A passive unsupervised domain adaptation image classification method

Abstract

The application discloses a kind of passive unsupervised domain adaptation image classification method, using source domain image and its label to train a source domain pre-training model, and using the pre-training model initialization a target model;Using the initialized target model to calculate the prediction score of model structure level and data structure level of target domain image, combine the two as the confidence score of target domain sample, and be used for the cross-entropy loss weighting of target model, introduce information maximization loss to assist target model training;The image of target domain sample, pseudo label, confidence score are mixed;The certainty difference distance of double classifier is calculated to be used for the maximum minimization training of target model;Finally, train model, calculate corresponding loss and sequentially iteratively update and optimize corresponding model parameters, to clearly identify some uncertain classification target sample, then introduce weight mixing strategy to make full use of target domain knowledge, to further improve target domain classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of domain adaptation technology in transfer learning, and in particular to a passive unsupervised domain adaptation image classification method. Background Technology

[0002] Unsupervised domain adaptation leverages knowledge from a labeled source domain to help learn a discriminative model for an unlabeled target domain. However, increasing concerns about data privacy present new challenges to this task. Specifically, source and target domain data are often captured and stored on different devices and contain private information; disclosing source data to the target domain is risky. Because unsupervised domain adaptation requires access to source domain data during training, this limits the application of mainstream unsupervised domain adaptation methods in many real-world scenarios.

[0003] The goal of passive unsupervised domain adaptation is to adapt a pre-trained model in the source domain to an unlabeled target domain when the labeled source domain is not visible. It has broad application value in privacy protection and data transmission, attracting the attention of many scholars both domestically and internationally. Based on whether pseudo-source data is generated during the adaptation process, it can be further divided into data-generation-based passive unsupervised domain adaptation and feature-based passive unsupervised domain adaptation.

[0004] Most existing passive unsupervised domain adaptation methods use only one classifier for decision-making. However, due to the distributional differences between the source and target domain data, some target instances often appear near the decision boundary, leading to uncertain classifications. Furthermore, some passive unsupervised domain adaptation methods based on confidence scoring mechanisms only consider the source model's predictive knowledge of target domain samples, neglecting the intrinsic structural knowledge of the target domain samples, resulting in the incomplete extraction and utilization of target domain information. Summary of the Invention

[0005] To address the above technical problems, this invention provides a passive, unsupervised, domain-adaptive image classification method, comprising the following steps:

[0006] S1. Train a source domain pre-trained model using the source domain image and its label. The source domain pre-trained model includes a feature extractor and a dual classifier. Use the source domain pre-trained model to initialize a target model for target domain sample recognition training. During the training process, the source domain image will be inaccessible.

[0007] S2. Calculate the model structure-level prediction score of the target domain sample using the initialized target model, and calculate the data structure-level prediction score of the target domain image based on the Gaussian mixture model. Combine the two prediction scores as the confidence score of the target domain sample and use it for cross-entropy loss weighting of the target model. Introduce information maximization loss to assist in the training of the target model.

[0008] S3. Use the Mixup technique to weight the samples in the target domain.

[0009] S4. Calculate the deterministic difference distance of the two classifiers and use this deterministic difference distance for the minimization training of the target model;

[0010] S5. Perform overall training on the target model. Overall training includes steps S5.1, S5.2, and S5.3, and iteratively update and optimize the corresponding model parameters in sequence. The training of the target model is completed by iteratively alternating steps S5.1, S5.2, and S5.3, and finally, the target domain samples are classified.

[0011] The technical solution further defined in this invention is:

[0012] Furthermore, step S1 specifically includes the following sub-steps.

[0013] S1.1, Assume (x) s ,y s ) represents labeled samples in the source domain, x t For unlabeled samples in the target domain, the source domain sample set. target domain sample set Where, n s n is the number of samples in the source domain. t The number of samples in the target domain. For the i-th sample in the source domain, This is the pseudo-label corresponding to the source domain sample. Let be the i-th sample in the target domain, where the source and target domains have the same number of classes K.

[0014] S1.2, Construct a source domain pre-trained model F s Source domain pre-trained model F s It includes a feature extractor G, a classifier C1, and a classifier C2; using supervised data from the source domain combined with cross-entropy loss, the classifiers classify the samples from the source domain. The resulting classification loss is shown in the following formula:

[0015]

[0016]

[0017] In the formula, This indicates the smoothing label, where α is the smoothing coefficient;

[0018] S1.3, Using the source domain pre-trained model F s Initialize the target model F t And the target model F t Used for training the recognition of samples in the target domain.

[0019] The passive unsupervised domain-adaptive image classification method described above, in step S2, calculating the confidence score of the target domain samples includes the following steps.

[0020] S2.1. Obtain the data structure-level predicted probability of the target domain samples based on the Gaussian mixture model, using the following formula:

[0021]

[0022]

[0023] In the formula, π c μ c ,∑ c Let c, c′ ∈ {1, 2, ..., K}, and c′ ∈ {1, 2, ..., K}, respectively.

[0024] S2.2 Calculate the pseudo-labels of the target domain samples. The model structure-level prediction scores, or MP scores, for the target domain samples are calculated on classifiers C1 and C2, respectively.

[0025]

[0026]

[0027] In the formula, m∈{1,2} represent classifiers C1 and C2, respectively. in and These represent the MP scores calculated by classifier C1 and classifier C2, respectively.

[0028] S2.3 Calculate the secondary pseudo-labels of the target domain samples:

[0029]

[0030] In the formula, 'a' indicates that the pseudo-label is the 'a'th category;

[0031] S2.4 Calculate the difference between pseudo-tags and secondary pseudo-tags:

[0032]

[0033] In the formula, These are the secondary pseudo-labels for the target domain samples;

[0034] S2.5. Normalize the difference between the pseudo-label and the secondary pseudo-label into a prediction score at the data structure level of the target domain samples, i.e., the DS score.

[0035]

[0036] In the formula, i,j∈{1,2,…,n t};

[0037] S2.6. Based on the MP score and DS score of the target domain samples, calculate the final confidence score of the target domain samples on classifiers C1 and C2 respectively:

[0038]

[0039]

[0040] in, and These represent the final confidence scores of the target domain samples calculated by classifiers C1 and C2, respectively.

[0041] In the previously described passive unsupervised domain-adaptive image classification method, step S2 involves weighting the cross-entropy loss of the target domain sample classification using the confidence scores of the target domain samples.

[0042]

[0043] In the formula, It is the cross-entropy loss for classifying samples in the target domain.

[0044] In the previously described passive unsupervised domain-adaptive image classification method, step S2 involves calculating the information maximization loss and the weighted cross-entropy loss for training.

[0045]

[0046] In the formula, Represents all n t The average predicted probability of a sample being classified as the k-th class. This represents the predicted probability that the j-th target domain sample is predicted to be of the k-th class.

[0047] In the previously described passive unsupervised domain-adaptive image classification method, step S3 involves mixing the target domain sample image, pseudo-label, and confidence score. The process of mixing the target domain samples is as follows:

[0048]

[0049]

[0050]

[0051]

[0052] In the formula, η ~ Beta(α,α) and α∈{0,+∞} are weighting parameters; and Given two random target domain samples and their pseudo-labels, the sample weighted mixed loss is calculated as follows:

[0053]

[0054] Among them, L mixup This represents the weighted loss of the samples.

[0055] In the previously described passive unsupervised domain-adaptive image classification method, step S4 involves training classifiers C1 and C2 in an adversarial manner to learn transferable features and discriminative decision boundaries in the target domain. The output difference between the two classifiers is defined as follows:

[0056]

[0057] In the formula, 0(.,.) represents the Euclidean distance.

[0058] The passive unsupervised domain adaptive image classification method described above, in step S5, involves overall training of the target model, including the following steps.

[0059] S5.1. Using the loss obtained in step S2, train the feature extractor G, classifier C1, and classifier C2. The overall training loss is shown in the following formula:

[0060]

[0061] In the formula, β is the loss L im hyperparameters, θ G , as well as Let G, C1, and C2 represent the network parameters of the feature extractor G, classifier C1, and classifier C2, respectively.

[0062] S5.2. Fix the feature extractor G, and train classifiers C1 and C2 using the losses obtained in steps S3 and S4. The overall training loss is shown in the following formula:

[0063]

[0064] In the formula, λ is the hyperparameter of the weighted mixing loss, and λ is the hyperparameter of the difference loss of the bi-classifier output;

[0065] S5.3. Fix classifiers C1 and C2, and train the feature extractor G using the loss obtained in step S4. The overall training loss is shown in the following formula:

[0066]

[0067] In the formula, λ is the hyperparameter of the difference loss of the dual classifier output.

[0068] The beneficial effects of this invention are:

[0069] In this invention, the cross-entropy loss of target domain sample classification is weighted by calculating confidence scores that include knowledge of both the source and target domains, preventing the problem that the confidence scores only include source domain knowledge and ignore target domain knowledge due to the lack of target domain knowledge. Target domain knowledge is fully utilized by mixing the images, pseudo-labels, and confidence scores of target domain samples. Adversarial training with a dual classifier is used to detect target samples near the decision boundary that are uncertain in classification and to complete the correct classification. The corresponding loss is calculated through three-step training, and the model parameters are alternately updated in an adversarial manner to align the distribution of the source and target domains and achieve accurate classification of target domain samples. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0071] Figure 2 This is a schematic diagram of the framework for calculating confidence scores in this invention;

[0072] Figure 3 This is a schematic diagram of the model structure of the present invention. Detailed Implementation

[0073] This embodiment provides a passive, unsupervised, domain-adaptive image classification method, such as... Figures 1 to 3 As shown, it includes the following steps

[0074] S1. Train a source domain pre-trained model using the source domain image and its labels. The source domain pre-trained model includes a feature extractor and a dual classifier. Use the source domain pre-trained model to initialize a target model for target domain sample recognition training. During the training process, the source domain image will be inaccessible.

[0075] Step S1 specifically includes the following sub-steps.

[0076] S1.1 In passive unsupervised domain adaptation, it is assumed that (x s ,y s ) represents labeled samples in the source domain, x t For unlabeled samples in the target domain, the source domain sample set. target domain sample set Where, n s n is the number of samples in the source domain. t The number of samples in the target domain. For the i-th sample in the source domain, This is the pseudo-label corresponding to the source domain sample. Let be the i-th sample in the target domain, where the source and target domains have the same number of classes K.

[0077] S1.2, Construct a source domain pre-trained model F s Source domain pre-trained model F s It includes a feature extractor G, a classifier C1, and a classifier C2; using supervised data from the source domain combined with cross-entropy loss, the classifiers classify the samples from the source domain. The resulting classification loss is shown in the following formula:

[0078]

[0079] In the formula, This represents the smoothing label, where α is the smoothing coefficient. Based on experience, we set α to 0.1.

[0080] S1.3, Using the source domain pre-trained model F s Initialize the target model F t And the target model F t Used for training the recognition of samples in the target domain.

[0081] S2, such as Figure 2 As shown, the model structure level prediction score of the target domain sample is calculated using the initialized target model, and the data structure level prediction score of the target domain image based on the Gaussian Mixture Model (GMM) is calculated. The two prediction scores are combined as the confidence score of the target domain sample and used for cross-entropy loss weighting of the target model. Information maximization loss is introduced to assist in the training of the target model.

[0082] Calculating the confidence score for the target domain samples includes the following steps:

[0083] S2.1. Obtain the data structure-level predicted probability of the target domain samples based on the Gaussian mixture model, using the following formula:

[0084]

[0085]

[0086] In the formula, π c μ c ,∑ c Let c, c′ ∈ {1, 2, ..., K}, and c′ ∈ {1, 2, ..., K}, respectively.

[0087] S2.2 Calculate the pseudo-labels of the target domain samples. The model structure-level prediction scores (MP scores) for the target domain samples are calculated on classifiers C1 and C2, respectively.

[0088]

[0089]

[0090] In the formula, m∈{1,2} represent classifiers C1 and C2, respectively. in and These represent the MP scores calculated by classifier C1 and classifier C2, respectively.

[0091] S2.3 Calculate the secondary pseudo-labels of the target domain samples:

[0092]

[0093] In the formula, 'a' indicates that the pseudo-label is the 'a'th category;

[0094] S2.4 Calculate the difference between pseudo-tags and secondary pseudo-tags:

[0095]

[0096] In the formula, These are the secondary pseudo-labels for the target domain samples;

[0097] S2.5. Normalize the difference between the pseudo-label and the secondary pseudo-label into a prediction score at the data structure level of the target domain samples (referred to as the DS score).

[0098]

[0099] In the formula, i,j∈{1,2,…,n t};

[0100] S2.6. Based on the MP score and DS score of the target domain samples, calculate the final confidence score of the target domain samples on classifiers C1 and C2 respectively:

[0101]

[0102]

[0103] in, and These represent the final confidence scores of the target domain samples calculated by classifiers C1 and C2, respectively.

[0104] The cross-entropy loss for classifying target domain samples is weighted using the confidence scores of the target domain samples:

[0105]

[0106] In the formula, It is the cross-entropy loss for classifying samples in the target domain.

[0107] The computational information maximization loss is used in auxiliary weighted cross-entropy loss training, which makes the target output individually deterministic and globally diverse.

[0108]

[0109] In the formula, Represents all n t The average predicted probability of a sample being classified as the k-th class. This represents the predicted probability that the j-th target domain sample is predicted to be of the k-th class.

[0110] S3. Use the Mixup technique to weight the samples in the target domain:

[0111] Some samples in the target domain can be classified well after model training; these are called high-confidence samples. The remaining samples, after being trained with the same model, have poor classification results and contain a lot of noise; these are called low-confidence samples. Although the labels of low-confidence samples contain significant noise, they also contain a lot of important target domain knowledge.

[0112] By mixing the images, pseudo-labels, and confidence scores of the target domain samples, a more robust target model can be trained by fully utilizing target domain knowledge. The process of mixing target domain samples is as follows:

[0113]

[0114]

[0115]

[0116]

[0117] In the formula, η ~ Beta(α,α) and α∈{0,+∞} are weighting parameters; and Given two random target domain samples and their pseudo-labels, the sample weighted mixed loss is calculated as follows:

[0118]

[0119] Among them, L mixup This represents the weighted loss of the samples.

[0120] S4. Calculate the deterministic discrepancy distance of the two classifiers and use this deterministic discrepancy distance for minimizing the target model:

[0121] Classifiers C1 and C2 are trained adversarially to learn transferable features and discriminative decision boundaries in the target domain. The output difference between the two classifiers is defined as follows:

[0122]

[0123] In the formula, d(.,.) is the Euclidean distance.

[0124] First, fix the feature extractor G, and maximize L. cdd We train classifiers C1 and C2 to promote diverse predictions so that we can detect target samples with uncertain classifications that are near the decision boundary and have different predictions; then we fix classifiers C1 and C2 and minimize L. cdd To train the feature extractor G, we encourage prediction consistency and promote alignment between the source domain distribution and the target domain distribution.

[0125] S5. Perform overall training on the target model. Overall training includes steps S5.1, S5.2, and S5.3, and iteratively update and optimize the corresponding model parameters in sequence. The training of the target model is completed by iteratively alternating steps S5.1, S5.2, and S5.3, and finally, the target domain samples are classified.

[0126] S5.1. Using the loss obtained in step S2, train the feature extractor G, classifier C1, and classifier C2 to ensure that the generalization bound still exists, thereby producing different outputs. The overall training loss is shown in the following formula:

[0127]

[0128] In the formula, β is the loss L im hyperparameters, θ G , as well as Let G, C1, and C2 represent the network parameters of the feature extractor G, classifier C1, and classifier C2, respectively.

[0129] S5.2. Fix the feature extractor G, and use the loss obtained in steps S3 and S4 to train classifiers C1 and C2 to detect target samples with uncertain classifications that are located near the decision boundary and have different predictions. This makes the target model robust. The overall training loss is shown in the following formula:

[0130]

[0131] In the formula, Let λ be the hyperparameter of the weighted mixing loss, and let λ be the hyperparameter of the difference loss of the bi-classifier output. Based on experience, we set λ to 0.01.

[0132] S5.3. Fix classifiers C1 and C2, and train the feature extractor G using the loss obtained in step S4. This encourages prediction consistency and promotes the alignment of the source domain distribution with the target domain distribution. The overall training loss is shown in the following formula:

[0133]

[0134] In the formula, λ is the hyperparameter of the difference loss of the dual classifier output.

[0135] To verify the effectiveness of the image classification method of this invention, experiments were conducted on the Office-Home dataset and the Office-31 dataset. The Office-Home dataset contains 65 categories and 15,500 images, which come from four visual domains: Art (art images in the form of sketches, paintings, decorations, etc.), Clipart (collections of clipart images), Product (images of objects without background), and Real-World (images of objects taken by ordinary cameras). The effectiveness of the image classification method of this invention was evaluated in twelve domain adaptation tasks: Ar->Cl, Ar->Pr, Ar->Rw, Cl->Ar, Cl->Pr, Cl->Rw, Pr->Ar, Pr->Cl, Pr->Rw, Rw->Ar, Rw->Cl, and Rw->Pr.

[0136] The Office-31 dataset contains 4,652 images across 31 categories. These images come from three visual domains: Amazon (e-commerce images), Webcam (images taken by webcams), and DSLR (images taken by DSLR cameras). The effectiveness of the image classification method of this invention is evaluated in six domain adaptation tasks: A->W, A->D, D->A, D->W, W->A, and W->D.

[0137] The experimental results are shown in Tables 1 and 2, respectively. BCWA represents the method of this invention and is indicated in bold. For ease of display, we have abbreviated the names of each task on the Office-Home dataset, using only the first letter.

[0138] Table 1. Validation of the effectiveness of the method of the present invention on the Office-Home dataset.

[0139]

[0140]

[0141] As shown in Table 1, the BCWA method proposed in this invention achieves an average classification accuracy of 74.0%. Except for a few tasks where its performance is slightly lower than other methods, the accuracy of the Ar->Pr task reaches 80.5%, the accuracy of the Pr->Ar task reaches 68.7%, and the accuracy of the Rw->Ar task reaches 75.3%. The average accuracy on these tasks is significantly improved compared to other methods, and the overall average classification accuracy of the BCWA method is significantly better than other classification methods.

[0142] Table 2. Validation of the method of the present invention on the Office-31 dataset.

[0143] Method DMCD SHOT DIPE NRC BAIT ASL CoWA BCWA A->D 94.1 94.0 96.6 96.0 92.0 93.4 94.2 95.6 A->W 93.5 90.1 93.1 90.8 94.6 94.1 95.7 94.7 D->A 76.4 74.7 75.5 75.3 74.6 76.0 77.3 78.7 D->W 98.8 98.4 98.4 99.0 98.1 98.4 98.4 99.6 W->A 76.4 74.3 77.2 75.0 75.2 75.0 78.0 78.8 W->D 100.0 99.9 99.6 100.0 100.0 99.8 99.8 100.0 Average 89.9 88.6 90.1 89.4 89.1 89.5 90.6 91.2

[0144] As shown in Table 2, the average classification accuracy of the BCWA method of this invention reached 91.2%, which is significantly better than the accuracy of CoWA (90.6%). The BCWA method achieved 100% classification accuracy on the W->D task. In addition, in the D->A, D->W, and W->A tasks, BCWA achieved classification accuracy of 78.7%, 99.6%, and 78.8%, respectively, which is significantly better than the accuracy of CoWA on the three tasks (77.3%, 98.4%, and 78.0%).

[0145] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A passive, unsupervised, domain-adaptive image classification method, characterized in that: Includes the following steps S1. Train a source domain pre-trained model using the source domain image and its label. The source domain pre-trained model includes a feature extractor and a dual classifier. Use the source domain pre-trained model to initialize a target model for target domain sample recognition training. During the training process, the source domain image will be inaccessible. S2. Calculate the model structure-level prediction score of the target domain sample using the initialized target model, and calculate the data structure-level prediction score of the target domain image based on the Gaussian mixture model. Combine the two prediction scores as the confidence score of the target domain sample and use it for cross-entropy loss weighting of the target model. Introduce information maximization loss to assist in the training of the target model. S3. Use the Mixup technique to perform weighted mixing on the target domain samples; mix the images, pseudo-labels, and confidence scores of the target domain samples. The process of mixing the target domain samples is as follows: In the formula, η ~ Beta(α, α) and α∈{0, +∞}, is the weighting parameter; and Given two random target domain samples and their pseudo-labels, the sample weighted mixed loss is calculated as follows: Among them, L mixup This represents the weighted loss of the samples. S4. Calculate the deterministic difference distance of the two classifiers and use this deterministic difference distance for the minimization training of the target model; S5. Perform overall training on the target model. Overall training includes steps S5.1, S5.2, and S5.3, and iteratively update and optimize the corresponding model parameters in sequence. The training of the target model is completed by iteratively alternating steps S5.1, S5.2, and S5.3, and finally, the target domain samples are classified.

2. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: Step S1 specifically includes the following sub-steps. S1.1, Assume (x) s y s ) represents labeled samples in the source domain, x t For unlabeled samples in the target domain, the source domain sample set. target domain sample set Where, n s n is the number of samples in the source domain. t The number of samples in the target domain. For the i-th sample in the source domain, This is the pseudo-label corresponding to the source domain sample. Let K be the i-th sample in the target domain, and let the source domain and the target domain have the same number of classes K. S1.2, Construct a source domain pre-trained model F s Source domain pre-trained model F s It includes a feature extractor G, a classifier C1, and a classifier C2; using supervised data from the source domain combined with cross-entropy loss, the classifiers classify the samples from the source domain. The resulting classification loss is shown in the following formula: In the formula, This indicates the smoothing label, where α is the smoothing coefficient; S1.3, Using the source domain pre-trained model F s Initialize the target model F t And the target model F t Used for training the recognition of samples in the target domain.

3. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: In step S2, calculating the confidence score of the target domain sample includes the following steps: S2.

1. Obtain the data structure-level predicted probability of the target domain samples based on the Gaussian mixture model, using the following formula: In the formula, π c μ c ,∑ c Let c and c' represent the mixing coefficient, mean vector, and covariance matrix of category c, respectively, and c, c' ∈ {1, 2, ..., K}; S2.2 Calculate the pseudo-labels of the target domain samples. The model structure-level prediction scores, or MP scores, for the target domain samples are calculated on classifiers C1 and C2, respectively. In the formula, They represent classifier C1 and classifier C2, respectively. in and These represent the MP scores calculated by classifier C1 and classifier C2, respectively. S2.3 Calculate the secondary pseudo-labels of the target domain samples: In the formula, 'a' indicates that the pseudo-label is the 'a'th category; S2.4 Calculate the difference between pseudo-tags and secondary pseudo-tags: In the formula, These are the secondary pseudo-labels for the target domain samples; S2.

5. Normalize the difference between the pseudo-label and the secondary pseudo-label into a prediction score at the data structure level of the target domain samples, i.e., the DS score. In the formula, i, j∈{1, 2, ..., n} t }; S2.

6. Based on the MP score and DS score of the target domain samples, calculate the final confidence score of the target domain samples on classifiers C1 and C2 respectively: in, and These represent the final confidence scores of the target domain samples calculated by classifiers C1 and C2, respectively.

4. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: In step S2, the cross-entropy loss for classifying the target domain samples is weighted using the confidence scores of the target domain samples: In the formula, It is the cross-entropy loss for classifying samples in the target domain.

5. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: In step S2, the information maximization loss is calculated and weighted cross-entropy loss is used for training. In the formula, Represents all n t The average predicted probability of a sample being classified as the k-th class. This represents the predicted probability that the j-th target domain sample is predicted to be of the k-th class.

6. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: In step S4, classifiers C1 and C2 are trained adversarially to learn transferable features and discriminative decision boundaries in the target domain. The output difference between the two classifiers is defined as follows: In the formula, d(.,.) is the Euclidean distance.

7. The passive unsupervised domain-adaptive image classification method according to claim 1, characterized in that: In step S5, the target model is trained as a whole, including the following steps: S5.

1. Using the loss obtained in step S2, train the feature extractor G, classifier C1, and classifier C2. The overall training loss is shown in the following formula: In the formula, β is the loss L im hyperparameters, θ G , as well as Let G, C1, and C2 represent the network parameters of the feature extractor G, classifier C1, and classifier C2, respectively. S5.

2. Fix the feature extractor G, and train classifiers C1 and C2 using the losses obtained in steps S3 and S4. The overall training loss is shown in the following formula: In the formula, λ is the hyperparameter of the weighted mixing loss, and λ is the hyperparameter of the difference loss of the bi-classifier output; S5.

3. Fix classifiers C1 and C2, and train the feature extractor G using the loss obtained in step S4. The overall training loss is shown in the following formula: In the formula, λ is the hyperparameter of the difference loss of the dual classifier output.

Images