Semi-supervised efficient video SAR shadow tracking method based on small samples

Abstract

The invention discloses a semi-supervised efficient video SAR shadow tracking method based on small samples, and belongs to the field of radar target recognition. The method provided by the invention solves the limitation of a traditional shadow tracking technology, and improves the detection precision and robustness of the shadow region through an improved segmentation model and a semi-supervised strategy; on the basis of the designed new mixed training strategy, labeled data and unlabeled data are effectively integrated, and the learning efficiency and generalization ability of the model are remarkably improved. The deep learning technology and the semi-supervised learning strategy are combined, limited annotation data and rich non-annotation data are fully utilized, the defects of a traditional method under the small sample condition are overcome, and the detection and tracking precision of the shadow area in the SAR image is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of radar target recognition, and in particular to a small sample-based semi-supervised efficient video SAR shadow tracking method. Background Art

[0002] With the rapid development of synthetic aperture radar (SAR) technology, SAR images are increasingly used in areas such as surface monitoring, disaster assessment, and military reconnaissance. Especially in complex environments, accurate tracking of shadow areas is crucial for target recognition and object classification. However, in SAR images, shadow areas usually cause signal attenuation due to imaging geometry and reflection characteristics, appearing as low-brightness areas, which poses a challenge to subsequent target detection and classification tasks. Therefore, accurately identifying and tracking shadow areas has become the key to improving the performance of SAR image processing.

[0003] Currently, most shadow tracking algorithms rely on a large amount of labeled data for training, which becomes particularly difficult in small sample scenes. Therefore, how to effectively use limited labeled samples and a large number of unlabeled samples to improve the accuracy of shadow tracking has become an urgent problem to be solved.

[0004] In recent years, semi-supervised learning technology has gradually attracted the attention of researchers due to its advantage of combining labeled and unlabeled data. At the same time, image processing technology based on semi-supervised learning has made significant progress in many fields. However, shadow tracking methods for video SAR images are still relatively scarce, especially in application scenarios with small sample conditions. Summary of the invention

[0005] In order to solve the problem that existing traditional shadow tracking algorithms rely on a large amount of labeled data for training, this application provides a semi-supervised and efficient video SAR shadow tracking method based on small samples. By optimizing the model structure and introducing advanced semi-supervised learning technology, the tracking accuracy and robustness of the shadow area are improved to meet the actual application needs in various complex environments.

[0006] The technical solution adopted in this application is:

[0007] A small sample-based semi-supervised efficient video SAR shadow tracking method includes the following steps:

[0008] Step 1: Obtain a SAR image and perform image preprocessing on it (to adapt to the set semantic segmentation model, such as the DeepLabV3+ model), and obtain an original data set based on a number of SAR images after image preprocessing; wherein the image preprocessing includes: image binarization processing, and image size normalization processing of the binarized image;

[0009] In this step, the binarization process is as follows: the background pixels are set to 0, the target pixels are set to 1, and the uniform image size can be set to 512×512.

[0010] Step 2, divide the original data set into a labeled original data set and an unlabeled original data set;

[0011] Step 3, dividing the labeled original data set into a first training set and a first validation set, and then performing supervised training on the set semantic segmentation model based on the first training set to optimize the network parameters of the model (such as weights and bias terms), and tuning the hyperparameters of the semantic segmentation model based on the first validation set; and the loss function during supervised training adopts consistency loss, such as cross entropy loss;

[0012] The semantic segmentation model is used to identify multiple targets and perform segmentation, and its output layer is used to perform N+1 classification and segmentation of the input data, which includes N required target categories and one unknown category;

[0013] Step 4: Perform semi-supervised training on the supervised semantic segmentation model based on the unlabeled original dataset;

[0014] Performing image weak enhancement processing on each unlabeled original data sample of the unlabeled original data set to obtain at least one weakly enhanced sample of each unlabeled original data sample;

[0015] Input the weakly enhanced sample into the supervised trained semantic segmentation model, and obtain the pseudo label of the current weakly enhanced sample based on the predicted category output by it;

[0016] Performing image strong enhancement processing on each unlabeled original data sample of the unlabeled original data set to obtain at least one strongly enhanced sample of each unlabeled original data sample;

[0017] For each unlabeled original data sample j, a pair of weak and strong sample pairs corresponding to j is obtained based on a weakly enhanced sample and a strongly enhanced sample;

[0018] A mixed set is formed based on the labeled original data set and all weak and strong sample pairs, and then the mixed set is used as the second training set to perform secondary training on the supervised semantic segmentation model;

[0019] The total loss function used in the secondary training is: the weighted sum of the first consistency loss corresponding to the labeled original data set and the second consistency loss based on all weak-strong sample pairs; for the second consistency loss, the strong enhanced sample in the weak-strong sample pair is used as the model input, and the corresponding pseudo label of the weak enhanced sample is regarded as the true label;

[0020] Step 5: extract a fine-tuning dataset from the mixed set to fine-tune the semantic segmentation model after the second training. The loss function during fine-tuning adopts consistency loss. In the fine-tuning dataset, each weak and strong sample pair is regarded as a model input sample, and the corresponding pseudo label is regarded as the true label of each sample.

[0021] Based on the fine-tuned semantic segmentation model, a segmentation model for SAR shadow tracking is obtained, and the shadow area prediction of the SAR image to be tracked is realized based on the segmentation model. That is, the SAR image to be tracked needs to be preprocessed to adapt to the input of the semantic segmentation model.

[0022] Furthermore, step 4 also includes:

[0023] The pseudo-labels whose prediction confidence (output of semantic segmentation model) is greater than the preset threshold are regarded as reliable pseudo-labels;

[0024] For each original data sample with a reliable pseudo-label, a weak-strong sample pair of the current original data sample j is formed based on a strong enhancement sample and a weak enhancement sample corresponding to a reliable pseudo-label.

[0025] Furthermore, the weak-strong sample pairs in the fine-tuning dataset extracted in step 5 are weak-strong sample pairs with reliable pseudo labels.

[0026] Furthermore, in step 2, the ratio of labeled data to unlabeled data is 2:8 to form a small sample data set.

[0027] Furthermore, in step 3, the first training set and the first validation set account for 80% and 10% respectively.

[0028] Furthermore, the semantic segmentation model uses the DeepLabV3+ model, whose backbone network is MobileNet, and the optimizer selects SGD (stochastic gradient descent). Based on the set initial learning rate and minimum learning rate lr min , maximum learning rate lr max The learning rate reduction method used for the total number of training rounds T is cosine annealing.

[0029] Furthermore, in step 4, the weak image enhancement processing includes: randomly flipping the data (eg, with a probability of 50%), randomly cropping, and randomly rotating the data (eg, 10°).

[0030] Furthermore, in step 4, the image enhancement process includes: randomly blocking the image, generating a new image by linearly combining two images in a specified ratio, and adding noise.

[0031] Furthermore, in step 4, the second consistency loss is a classification cross entropy loss in which the strongly enhanced image is used as input and the corresponding pseudo-label / reliable pseudo-label is regarded as the true label.

[0032] Furthermore, in step 4, the total loss function is: Total Loss = Ls + λ*Lu, wherein Ls and Lu are the first and second consistency losses, respectively, and λ is a preset weight coefficient, indicating the contribution of the unlabeled loss to the total loss. Preferably, it can be set to λ = 0.67.

[0033] Furthermore, step 5 also includes calculating the intersection-over-union ratio between the shadow area predicted by the fine-tuned semantic segmentation model and the real shadow area, and when it is greater than 0.5, it is considered that the SAR shadow tracking is successful. Then, the tracking rate can be calculated to evaluate whether the model has achieved the purpose.

[0034] The technical solution provided by this application brings at least the following beneficial effects:

[0035] This application can improve the accuracy of shadow tracking while reducing labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0037] Figure 1 A flowchart of a small sample-based semi-supervised efficient video SAR shadow tracking method provided in an embodiment of the present application;

[0038] Figure 2 Schematic diagram of SAR shadow tracking in an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions of the embodiments of the present application will be described in detail and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the embodiments described with reference to the drawings are exemplary and are intended to be used to explain the present application, and cannot be understood as limiting the present application.

[0040] The embodiment of the present application provides a semi-supervised efficient video SAR shadow tracking method based on small samples, which combines deep learning and semi-supervised learning in a new way under small sample conditions to solve the limitations of traditional shadow tracking technology. Through improved segmentation models (such as DeepLabV3+ models) and semi-supervised strategies (such as FixMatch strategies), the detection accuracy and robustness of shadow areas are improved, and based on the designed new hybrid training strategy, labeled and unlabeled data are effectively integrated, significantly improving the learning efficiency and generalization ability of the model. By combining deep learning technology and semi-supervised learning strategies, limited labeled data and abundant unlabeled data are fully utilized, the shortcomings of traditional methods under small sample conditions are solved, and the detection and tracking accuracy of shadow areas in SAR images is effectively improved.

[0041] In one embodiment, see Figure 1 The present application provides a small sample-based semi-supervised efficient video SAR shadow tracking method comprising the following steps:

[0042] S1. Obtain a SAR remote sensing image dataset and perform image preprocessing on it to match the input of the segmentation model; wherein the segmentation model is used to indirectly achieve simultaneous recognition and segmentation of multiple targets by distinguishing different categories, and classify the targets, that is, the segmentation model is used to recognize and segment multiple targets.

[0043] In this embodiment, the image preprocessing of each SAR image in the acquired data set includes: setting the background pixel points to 0, setting the target pixel points to 1, and cutting all data set images (composed of several acquired SAR remote sensing images) into pixel blocks of 512×512 pixels to ensure that they can adapt to network training. At the same time, the data is enhanced, randomly cropped, rotated by 15° and horizontally flipped to increase the diversity of samples, thereby enhancing the generalization ability of the model. In this embodiment, the segmentation model uses an improved DeepLabV3+ model.

[0044] S2, dividing the original data set (the data set composed of each pixel block after segmentation) into labeled data and unlabeled data;

[0045] In this embodiment, the original data set is divided into labeled data and unlabeled data, where the labeled data is divided into training set, validation set and test set in a ratio of 80%:10%:10%, so as to ensure the representativeness of each subset. The ratio of labeled data to unlabeled data is 2:8, and a small sample data set is constructed.

[0046] The training set is the data sample for model fitting; the validation set is a sample set set aside during model training to adjust the model's hyperparameters and to make a preliminary assessment of the model's capabilities. It is usually used to verify the current model's generalization ability (accuracy, recall, etc.) during model iterative training to decide whether to stop training. The test set is used to evaluate the generalization ability of the final model (trained). However, it cannot be used as a basis for algorithm-related choices such as parameter adjustment and feature selection.

[0047] S3, the training phase is divided into supervised training and semi-supervised training. The DeepLabV3+ model is trained using a labeled data set, and relevant parameters are set to obtain the consistency loss Ls of the data set.

[0048] In this embodiment, the DeepLabV3+ model is trained with labeled data in the full supervision stage, and the relevant parameters of the model are configured. The backbone network of the model selects MobileNet, the number of classes to be distinguished is 6, the downsampling multiple is set to 8, the optimizer selects SGD (stochastic gradient descent), the initial learning rate is set to 0.02, and the weight decay is set to 0.0005 to prevent overfitting and increase the generalization ability of the model. The total number of training rounds T is set to 500, the minimum learning rate is 0.002, the maximum learning rate is 0.02, and the learning rate reduction method adopts cosine annealing (Cosine Annealing):

[0049]

[0050] Among them, t is the number of rounds of current training, T is the total number of rounds of training, lr min Represents the minimum learning rate, lr max Represents the maximum learning rate.

[0051] The cross entropy loss Ls of the test set and validation set is obtained by training, and its calculation formula is:

[0052]

[0053] Among them, y i is the true label, is the model prediction output, N is the total number of samples, and i is the sample number.

[0054] The best model weights are retained and the model is validated using the test set to ensure the model's generalization ability on unseen data. That is, the segmentation model is fully supervised based on the training set and the validation set, and the loss function in the fully supervised training process is the cross loss.

[0055] Then, the unlabeled data is weakly enhanced (each original image has at least one weakly enhanced image), and this set of data is trained with the trained model to generate pseudo labels. At the same time, a threshold is set. When the model's prediction confidence for the data is greater than the threshold, the set of pseudo labels is trusted, that is, the pseudo labels greater than the threshold are reliable pseudo labels; then the same set of unlabeled data is strongly enhanced (at least one strong enhancement result), that is, the original image with reliable pseudo labels is strongly enhanced, and the enhanced unlabeled data is trained with the network to obtain consistency loss; for each original image with a reliable pseudo label, a weak-strong sample pair of the current original image can be formed based on a strongly enhanced image (if there are multiple images, one is randomly selected) and a weakly enhanced image corresponding to a reliable pseudo label.

[0056] In this embodiment, the weakly enhanced image processing is specifically set as follows: randomly flipping the data (with a probability of 50%), randomly cropping and randomly rotating (10°), and then using the trained DeepLabV3+ model to predict these weakly enhanced images to obtain pseudo labels. At the same time, a confidence threshold of 0.95 is set to screen the reliability of the pseudo labels.

[0057] The confidence calculation formula is:

[0058]

[0059] in, The class probabilities predicted by the model.

[0060] Only when the current model's prediction confidence for these unlabeled samples is greater than this threshold, the pseudo-label of the data is considered reliable.

[0061] The same set of unlabeled data is strongly enhanced by RandAugment, with the number of transformations set to 2, the intensity parameter set to 7, 25% of the image randomly blocked, and two images are linearly combined at a ratio of 0.5 to generate a new image. In addition, random noise is added to the image, and the noise standard deviation is set to 0.13. The trained DeepLabV3+ model is also used to predict the images after strong enhancement to obtain the prediction results of this set of data.

[0062] S4. Calculate the consistency loss L between the prediction result of the strongly enhanced image and the corresponding weakly enhanced image pseudo label u ;

[0063] At this time, the strong enhanced sample in each weak-strong sample pair can be used as the input of the semantic segmentation model, and the reliable pseudo-label of the corresponding weakly enhanced sample can be regarded as the true label of the strong enhanced sample. The corresponding consistency loss can be calculated based on the classification prediction output of the model output, which is denoted as L u , classification such as cross entropy loss.

[0064] In this embodiment, the calculation formula of the cross entropy loss Lu between the two is:

[0065]

[0066] Among them, y j is the pseudo label of the weakly enhanced image, is the predicted output of the strongly enhanced image, M is the total number of samples (corresponding unlabeled data), j represents the number of the unlabeled data (specifically, it may refer to the number of the unlabeled data with reliable pseudo-labels), and weak enhancement and strong enhancement may be pre-set.

[0067] S5. Combine the unlabeled data and the labeled data to form a new mixed data set, and then use the loss Ls of the labeled data obtained in S3 and the consistency loss Lu of the unlabeled data to form a total loss function to update the DeepLabV3+ model.

[0068] A mixed set is formed based on the labeled original data set and all weak and strong sample pairs, and then the mixed set is used as the second training set to perform secondary training on the supervised semantic segmentation model;

[0069] The total loss function used in the secondary training is: the weighted sum of the first consistency loss corresponding to the labeled original data set and the second consistency loss based on all weak-strong sample pairs; for the second consistency loss, the strong enhanced sample in the weak-strong sample pair is used as the model input, and the reliable pseudo-label of the corresponding weak enhanced sample is regarded as the true label;

[0070] In this step, you can also directly take out a strong enhancement and a weak enhancement data as a pair of data for each unlabeled data, and use the strong enhancement in the pair of data as the model input, and regard the pseudo label corresponding to the weak enhancement data as the true label to perform the calculation of the consistency loss Lu of all unlabeled data.

[0071] In this embodiment, the unlabeled data and the labeled data are combined to form a new mixed data set, and then the loss Ls of the labeled data obtained in S32 and the consistency loss Lu of the unlabeled data obtained in S35 are combined to form a total loss function, which is calculated as follows:

[0072] Total Loss=Ls+λ*Lu

[0073] Among them, λ is the weight coefficient, which represents the contribution of the unlabeled loss to the total loss. Preferably, λ=0.67 can be set.

[0074] S6. Train the updated model on the mixed set and verify it. Fine-tune the model based on the verification results to get a converged result. Finally, predict the adjusted model on an independent labeled data set and compare it with the real label to get the tracking effect.

[0075] At this time, for unlabeled data, only those unlabeled data whose confidence level corresponding to the pseudo-label is not less than 0.95 can be screened out from the mixed set, and then these pseudo-labels are regarded as the true labels of the samples. These unlabeled samples are respectively a weakly enhanced sample and a strongly enhanced sample. The fine-tuning dataset of the semantic segmentation model trained by S5 is formed based on these weakly enhanced samples, strongly enhanced samples and labeled datasets. The semantic segmentation model trained by S5 is fine-tuned again based on the true label and the regarded true label, as well as the consistency loss. When the training converges, it stops. Thus, a segmentation model for SAR shadow tracking is obtained based on the fine-tuned semantic segmentation model, and the shadow area of ​​the target object is predicted based on the segmentation model.

[0076] That is, in this application, the fine-tuned model is finally used to predict the independent labeled data set and compare it with the real label to obtain the tracking effect.

[0077] In this embodiment, the updated model is used to train the mixed set, the number of training rounds is set to 500, and verification is performed every 50 rounds. The batch size is set to 32, and the model is fine-tuned according to the verification results. Finally, the independent labeled data set is predicted and then compared with the real label to obtain the tracking effect, such as Figure 2 As shown. Early Stopping can be used during training to prevent overfitting. The IoU (intersection over union) value is used to determine whether the target is tracked successfully. If IoU is greater than 0.5, it is considered to be tracked successfully. The calculation formula is: Where A is the predicted shadow area and B is the actual shadow area.

[0078] The tracking rate is then calculated to evaluate whether the model has achieved its purpose.

[0079] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0080] In addition, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first", "second", etc. may explicitly or implicitly include at least one of the features.

[0081] Any process or method description described in a flowchart or otherwise in this specification may be understood to represent a module, fragment or portion of code that includes one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0082] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0083] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

[0085] The above are only some embodiments of the present application. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A small sample-based semi-supervised efficient video SAR shadow tracking method, characterized in that: The following steps are involved: Step 1, acquiring a SAR image and performing image preprocessing on it, and obtaining an original data set based on a plurality of SAR images after image preprocessing; wherein the image preprocessing includes: image binarization processing, and image size normalization processing of the binarized image; Step 2, divide the original data set into a labeled original data set and an unlabeled original data set; Step 3, dividing the labeled original data set into a first training set and a first validation set, and then performing supervised training on the set semantic segmentation model based on the first training set to optimize the network parameters of the model, and tuning the hyperparameters of the semantic segmentation model based on the first validation set; and the loss function during supervised training adopts consistency loss; The semantic segmentation model is used to identify multiple targets and perform segmentation, and its output layer is used to perform N+1 classification and segmentation of the input data, which includes N required target categories and one unknown category; Step 4: Perform semi-supervised training on the supervised semantic segmentation model based on the unlabeled original dataset; Performing image weak enhancement processing on each unlabeled original data sample of the unlabeled original data set to obtain at least one weakly enhanced sample of each unlabeled original data sample; Input the weakly enhanced sample into the supervised trained semantic segmentation model, and obtain the pseudo label of the current weakly enhanced sample based on the predicted category output by it; Performing image strong enhancement processing on each unlabeled original data sample of the unlabeled original data set to obtain at least one strongly enhanced sample of each unlabeled original data sample; For each unlabeled original data sample j, a pair of weak and strong sample pairs corresponding to j is obtained based on a weakly enhanced sample and a strongly enhanced sample; A mixed set is formed based on the labeled original data set and all weak and strong sample pairs, and then the mixed set is used as the second training set to perform secondary training on the supervised semantic segmentation model; The total loss function used in the secondary training is: the weighted sum of the first consistency loss corresponding to the labeled original data set and the second consistency loss based on all weak-strong sample pairs; for the second consistency loss, the strong enhanced sample in the weak-strong sample pair is used as the model input, and the corresponding pseudo label of the weak enhanced sample is regarded as the true label; Step 5: extract a fine-tuning dataset from the mixed set to fine-tune the semantic segmentation model after the second training. The loss function during fine-tuning adopts consistency loss. In the fine-tuning dataset, each weak and strong sample pair is regarded as a model input sample, and the corresponding pseudo label is regarded as the true label of each sample. Based on the fine-tuned semantic segmentation model, a segmentation model for SAR shadow tracking is obtained, and the shadow area prediction of the SAR image to be tracked is achieved based on the segmentation model.

2. The method according to claim 1, characterized in that In step 1, the binarization process is as follows: the background pixels are set to 0, and the target pixels are set to 1.

3. The method according to claim 1, characterized in that In step 2, the ratio of labeled data to unlabeled data is 2:

8.

4. The method according to claim 1, characterized in that In step 3, the first training set and the first validation set account for 80% and 10% respectively.

5. The method according to claim 1, characterized in that The semantic segmentation model uses the DeepLabV3+ model, whose backbone network is MobileNet, and the optimizer selects SGD. Based on the set initial learning rate and minimum learning rate lr min , maximum learning rate lr max The learning rate reduction method used for the total number of training rounds T is cosine annealing.

6. The method according to claim 1, characterized in that In step 4, the weak image enhancement processing includes: randomly flipping the data, randomly cropping and randomly rotating the data; the strong image enhancement processing includes: randomly blocking the image, generating a new image by linearly combining two images in a specified ratio, and adding noise.

7. The method according to claim 1, characterized in that In step 4, the second consistency loss is a classification cross entropy loss in which the strongly enhanced image is used as input and the corresponding pseudo-label / reliable pseudo-label is regarded as the true label.

8. The method according to claim 1, characterized in that In step 4, the total loss function is: Total Loss = Ls + λ*Lu, where Ls and Lu are the first and second consistency losses respectively, and λ is a preset weight coefficient, indicating the contribution of the unlabeled loss to the total loss. Preferably, it can be set to λ = 0.

67.

9. The method according to claim 1, characterized in that Step 4 also includes: The pseudo-labels whose prediction confidence is greater than the preset threshold are regarded as reliable pseudo-labels; For each original data sample with a reliable pseudo-label, a weak-strong sample pair of the current original data sample j is formed based on a strong enhancement sample and a weak enhancement sample corresponding to a reliable pseudo-label.

10. The method according to claim 1, characterized in that The weak-strong sample pairs in the fine-tuning dataset extracted in step 5 are weak-strong sample pairs with reliable pseudo labels, wherein a reliable pseudo label refers to a pseudo label whose prediction confidence of a weakly enhanced sample is greater than a preset threshold.

Images