A black-box source domain unsupervised domain adaptation method based on hybrid enhanced contrast

Abstract

The application belongs to the technical field of transfer learning under machine learning, and discloses a black-box source domain unsupervised field adaptive method based on mixed enhancement contrast, which comprises knowledge distillation initialization target model, mixed enhancement feature contrast learning and random mixed enhancement correction. The black-box source domain model is used to initialize the target model by using the knowledge distillation, mutual information entropy maximization and early learning regularization method, obtain the category prototype and the category learning difficulty threshold, and select the target domain sample according to the category prototype and the category learning difficulty threshold, minimize the mixed feature contrast loss, and calculate the cross-entropy correction optimization model by using the random mixed target domain sample and the pseudo label as the enhancement sample. The method can perform knowledge transfer between fields without using the source domain data and model parameters, can classify the unlabeled target domain data, can optimize the target domain inter-class structure from a more detailed angle to accurately identify the data characteristics, and can guarantee good generalization performance, and has strong security and privacy protection ability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of transfer learning under machine learning, and relates to a domain adaptation model method, in particular to a black-box unsupervised domain adaptation method based on mixed enhanced contrast. BACKGROUND

[0002] With the advent of the big data era, data is being generated at an ever-increasing rate, and the scale of data is growing explosively, which has attracted much attention to machine learning techniques capable of handling large data sets. A large amount of data provides more opportunities for training and optimization for machine learning and deep learning, thereby improving the performance and applicability of models. Although machine learning has achieved remarkable success in many fields, there are still some limitations in real-world scenarios. Traditional machine learning methods usually rely on a large amount of labeled data to build models to achieve high classification performance. However, it is not always easy or practically feasible to obtain large-scale labeled data. This brings a new challenge, i.e., how to train a model with good generalization ability under the condition of limited labeled data and accurately predict unlabeled data.

[0003] Transfer learning aims to utilize the knowledge and features in a well-trained model and apply it to a new task or domain to improve performance. Through transfer learning, we can transfer knowledge and experience from one domain to another related or similar domain, thereby saving a lot of time and resources. Domain adaptation is a branch of transfer learning, which focuses on the transfer of knowledge between different domains. In real-world scenarios, data from different domains may have inter-domain differences, such as image shooting environments, text language styles, etc., which will affect the performance of the model in the target domain. Domain adaptation aims to reduce the differences between different domains so that the model can have good generalization ability in the target domain. One common domain adaptation method is unsupervised domain adaptation, which utilizes unlabeled data in the target domain for training. Unsupervised domain adaptation usually achieves this by learning shared features between domains or adversarial learning without labeled data in the target domain.

[0004] Although unsupervised domain adaptation has achieved remarkable success, increasing concerns about data privacy have brought new challenges to this task. The data of the source domain and the target domain are usually stored on different devices and contain private information, so exposing the source domain data to the target domain poses certain risks. In other words, the already labeled source domain data may not be available for the target model, which makes some existing unsupervised domain adaptation methods no longer applicable, so there is a source-free domain adaptation method to facilitate model transfer and protect the privacy and security of source data. Source-free domain adaptation provides a well-trained source model to the unlabeled target domain instead of already labeled source domain data, so source-free domain adaptation is also called white-box domain adaptation.

[0005] However, in practical applications, white-box source domain models are not always available. Common cloud service models such as Google Cloud and Tencent Cloud are encapsulated as application programming interfaces provided to users, where only the input and output interfaces of the model are available, and the model itself is saved as a black-box interface, which makes a large number of source-free domain adaptation methods unusable in practice. The black-box domain adaptation method can only use the interface of the source domain model to access, which improves security but also brings some challenges to the domain adaptation task. The difficulty of solving the domain bias problem is caused by the inability to obtain the sample features of the source model output, and the uncertainty of the signal-to-noise ratio of the source model interface also makes the pseudo-label unreliable. SUMMARY

[0006] To solve the above technical problems, the present application provides a black-box source domain unsupervised domain adaptation method based on mixed enhanced contrast, which adds an improved mixed feature contrast module, an early learning regularization module and a random mixed enhancement module based on a knowledge distillation model, helps to learn the knowledge of shared classes and the knowledge of private classes in the target domain between the source domain and the target domain, and effectively improves the prediction accuracy of the target model.

[0007] In order to achieve the above purpose, the present application is realized by the following technical scheme:

[0008] The present application is a black-box unsupervised domain adaptation method based on mixed enhanced contrast, comprising the following steps:

[0009] Step 1, input each target domain sample into the black-box source domain to obtain the source domain prediction, which represents the probability of the sample belonging to each class in the source domain. Calculate the prototype sample of each class and the learning difficulty threshold according to the source domain prediction;

[0010] Step 2, input each target domain sample into the target model to calculate the mutual information entropy and relative entropy of the target model output as the distillation loss;

[0011] Step 3, calculate and store the distance between each sample and the class prototype sample feature as a nonlinear prediction, add an early learning regularization term, initialize the target model with the distillation loss, and iteratively update the sample feature to retain the easy-to-learn features in the early stage of model training;

[0012] Step 4, calculate the pseudo-label according to the source domain prediction obtained in step 1 and the nonlinear prediction obtained in step 3, and filter the confident non-class samples for the target samples according to the learning difficulty threshold obtained in step 1, and then mix and enhance the two samples in equal proportion to obtain the feature as a mixed negative sample;

[0013] Step 5, calculate the mixed enhanced contrastive loss according to the class prototype feature and the mixed negative sample feature obtained in step 4, the purpose is to make each target sample close to the class prototype and far away from other class prototypes and mixed negative samples;

[0014] Step 6, randomly select a pair of target samples to perform data augmentation by mixing at a ratio of 0.25 and 0.75, and calculate the cross-entropy according to the output of the target model and the pseudo label after mixing;

[0015] Step 7, calculate the gradient of the overall loss, backpropagate, and iteratively update the network parameters, class prototype features, and learning difficulty threshold until the loss converges. Predict the prediction label of the target domain data sample, compare it with the true label of the target domain data sample, and calculate the average classification accuracy of each class as the measurement result.

[0016] Further, in step 1, the prototype sample and the learning difficulty threshold of each class are calculated by the output of the black-box source domain model, as follows,

[0017]

[0018]

[0019] wherein represents the target domain sample; f s k represents the probability of the kth class predicted by the source domain model; is a hyperparameter.

[0020] Further, in step 2, a distillation loss is constructed, and the target model is updated by minimizing the distillation loss, which is composed of relative entropy and mutual information entropy, defined as follows:

[0021]

[0022]

[0023] L warm =L kd -L im

[0024] wherein D kl represents the relative entropy, f t represents the target model, h(p)=-∑ i p i logp i represents the self-information entropy.

[0025] Further, in step 3, the model training process is regularized by early learning regularization term, which retains the clean samples with correct labels in the early memory of the model and prevents the influence of noisy data. The storage The nonlinear prediction of each sample is recorded, and the new prediction is updated based on the momentum strategy. The nonlinear prediction, momentum strategy, and early learning regularization term are defined as follows:

[0026]

[0027]

[0028]

[0029] where l2() is the L2 norm, and σ is the softmax function, represents the class prototype sample, o i represents the sample The nonlinear prediction in the current model, and β is a hyperparameter.

[0030] Further, both the linear and nonlinear predictions of the target model have their limitations. In step 4, both are considered to obtain pseudo-labels, and the confident non-same-class samples are screened for the target sample for mixed enhancement. The definition is as follows:

[0031]

[0032]

[0033] where X i represents the set of target samples with the same pseudo-label as the i-th sample and a confidence greater than the learning difficulty threshold. The i-th sample is mixed with each sample in the set for mixed enhancement, and the enhanced sample features are obtained as mixed enhancement contrast negative samples. The mixed enhancement is defined as follows:

[0034]

[0035] Mix λ (a,b)=λa+(1-λ)b

[0036] Further, in step 5, the infoNCE loss function of the target domain sample is minimized, where the sample feature is used as the anchor point, the class prototype feature is used as the positive sample, and the other class prototype features and mixed enhancement sample features are used as negative samples. In order to reduce the impact of uncertain data on model training, the confidence is used as a weight coefficient. The definition is as follows:

[0037]

[0038] where z i represents the anchor sample feature, z d represents the current class prototype feature, z k represents each class prototype feature, and z ukrepresents the mixed negative sample feature, K represents the number of categories, represents the number of mixed negative samples corresponding to the sample, w i represents the confidence, tau represents the temperature coefficient, and z i , z d , z k , and z uk are normalized.

[0039] Further, in step 6, the target domain data sample pair is randomly selected for mixed data enhancement, the cross entropy is calculated according to the pseudo label after mixing, the target model is optimized again by minimizing the cross entropy, and the cross entropy is defined as follows:

[0040]

[0041] Wherein, l ce represents the cross entropy loss.

[0042] Further, in step 7, the overall loss is calculated, back propagation is performed, and the network parameters are iteratively updated, and the overall loss is defined as follows:

[0043] L = L kd -L im + alpha L elr + gamma L UniCon + delta L mix

[0044] Wherein, alpha, gamma and delta are hyperparameters.

[0045] The beneficial effects of the present application are: the present application proposes a black box source domain unsupervised field self-adaptive method based on mixed enhancement contrast, for the case that the source domain data and model parameters cannot be obtained, the model output is close to the source model output by constructing the distillation loss, and the label distribution is encouraged to be uniform, preventing the occurrence of class imbalance phenomenon; by adding early regularization term, effectively prevent the accumulation of errors caused by noise data; by mixed enhancement contrast and random mixed data enhancement, the inter-class and intra-class structure of the target data is optimized, and the data features are more accurately identified. Compared with other models, the present application has better generalization performance and strong security and privacy protection ability. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of the method of the present application.

[0047] Figure 2 is a network model overall architecture diagram of the present application. DETAILED DESCRIPTION

[0048] Embodiments of the present application will be described below with reference to drawings. Many practical details will be described in the following description for the purpose of providing a thorough understanding of the present application. However, it will be apparent to those skilled in the art that these practical details are not intended to limit the present application. That is, in some embodiments of the present application, these practical details are not necessary.

[0049] As shown in Figure 1 , the present application is a black-box source domain unsupervised domain adaptation method based on mixed augmented contrast. First, the unlabeled target picture data is preprocessed into a unified format required by the network model through size adjustment, random cropping and other methods. Model training is divided into three parts: model initialization, mixed feature contrast and random mixed augmented correction. The initialization of the model is to learn as much knowledge as possible in the black-box source domain and accurately identify the features of the target data for feature contrast in formal training. Model initialization is divided into three parts: constructing relative entropy loss to minimize the difference between source domain prediction and target prediction output, constructing mutual information entropy to encourage the prediction label distribution of the target source model to be uniform, and using the source domain prediction to select class prototype samples, calculate the class learning difficulty threshold and construct an early regularization term to optimize the feature distribution of the target model.

[0050] Specifically, the black-box source domain unsupervised domain adaptation method of the present application includes the following steps:

[0051] Step 1, input each target domain sample into the black-box source domain to obtain source domain prediction, which represents the probability of the sample belonging to each class in the source domain. According to the source domain prediction, the prototype sample and the learning difficulty threshold of each class are calculated.

[0052] In this step, the prototype sample and the learning difficulty threshold of each class are calculated according to the source domain prediction, which is specifically represented as:

[0053]

[0054]

[0055] Wherein represents the target domain sample, f s k represents the probability of the source domain model predicting the kth class, is a hyperparameter, C k represents the prototype sample, Φ k represents the learning difficulty threshold.

[0056] Step 2, input each target domain sample into the target model to calculate the mutual information entropy and the relative entropy of the source domain prediction as the distillation loss. In this step, the distillation loss is constructed, and the target model is updated by minimizing the distillation loss. The distillation loss is composed of the relative entropy and the mutual information entropy, which is defined as follows:

[0057]

[0058]

[0059] L warm =L kd -L im

[0060] where D kl denotes the relative entropy, f t denotes the target model, h(p) = -∑ i p i logp i denotes the self-information entropy, L kd denotes the knowledge distillation loss, E denotes the empirical risk, X t denotes all target domain samples, x t denotes the current target domain sample, L im denotes the mutual information maximization loss, L warm denotes the total distillation loss.

[0061] Step 3, calculate and store the distance between each target domain sample and the class prototype sample feature as a nonlinear prediction, add an early learning regularization term, initialize the target model with the distillation loss, and iteratively update all target domain sample features to retain the easy-to-learn features of the early training of the target model.

[0062] In this step, the model training process is regularized by the early learning regularization term, clean samples with correct labels that retain the early memory of the model are retained, noise data is prevented from affecting, and the storage is used to record the nonlinear prediction of each sample, and is updated based on the new prediction through a momentum strategy, the nonlinear prediction, momentum strategy, and early learning regularization term are defined as follows:

[0063]

[0064]

[0065]

[0066] where l2() is the L2 norm, σ is the softmax function, denotes the class prototype sample, β is a hyperparameter, L elr denotes the early learning regularization term, o i denotes the i-th sample the nonlinear prediction in the current model, denotes the nonlinear prediction recorded in the storage .

[0067] Step 4, calculate pseudo labels according to the source domain prediction obtained in step 1 and the nonlinear prediction obtained in step 3, filter confident non-same-class samples for the target sample according to the learning difficulty threshold obtained in step 1, and re-obtain the features after mixing and enhancing them in equal proportions to serve as mixed negative sample features.

[0068] Both linear and nonlinear predictions of the target model have their limitations. In step 4, both are integrated to obtain pseudo labels, and confident non-same-class samples for the target sample are filtered and mixed and enhanced through the pseudo labels, which are defined as follows:

[0069]

[0070]

[0071] where X i represents a set of target samples with the same pseudo label as the ith sample and a confidence greater than the learning difficulty threshold. The ith sample is mixed and enhanced with each sample in the set, and the enhanced sample features are obtained as mixed enhanced contrast negative samples. The mixed enhancement is defined as follows:

[0072]

[0073] Mix λ (a,b) = lambda a + (1-lambda) b

[0074] where a and b represent two data of the same dimension and shape, and lambda represents the mixing ratio.

[0075] Step 5, calculate the mixed enhanced contrast loss according to the class prototype sample features and the mixed negative sample features obtained in step 4, so that each target sample is close to the class prototype and far from other class prototypes and mixed negative samples.

[0076] In this step, the infoNCE loss function of the target domain sample is minimized, where the sample feature is used as the anchor point, the class prototype feature is used as the positive sample, and the other class prototype features and the mixed enhanced sample features are used as the negative samples. In order to reduce the impact of uncertain data on model training, the confidence is used as a weight coefficient, which is defined as follows:

[0077]

[0078] where z i represents the anchor sample feature, z d represents the current class prototype feature, z k represents each class prototype feature, z uk represents the mixed negative sample feature, K represents the number of classes, represents the number of mixed negative samples corresponding to the sample, w i represents the confidence, tau represents the temperature coefficient, zi, zd , z k and z uk are normalized.

[0079] Step 6, randomly selecting target sample pairs, mixing data enhancement according to the proportion of 0.25 and 0.75, calculating cross-entropy according to the output of the target model and the pseudo label after mixing.

[0080] In this step, randomly selecting target domain sample pairs, calculating cross-entropy according to the pseudo label after mixing, optimizing the target model again by minimizing the cross-entropy, and the cross-entropy is defined as follows:

[0081]

[0082] wherein l ce represents the cross-entropy loss, and represent two randomly selected samples in the target domain, y i and y j represent the pseudo label vectors of and respectively.

[0083] Step 7, calculating the gradient of the overall loss, backpropagation, iteratively updating the network parameters, class prototype features, and learning difficulty threshold until the loss converges, predicting the target domain samples to obtain the predicted label, comparing the predicted label with the true label of the target domain samples, and calculating the average classification accuracy of each class as the measurement result.

[0084] In this step, the overall loss is calculated, backpropagation is performed, and the network parameters are iteratively updated, and the overall loss is defined as follows:

[0085] L = L kd -L im + aL elr + gL UniCon + dL mix

[0086] wherein a, g and d are hyperparameters, L kd represents the knowledge distillation loss, L im represents the mutual information maximization loss, L elr represents the early learning regularization term, L UniCon represents the weight coefficient, and L mix represents the cross-entropy loss.

[0087] The following takes the Office-31 dataset as an example to illustrate the processing flow of the method of the application:

[0088] The source domain data shares the category with the target data, and no assumption is made on the source domain data and the source domain model, and no source domain information is used in the processing procedure except for the output of the source domain model.

[0089] 1. The target domain data set is converted into an RGB image with a pixel size of 256*256 after being enhanced by random cropping, random horizontal flipping, etc.; a target model composed of a ResNet model pre-trained by ImageNet, a bottleneck layer, and a classifier is used; and a nonlinear prediction memory is initialized.

[0090] 2. All target data are input into the source model to obtain source domain predictions, and prototype samples C of each class are calculated. k and a learning difficulty threshold Φ k ;

[0091] 3. The target data set is divided into multiple batches of data, each batch containing 64 target samples. A batch of data is input into the target model to obtain target model predictions, and relative entropy L kd and mutual information entropy L im are calculated. The target samples and class prototype samples are input into the target model feature extractor to obtain sample features and prototype features, nonlinear predictions are calculated, and recorded in the memory. If it is the first time to process the sample, it is directly recorded in the memory. Otherwise, the memory is updated according to the momentum strategy, and an early learning regularization term L elr is calculated. Set α to 1 and β to 0.6, update the network by back propagation, update the source domain prediction according to the EMA strategy, initialize the iteration number to 30% of the total process iteration number, and the three loss terms are still retained in the subsequent process.

[0092] 4. Calculate the sample pseudo-label vector through the source domain prediction and the nonlinear prediction, select the non-same-class sample set with a confidence greater than the class learning difficulty threshold for each sample, and perform hybrid enhancement with a λ of 0.5 with the samples in the set.

[0093] 5. For each target data in each batch, input it as an anchor sample into the target model with the class prototype samples and the hybrid enhanced samples obtained in the previous step to obtain features z i , z d , and z uk , respectively. Set τ to 0.1 to calculate L UniCon .

[0094] 6. Match the batch data with the randomly shuffled data of the same batch one by one to form sample pairs for hybrid data enhancement with a λ of 0.25, input the enhanced samples into the target model to obtain predictions, and calculate the cross-entropy loss L mix according to the pseudo-label vector after hybrid enhancement.

[0095] 7. Set gamma as 8 and delta as 1, calculate the total loss L to update the network parameters by back propagation. All target samples pass through the network once for training 1 generation, calculate the average accuracy of target data every 5 generations, save the optimal model, and train for a total of 50 generations.

[0096] The following Table 1 is the comparison data of the results of the present application and other algorithms.

[0097] Table 1

[0098]

[0099] In the above Table 1, each row in the table gives the performance of a specific method on each task. No Adapt is a ResNet50 trained using cross-entropy only, LNL-KL and LNL-OT are noise label learning methods with KL divergence and optimal transport, respectively. HD-SHOT and SD-SHOT use pseudo labels to obtain models and apply self-training and weighted cross-entropy loss on SHOT unsupervised domain adaptation methods, respectively. DINE is a black-box domain adaptation method using KL divergence and mutual information maximization, and OURS is the complete method proposed in the present application. The present application improves by 1.1 percentage points compared to DINE, proving the effectiveness of the present application.

[0100] The method of the present application performs inter-domain knowledge transfer without using source domain data and model parameters, classifies unlabeled target domain data, optimizes the structure between target domain classes from a more detailed perspective to accurately identify data features, and has good generalization performance while having strong security and privacy protection capabilities.

[0101] The above description is only the preferred scheme of the present application, and is not intended to further limit the present application. Any equivalent changes made according to the content of the specification and drawings of the present application are within the scope of protection of the present application.

Claims

1. A black-box source domain unsupervised domain adaptation method based on hybrid enhanced contrast, characterized in that: The specific steps are as follows: Step 1, obtain the target domain data set, the target domain data set is an image data set, input each target domain sample in the target domain data set into the black box source domain, obtain the source domain prediction, which represents the probability of the sample belonging to each class in the source domain, calculate the prototype sample and the learning difficulty threshold of each class according to the source domain prediction; Step 2, input each target domain sample into the target model, calculate the mutual information entropy and the relative entropy of the source domain prediction output by the target model as the distillation loss; Step 3, calculate and store the distance between each target domain sample and the class prototype sample feature as the nonlinear prediction, add the early learning regularization term, initialize the target model with the distillation loss, and iteratively update all target domain sample features to retain the easy-to-learn features of the early training of the target model; Step 4, calculate the pseudo label according to the source domain prediction obtained in step 1 and the nonlinear prediction obtained in step 3, and select the confident non-same class samples for the target sample according to the learning difficulty threshold obtained in step 1, mix and enhance them in equal proportion to obtain the feature as the mixed negative sample feature; Step 5, calculate the mixed enhanced contrastive loss according to the class prototype sample feature and the mixed negative sample feature obtained in step 4, so that each target sample is close to the class prototype and far from other class prototypes and mixed negative samples; Step 6, randomly select target sample pairs for mixed data enhancement in the proportions of 0.25 and 0.75, and calculate the cross entropy according to their output in the target model and their mixed pseudo labels; Step 7, calculate the gradient of the overall loss, back propagation, and iteratively update the network parameters, class prototype features, and learning difficulty threshold until the loss converges, and predict the target domain samples to obtain the predicted label, compare it with the true label of the target domain sample, and calculate the average classification accuracy of each class as the measurement result, wherein, in step 5, the infoNCE loss function of the target domain sample is minimized, wherein the sample feature is used as the anchor point, the class prototype feature is used as the positive sample, and the other class prototype features and the mixed enhanced sample features are used as the negative samples, and in order to reduce the influence of uncertain data on model training, the confidence is used as the weight coefficient, which is defined as follows: ， wherein denotes the anchor sample feature, denotes the current class prototype feature, denotes each class prototype feature, denotes the mixed negative sample feature, K denotes the number of classes, denotes the number of mixed negative samples corresponding to the sample, denotes the confidence, denotes the temperature coefficient, , , and are all normalized.

2. The method of claim 1, wherein the method is a hybrid enhancement contrast-based black-box source domain unsupervised domain adaptation method. In step 1, the prototype sample and the learning difficulty threshold of each class are calculated according to the source domain prediction, which is specifically represented as: ， ， wherein denotes a target domain sample, denotes a probability of the source domain model predicting the k-th class, is a hyper-parameter, denotes a prototype sample, denotes a learning difficulty threshold. 3.The method of claim 1, wherein: In step 2, the distillation loss is constructed, and the target model is updated by minimizing the distillation loss, which is composed of the relative entropy and the mutual information entropy, and is defined as follows: ， ， ， wherein denotes relative entropy, denotes a target model, denotes self-information entropy, denotes a knowledge distillation loss, denotes an empirical risk, denotes all target domain samples, denotes a current target domain sample, denotes a mutual information maximization loss, denotes a total distillation loss.

4. The method of claim 1, wherein the method is a hybrid enhancement contrast-based black-box source domain unsupervised domain adaptation method. The model training process is regularized by early learning regularization term in step 3, clean samples with correct labels that preserve the early memory of the model, prevent noise data from affecting, reservoir For recording the nonlinear prediction of each sample, and updating based on the new prediction by momentum strategy, the nonlinear prediction, momentum strategy and early learning regularization term are defined as follows: ， ， ， wherein is the L2 norm, is the softmax function, denotes a class prototype sample, is a hyperparameter, denotes an early learning regularization term, denotes the non-linear prediction of the th sample in the current model, denotes the non-linear prediction of the th sample recorded in the reservoir.

5. The method of claim 1, wherein the method is a hybrid enhancement contrast-based black-box source domain unsupervised domain adaptation method. In step 4, the pseudo label is obtained by combining the linear and nonlinear predictions of the target model, and the confident non-same class samples are selected for the target sample through the pseudo label for mixed enhancement, which is defined as follows: ， ， wherein represents a set of target samples with the same pseudo-label as the i-th sample and a confidence greater than a learning difficulty threshold, the i-th sample is mixed with each sample in the set for mixed augmentation, and the feature of the sample after augmentation is obtained as a mixed augmentation contrast negative sample. ， ， wherein , represents two data of the same dimension and shape, represents a mixing ratio.

6. The method of claim 1, wherein the method is a hybrid enhancement contrast-based black-box source domain unsupervised domain adaptation method. In step 6, randomly select target domain sample pairs for mixed data enhancement, calculate the cross entropy according to their mixed pseudo labels, and optimize the target model again by minimizing the cross entropy, which is defined as follows: ， wherein denotes the cross-entropy loss, and denotes two randomly selected samples in the target domain, and denote the pseudo label vectors of and respectively.

7. The method of claim 1, wherein the method is a hybrid enhancement contrast-based black-box source domain unsupervised domain adaptation method. In step 7, the overall loss is calculated, back propagation is performed, and the network parameters are iteratively updated, and the overall loss is defined as follows: ， wherein 、 and are hyperparameters, denotes a knowledge distillation loss, denotes a mutual information maximization loss, denotes an early learning regularization term, denotes a weight coefficient, denotes a cross-entropy loss.

Images