A domain generalization method based on improved meta-learning adversarial domain augmentation

Abstract

The application relates to the field of machine learning domain generalization, and particularly discloses a domain generalization method based on improved meta-learning adversarial domain enhancement, wherein the method can achieve better effect of the model trained in a single training domain generalizing to an unknown field by reconstructing a feature extraction network structure, introducing a WGAN-GP generative adversarial network, and using self-normalization and cross-normalization methods for data standardization. In addition, the model has high correct recognition rate and generalization ability, and the model volume is small, so that the model can stably and efficiently run in a limited resource computing environment. Therefore, the method is suitable for being quickly and widely deployed in the domain generalization application, and effectively improves the performance and work efficiency of the domain generalization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning domain generalization, and particularly relates to a domain generalization method based on improved meta-learning adversarial domain enhancement. BACKGROUND

[0002] Domain generalization is an important research direction in machine learning, aiming to improve the performance of models on unseen data distributions (i.e., target domains). Traditional machine learning methods often train on a single dataset, resulting in insufficient generalization ability of models in new environments. Therefore, it is of great practical significance to study how to improve the domain generalization ability of models.

[0003] Currently, domain generalization methods are mainly divided into two categories: data augmentation methods and model adjustment methods. Data augmentation methods generate new training samples by transforming training data (such as rotation, scaling, cropping, color jittering, etc.) to improve the adaptability of models to diversity. Model adjustment methods modify the structure or training process of the model to improve its generalization ability, such as using meta-learning (Meta-Learning) technology to train the model to quickly adapt to new environments.

[0004] Although existing domain generalization methods have made some progress, they often require a large amount of data and complex network structures, making it difficult to deploy models in practical applications. In addition, many methods sacrifice the performance of models in new fields while improving their generalization ability, or result in excessive model size, which is not suitable for running in resource-limited computing environments.

[0005] In practical applications, models need to maintain good performance when facing diverse, unseen data distributions. However, existing domain generalization methods have certain limitations in dealing with this challenge. In today's machine learning and artificial intelligence field, the generalization ability of models is a core research topic. With the continuous progress of technology, we increasingly rely on the performance of models in handling new situations and new data. In practical applications, this generalization ability is crucial because models often need to encounter diverse, unseen data distributions after the training phase and maintain stable performance in these new situations.

[0006] Therefore, there is an urgent need for a domain generalization method that can train a model in a single training domain, effectively generalize to multiple unseen domains, and improve performance. SUMMARY

[0007] In order to solve the problems in the prior art, the application provides a domain generalization method based on improved meta-learning adversarial domain enhancement, which realizes better effect of the model trained in a single training domain generalizing to an unknown field by reconstructing a feature extraction network structure, introducing a WGAN-GP generative adversarial network, and using self-normalization and cross-normalization methods for data standardization, and solves the problems mentioned in the background art.

[0008] In order to achieve the above object, the application provides the following technical scheme: a domain generalization method based on improved meta-learning adversarial domain enhancement, comprising the following steps:

[0009] S1, data set acquisition and data set division;

[0010] S2, data preprocessing of the data set;

[0011] S3, inputting the data after data preprocessing into the replaced generative adversarial network WGAN-GP, and using the adversarial training of WGAN-GP to generate a hypothetical data set;

[0012] S4, splicing the data set after data preprocessing and the generated hypothetical data set, standardizing the spliced data using self-normalization and cross-normalization methods, obtaining a merged data set and inputting it into the improved feature extraction network VGG16 network for training;

[0013] S5, inputting unknown domain data into the model trained in step S4 to complete detection of the unknown domain data.

[0014] Preferably, in step S1, the following is specifically included: selecting a CIFAR-10 data set, which includes 10 categories of 32x32 pixel color images, and the 10 categories are airplane, car, bird, cat, deer, dog, frog, horse, ship and truck; dividing the images in the CIFAR-10 data set into a training set and a test set in a ratio of 5:1.

[0015] Preferably, in step S2, the data set is preprocessed, specifically including: first converting the data format of the data set to an input form suitable for the optimized VGG16 network model; then pre-processing the training set and the test set, including rotation, scaling, cropping and color jittering operations, to improve the generalization ability of the model.

[0016] Preferably, in step S3, the replaced generative adversarial network WGAN-GP is specifically a generative adversarial network replacing the original autoencoder WAE; in the WGAN-GP, the Wasserstein distance is used to measure the distance between the generator G and the discriminator D, so that the generator G generates data closer to the real data distribution; and a gradient penalty term is introduced to ensure the continuity of the gradient of the discriminator D in space.

[0017] Preferably, the formula of the Wasserstein distance is:

[0018]

[0019] Where P and are the generated data distribution and the real data distribution respectively, Z is the latent space, X is the original data space, P z is the mapping from the generated data distribution P to the latent space Z, G(z) is the mapping of the generator to generate data from the latent space, D(x) is the output of the discriminator to the data x, and D(G(z)) is the output of the discriminator to the generated data G(z);

[0020] The gradient penalty term is introduced, specifically including: an additional penalty term is introduced to punish the gradient of the discriminator output, and the penalty term is realized by calculating the gradient constant Lipchitz of the discriminator output. The Lipchitz constant is an index to measure the rate of function change, and the formula is:

[0021]

[0022] Where D(x) is the output of the discriminator to the input x, and x and x' are two different data points in the input space.

[0023] The loss function of the gradient penalty term is:

[0024]

[0025] Where λ is a hyperparameter used to control the influence degree of the gradient penalty term.

[0026] Preferably, in step S3, the adversarial training of the WGAN-GP generates the imaginary data set, specifically including:

[0027] Using the generator G to randomly generate an imaginary data set of a given size;

[0028] Using the discriminator D to discriminate the similarity between the generated imaginary data set and the original data set;

[0029] The hypothetical data set is regenerated according to the similarity until the similarity of the hypothetical data set to the original data set is close.

[0030] Preferably, the improved feature extraction network VGG16 network refers to: improving three layers of convolution and three layers of full connection of the original feature extraction network model into eleven layers of convolution and one layer of full connection; and using SeLU activation function in the convolution layer instead of the original ReLU activation function, and using LeakyReLU activation function in the full connection layer instead of the original ReLU activation function.

[0031] Preferably, in step S4, the improved feature extraction network VGG16 network includes five convolution blocks C1, C2, C3, C4 and C5, which are connected in series in the manner that the output of C1 is the input of C2, the output of C2 is the input of C3, the output of C3 is the input of C4, and the output of C4 is the input of C5.

[0032] Among them, the C1 and C2 convolution blocks respectively include two convolution layers B, two activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output after the second convolution layer B is taken as the input of P.

[0033] The C3 and C4 convolution blocks respectively include three convolution layers B, three activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output after the third convolution layer B is taken as the input of P.

[0034] The C5 convolution block includes one convolution layer B, one activation function and one pooling layer P; specifically, an activation function is added after the convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output after the convolution layer B is taken as the input of P.

[0035] The convolution kernel size of the convolution layer B is 3*3, and the moving step is 1; the pooling layer P uses the maximum pooling method, and the pooling window size is set to 2*2 and the moving step is 2.

[0036] Preferably, the activation function SeLU has the following calculation formula:

[0037]

[0038] In the formula, x is the data to be processed, f(x) is the result after the activation function processing, and a and μ are parameters.

[0039] The activation function LeakyReLU has the following calculation formula:

[0040]

[0041] In the formula, x is the data to be processed, f(x) is the result after the activation function processing, and b is the angle of the negative slope.

[0042] Preferably, in step S4, the obtained merged data set is input into the improved feature extraction network VGG16 network for training, and the training parameters are set as follows: the initial learning rate is set to 0.001, the batch_size is set to 64, the optimizer uses SGD, the weight_decay is set to 0.0005, the momentum is set to 0.9, and the training rounds are set to 9001 rounds.

[0043] The beneficial effects of the present application are:

[0044] 1) The present application aims at the domain generalization problem, and the feature extraction network model uses the optimized VGG16 to replace the original feature extraction network model, improves the original three layers of convolution and three layers of full connection to 11 layers of convolution and one layer of full connection, replaces the activation function of the convolution layer to SeLU, and replaces the activation function of the full connection layer to LeakyRelu, thereby improving the generalization effect of the model; the data set is preprocessed, the data after data preprocessing is input into the replaced WGAN-GP generative adversarial network for training to generate a hypothetical data set, so as to improve the generalization ability of the model;

[0045] 2) The model proposed in the present application has high correct recognition rate and generalization ability, and the model volume is small, which can stably and efficiently run in a resource-limited computing environment. Therefore, the method is suitable for rapid and wide deployment in domain generalization applications, effectively improving the performance and work efficiency of domain generalization. The present application can not only achieve high recognition accuracy on known source domains and unknown target domains, but also has no limitation on the selection of data sets, so it can be applied to various types of data sets and has universal applicability. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a step flow diagram of the domain generalization method based on improved meta-learning adversarial domain enhancement of the embodiment of the present application.

[0047] Figure 2 It is a schematic diagram of the improved feature extraction deep convolutional neural network structure of the present application.

[0048] Figure 3 It is a work flow diagram of the adversarial generative network WGAN-GP introduced by the present application.

[0049] Figure 4 It is a schematic diagram of the detection result in the image of the embodiment of the present application. DETAILED DESCRIPTION

[0050] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0051] The embodiment of the present application provides a technical solution: a domain generalization method based on improved meta-learning adversarial domain enhancement. By replacing and optimizing the feature extraction network structure, introducing the WGAN-GP generative adversarial network, and using self-normalization and cross-normalization methods for data standardization, the model trained in a single training domain can achieve good results in unknown fields. Among them, the SeLU activation function is used in the convolutional layer instead of the original ReLU activation function, which helps to train the deep network. The LeakyReLU activation function is used in the fully connected layer instead of the original ReLU activation function, which helps to improve the robustness of the network. The WGAN-GP adversarial generation network is used instead of the original autoencoder to generate the hypothetical data set, which helps to enhance the model generalization effect. The embodiment of the present application is based on the improved meta-learning adversarial domain enhancement domain generalization method, as shown in the figure, which specifically includes the following steps: Figure 1

[0052] S1, data set acquisition and data set division;

[0053] S2, data preprocessing is performed on the data set;

[0054] S3, input the data after data preprocessing into the replaced generative adversarial network WGAN-GP, and use the adversarial training of WGAN-GP to generate a hypothetical data set;

[0055] S4, splice the data set after data preprocessing and the generated hypothetical data set, and use self-normalization and cross-normalization methods to standardize the spliced data, obtain a merged data set and input it into the improved feature extraction network VGG16 network for training;

[0056] S5, input the unknown domain data into the model trained in step S4 to complete the detection of the unknown domain data.

[0057] ​Furthermore, the CIFAR-10 dataset was selected. This dataset contains 60,000 color images of 32x32 pixels each across 10 categories: airplane, car, bird, cat, deer, dog, frog, horse, boat, and truck. The images in the CIFAR-10 dataset were divided into training and test sets in a 5:1 ratio, with 500 training images and 100 test images per category. The training set contained 50,000 images, and the test set contained 10,000 images, used for model training and performance evaluation.

[0058] By acquiring and partitioning the dataset as described above, we can effectively improve the model's performance in the generalization domain and enhance its generalization ability and robustness.

[0059] Furthermore, in step S2, the data preprocessing of the dataset specifically includes: firstly, converting the data format of the CIFAR-10 dataset into an input format suitable for the optimized VGG16 network model; then, performing data preprocessing on the training and test sets, including rotation, scaling, cropping, and color jittering operations, to improve the generalization ability of the model.

[0060] The above data preprocessing helps improve the model's generalization ability, learning efficiency, and accuracy, while reducing the risk of overfitting and improving the model's robustness.

[0061] Furthermore, in step S3, the replaced generative adversarial network WGAN-GP specifically replaces the generative adversarial network in the original Wasserstein Autoencoder (WAE) with Wasserstein GAN with Gradient Penalty (WGAN-GP). In WGAN-GP, Wasserstein distance is used to measure the distance between the generator G and the discriminator D to solve the mode collapse problem in GAN training, enabling the generator G to generate data that more closely approximates the real data distribution. A gradient penalty term is also introduced to ensure the spatial continuity of the gradient of the discriminator D, thereby improving the quality of samples generated by the generator G.

[0062] like Figure 3 As shown, adversarial training using WGAN-GP generates a hypothetical dataset, enabling the model to establish a more stable mapping between the source and target domains, thus improving the model's domain generalization ability. These improvements effectively enhance the model's performance in the target domain, reduce overfitting, and improve its generalization ability and robustness.

[0063] The Wasserstein distance is used to measure the distance between the generator G and the discriminator D, which includes comparing the difference between the data distribution generated by the generator G and the real data distribution.

[0064] The formula of Wasserstein distance is:

[0065]

[0066] where P and are the generated data distribution and the real data distribution, Z is the latent space, X is the original data space, P z is the mapping from the generated data distribution P to the latent space Z, G(z) is the mapping of the generator to generate data from the latent space, D(x) is the output of the discriminator for data x, and D(G(z)) is the output of the discriminator for generated data G(z). By using the Wasserstein distance, it helps the generator G to generate data closer to the real data distribution, and improves the stability of the training process.

[0067] The gradient penalty term is introduced, specifically including: an additional penalty term is introduced to punish the gradient of the discriminator output, and the penalty term is realized by calculating the gradient constant Lipchitz of the discriminator output. The Lipchitz constant is an index to measure the rate of function change, and the formula is:

[0068]

[0069] where D(x) is the output of the discriminator for input x, and x and x' are two different data points in the input space;

[0070] The loss function of the gradient penalty term is:

[0071]

[0072] where λ is a hyperparameter used to control the influence degree of the gradient penalty term. To ensure that the decision boundary of the discriminator is not too flat.

[0073] In the above manner, WGAN-GP can improve the training stability and generate more diverse and high-quality samples.

[0074] Further, the WGAN-GP is used to generate the hypothetical data set through adversarial training, so that the model can establish a more stable mapping relationship between the source domain and the target domain, and improve the domain generalization ability of the model, specifically including:

[0075] Train a generator network that aims to generate hypothetical data similar to the real data set but with different attributes. Use the generator G to randomly generate a hypothetical data set of a given size;

[0076] training a discriminator network that can distinguish between real data and fake data generated by the generator network. The discriminator D is used to determine the similarity of the generated fake data set to the original data set;

[0077] re-generating the fake data set according to the similarity until the similarity of the fake data set to the original data set is close.

[0078] Under the WGAN-GP generative adversarial network, through adversarial training, the generator network continuously adjusts its parameters to deceive the discriminator network, while the discriminator network continuously learns how to distinguish between real data and fake data; the generator network generates a high-quality fake data set by learning the distribution of real data, which contains data similar to real data but with different attributes; the generated fake data set is used to fine-tune the model to improve the performance and robustness of the model in the target domain.

[0079] Further, the feature extraction network model is improved by increasing the number of convolution layers to reduce information loss. The improved feature extraction network VGG16 network refers to: the original feature extraction network model is improved from three convolution layers and three fully connected layers to eleven convolution layers and one fully connected layer; and the SeLU activation function is used in the convolution layer instead of the original ReLU activation function, and the LeakyReLU activation function is used in the fully connected layer instead of the original ReLU activation function.

[0080] As shown in Figure 2 The improved feature extraction network VGG16 network includes five convolution blocks C1, C2, C3, C4, and C5, which are connected in series in the manner of C1 output as C2 input, C2 output as C3 input, C3 output as C4 input, and C4 output as C5 input.

[0081] Among them, C1 and C2 convolution blocks respectively contain two convolution layers B, two activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the second convolution layer B is input to P.

[0082] C3 and C4 convolution blocks respectively contain three convolution layers B, three activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the third convolution layer B is input to P.

[0083] The C5 convolution block contains one convolution layer B, one activation function and one pooling layer P; specifically, an activation function is added after the convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the convolution layer B is input to P.

[0084] The new convolution block is used to replace the convolution kernel size 5*5 of the convolution block in the original network, so that the calculation amount and information loss are reduced.

[0085] The convolution kernel size of the convolution layer B is 3*3, and the moving step is 1; the pooling layer P uses the maximum pooling method, and the pooling window size is set to 2*2 and the moving step is 2.

[0086] After each convolution block completes convolution, the channels are stacked to obtain the final convolution result; the number of convolution layers is increased to reduce information loss, thereby improving the efficiency and quality of feature extraction.

[0087] The activation function SeLU has the following calculation formula:

[0088]

[0089] In the formula, x is the data to be processed, f(x) is the result after the activation function processing, a and mu are parameters; in the application, the parameters are set as a=1.6732632423543772848170429916717, mu=1.0507009873554804934193349852946.

[0090] The activation function LeakyReLU has the following calculation formula:

[0091]

[0092] In the formula, x is the data to be processed, f(x) is the result after the activation function processing, b is the angle of the negative slope, and the parameter in the application is set to 0.01.

[0093] Further, the data set after data preprocessing is spliced with the generated hypothetical data set, and the spliced data is standardized by using self-normalization and cross-normalization method, to obtain a merged data set. The self-normalization reduces the internal covariant offset, the cross-normalization reduces the correlation between features, and helps to improve the learning process of the model and improve the model accuracy.

[0094] The obtained merged data set is input into the improved feature extraction network VGG16 network for training, and the improved VGG16 network is used to improve the speed and accuracy of feature extraction. The merged data set is used as the input of the data, and the merged data set includes the preprocessed data set and the generated hypothetical data set;

[0095] The input end uses image change to convert the images in the data set into the input requirements of the improved VGG16 network model;

[0096] The training parameter settings are as follows: the initial learning rate is set to 0.001, the batch_size is set to 64, the optimizer uses SGD, the weight_decay is set to 0.0005, and the momentum is set to 0.9.

[0097] The activation function of the convolutional layer uses SeLU, and the activation function of the fully connected layer uses LeakyReLU.

[0098] Adjust the model parameters through the optimization algorithm, so that the model can establish an effective mapping relationship between the source domain and the target domain, and improve the domain generalization ability of the model.

[0099] Training rounds: the entire training process is divided into 9001 rounds, and performance evaluation is performed after each round of training. The learning rate or training strategy is adjusted according to the evaluation results. The precision rate obtained after each round of training is output, and the model with the best training effect is finally saved. Through the above steps, the performance of the model in the new field can be effectively improved, and the generalization ability and robustness of the model can be improved.

[0100] Further, the unknown domain data is input into the trained model to complete the detection of the unknown domain data, as shown in Figure 4 , including:

[0101] The model is selected from the model trained on the source domain data set and the hypothetical data set generated by the generative adversarial network, ensuring that the model has generalization ability and robustness;

[0102] Obtain the data set of the unknown field, which is different from the source domain and is used to test the domain generalization ability of the model, including: selecting CIFAR-10-C data set as the generalization domain, which is further processed based on CIFAR-10 data set and can simulate image damage in real world. The processing of CIFAR-10-C includes Gaussian noise, shot noise, impulse noise, defocus blur, ground glass blur, motion blur, zoom blur, simulated snow, simulated icing, simulated fog, overexposure, low contrast, deformation, pixelization, and JPEG compression.

[0103] Perform necessary preprocessing on the data in the new field, such as normalization, cropping, rotation, etc., to adapt to the input requirements of the model;

[0104] Use the test data set in the new field to evaluate the performance, and measure the generalization performance of the model in the new field according to the accuracy of the model. The trained model can be effectively used in unknown fields, showing the adaptability and generalization level of the unknown field data.

[0105] Although the present application has been described in detail with reference to the foregoing embodiments, the technical solutions recorded in the foregoing embodiments can be modified, or some of the technical features can be replaced by equivalent features, by those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A domain generalization method based on improved meta-learning adversarial domain augmentation, characterized in that, Comprise the following steps: S1, data set acquisition and data set division; select CIFAR-10 data set, the data set includes 10 categories for 32x32 pixel color images, and 10 categories are airplane, car, bird, cat, deer, dog, frog, horse, ship and truck respectively; the images in the CIFAR-10 data set are divided into training set and test set according to the proportion of 5:1; S2, data preprocessing is carried out on the data set; S3, the data after data preprocessing is input into the replaced generative adversarial network WGAN-GP, and the imaginary data set is generated by using the adversarial training of WGAN-GP; S4, the data set after data preprocessing is spliced with the generated imaginary data set, the spliced data is standardized by using self-normalization and cross-normalization method, and the merged data set is input into the improved feature extraction network VGG16 network for training; the improved feature extraction network VGG16 network refers to that the three layers of convolution and three layers of full connection of the original feature extraction network model are improved to eleven layers of convolution and one layer of full connection; and the SeLU activation function is used in the convolution layer to replace the original ReLU activation function, and the LeakyReLU activation function is used in the full connection layer to replace the original ReLU activation function; The improved feature extraction network VGG16 network comprises five convolution blocks C1, C2, C3, C4 and C5, which are connected in series in the manner that the output of C1 is input into C2, the output of C2 is input into C3, the output of C3 is input into C4, and the output of C4 is input into C5; Wherein, the C1 and C2 convolution blocks respectively comprise two convolution layers B, two activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the second convolution layer B is input into P; The C3 and C4 convolution blocks respectively comprise three convolution layers B, three activation functions and a pooling layer P; specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the third convolution layer B is input into P; The C5 convolution block comprises a convolution layer B, an activation function and a pooling layer P; specifically, an activation function is added after the convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the convolution layer B is input into P; The convolution kernel size of the convolution layer B is 3*3, and the moving step is 1; the pooling layer P uses the maximum pooling method, and the pooling window size is set to 2*2 and the moving step is 2; S5, input unknown domain data into the model trained in step S4 to complete the detection of unknown domain data.

2. The domain generalization method based on improved meta-learning adversarial domain augmentation according to claim 1, characterized in that: In step S2, the data set is preprocessed, specifically including: first, converting the data format of the data set into an input form suitable for the optimized VGG16 network model; then, data preprocessing is carried out on the training set and test set, including rotation, scaling, cropping and color jittering operation, to improve the generalization ability of the model.

3. The domain generalization method based on improved meta-learning adversarial domain augmentation according to claim 1, characterized in that: In step S3, the replaced generative adversarial network WGAN-GP is specifically a generative adversarial network replacing the original autoencoder WAE; in the WGAN-GP, the Wasserstein distance is used to measure the distance between the generator G and the discriminator D, so that the generator G generates data closer to the real data distribution; And a gradient penalty term is introduced to ensure the continuity of the gradient of the discriminator D in space.

4. The domain generalization method based on improved meta-learning adversarial domain augmentation of claim 1, wherein: In step S3, the adversarial training of the WGAN-GP specifically includes: Using the generator G to randomly generate a hypothetical data set of a given size; Using the discriminator D to determine the similarity of the generated hypothetical data set to the original data set; According to the similarity, the hypothetical data set is regenerated until the similarity of the hypothetical data set to the original data set is close.

5. The domain generalization method based on improved meta-learning adversarial domain augmentation according to claim 1, characterized in that: The activation function SeLU has the following formula: Where x is the data to be processed, f(x) is the result after the activation function processing, and a and μ are parameters. The activation function LeakyReLU has the following formula: Where x is the data to be processed, F(x) is the result after the activation function processing, and b is the angle of the negative slope.

6. The domain generalization method based on improved meta-learning adversarial domain augmentation of claim 1, wherein: In step S4, the obtained merged data set is input into the improved feature extraction network VGG16 network for training, and the training parameters are set as follows: the initial learning rate is set to 0.001, the batch_size is set to 64, the optimizer uses SGD, the weight_decay is set to 0.0005, the momentum is set to 0.9, and the training rounds are set to 9001 rounds.

Images