Label Noisy Image Classification Method Based on Graph Consistency and Semi-Supervised Model

Abstract

The present invention relates to the field of computer vision and artificial intelligence technologies. More specifically, it relates to a method for classifying label-noisy images based on graph consistency and semi-supervised models. S1: Train the full amount of noisy label data in the training set to initialize the model; S2: Obtain the distribution of samples; S3: Classify and screen the samples; S4: Perform different data augmentations on the images; S5: Feed the augmented images into the semi-supervised model for training; S6: Perform graph encoding to obtain a consistency graph; S7: Jointly optimize the model and update the sample labels; S8: Use the semi-supervised model as the inference model. The present invention mainly solves the problem that the existing semi-supervised methods based only on classification consistency face very serious confidence bias problems and remember a large number of noisy labels, resulting in error accumulation and harm to the model. The graph consistency model of the present invention can effectively combat the memory of noisy labels and significantly improve the model performance.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer vision and artificial intelligence, and more particularly, to a method for classifying label-noisy images based on Figure 1 consistency and semi-supervised models. Background Art

[0002] Nowadays, the great success of deep learning is inseparable from a large amount of correctly labeled data. However, the annotation of these large-scale data sets often requires huge human, financial, and time costs. Even for some data (such as medical images), it is difficult for experts to correctly classify the data. Moreover, there are many scenarios in current computer vision research, and such a high cost of data set collection brings great challenges to model training in specific scenarios.

[0003] To solve the problem of excessive cost when collecting accurately labeled data, two low-cost solutions, crowdsourcing and web query, have become the preferred choices for enterprises and large institutions. The crowdsourcing solution hires a large number of cheap workers, which is more efficient and has lower costs compared to the expert annotation team. However, the consequence of high efficiency is that the annotation quality is worse than that of the expert team. The latter relies on the massive data and powerful search capabilities of web search engines and can obtain a large amount of image data labeled by users from public platforms such as Google and Baidu in a short time. Although the two methods have significantly improved the efficiency of image annotation, there are inevitably a large number of errors in image annotation. Such labels that are inconsistent with the ground-truth label are called noisy labels, and the labels of these images do not match their true semantic information. In recent years, noisy label learning has been widely studied as an important topic.

[0004] The Learning from noisy labels algorithm refers to effectively using data with noisy labels for training in the case of a certain amount of incorrect labels in the dataset, eliminating its negative impact to better complete relevant tasks. By learning from the noisy label data, the accuracy of the model can be improved. On the other hand, due to its huge capacity, a deep neural network can fit all training labels. However, when a deep neural network is trained under noisy supervision, the model will inevitably fit all the noisy labels, which will cause great damage to the generalization of the model. The model tends to remember the noisy labels during the training process, resulting in serious damage to the performance of the deep learning model. Incorrect data labels not only fail to provide useful information but also interfere with the entire task. Therefore, the more accurately the noisy labels are learned, the higher the accuracy of task execution will be, which will also have a positive impact on task execution. In the actual application process, label noise is a common problem in the dataset. There are always incorrect labels in the data and the quantity is very large. It is difficult to avoid the risks brought by incorrect labels in a unified mode. Therefore, how to effectively use noise for training and eliminate its negative impact is a very meaningful direction.

[0005] The most prominent methods in the field of noisy label learning currently are to combine specific sample selection strategies with specific semi-supervised learning models, effectively preventing the model from overfitting to noisy data and further improving the generation performance. The SELF model (Learning to filter noisy labels with self-ensembling) combines with semi-supervised learning methods to gradually filter out mislabeled examples from noisy data. By maintaining a running average model called mean-teacher (Mean teachers are better role models) as the backbone, it obtains self-ensemble predictions for all training examples, and then gradually deletes the examples whose ensemble predictions are inconsistent with their annotated labels. This method further utilizes the unsupervised loss of the examples not included in the selected clean set. The RoCL model (Robust curriculum learning) adopts a two-stage learning strategy: supervised training on the selected clean examples, and then semi-supervised learning on the relabeled noisy examples through self-supervision. For selection and relabeling, it calculates the exponential moving average of the loss in the training iterations. The SELFIE model (Refurbishing unclean samples for robust deep learning) is a hybrid method of sample selection and loss correction, correcting the loss of the refurbishable samples (i.e., loss correction), and then using it together with the loss of the samples with small losses (i.e., sample selection). Therefore, more training samples need to be considered to update the deep neural network. However, these semi-supervised methods based only on classification consistency training face very serious confidence bias problems and memorize a large number of noisy labels, with errors accumulating and hurting the performance of the model. Summary of the Invention

[0006] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] Based on Figure 1 A method for classifying label-noisy images based on consistency and semi-supervised models, comprising the following steps:

[0008] S1: Train all the noisy label data in the training set and initialize the model;

[0009] S2: Obtain the distribution of the samples according to the initialized model in step S1, and then use a Gaussian mixture model to divide the samples into correctly labeled samples and incorrectly labeled samples according to the distribution of the samples;

[0010] S3: Classify and screen the samples obtained in step S2: Retain the labels for the samples with correct labels as labeled images, and erase the labels for the samples with incorrect labels as unlabeled images; Transmit both the labeled images and the unlabeled images to the semi-supervised model;

[0011] S4: Perform different data augmentations on the labeled images and the unlabeled images;

[0012] S5: Feed the augmented labeled images and unlabeled images into the semi-supervised model for training;

[0013] S6: Perform graph encoding on the results obtained from the semi-supervised model training in step S5 to obtain a consistency graph;

[0014] S7: Use class distribution consistency and Figure 1 consistency regularization to jointly optimize the model and update the sample labels;

[0015] S8: Use the semi-supervised model obtained by constraining the model in step S7 as an inference model, and input the image data of the test set into it to obtain the result of classifying the noisy label data.

[0016] Preferably, use the full amount of data containing noisy labels to train the ResNet50 model; In the first epoch of the model training stage, read the data of the samples with correct labels and the samples with incorrect labels; Perform augmentations on the data, and the augmentations are flipping, cropping, and shifting data augmentations; Feed the augmented data into the ResNet50 model, and train two ResNet50 models simultaneously; The model outputs a probability distribution, calculate the cross-entropy loss function according to the probability distribution, and perform backpropagation to train the model and save the model parameters.

[0017] Preferably, input the probability distribution output by the model into a Gaussian mixture model, and perform distribution calculation through the Gaussian mixture model. The Gaussian mixture model can be expressed by the following formula:

[0018]

[0019] Use two two-dimensional Gaussian distributions for two clusters, with the number of components K = 2, and satisfying:

[0020]

[0021] Among them, N(x|μ k , ∑ k ) is: the k-th component in the mixture model; π k is: the weight of each component N(x|μ k , ∑ k ).

[0022] Preferably, the step S3 includes:

[0023] S301. For p(x) obtained in step three, if p(x) is greater than or equal to 0.5, its label is correct, and the original label is saved; if p(x) is less than 0.5, its label is incorrect, and the label is not saved.

[0024] S302. Send the labeled images and unlabeled images together into the semi-supervised model for training preparation.

[0025] Preferably, the step S4 includes:

[0026] S401. Perform weak augmentation operations on the labeled images, and perform data augmentation on the unlabeled images using weak augmentation techniques and AutoAugment techniques respectively. The weak augmentation operations include flipping, cropping, and shifting data augmentation.

[0027] S402. Perform mixup operations on the labeled images and the unlabeled images simultaneously; Mixup can be represented by the following formula:

[0028]

[0029]

[0030] where, x i and x j represent the input images, and the two pictures are fused; y i and y j represent the labels of the input images, and the label form is one-hot encoding format, and their labels are fused; λ is a fusion ratio with a value range of 0 to 1.

[0031] Preferably, the step S5 includes:

[0032] S501. For the augmented labeled images, use the ResNet50 model for training as the labeled module of the semi-supervised model, and use a dual-network structure.

[0033] S502. For the augmented unlabeled images, use the ResNet50 model for training as the unlabeled module of the semi-supervised model, and use a dual-network structure. Among them, after a group of unlabeled images enhanced by weak augmentation are trained using the ResNet50 model, a probability distribution result is generated. After the unlabeled images enhanced by the AutoAugment technique are trained using the ResNet50 model, a probability distribution representation, an embedding representation, and a feature map-level representation are generated.

[0034] Preferably, the step S6 includes:

[0035] S601. For the multi-scale results obtained from training the unlabeled images enhanced by the AutoAugment technology in step 5, perform autocorrelation calculations on all of them. First, use the probability distribution representation C U and its transpose matrix to calculate the similarity, obtaining the autocorrelation map

[0036]

[0037] Use the embedding representation E U and its transpose matrix to calculate the similarity, obtaining the autocorrelation map

[0038]

[0039] Use the feature map-level representation F U and its transpose matrix to calculate the similarity, obtaining the autocorrelation map

[0040]

[0041] where τ is the temperature hyperparameter;

[0042] S602. Perform normalization on :

[0043]

[0044]

[0045]

[0046] where represents the similarity degree between the i-th sample and the j-th sample in terms of class distribution, embedding, and feature map dimensions, and B represents the Batchsize;

[0047] S603. For the probability distribution representation obtained from training another batch of weakly enhanced unlabeled images in step 5, perform autocorrelation calculations on all of them. First, use the probability distribution representation C U2 and its transpose matrix to calculate the similarity, obtaining the autocorrelation map

[0048]

[0049] Also perform normalization on it:

[0050]

[0051] At this time, As a pseudo-label graph.

[0052] Preferably, the step S7 includes:

[0053] S701. For labeled samples, calculate their class consistency using cross-entropy loss The calculation method of cross-entropy loss is:

[0054]

[0055] where y is the true label; is the one-hot encoding of the predicted label; N is the number of samples; M is the number of classes.

[0056] S702. Calculate its class consistency using L2 loss, and the calculation method of L2 loss is as follows:

[0057]

[0058] where p0(y i |Aug2(u i )) represents the class distribution probability parameter θ after the model passes through softmax;

[0059] S703. At the same time, perform Figure 1 consistency calculation on unlabeled samples, and the calculation methods of the probability distribution graph and the pseudo-label Figure 1 consistency are as follows:

[0060]

[0061] The calculation methods of the embedding graph and the pseudo-label Figure 1 consistency are as follows:

[0062]

[0063] The calculation methods of the feature map and the pseudo-label Figure 1 consistency are as follows:

[0064]

[0065] Finally, add them up to obtain Figure 1 the consistency regularization term:

[0066]

[0067] S704. The final loss is the sum of the above several losses:

[0068]

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] The present invention provides a method for classifying label-noisy images based on Figure 1 consistency and semi-supervised models. First, the present invention uses the full amount of data containing noisy labels for model training in the first epoch. Using the full amount of data containing noisy labels to train the ResNet50 model in the first epoch can learn the features of the full amount of data to obtain better initial model parameters, which is beneficial to the training of subsequent epochs. Secondly, the present invention uses a Gaussian Mixture Model (GMM) to divide whether the sample labels are correct. The samples with correct labels retain their labels, otherwise the labels are erased, and then a semi-supervised model is used to learn the labeled samples and unlabeled samples. The Gaussian Mixture Model can make a relatively accurate division according to the distribution of the samples, erase the labels of the samples with wrong labels, and adopt a semi-supervised learning method for training, which can give full play to the role of the data, reduce the model's memory of wrong labels, and improve the model performance. Finally, the present invention designs Figure 1 a consistency model. The semi-supervised method based only on classification consistency training faces very serious confidence bias problems and remembers a large number of noisy labels, and the errors accumulate and damage the model performance. Our Figure 1 consistency model can effectively combat the memory of noisy labels and significantly improve the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is the overall flowchart of the present invention;

[0072] Figure 2 is the model structure diagram of the unlabeled sample training in the semi-supervised training module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0074] As Figure 1-2 shown, the method for classifying label-noisy images based on Figure 1 consistency and semi-supervised models includes the following steps:

[0075] S1. Use the full amount of data containing noisy labels to train the ResNet50 model;

[0076] S101. In the first epoch of the model training stage, read all the training set data, where the training set data includes correctly labeled samples and mislabeled samples.

[0077] S102. Perform data augmentation, where the augmentation here refers to performing weak augmentation operations such as flipping, cropping, and shifting data augmentation.

[0078] S103. Feed the augmented data into the ResNet50 model. We train two ResNet50 models simultaneously because for the problem of learning with noisy labels, the dual-network structure can effectively resist the memory of incorrect labels, prevent overfitting, and the model training structure will be more stable.

[0079] S104. The model outputs a probability distribution. We calculate the cross-entropy loss function based on the probability distribution and perform backpropagation to train the model. The calculation method of the cross-entropy loss function is as follows:

[0080]

[0081] where y is the true label; is the one-hot encoding of the predicted label; N is the number of samples; M is the number of classes.

[0082] S105. Save the model parameters.

[0083] S2. Use the Gaussian mixture model for sample partitioning to obtain correctly labeled samples and mislabeled samples;

[0084] S201. For the probability distribution output by the model, we input it into the Gaussian mixture model.

[0085] S202. Perform distribution calculation through the Gaussian mixture model. The mixture Gaussian model can be expressed by the following formula:

[0086]

[0087] where N(x|μ k , ∑ k ) is called the k-th component in the mixture model. To partition correctly labeled samples and mislabeled samples, we use two two-dimensional Gaussian distributions for two clusters, so the number of components K = 2, π k is the mixture coefficient, and it satisfies:

[0088]

[0089] It can be considered that π k is the probability of each component N(x|μ k , ∑k )'s weight.

[0090] S3. Divide the samples partitioned by the Gaussian mixture model. Keep the labels of the samples with correct labels as labeled data, erase the labels of the samples with incorrect labels as unlabeled data, and send the labeled data and unlabeled data into the semi-supervised model together;

[0091] S301. For p(x) obtained in step three, if p(x) is greater than or equal to 0.5, we consider its label to be correct and save the original label; if p(x) is less than 0.5, we consider its label to be incorrect and do not save the label;

[0092] S302. Send the labeled images and unlabeled images into the semi-supervised model for training preparation;

[0093] S4. Use the AutoAugment technique for image enhancement and perform the mixup operation on the data simultaneously;

[0094] S401. We perform different data augmentations on the labeled data and unlabeled data. Perform weak augmentation operations on the labeled images. Here, the augmentation refers to weak augmentation operations such as flipping, cropping, and shifting data augmentation. For the unlabeled images, use the above-mentioned weak augmentation technique and the AutoAugment technique for data augmentation respectively. At this time, we obtain the weakly augmented labeled images, weakly augmented unlabeled images, and unlabeled images augmented by the AutoAugment technique.

[0095] S402. To alleviate overfitting to mislabeled images, we apply the mixup operation. Mixup is also a way of data augmentation. It can make the discrete sample space continuous and improve the smoothness within the neighborhood. Mixup can be expressed by the following formula:

[0096]

[0097]

[0098] where x i and x j represent the input images, and fuse the two pictures; y i and y j represent the labels of the input images. The label form is one-hot encoding format, and fuse their labels; λ is a fusion ratio with a value range of 0 to 1, and is calculated according to the beta distribution (α and β are equal when calculating, that is, both take α).

[0099] S5. Use the semi-supervised model to train the augmented data;

[0100] S501. For the enhanced labeled images, the labeled module of the semi-supervised model is trained using the ResNet50 model with a dual-network structure. The dual-network structure can effectively enhance the robustness of the model to noisy labels.

[0101] S502. For the enhanced unlabeled images, the unlabeled module of the semi-supervised model is trained using the ResNet50 model with a dual-network structure. Among them, after a group of weakly enhanced unlabeled images are trained using the ResNet50 model, a probability distribution result is generated. After the unlabeled images enhanced by the AutoAugment technology are trained using the ResNet50 model, a probability distribution representation, an embedding representation, and a feature map-level representation are generated. Generating results in more dimensions can provide us with more hierarchical information for classification, which is beneficial for us to better avoid memorizing incorrect labels and making incorrect fittings to them in the case of noisy images, thus affecting the model effect.

[0102] S6. Perform graph encoding on the results obtained by semi-supervised learning to obtain our consistency graph. Please refer to Figure 2 ;

[0103] S601. For the multi-scale results obtained from training a batch of unlabeled images enhanced by the AutoAugment technology in step five, we perform autocorrelation calculations on all of them. First, use the probability distribution representation C U and its transpose matrix to calculate the similarity, obtaining the autocorrelation graph

[0104]

[0105] Use the embedding representation F U and its transpose matrix to calculate the similarity, obtaining the autocorrelation graph

[0106]

[0107] Use the feature map-level representation F U and its transpose matrix to calculate the similarity, obtaining the autocorrelation graph

[0108]

[0109] where τ is the temperature hyperparameter.

[0110] S602. We perform normalization on :

[0111]

[0112]

[0113]

[0114] Among them, represents the similarity degree between the i-th sample and the j-th sample in terms of class distribution, embedding, and feature map dimension, and B represents the Batchsize.

[0115] S603. For the probability distribution representations obtained from training another batch of weakly augmented unlabeled images in Step 5, we perform autocorrelation calculations on all of them. First, we use the probability distribution representation C U and its transposed matrix to calculate the similarity, obtaining the autocorrelation map

[0116]

[0117] We also perform normalization on it:

[0118]

[0119] At this time, is used as the pseudo-label map.

[0120] S7. Adopt class consistency and Figure 1 consistency regularization to jointly optimize the model;

[0121] S701. For labeled samples, apply cross-entropy loss to calculate their class consistency The calculation method of cross-entropy loss is shown in S104.

[0122] S702. We use L2 loss to calculate their class consistency. The calculation method of L2 loss is as follows:

[0123]

[0124] In the above formula, p θ (y i |Aug2(u i )) represents the class distribution probability parameter θ after the model passes through softmax;

[0125] S703. At the same time, perform Figure 1 consistency calculation on unlabeled samples. The calculation methods of the probability distribution map and the pseudo-label Figure 1 consistency are as follows:

[0126]

[0127] Embedding diagram and pseudo labels Figure 1 The calculation method of consistency is as follows:

[0128]

[0129] Feature map and pseudo labels Figure 1 The calculation method of consistency is as follows:

[0130]

[0131] Finally, add them up to get Figure 1 Consistency regularization term:

[0132]

[0133] S704. The final loss is the sum of the above losses:

[0134]

[0135] S8. Use the obtained semi-supervised model as the inference model, input the image data of the test set into it, and obtain the result of noise label image classification.

[0136] The above only elaborates on the preferred embodiments of the present invention in detail. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the purpose of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. A method for classifying label-noisy images based on graph consistency and semi-supervised models, characterized in that: It includes the following steps: S1: Train all the noisy label data in the training set to initialize the model; S2: Obtain the distribution of the samples according to the initialized model in step S1, and then use the Gaussian mixture model to divide the samples into correctly labeled samples and incorrectly labeled samples according to the distribution of the samples; S3: Classify and screen the samples obtained in step S2: The samples with correct labels retain their labels and serve as labeled images, and the samples with incorrect labels have their labels erased and serve as unlabeled images; Transmit both the labeled images and the unlabeled images to the semi-supervised model; S4: Perform different data augmentations on the labeled images and the unlabeled images; S5: Feed the augmented labeled images and unlabeled images into the semi-supervised model for training; S6: Perform graph encoding on the results obtained from training the semi-supervised model in step S5 to obtain a consistency graph, where: Step S6 includes: S601. For the multi-scale results obtained from the unlabeled images enhanced by the AutoAugment technology in the step S5, perform autocorrelation calculations on all of them. First, use a probability distribution to represent C U and its transpose matrix to calculate the similarity and obtain an autocorrelation map Use the embedding to represent E U and its transposed matrix Perform similarity calculation to obtain the autocorrelation graph Using the feature map-level representation F U and its transposed matrix to calculate the similarity, obtaining the autocorrelation graph where τ is the temperature hyperparameter; S602. Normalize as follows: Among them, represents the similarity degree between the $i$-th sample and the $j$-th sample in terms of class distribution, embedding, and feature map dimension, and $B$ represents the Batchsize; S603. For the probability distribution representations obtained by training another batch of weakly augmented unlabeled images in step S5, perform autocorrelation calculations on all of them. First, use the probability distribution representation C U2 and its transpose matrix to calculate the similarity and obtain the autocorrelation map Also perform normalization on it: At this time, serve as a pseudo-label map; S7: Jointly optimize the model using class distribution consistency and graph consistency regularization, and update the sample labels; S8: Use the semi-supervised model obtained by constraining the model in step S7 as the inference model, and input the image data of the test set into it to obtain the result of classifying the noisy label data.

2. The method for classifying label-noisy images based on graph consistency and semi-supervised models according to claim 1, characterized in that: Use the full amount of data containing noisy labels to train the ResNet50 model; In the first epoch of the model training stage, read the correctly labeled sample and incorrectly labeled sample data; Augment the data, and the augmentation is to perform flipping, cropping, and shifting data augmentations; Feed the augmented data into the ResNet50 model, and train two ResNet50 models simultaneously; The model outputs a probability distribution, calculate the cross-entropy loss function according to the probability distribution, and perform backpropagation to train the model and save the model parameters.

3. The method for classifying label-noisy images based on graph consistency and semi-supervised models according to claim 2, characterized in that: Input the probability distribution output by the model into the Gaussian mixture model, and perform distribution calculation through the Gaussian mixture model. The mixture Gaussian model can be expressed by the following formula: Use two two-dimensional Gaussian distributions for two clusters, with the number of components K = 2, and satisfy: where, N(x∣μ k ,Σ k ) is the k-th component in the mixture model; π k is the weight of each component N(x∣μ k ,Σ k ).

4. The method for classifying label-noisy images based on graph consistency and semi-supervised models according to claim 3, characterized in that: Step S3 includes: S301: For p(x) obtained in step three, if p(x) is greater than or equal to 0.5, its label is correct, and save the original label; If p(x) is less than 0.5, its label is incorrect, and do not save the label; S302: Feed the labeled images and unlabeled images together into the semi-supervised model for preparation for training; 5. The method for classifying label-noisy images based on graph consistency and semi-supervised model according to claim 1, wherein: Step S4 includes: S401: Perform weak augmentation operations on the labeled images, and perform data augmentations on the unlabeled images using weak augmentation techniques and AutoAugment techniques respectively. The weak augmentation operations include performing flipping, cropping, and shifting data augmentations; S402: Perform mixup operations on the labeled images and the unlabeled images simultaneously; Mixup can be expressed by the following formula: Among them, x i and x j represent the input images, and the two images are fused; y i and y j represent the labels of the input images, and the label format is one-hot encoding format. Their labels are fused; λ is a fusion ratio with a value range from 0 to 1.

6. The method for classifying label-noisy images based on graph consistency and semi-supervised model according to claim 1, wherein: Step S5 includes: S501: For the augmented labeled images, use the ResNet50 model as the labeled module of the semi-supervised model for training, and use a dual-network structure; S502. For the enhanced unlabeled images, the unlabeled module of the semi-supervised model uses the ResNet50 model for training and adopts a dual-network structure. Among them, after a group of weakly enhanced unlabeled images are trained using the ResNet50 model, a probability distribution result is generated. After the unlabeled images enhanced by the AutoAugment technique are trained using the ResNet50 model, a probability distribution representation, an embedding representation, and a feature map-level representation are generated.

7. The method for classifying label-noisy images based on graph consistency and semi-supervised model according to claim 1, characterized in that: The step S7 includes: S701. For labeled samples, calculate their class consistency using cross-entropy loss The calculation method of cross-entropy loss is as follows: Among them, y is the true label; is the one-hot encoding of the predicted label; N is the number of samples; M is the number of classes; S702. Calculate its class consistency using the L2 loss. The calculation method of the L2 loss is as follows: where p θ (y i ∣Aug2(u i )) represents the class distribution probability parameter θ after the model passes through softmax; S703. At the same time, calculate the graph consistency of the unlabeled samples. The calculation methods of the consistency between the probability distribution graph and the pseudo-label graph are as follows: The calculation method of the consistency between the embedding graph and the pseudo-label graph is as follows: The calculation method of the consistency between the feature map and the pseudo-label graph is as follows: Finally, add them up to obtain the graph consistency regularization term: S704. The final loss is the sum of the above several losses:

Images