SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo label balancing
By employing semi-supervised prototype representation learning and pseudo-label balancing methods, and utilizing high-dimensional feature mapping and prototype consistency regularization, the problems of insufficient label data and pseudo-label noise in SAR image road segmentation are solved, thereby improving segmentation accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2025-02-20
- Publication Date
- 2026-05-19
AI Technical Summary
Existing semi-supervised learning methods face problems such as insufficient label data, class imbalance, and excessive pseudo-label noise in SAR image road segmentation, resulting in poor model training performance.
We employ a semi-supervised prototype representation learning and pseudo-label balancing approach. By using high-dimensional feature mapping and prototype consistency regularization, we update the feature prototype with labeled data, dynamically adjust the pseudo-label frequency, and perform consistency regularization to improve segmentation performance.
It effectively improves the accuracy of road segmentation in SAR images, reduces the dependence on labeled data, solves the problems of class imbalance and pseudo-label noise, and improves the segmentation effect of the model.
Smart Images
Figure CN120032130B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, specifically relating to a SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing. Background Technology
[0002] With the rapid development of remote sensing technology, remote sensing images have experienced explosive growth. Due to the widespread applications of road extraction in real life, such as urban management, map updating, and traffic planning, road extraction from remote sensing images has become a popular research topic. Among these technologies, Synthetic Aperture Radar (SAR) is capable of imaging the Earth's surface around the clock, making it an important tool for acquiring road network information.
[0003] In recent years, with the rapid development of deep learning, deep learning methods have become dominant in road extraction from remote sensing images. Compared with traditional methods that extract road features manually, deep learning-based road extraction models offer better performance and stronger generalization capabilities. However, while remote sensing images are abundant, labeled images are scarce, primarily due to their high resolution, wide coverage, and complex land cover types. Labeling SAR images is more complex than labeling optical remote sensing images, typically requiring lengthy and time-consuming work by professionals. Therefore, deep learning-based road extraction methods often face the problem of insufficient training data.
[0004] To address the problem of lacking high-quality annotations in remote sensing images, semi-supervised learning utilizes a small amount of labeled data for training and a large amount of unlabeled data for auxiliary training. Due to the success of semi-supervised learning in image classification, many researchers have attempted to apply it to image segmentation tasks, which can be broadly categorized into two main strategies. One is based on consistency regularization. This method aims to enforce consistency between the original image, features, or model and the predictions made with perturbations, which can promote the model to learn robust feature representations and predictions free from noise and perturbations. The other is based on self-training. This type of method attempts to assign pseudo-labels to unlabeled data and use them for pseudo-supervised training of the model. Since pseudo-labels generally contain a lot of noise, a threshold is usually set to allow only high-confidence unlabeled data to be used for training, in order to reduce the impact of erroneous pseudo-labels on model performance.
[0005] While the aforementioned semi-supervised semantic segmentation methods have made significant progress in image interpretation, most have only been experimentally validated on natural scene images. Due to the substantial differences between SAR images and natural scene images, most existing methods cannot be directly transferred to SAR images. First, the pixel frequency of road categories in SAR images is much lower than that of non-road categories, exhibiting a severe class imbalance problem. In semi-supervised learning, this imbalance often leads to an unbalanced class distribution when generating pseudo-labels, thus affecting the model's training process. If these imbalanced pseudo-labels are used for supervised training, the model may produce segmentation results biased towards the majority class, exacerbating the impact of class imbalance on the final prediction. Second, road samples are relatively scarce and the land cover categories are complex in SAR images, making it difficult for the model to learn discriminative features during training, especially with insufficient effective representation of roads. This leads to the model easily generating pseudo-labels containing a large amount of noise, affecting the model's segmentation results. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies, this invention aims to provide a SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing. By utilizing a small number of labeled samples to provide discriminative information, a high-dimensional feature mapping is introduced. During training, the feature mean (i.e., the feature prototype) is progressively updated, and the predicted probability of samples is perturbed to obtain a fairer feature space distribution. Consistency regularization supervises the unlabeled samples, thereby achieving better segmentation performance. This method utilizes a small amount of labeled data and a large amount of unlabeled data, improving segmentation accuracy even with a limited number of labeled samples.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing includes the following steps;
[0009] Step 1: Divide the SAR road dataset captured by the GF-3 satellite into labeled data and unlabeled data, and input the labeled branch and unlabeled branch respectively;
[0010] Step 2: In the labeled branch, the labeled data is processed by the classification head to obtain feature predictions. High-dimensional feature maps are obtained through the mapping head. Predict based on features and real labels Calculate the cross-entropy loss and optimize the segmentation network during backpropagation; feature predictions will generate pseudo-labels in subsequent steps; based on high-dimensional feature mapping... and real labels Calculate the labeled feature prototype, based on the labeled feature prototype and the high-dimensional feature mapping. Calculate the prototype contrast loss;
[0011] Step 3: In the unlabeled branch, the unlabeled data is processed by the classification head to obtain feature predictions. The feature map is obtained through the mapping head. ;
[0012] Step 4: Generate pseudo-labels based on the segmentation results. The feature prediction in step 3 is dynamically adjusted based on the class frequency in the pseudo-labels. ;
[0013] Step 5: Predict the adjusted features With pseudo-tags Perform cross-entropy loss calculation;
[0014] Step 6: Based on the feature mapping With unlabeled feature prototype The results of the cross-entropy loss calculation are used to calculate the prototype loss of unlabeled data.
[0015] Step 7: Replace the obtained unlabeled prototype with a labeled prototype, use the labeled prototype to supervise the unlabeled segmentation results, calculate the cross-supervision loss, and ensure that the unlabeled prototype moves closer to the labeled prototype.
[0016] Step 8: Perform prototype consistency regularization on the feature map and the unlabeled prototype;
[0017] Step 9: Update the prototype using an exponential moving average update based on pixel percentage.
[0018] Step 10: Iteratively train the above steps until optimization is complete and the final model M is obtained. Use model M to complete the road segmentation task.
[0019] Step 1 specifically involves:
[0020] Given a finite number of labeled datasets and a set of unlabeled datasets ;
[0021] in Representing the Zhang has labeled input images, It is its corresponding pixel-level real label. Represents background, Represents roads;
[0022] Similarly, Indicates the first Zhang input image without label, and These represent the number of labeled images and the number of unlabeled images, respectively. In other words, the amount of unlabeled data is much greater than the amount of labeled data. It represents the real number field.
[0023] Step 2 specifically involves:
[0024] Labeled data is fed into a neural network, and feature predictions are obtained through a classification head. Then, and the real label Perform cross-entropy loss;
[0025]
[0026] in, This represents the output of the classification head after the i-th labeled image is input into the network, i.e., the category score of each pixel. This represents the cross-entropy loss function.
[0027] Labeled data is fed into a neural network, and a high-dimensional feature map is obtained through a mapping head. Then based on the real labels Calculate the high-dimensional feature center, or feature prototype, for each category based on the pixel. After obtaining the feature prototype, calculate the high-dimensional feature mapping. With feature prototype similarity between And use it as the predicted probability and the true label Apply cross-entropy loss:
[0028] ;
[0029]
[0030]
[0031] in, This represents the output of the mapping head after the i-th labeled image is input into the network, i.e., the high-dimensional feature map corresponding to each pixel. Let represent the actual label of the i-th labeled image.
[0032] Step 3 specifically involves:
[0033] The unlabeled data is then subjected to weak and strong augmentation to obtain the input. and Feature predictions are obtained through the classification head. and The feature maps are obtained through the mapping head. and .
[0034] Step 4 specifically involves:
[0035] Through feature prediction After calculating the pseudo-label mask, calculate the proportion of road pixels. Then, feature prediction is performed based on the ratio. Make the corresponding adjustments:
[0036]
[0037]
[0038] in This represents the output of the classification head after the j-th unlabeled image is input into the network, i.e., the class score of each pixel. This represents a Gaussian distribution.
[0039] Step 5 specifically involves:
[0040] Obtain new feature predictions Then combine with pseudo-label mask Perform cross-entropy loss calculation:
[0041]
[0042] in Represents the cross-entropy loss function. This indicates the number of unlabeled images.
[0043] Step 6 specifically involves:
[0044] Similar to step 2, a high-dimensional feature map is obtained. Then, based on the pseudo-label mask obtained in step 5... Based on the pixels, find the high-dimensional feature center, i.e., the feature prototype, for each category. After obtaining the prototype, find and similarity between And use it as the predicted probability and pseudo-label mask. Apply cross-entropy loss:
[0045]
[0046]
[0047]
[0048] in, This represents the output of the mapping head after the j-th unlabeled image is input into the network, i.e., the high-dimensional feature map corresponding to each pixel. This represents the pseudo-label mask of the j-th unlabeled image after weak enhancement. Represents the cross-entropy loss function;
[0049] Step 7 specifically involves:
[0050] Similar to the calculation method in step 6, the unlabeled prototype will be... Replace with a tagged prototype Then, the same prototype loss calculation is performed.
[0051] Step 8 specifically involves:
[0052] Following steps 3 and 6, we obtain the feature maps for weak and strong augmentation of the unlabeled data. and and unlabeled prototype ,Will and To each Similarity matrix is obtained by performing similarity calculation. and Then, calculate using KL loss. and Distribution similarity between them:
[0053]
[0054]
[0055]
[0056] in It is the Kullback-Leibler divergence loss function.
[0057] Step 9 specifically involves:
[0058] To ensure the prototypes obtained in steps 2 and 6 It can adapt to the updates of the segmentation network and uses an exponential moving average update method based on pixel proportion to update the trained prototype. First, it calculates the proportion r of the road and background categories in the labels of the features participating in the prototype calculation each time, and then updates according to the following formula:
[0059]
[0060] in This represents the proportion of pixels of class c in the total number of historical pixels. This represents the prototype calculated in the current round for class c. This represents the prototype of class c in historical training.
[0061] Step 10 specifically involves:
[0062] Based on iterative optimization of the above steps, the final model M is obtained. For the input image... Directly predicting features by passing through the labeled branch. And use argmax to generate a binary map, i.e., the road segmentation result. ;
[0063]
[0064] .
[0065] The beneficial effects of this invention are:
[0066] 1. This invention proposes a novel semi-supervised SAR road segmentation method that solves the problem of insufficient labeled samples. Specifically, it employs supervised training for labeled data and pseudo-label training and consistency regularization strategy for unlabeled data.
[0067] 2. This invention proposes a prototype representation learning and prototype-based consistency regularization method, which solves the problem of excessive noise in pseudo-labels. It uses the prototype as a classifier to directly supervise unlabeled data, thereby avoiding the problem of excessive noise when using pseudo-label supervision.
[0068] 3. This invention proposes a probability-based pseudo-label balancing method, which largely solves the problem of model prediction bias towards the background caused by the small proportion of road pixels. By utilizing the different class frequencies in the pseudo-labels, the prediction probability of the model for different classes is dynamically adjusted. The smaller the class frequency, the greater the prediction probability, thereby alleviating the problem of model prediction imbalance.
[0069] This invention effectively reduces the dependence on labeled data and solves the problems of complex background and low target pixel ratio in SAR images, thereby improving the accuracy of road segmentation in SAR images. Attached Figure Description
[0070] Figure 1 This is a flowchart illustrating the tagged branching process of this invention.
[0071] Figure 2 This is a flowchart illustrating the label-free branching process of this invention.
[0072] Figure 3 This is the GF-3 image used in the simulation of this invention.
[0073] Figure 4 This is the segmentation result obtained by the present invention and existing methods when dividing different datasets.
[0074] Figure 5 This is a schematic diagram of the framework of the present invention. Detailed Implementation
[0075] The present invention will now be described in further detail with reference to the accompanying drawings.
[0076] Reference Figure 1 , Figure 2 as well as Figure 5 This invention relates to a SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing. The specific implementation steps include:
[0077] Step 1: Divide the dataset into labeled samples and unlabeled samples according to certain partitioning rules;
[0078] Given a finite number of labeled datasets and a set of unlabeled datasets ;
[0079] in Representing the Zhang has labeled input images, It is its corresponding pixel-level real label. Represents background, Represents roads;
[0080] Similarly, Indicates the first Zhang input image without label, and These represent the number of labeled images and the number of unlabeled images, respectively. In other words, the amount of unlabeled data is much greater than the amount of labeled data. It represents the real number field.
[0081] Step 2: The labeled data is fed into the neural network, and feature predictions are obtained through the classification head and mapping head. and feature mapping The following will be utilized Perform cross-entropy loss calculation, using Calculate the prototype loss;
[0082] Step 3, predict based on features and real labels Calculate the cross-entropy loss, and then optimize the network through backpropagation;
[0083]
[0084] Step 4, based on feature mapping and real labels Calculate the high-dimensional feature center, or feature prototype, for each category based on the pixel. And update it, and then calculate prototype loss and prototype consistency loss based on the prototype of the desired feature;
[0085]
[0086]
[0087] Step 5: Calculate the prototype contrast loss based on the labeled prototype and feature map. First, calculate... and similarity between And use it as the predicted probability and the true label Apply cross-entropy loss:
[0088]
[0089]
[0090] Step 6: First, perform weak enhancement on the unlabeled data by randomly cropping and flipping the images. Then, perform strong enhancement by color dithering, random grayscale transformation, and Gaussian blur to obtain the enhanced data. and Subsequently, consistency regularization processing is performed based on the weak and strong enhancement data to make full use of the unlabeled data;
[0091] Step 7: Input the enhanced data into the network, and obtain feature predictions after passing through the classification head and mapping head. , and feature mapping , ;
[0092] Step 8: Generate pseudo-labels based on feature predictions from the weakly augmented data, and then update the feature predictions based on the predicted ratio r of road pixels to background pixels in the pseudo-labels. Obtain new feature predictions;
[0093]
[0094]
[0095] in This represents the output of the classification head after the j-th unlabeled image is input into the network, i.e., the class score of each pixel. This represents a Gaussian distribution.
[0096] Step 9, based on the updated weakly enhanced feature mapping and pseudo-tags Calculate and update the unlabeled prototype;
[0097]
[0098]
[0099] Step 10: Calculate the loss function based on the obtained unlabeled prototype, pseudo-label, strongly enhanced feature mapping, strongly enhanced feature prediction, and labeled prototype. The following calculation can optimize the network to make the features more representative and make full use of the unlabeled data.
[0100] 10a) Based on strong enhancement prediction and pseudo-tags Calculate the cross-entropy loss;
[0101]
[0102] 10b) Based on the unlabeled prototype and strongly enhanced feature mapping Calculate the prototype contrast loss;
[0103]
[0104]
[0105] 10c) Based on the labeled prototype and strongly enhanced feature mapping Calculate the prototype contrast loss;
[0106]
[0107]
[0108] 10d) Based on the unlabeled prototype and strongly enhanced feature mapping Weakly enhanced feature mapping Calculate the prototype consistency regularization loss; and To each Similarity matrix is obtained by performing similarity calculation. and Then, calculate using KL loss. and Distribution similarity between them:
[0109]
[0110]
[0111]
[0112] in It is the Kullback-Leibler divergence loss function.
[0113] Since weakly and strongly augmented data originate from the same source, the calculated distribution similarity should be similar. By performing KL loss calculation, consistency regularization of unlabeled data can be achieved, thereby improving its utilization.
[0114] The effects of this invention are further illustrated by the following simulation experiments:
[0115] Simulation conditions:
[0116] The simulation experiment uses SAR images taken by the Gaofen-3 satellite for experimental verification, such as... Figure 3 As shown, the image size is 512. 512, a total of 1992 images, with 192 randomly selected as the test set and the remaining 1800 images used as the training set.
[0117] The simulation experiments were conducted on a Linux system with an Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz and an NVIDIA GeForce RTX 3090 GPU. The experimental code was based on the PyTorch framework.
[0118] Table 1. Pixel percentage under different data divisions for two types in GF-3 SAR images
[0119]
[0120] Simulation content and analysis:
[0121] This invention, along with several existing methods, is used to classify GF-3 SAR images. These existing methods include UniMatch, FixMatch, U2PL, ST++, PS-MT, AdaptMatch, and RanPaste. The method proposed in this invention... Abbreviated representation.
[0122] Each segmentation method uses Deeplabv3plus and ResNet50 as the base segmentation networks. The number of training epochs for each model is set to 80, the batch size is set to 4, and the backbone network is optimized by stochastic gradient descent (SGD) with a learning rate of 0.001. The weak image enhancement method in this application uses random image cropping and flipping, while the strong image enhancement method uses color dithering, random grayscale transformation, and Gaussian blur.
[0123] A. Table 2 shows the performance of the proposed method and other state-of-the-art methods on the GF-3 SAR dataset. Combined with... Figure 4As shown, since the comparative methods were all tested on optical or optical remote sensing datasets, the conclusions in the table were obtained by following the settings in the paper and training according to the characteristics of the dataset. It can be seen that the method proposed in this application achieves consistent performance improvements compared to current state-of-the-art methods. In particular, the method proposed in this application outperforms the suboptimal algorithm by 2.72%, 2.45%, 1.34%, and 1.09% on each segmentation, demonstrating its effectiveness and better suitability for SAR images. Meanwhile, it can be seen that the ST++ method achieves better performance compared to UniMatch, proving that the complex background of SAR images and excessive feature perturbations can lead to poor model learning. Furthermore, ST++ achieves better performance by obtaining high-quality pseudo-labels. The method in this application effectively addresses these issues through class feature prototypes and prediction probability perturbations.
[0124] Table 2. Performance comparison of this invention with other methods on the GF-3 dataset.
[0125]
[0126] B. First, in Table 3... Ablation studies were conducted on different components to demonstrate their effectiveness. Using FixMatch as the backbone alone, the calculated IoU was 60.25%, 62.44%, 62.48%, and 65.14% for 1 / 16, 1 / 8, 1 / 4, and 1 / 2 partitions, respectively. By using Prototype Representation Learning (PRL) to supervise high-dimensional features, performance was improved by 1.61%, 0.25%, 1.59%, and 0.21% under different dataset partitioning strategies. Similarly, by perturbing the predicted probabilities using Probability-Based Pseudo-Label Balancing (PPB), performance was improved by 3.68%, 2.09%, 2.58%, and 1.17% under different dataset partitioning strategies. These results demonstrate the effectiveness of the proposed PRL and PPB. Furthermore, when PRL and PPB are used simultaneously, they further improve performance, proving that the proposed PRL and PPB are not mutually exclusive and can be effective simultaneously. Similarly, performance improvements were achieved in all experiments with the help of prototype-based consistent regularization (PRLC).
[0127] In summary, this invention segments SAR images for roads based on semi-supervised feature prototypes and probabilistic perturbations, making full use of both labeled and unlabeled information. This reduces reliance on labeled data and achieves higher performance, giving it certain advantages over existing methods.
Claims
1. A SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing, characterized in that, Includes the following steps; Step 1: Divide the SAR road dataset into labeled data and unlabeled data, and input the labeled branch and the unlabeled branch respectively; Step 2: In the labeled branch, the labeled data is processed by the classification head to obtain feature predictions. High-dimensional feature maps are obtained through the mapping head. Predict based on features and real labels Calculate the cross-entropy loss and optimize the segmentation network during backpropagation; feature predictions will generate pseudo-labels in subsequent steps; based on high-dimensional feature mapping... and real labels Calculate the labeled feature prototype, based on the labeled feature prototype and the high-dimensional feature mapping. Calculate the prototype contrast loss; Step 3: In the unlabeled branch, the unlabeled data is processed by the classification head to obtain feature predictions. The feature map is obtained through the mapping head. ; Step 4: Generate pseudo-labels based on the segmentation results. The feature prediction in step 3 is dynamically adjusted based on the class frequency in the pseudo-labels. ; Step 5: Predict the adjusted features With pseudo-tags Perform cross-entropy loss calculation; Step 6: Based on the feature mapping With unlabeled feature prototype The results of the cross-entropy loss calculation are used to calculate the prototype loss of unlabeled data. Step 7: Replace the obtained unlabeled prototype with a labeled prototype, use the labeled prototype to supervise the unlabeled segmentation results, calculate the cross-supervision loss, and ensure that the unlabeled prototype moves closer to the labeled prototype. Step 8: Perform prototype consistency regularization on the feature map and the unlabeled prototype; Step 9: Update the prototype using an exponential moving average update based on pixel percentage. Step 10: Iteratively train the above steps until optimization is complete and the final model M is obtained. Use model M to complete the road segmentation task.
2. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 1, characterized in that, Step 1 specifically involves: Given a finite number of labeled datasets and a set of unlabeled datasets ; in Representing the Zhang has labeled input images, It is its corresponding pixel-level real label. Represents background, Represents roads; Similarly, Indicates the first Zhang input image without label, and These represent the number of labeled images and the number of unlabeled images, respectively. In other words, the amount of unlabeled data is much greater than the amount of labeled data. It represents the real number field.
3. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 2, characterized in that, Step 2 specifically involves: Labeled data is fed into a neural network, and feature predictions are obtained through a classification head. Then, and the real label Perform cross-entropy loss; in, This represents the output of the classification head after the i-th labeled image is input into the network, i.e., the category score of each pixel. This represents the cross-entropy loss function.
4. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 3, characterized in that, In step 2: Labeled data is fed into a neural network, and a high-dimensional feature map is obtained through a mapping head. Then based on the real labels Calculate the high-dimensional feature center, or feature prototype, for each category based on the pixel. After obtaining the feature prototype, calculate the high-dimensional feature mapping. With feature prototype Similarity between And use it as the predicted probability and the true label Apply cross-entropy loss: ； in, This represents the output of the mapping head after the i-th labeled image is input into the network, i.e., the high-dimensional feature map corresponding to each pixel. Let represent the actual label of the i-th labeled image.
5. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 4, characterized in that, Step 3 specifically involves: The unlabeled data is then subjected to weak and strong augmentation to obtain the input. and Feature predictions are obtained through the classification head. and The feature maps are obtained through the mapping head. and ; Step 4 specifically involves: Through feature prediction After calculating the pseudo-label mask, calculate the proportion of road pixels. Then, feature prediction is performed based on the ratio. Make the corresponding adjustments: in This represents the output of the classification head after the j-th unlabeled image is input into the network, i.e., the class score of each pixel. This represents a Gaussian distribution.
6. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 5, characterized in that, Step 5 specifically involves: Obtain new feature predictions Then combine with pseudo-label mask Perform cross-entropy loss calculation: in Represents the cross-entropy loss function. This indicates the number of unlabeled images.
7. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 6, characterized in that, Step 6 specifically involves: Obtain high-dimensional feature mapping Then, based on the pseudo-label mask obtained in step 5... Calculate the high-dimensional feature center, or feature prototype, for each category based on the pixel. After obtaining the prototype, find and Similarity between And use it as the predicted probability and pseudo-label mask. Apply cross-entropy loss: in, This represents the output of the mapping head after the j-th unlabeled image is input into the network, i.e., the high-dimensional feature map corresponding to each pixel. This represents the pseudo-label mask of the j-th unlabeled image after weak enhancement. This represents the cross-entropy loss function.
8. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 7, characterized in that, Step 7 specifically involves: Unlabeled prototype Replace with a tagged prototype Then, the same prototype loss calculation is performed. Step 8 specifically involves: Obtain feature maps for weak and strong augmentation of unlabeled data. and and unlabeled prototype ,Will and To each Similarity matrix is obtained by performing similarity calculation. and Then, calculate using KL loss. and Distribution similarity between them: in It is the Kullback-Leibler divergence loss function.
9. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 8, characterized in that, Step 9 specifically involves: The trained prototype is updated using an exponential moving average method based on pixel proportion. First, the proportion r of the road and background categories in the labels of the features participating in the prototype calculation is calculated, and then updated according to the following formula: in This represents the proportion of pixels of class c in the total number of historical pixels. This represents the prototype calculated in the current round for class c. This represents the prototype of class c in historical training.
10. The SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo-label balancing according to claim 9, characterized in that, Step 10 specifically involves: Based on iterative optimization of the above steps, the final model M is obtained for the input image. Directly predicting features by passing through the labeled branch. And use argmax to generate a binary map, i.e., the road segmentation result. ; 。