network.

[0019] Optionally, the data enhancement processing on the images in the training data set to obtain a processed training data set specifically comprises:

[0020] randomly crop the images in the training data set to obtain a cropped training data set;

[0021] randomly horizontally flip the cropped images in the cropped training data set to obtain a processed training data set.

[0022] Optionally, the loss function is a cross-entropy loss function or a mean square error loss function.

[0023] An image classification system comprises:

[0024] an image acquisition module configured to acquire an image to be classified;

[0025] an image classification module configured to classify the image to be classified by using an image classification network to determine a class of the image to be classified, wherein the image classification network is obtained by training a residual network by using a training data set, and the training data set is a training set in a CIFAR100 data set.

[0026] An electronic device comprises a memory configured to store a computer program and a processor configured to execute the computer program to cause the electronic device to perform the image classification method.

[0027] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the image classification method.

[0028] According to the embodiments of the present application, the following technical effects are provided:

[0029] The image classification method, system, electronic device and storage medium provided by the present application first acquire an image to be classified, and determine a class of the image to be classified by using an image classification network. The image classification network is obtained by training a residual network by using a training data set. The present application fuses deep and shallow feature information on the basis of a backbone network of the residual network, and reconstructs a teacher network by using the fused feature information, so that the shallow layer can learn deeper semantic information, and the classification precision of the model is enhanced. In addition, by distilling the reconstructed teacher feature information and prediction result information, the dark knowledge contained in the teacher can be more effectively utilized, and the accuracy of the target image classification is improved. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 A flowchart of the image classification method provided by the present invention;

[0032] Figure 2 The residual network processing flowchart provided by this invention;

[0033] Figure 3 The network training flowchart provided by this invention;

[0034] Figure 4 A test flowchart for the test set provided by this invention;

[0035] Figure 5 This is a diagram of the residual network structure used in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] The purpose of this invention is to provide an image classification method, system, electronic device, and storage medium to improve the accuracy of image classification.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1

[0040] like Figure 1 As shown, the image classification method provided by the present invention includes:

[0041] Step 101: Obtain the image to be classified.

[0042] Step 102: Classify the image to be classified using an image classification network to determine the category of the image to be classified; in practical applications, image categories are divided into aquatic animals (e.g., fish), flowers, food containers, fruits, vegetables, household appliances, furniture, mammals (including large mammals, medium mammals, and small mammals), large cats, invertebrates (e.g., insects), people, buildings, fog, clouds, fire, plains, hills, and sky (including urban sky, high sky, and street garden sky); wherein, the image classification network is obtained by training a residual network using a training dataset; the training dataset is the training set in the CIFAR100 dataset.

[0043] In practical applications, a residual network is constructed, the residual network is used as a backbone network, deep features and shallow features are fused, semantic information richer features are fused from top to bottom, and then the features of the teacher network are reconstructed from bottom to top to give a prediction result. The prediction result and the features are both used to guide the student network learning.

[0044] The residual network, as shown in Figure 5 includes a convolutional layer, four cascaded residual blocks, a pooling layer, and a fully connected layer connected in sequence. The residual block includes two 3x3 convolutional layers and a 1x1 convolutional layer connected by a skip connection. The features 1 are obtained by inputting the image to be classified into the convolutional layer for convolution calculation, the features 2 are obtained by inputting the features 1 into the residual block 1, the features 3 are obtained by inputting the features 2 into the residual block 2, the features 4 are obtained by inputting the features 3 into the residual block 3, the features 5 are obtained by inputting the features 4 into the residual block 4, the features 6 are obtained by inputting the features 5 into the pooling layer, and the final classification result is obtained by inputting the features 6 into the fully connected layer.

[0045] 1. Convolutional layer: the residual network starts with three convolutional layers, which are responsible for extracting features from the input image; 2. Residual block: each residual block consists of three convolutional layers and realizes "residual" learning of the network through a skip connection; 3. Pooling layer: the pooling layer in the residual network is used to reduce the size and number of parameters of the feature map, thereby improving the computational efficiency; 4. Fully connected layer: the output of the average pooling layer is connected to the predefined number of categories, which is used to perform the final classification task. The deep feature and the current layer feature weight are calculated from top to bottom, and the calculation process is as follows: the classification image is divided into four layers of different depths by the four residual blocks, and the more residual blocks the image passes through, the deeper the feature depth. After the i+1 layer feature is up-sampled to the same size, it is stacked with the current i layer feature by channel, and then input into the convolutional layer. The input channel number is 2, the first channel is the deep layer weight, and the second channel is the current layer weight. The weight can be taken as 1, 2, 3, or 4 according to the residual block division. The fused feature is obtained by multiplying the weight and the feature; then the i layer fused feature and the i layer original convolutional feature are stacked from bottom to top, and a 1x1 convolutional layer is used as a calculation unit to output the fused reconstructed feature. The convolutional calculation forms the reconstructed feature, and finally the prediction result is obtained. The prediction result and the reconstructed feature are both used to guide the student network learning.

[0046] As an optional implementation, the residual network is trained using a training data set, which specifically includes:

[0047] S1: Obtain a training data set. In practical applications, the CIFAR data set is obtained, and the CIFAR10 and CIFAR100 data sets are divided according to a ratio of five to one, into a training set and a test set. CIFAR100 contains images of various types, from plants and animals in nature, to vehicles and household items in daily life, to abstract art, and the like. The images in the test set of CIFAR100 are classified using a trained neural network and weight file, and the accuracy of the classification is detected.

[0048] S2: Perform data enhancement processing on the images in the training data set to obtain a processed training data set. In practical applications, random cropping and random horizontal flipping data enhancement methods are used for the size of the target in the data set, and the cropping padding size is 4.

[0049] S3: Input the processed images in the processed training data set into the current residual network to obtain a prediction result.

[0050] In practical applications, as shown in Figure 2 S3 includes the following steps:

[0051] 1) Four different deep features are output by four cascaded residual blocks of the residual network, the deep features are upsampled to align with the scale of shallow features, then stacked according to the channel and the respective attention weight is calculated by 1x1 convolution, multiplied by the corresponding feature from top to bottom to obtain the fusion feature, and the network learns the important features.

[0052] 2) The fusion feature is used to calculate the reconstruction feature from bottom to top. The bottom layer feature is superimposed with the fusion feature, and the reconstruction feature from bottom to top is obtained by a convolution layer.

[0053] 3) Finally, the FC layer is used to predict the extracted features. After layer-by-layer calculation, the prediction result of the reconstruction feature is obtained through the last fully connected layer, the fusion feature of the intermediate layer is used to guide the semantic learning of the student network, and the prediction result is used to guide the classification learning of the student network.

[0054] S4: Use a loss function to calculate the loss function value of the prediction result and the class label corresponding to the processed image in the processed training data set.

[0055] S5: According to the loss function value, the gradient of each parameter in the current residual network is calculated using a back propagation algorithm.

[0056] S6: according to the gradient and learning rate, the parameters and weights of the current residual network are updated by using the Adam optimization algorithm, and the step of inputting the processed image in the processed training data set into the current residual network to obtain the prediction result is returned until the processed training data set is traversed or the loss function value is less than the preset value, and the current residual network is used as an image classification network. In actual application, the method of random gradient descent is used for optimization, and the learning rate is attenuated twice, from the initial value, so that the neural network can achieve better distillation results.

[0057] In actual application, the neural network is trained, and each iteration period includes the following steps: 1, randomly selecting a batch of training samples with a size of 128; 2, inputting the selected batch into the residual network, and calculating the output of the network; 3, using the selected loss function (such as mean square error or cross entropy), calculating the loss between the network output and the target output; 4, according to the loss value, the gradient of each parameter in the network is calculated by the back propagation algorithm; 5, using the Adam optimization algorithm, according to the learning rate and the direction of the gradient, the parameters and weights of the network are updated. Repeat the steps until the entire training data set is traversed.

[0058] As shown in Figure 3 Before training the network, the mean and variance of the data set are recalculated, the data is normalized, random weights are used as initial weights, the learning rate, the number of iterations, batch_size and the like are set, and the learning rate is attenuated from the initial value at 100 and 150 rounds, so that the neural network can achieve better detection results; the input image is augmented, and the training is performed, and when the loss function converges or reaches the maximum number of iterations, the training is stopped to obtain the weight file after self-distillation.

[0059] As shown in Figure 4 The images of the test set are input into the improved residual network (trained residual network) backbone network, the predictions of different stages are fused, and the prediction result of the last layer for image classification is obtained.

[0060] Training and testing are performed on a TITAN XP graphics card, the distillation temperature is set to 4.0 during distillation, the weight attenuation in the random gradient descent algorithm is set to 0.0001, the value of the loss function is output in the terminal during each round of training, which is convenient for observing the overall convergence, and the test set is used for verification at the end of each round, the classification result of the fourth layer of the residual network is compared when verifying the accuracy, and if the current verification result is greater than the historical optimal accuracy, the weight is updated. It is verified that the present application can achieve a classification accuracy of 77.50% on CIFAR100.

[0061] The image classification method of the present application firstly downloads the CIFAR dataset and divides and augments it; then, on the basis of the residual network network structure, teacher features are reconstructed using four-stage feature cascading, and a prediction result is given; the divided dataset is sent into the residual network for training until the network converges, and a weight file is obtained; finally, the trained neural network and the weight file are used to detect test images, and a classification result is output. The present application well solves the problems of low accuracy of small models in the field of image classification and time-consuming pre-training of teacher networks in the distillation framework.

[0062] Embodiment two

[0063] In order to perform the method corresponding to the above-mentioned embodiment one, to realize the corresponding function and technical effect, the following provides an image classification system, comprising:

[0064] An image acquisition module is configured to acquire an image to be classified.

[0065] An image classification module is configured to classify the image to be classified by using an image classification network to determine the category of the image to be classified; wherein the image classification network is obtained by training a residual network by using a training dataset; and the training dataset is a training set in a CIFAR100 dataset.

[0066] Embodiment three

[0067] The present application provides an electronic device, comprising: a memory and a processor, the memory is used for storing computer programs, the processor runs the computer programs to make the electronic device execute the image classification method of embodiment one.

[0068] Embodiment four

[0069] The present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to realize the image classification method of embodiment one.

[0070] In the present specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0071] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. An image classification method, characterized by, The method comprises the following steps: obtaining an image to be classified; classifying the image to be classified by using an image classification network to determine the category of the image to be classified; wherein the image classification network is obtained by training a residual network using a training data set; the training data set is a training set in a CIFAR100 data set; the residual network comprises a convolutional layer, four cascaded residual blocks, a pooling layer and a fully connected layer connected in sequence; wherein the residual block comprises two 3*3 convolutional layers and a 1*1 convolutional layer connected by a skip connection; training the residual network using the training data set, specifically comprising: obtaining a training data set; performing data enhancement processing on the images in the training data set to obtain a processed training data set; inputting the processed images in the processed training data set into the current residual network to obtain a prediction result; inputting the processed images in the processed training data set into the current residual network to obtain a prediction result, specifically comprising: 1) using the four cascaded residual blocks of the residual network to output four different deep features, upsampling the (i+1)th deep feature to align with the scale of the ith deep feature, then stacking them according to the channel and calculating the respective attention weights through a 1*1 convolution, multiplying the weights with the corresponding features to obtain the ith fused feature, and letting the network learn the important features; i=1,2,3; 2) then calculating the reconstruction features from bottom to top using the fused features: superimposing the first layer of fused features with the features output by the first residual block in the residual network, and obtaining reconstruction feature 1 through a convolutional layer; superimposing the second layer of fused features with reconstruction feature 1, and obtaining reconstruction feature 2 through a convolutional layer; superimposing the third layer of fused features with reconstruction feature 2, and obtaining reconstruction feature 3 through a convolutional layer; 3) finally, using a fully connected layer to predict the extracted features: after layer-by-layer calculation, the prediction result of the reconstruction feature is obtained through the last fully connected layer, the fused features of the intermediate layer are used to guide the semantic learning of the student network, and the prediction result is used to guide the classification learning of the student network; the student network is the residual network; using a loss function to calculate the loss function value of the prediction result and the category label corresponding to the processed image in the processed training data set; using a backpropagation algorithm to calculate the gradient of each parameter in the current residual network according to the loss function value; using an Adam optimization algorithm to update the parameters and weights of the current residual network according to the gradient and learning rate, and returning to the step of "inputting the processed images in the processed training data set into the current residual network to obtain a prediction result" until the processed training data set is traversed or the loss function value is less than a preset value, and taking the current residual network as the image classification network.

2. The image classification method of claim 1, wherein, performing data enhancement processing on the images in the training data set to obtain a processed training data set, specifically comprising: randomly cropping the images in the training data set to obtain a cropped training data set; randomly horizontally flipping the cropped images in the cropped training data set to obtain a processed training data set.

3. The image classification method of claim 1, wherein, The loss function is a cross-entropy loss function or a mean square error loss function.

4. An image classification system characterized by, The method comprises the following steps: An image acquisition module is configured to acquire an image to be classified. An image classification module is configured to classify the image to be classified by using an image classification network to determine the category of the image to be classified. The image classification network is obtained by training a residual network by using a training data set. The residual network comprises a convolutional layer, four cascaded residual blocks, a pooling layer, and a fully connected layer connected in sequence. The training of the residual network by using the training data set comprises the following steps: A training data set is acquired. Data enhancement processing is performed on the images in the training data set to obtain a processed training data set. A processed image in the processed training data set is input into a current residual network to obtain a prediction result. The inputting of the processed image into the current residual network to obtain the prediction result comprises the following steps: 1) Four different deep features are output by the four cascaded residual blocks of the residual network. 2) Reconstruction features are calculated from bottom to top by using the fusion features. 3) The prediction result is obtained by using a fully connected layer to predict the extracted features. A loss function value is calculated by using a loss function.

5. An electronic device, comprising: The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The loss function value is calculated by using a loss function. The 6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the image classification method in any one of claims 1-3.

Images