Semi-supervised SAR target identification method and system based on contour enhancement and cooperative training

Abstract

The invention relates to a semi-supervised SAR target identification method and system based on contour enhancement and cooperative training. Specifically, according to the method, a to-be-recognized SAR image is input into a trained target recognition model to generate a corresponding target prediction result. The training process is divided into two stages: in the first stage, contour extraction and enhancement processing are carried out on an SAR image sample without a label, and then semi-supervised learning is carried out by using a FixMatch framework; and in the second stage, the weak augmented images of the label samples in the first stage are input into the cooperative training network for training, judgment results of different rounds are summarized, and the reliability of each judgment result is judged through an entropy value. The common judgment results of the cooperative training network form a sampling pool, a certain number of high-confidence samples are extracted from the sampling pool to serve as pseudo-label samples to expand a label training set of the first stage, and then the expanded label training set is used for continuing to train FixMatch. According to the method, the accuracy, the reliability and the robustness of SAR target recognition are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target detection, and in particular to a semi-supervised SAR target recognition method and system based on contour enhancement and collaborative training. Background Art

[0002] Synthetic Aperture Radar (SAR) is an imaging radar system that is widely deployed on mobile platforms such as drones, aircraft and satellites. By transmitting microwave pulses to the observed object and receiving the echo signal, SAR achieves accurate imaging observation of the ground and sea surface. Compared with optical imaging systems, SAR has the ability to acquire data all the time and in all weather conditions, high resolution and long-distance operation, and is therefore widely used in battlefield reconnaissance, topographic mapping, geological exploration, marine surveillance and other fields. Automatic Target Recognition (ATR) of SAR images is one of the key technologies for the application of SAR data, which significantly improves the efficiency and accuracy of surveillance and reconnaissance missions.

[0003] In recent years, artificial intelligence technologies represented by deep learning have enabled SAR target recognition and greatly improved target recognition performance. However, deep learning models require a large number of labeled training samples, and labeling samples in practical applications is time-consuming and costly.

[0004] Semi-supervised learning (SSL) is a machine learning method that effectively utilizes unlabeled samples. In semi-supervised learning, data augmentation and pseudo-label generation are two important methods. Through data augmentation, the diversity of training samples can be improved, thereby improving the generalization and robustness of deep learning models. By generating pseudo-labels for unlabeled samples, the labeled training samples can be effectively expanded, thereby improving the recognition accuracy of deep learning models. However, the existing SAR target recognition methods based on semi-supervised learning still have the following problems: 1) The main area of ​​interest in SAR image data is small and there are many background noise points, which may cause the model to spend a lot of energy on learning meaningless features, thereby making the accuracy of SAR target recognition poor. 2) In semi-supervised deep learning technology, when there are few labeled samples, the existing pseudo-label generation method is difficult to obtain highly reliable pseudo-labels, resulting in the model being unable to effectively learn reliable features, thereby limiting the recognition performance and generalization ability of the SAR target recognition model. Therefore, how to improve the accuracy, reliability and robustness of SAR target recognition is a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a semi-supervised SAR target recognition method based on contour enhancement and collaborative training, which effectively processes the region of interest and highlights the key features by performing contour extraction and enhancement processing on the SAR image, and at the same time enables the model to learn more extensive and more reliable SAR target features through collaborative training, thereby improving the accuracy, reliability and robustness of SAR target recognition.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] Semi-supervised SAR target recognition method based on contour enhancement and collaborative training, including:

[0008] S1: Acquire the SAR image to be identified;

[0009] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result;

[0010] The processing steps when training the object recognition model are as follows:

[0011] S201: Obtain labeled SAR image samples and unlabeled SAR image samples for model training;

[0012] S202: performing contour extraction and enhancement processing on the unlabeled SAR image sample to obtain an unlabeled contour enhanced image;

[0013] S203: performing strong and weak augmentation on the unlabeled contour enhanced image respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; performing weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples;

[0014] S204: using the unlabeled strong augmented samples, the unlabeled weak augmented samples, and the labeled weak augmented samples, and using the FixMatch framework to perform semi-supervised learning, to obtain a FixMatch-based object recognition model;

[0015] S205: Input the labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after a specified round, and take the category with the largest probability value as the prediction label of the judgment in this round for each judgment; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample;

[0016] S206: Preliminarily screening the judgment result distribution of all samples based on each sub-model through the characteristics of entropy; determining pseudo-label samples with high confidence through collaborative judgment of multiple sub-models; and randomly extracting a certain number of pseudo-label samples for each type;

[0017] S207: supervise the object recognition model based on FixMatch through the extracted pseudo-label samples, that is, expand the label training set of FixMatch to continue training, and obtain a trained object recognition model;

[0018] S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

[0019] Preferably, in step S202, contour extraction and enhancement processing are implemented through the following steps:

[0020] S2011: Perform contour extraction on the original SAR image to extract the target area of ​​interest;

[0021] S2012: generating a corresponding mask image based on the target region of interest;

[0022] The formula of the mask image is expressed as:

[0023]

[0024] Where: X mask represents the mask image; p represents the pixel value of the corresponding position of the original SAR image; contour represents the target area of ​​interest obtained by contour extraction;

[0025] S2013: superimposing the mask image with the original SAR image to obtain a contour enhanced image;

[0026] The formula for contour enhancement image is expressed as:

[0027] X new1 =(X mean1 X mask +X1)*0.6;

[0028] Where: X new1 represents the contour enhanced image; X1 represents the original SAR image; X mean1 Represents the image after smoothing and averaging the original SAR image X1.

[0029] Preferably, in step S203, weak augmentation of the SAR image is achieved by flipping and translation; strong augmentation of the SAR image is achieved by RandAugment and CTAugment strategies.

[0030] Preferably, in step S204, during the semi-supervised learning process, the loss function of the target recognition model includes a supervised loss L x and the unsupervised loss L u ;

[0031] The formula is as follows:

[0032]

[0033] Where: N x represents the number of labeled weakly augmented samples; H represents the cross entropy between two probability distributions; represents the target prediction result output by the target recognition model based on the input SAR image x under the image enhancement mode θ; Represents all categories and their corresponding predicted probabilities; N u Represents the number of unlabeled samples; Represents the target recognition model for the unlabeled strong augmentation sample U s The predicted probability of each category; y represents the category label corresponding to the label image; Represents the category label with the highest probability in the target prediction results of the unlabeled weakly augmented samples.

[0034] Preferably, step S205 specifically includes the following steps:

[0035] S2051: Construct a collaborative training network consisting of four sub-models; the four sub-models have different depths and widths;

[0036] S2052: Input the labeled weakly augmented samples into each sub-model for training, and obtain the target prediction result output by each sub-model for each sample;

[0037] S2053: After each preset round, the target prediction results of all samples are judged: for each sample, the category with the largest probability value among all target prediction results is selected as the prediction label of the judgment round;

[0038] The formula is:

[0039] label_X 1_n = argmax(y 1_n );

[0040] Where: label_X 1_n represents the category with the maximum probability in the judgment result of the first sub-model in the nth round, that is, the predicted label of the sample in the nth round; y 1_n Represents the probability distribution of the categories in all predicted values ​​of the samples of the first sub-model in round n;

[0041] The first sub-model's multiple rounds of judgment results for the same sample are expressed as follows:

[0042] label_X1={label_X 1_1 ,label_X 1_2 ,…,label_X 1_n}.

[0043] Preferably, step S206 specifically includes the following steps:

[0044] S2061: Calculate the entropy value of the sample based on the judgment results of multiple rounds of the same sample by each sub-model;

[0045] The formula is:

[0046]

[0047] Where: H(y) represents the entropy value of the sample; y i represents the multiple rounds of judgment results of the sub-model on the same sample; c is the number of categories in the judgment results;

[0048] S2062: Delete samples whose entropy values ​​are greater than a preset entropy value;

[0049] S2063: retain samples with consistent judgment results from each sub-model to build a sampling pool, and use the categories in the judgment results as pseudo labels for the samples;

[0050] The formula is:

[0051]

[0052] Where: label_X represents the pseudo label of the sample after collaborative screening by the four sub-models;

[0053] S2064: For each category, a certain number of samples are randomly selected from the sampling pool as pseudo-label samples.

[0054] Preferably, in step S207, when the accuracy of the target recognition model using FixMatch for semi-supervised learning reaches a preset value and cannot be further improved, the target recognition model based on FixMatch is further trained by expanding the pseudo-label samples to the label training set, i.e., collaborative training.

[0055] Preferably, in step S207, during the collaborative training process, the loss function of the target recognition model includes the collaborative supervised loss L x2 ;

[0056] The formula is:

[0057]

[0058] Where: X sa represents a pseudo-label sample; N sa Represents the number of pseudo-label samples.

[0059] The present invention discloses a semi-supervised SAR target recognition system based on contour enhancement and collaborative training, which is implemented based on the semi-supervised SAR target recognition method based on contour enhancement and collaborative training of the present invention, including:

[0060] An image input module is used to obtain a SAR image to be identified;

[0061] The target recognition module is used to input the SAR image to be recognized into the trained target recognition model and output the corresponding target prediction result;

[0062] A result output module is used to output the target prediction result output by the target recognition model as the target recognition result of the SAR image to be recognized;

[0063] In the target recognition module, the processing steps for training the target recognition model are as follows:

[0064] 1) Obtain labeled SAR image samples and unlabeled SAR image samples for model training;

[0065] 2) Perform contour extraction and enhancement processing on the unlabeled SAR image samples to obtain unlabeled contour enhanced images;

[0066] 3) Perform strong and weak augmentation on the unlabeled contour enhanced images respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; perform weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples;

[0067] 4) Using unlabeled strong augmented samples, unlabeled weak augmented samples, and labeled weak augmented samples, the FixMatch framework is used for semi-supervised learning to obtain a FixMatch-based object recognition model;

[0068] 5) Input the labeled weakly augmented samples into the collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after the specified rounds, and take the category with the largest probability value as the prediction label for the judgment of this round; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample;

[0069] 6) Based on the distribution of the judgment results of all samples of each sub-model, the entropy characteristics are used to perform a preliminary screening; multiple sub-models are used to collaboratively determine the pseudo-label samples with high confidence; for each type, a certain number of pseudo-label samples are randomly selected;

[0070] 7) The extracted pseudo-label samples are used to supervise the FixMatch-based target recognition model, that is, the label training set of FixMatch is expanded to continue training and a trained target recognition model is obtained.

[0071] Compared with the prior art, the semi-supervised SAR target recognition method based on contour enhancement and collaborative training in the present invention has the following beneficial effects:

[0072] The present invention aims at the problem that the main area of ​​interest in SAR image data is small and there are many background noise points, which may cause the model to spend a lot of energy on learning meaningless features, and proposes a data augmentation method based on contour extraction. Based on the characteristics of SAR images, it is found that the target area of ​​interest is usually concentrated in a bright spot area, and there are many noise points in the background of SAR images. Therefore, the present invention uses contour extraction to further highlight the key features of the target area of ​​SAR images, helping the model to better learn key information. This method expands the data enhancement method in SAR image preprocessing, can process the area of ​​interest more effectively, can highlight key features through contour extraction, and further guide the model to learn key information, thereby improving the accuracy of SAR target recognition.

[0073] Based on the semi-supervised learning based on the FixMatch framework, the present invention proposes a pseudo-label expansion mechanism based on a collaborative training network to address the problem that the performance will be significantly reduced due to the scarcity of label samples, aiming to introduce discriminant information outside the original FixMatch framework. The present invention adopts a collaborative training network to generate high-confidence pseudo-labels, and selects a certain number of high-reliability pseudo-label samples as label data, and adds them to the learning process of the target recognition model, thereby alleviating the problem of scarce label samples. The collaborative training network is composed of four WideResNets with inconsistent depth and width, and is trained using weakly augmented label samples generated by the FixMatch framework to learn SAR image features more comprehensively; after the training of each sub-model is completed, the collaborative training network makes collaborative judgments on the samples and summarizes the judgment results of different rounds, and then uses entropy to evaluate the reliability of the judgment results of each sub-model; a certain number of pseudo-label samples are extracted from each type of high-confidence judgment results, and the common judgment results of the four models form a sampling pool, from which a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for continued training. The present invention introduces a collaborative training network so that the target recognition model can learn more extensive and more reliable SAR target features, thereby improving the reliability and robustness of SAR target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0075] Figure 1 The overall block diagram of the semi-supervised SAR target recognition method based on contour enhancement and collaborative training;

[0076] Figure 2 This is the algorithm block diagram for contour enhancement;

[0077] Figure 3 Block diagram of pseudo-label selection algorithm based on co-training network;

[0078] Figure 4 The influence of different confidence thresholds on model training. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0080] The following is a further detailed description through specific implementation methods:

[0081] Example:

[0082] This embodiment discloses a semi-supervised SAR target recognition method based on contour enhancement and collaborative training.

[0083] like Figure 1 As shown, the semi-supervised SAR target recognition method based on contour enhancement and collaborative training includes:

[0084] S1: Acquire the SAR image to be identified;

[0085] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result; the target recognition model can output all categories and their corresponding prediction probabilities, and take the category with the largest prediction probability as the target prediction result.

[0086] The target prediction results include all categories and their corresponding predicted probabilities.

[0087] The processing steps when training the object recognition model are as follows:

[0088] S201: Obtain labeled SAR image samples and unlabeled SAR image samples for model training;

[0089] S202: performing contour extraction and enhancement processing on the unlabeled SAR image sample to obtain an unlabeled contour enhanced image;

[0090] S203: performing strong and weak augmentation on the unlabeled contour enhanced image respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; performing weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples;

[0091] S204: using the unlabeled strong augmented samples, the unlabeled weak augmented samples, and the labeled weak augmented samples, and using the FixMatch framework to perform semi-supervised learning, to obtain a FixMatch-based object recognition model;

[0092] S205: Input the labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after a specified round, and take the category with the largest probability value as the prediction label of the judgment in this round for each judgment; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample;

[0093] S206: In the first step, the judgment result distribution of all samples based on each sub-model is preliminarily screened by the entropy characteristics (screening out "simple" samples); in the second step, collaborative screening and extraction, multiple sub-models are used to collaboratively determine high-confidence pseudo-label samples; for each type, a certain number of pseudo-label samples are randomly extracted for the first-stage FixMatch framework training;

[0094] S207: supervised learning of the FixMatch-based target recognition model is performed through the extracted pseudo-label samples, that is, the FixMatch label training set is expanded to continue training, so as to obtain a well-trained (better performance and better robustness) target recognition model;

[0095] S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

[0096] The present invention aims at the problem that the main area of ​​interest in SAR image data is small and there are many background noise points, which may cause the model to spend a lot of energy on learning meaningless features, and proposes a data augmentation method based on contour extraction. Based on the characteristics of SAR images, it is found that the target area of ​​interest is usually concentrated in a bright spot area, and there are many noise points in the background of SAR images. Therefore, the present invention uses contour extraction to further highlight the key features of the target area of ​​SAR images, helping the model to better learn key information. This method expands the data enhancement method in SAR image preprocessing, can process the area of ​​interest more effectively, can highlight key features through contour extraction, and further guide the model to learn key information, thereby improving the accuracy of SAR target recognition.

[0097] Based on the semi-supervised learning based on the FixMatch framework, the present invention proposes a pseudo-label expansion mechanism based on a collaborative training network to address the problem that the performance will be significantly reduced due to the scarcity of label samples, aiming to introduce discriminant information outside the original FixMatch framework. The present invention adopts a collaborative training network to generate high-confidence pseudo-labels, and selects a certain number of high-reliability pseudo-label samples as label data, and adds them to the learning process of the target recognition model, thereby alleviating the problem of scarce label samples. The collaborative training network is composed of four WideResNets with inconsistent depth and width, and is trained using weakly augmented label samples generated by the FixMatch framework to learn SAR image features more comprehensively; after the training of each sub-model is completed, the collaborative training network makes collaborative judgments on the samples and summarizes the judgment results of different rounds, and then uses entropy to evaluate the reliability of the judgment results of each sub-model; a certain number of pseudo-label samples are extracted from each type of high-confidence judgment results, and the common judgment results of the four models form a sampling pool, from which a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for continued training. The present invention introduces a collaborative training network so that the target recognition model can learn more extensive and more reliable SAR features, thereby improving the reliability and robustness of SAR target recognition.

[0098] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0099] 1. Outline Enhancement

[0100] The applicant has found that the target region of interest in SAR images is usually small and contains a lot of noise in the background, which easily causes the model to waste a lot of energy on meaningless features. To this end, this embodiment performs targeted contour extraction on the SAR image to process the main target region of interest, and generates a mask image for the extracted target region of interest, referred to as X mask, the mask image performs data enhancement on the pixel grayscale values ​​of the region of interest, and the rest of the area remains black.

[0101] Specifically, Figure 2 As shown, contour extraction and enhancement processing are achieved through the following steps:

[0102] S2011: Perform contour extraction on the original SAR image to extract the target area of ​​interest;

[0103] In this embodiment, the target region of interest of the SAR image can be extracted by referring to the following steps:

[0104] 1) Image preprocessing: SAR images are preprocessed, including denoising (such as using filters) and contrast enhancement (such as histogram equalization) to improve image quality and make the target area more prominent.

[0105] 2) Binarization: Convert the preprocessed image into a binary image. Binarization is to divide the pixels in the image into two categories by setting a threshold: pixels above the threshold are set to white (or a specific color), and pixels below the threshold are set to black (or another specific color). In this way, the target area and the background in the image are clearly distinguished.

[0106] 3) Contour detection: Use contour detection algorithms (such as edge detection algorithms) on binary images to identify the contours of the target area.

[0107] 4) Contour screening: Since contour detection may identify multiple contours, including some areas of no interest (such as noise or clutter in the background), the detected contours are screened. This can be achieved by setting restrictions on the contour's area, perimeter, shape and other features, and only retaining contours that meet specific conditions.

[0108] 5) Extracting the target region of interest: After screening, the remaining contours represent the target region of interest, and the information of these regions, such as coordinates, size, shape, etc., can be extracted as needed.

[0109] S2012: generating a corresponding mask image based on the target region of interest;

[0110] The formula of the mask image is expressed as:

[0111]

[0112] Where: X mask represents the mask image; p represents the pixel value of the corresponding position of the original SAR image; contour represents the target area of ​​interest obtained by contour extraction;

[0113] S2013: superimposing the mask image with the original SAR image to obtain a contour enhanced image;

[0114] The formula for contour enhancement image is expressed as:

[0115] X new1 =(X mean1 X mask +X1)*0.6;

[0116] Where: X new1 represents the contour enhanced image; X1 represents the original SAR image; X mean1 It represents the image after smoothing and averaging the original SAR image X1. By multiplying the smoothed image with the mask, non-feature points can be effectively suppressed and key feature points can be highlighted.

[0117] In this embodiment, the mask image X obtained by contour extraction mask It will be superimposed with the original SAR image, corresponding to the mask image X mask The original SAR image will also be suppressed at the position where the pixel value is 0, and the corresponding mask image X mask The pixel value in the image is 1. By superimposing the original SAR image and the mask image, the contour enhancement effect can be achieved. The present invention can effectively highlight the contour features of the key areas of the SAR image through this data augmentation method and reduce the interference of noise points on model training.

[0118] 2. Weak Augmentation and Strong Augmentation

[0119] In this embodiment, weak augmentation of SAR images is achieved by flipping and translating; strong augmentation of SAR images is achieved by two strategies: RandAugment and CTAugment. Among them, strong and weak augmentation (augmentation) is a method in the FixMatch framework. Strong and weak augmentation: In the experiment, "flipping and translating" are used for weak augmentation; RandAugment and CTAugment are used for strong augmentation. The main random sampling strategy is to select a part from dozens of enhancement categories and a series of enhancement amplitudes for image enhancement. The core is to make the image after strong augmentation difficult to confirm, but still retain sufficient semantic information.

[0120] 3. Semi-supervised learning training

[0121] In this embodiment, the semi-supervised learning training of the target recognition model is implemented through the FixMatch framework. FixMatch combines pseudo-label generation and consistency regularization technology. This method performs supervised learning on labeled data, generates pseudo-labels using unlabeled data, and trains the model through consistency enhancement of strong and weak data. The pseudo-labels in FixMatch are generated by high-confidence predictions, thereby effectively utilizing unlabeled data and improving the generalization ability of the model.

[0122] In the semi-supervised learning process, the loss function of FixMatch includes the supervised loss L x and the unsupervised loss L u ;

[0123] The formula is as follows:

[0124]

[0125] Where: Cross entropy loss L x is the supervised model training loss for labeled samples, and the input data is labeled weakly augmented samples X1; N x represents the number of labeled weakly augmented samples; H represents the cross entropy between two probability distributions; represents the target prediction result output by the target recognition model based on the input SAR image x under the image enhancement mode θ; Represents all categories and their corresponding prediction probabilities. The category with the largest prediction probability is the target prediction result; cross entropy loss L u It is the unsupervised model prediction training loss for unlabeled samples, and the input data is the unlabeled strongly augmented sample U s and unlabeled weak augmented samples U w ; N u Represents the number of unlabeled samples; Represents the target recognition model for the unlabeled strong augmentation sample U s The predicted probability of each category; y represents the category label corresponding to the label image; Represents the category label with the highest probability in the target prediction result of the unlabeled strongly augmented sample. If the maximum probability is greater than the confidence threshold τ, the sample will participate in the loss function calculation of the unlabeled sample.

[0126] 4. Pseudo-label selection based on co-training network

[0127] The applicant found that when the label samples are extremely scarce, the performance of FixMatch will drop significantly. Different from the single pseudo-label selection method of FixMatch, this embodiment proposes a pseudo-label expansion mechanism based on a collaborative training network, which aims to introduce discriminant information outside the original FixMatch framework. Specifically, the model framework proposed in this embodiment is divided into two stages. The first stage is to enter the FixMatch training after the SAR data set is contour enhanced. The second stage is the training and screening of the collaborative training network. First, the weak augmentation method is adopted to expand the label training set in the first stage to 2 times the number and serve as the training set of the collaborative training network in the second stage. Since the simple model can easily correctly identify samples with certain obvious features, the collaborative training network uses four WideResNets with different depths and widths as sub-models for joint training. These four sub-models can judge the test samples from different angles, and their common judgment results are relatively reliable. After the training of each sub-model is completed, the judgment results of different rounds are summarized, and the reliability of each judgment result is judged according to the entropy value. Finally, the common judgment results of the four sub-models are combined into a sampling pool, and a certain number of high-confidence images are extracted from it as pseudo-label samples to expand the label training set of the first stage, and the expanded label training set is used to continue training FixMatch.

[0128] The collaborative training network is divided into two parts: multi-batch aggregation of data and data screening. Multi-batch aggregation aims to collect the judgment results of each sub-model on the test samples at different training stages. Through aggregation, the judgment distribution of each sub-model on each test sample can be obtained. Subsequently, the characteristics of entropy are used to screen the data and extract high-confidence samples. The following is a detailed analysis of these two parts.

[0129] 1. Multi-batch aggregation of data

[0130] Combination Figure 3 As shown, the specific steps include:

[0131] S2051: Construct a collaborative training network consisting of four sub-models; the four sub-models have different depths and widths.

[0132] In this embodiment, the collaborative training network uses four WideResNets with different depths and widths as sub-models for joint training. These four sub-models can judge the test samples from different angles, and their joint judgment results are relatively reliable. For some significant features of SAR images, simple models are easier to distinguish, and these judgment results should be consistent throughout the training process. Here, WideResNet uses a smaller depth and width, and only uses weak augmented data of labeled samples for training.

[0133] S2052: Input the labeled weakly augmented samples into each sub-model for training, and obtain the target prediction result output by each sub-model for each sample;

[0134] S2053: After each preset round (every 200 rounds in this embodiment), the target prediction results of all samples are judged: for each sample, the category with the largest probability value among all target prediction results is selected as its prediction label for this judgment round;

[0135] The formula is:

[0136] label_X 1_n = argmax(y 1_n );

[0137] Where: label_X 1_n represents the category with the maximum probability in the judgment result of the first sub-model in the nth round, that is, the predicted label of the sample in the nth round; y 1_n Represents the probability distribution of the categories in all predicted values ​​of the samples of the first sub-model in round n;

[0138] The first sub-model's multiple rounds of judgment results for the same sample are expressed as follows:

[0139] label_X1={label_X 1_1 ,label_X 1_2 ,…,label_X 1_n}.

[0140] Figure 3 This is a block diagram of the pseudo-label selection algorithm based on the collaborative training network. Four relatively small WideResNets with different depths and widths are used as sub-models of the collaborative training network. These four models can make simple judgments on the test samples from different angles, and their common judgment results are relatively reliable.

[0141] 2. Data screening

[0142] The first step of data screening is entropy-based screening. For multiple rounds of sub-model judgment results on the same sample, multiple different prediction results may appear. When these results are relatively scattered, it indicates that the sub-model's judgment on the sample is not certain enough; only when the prediction results are concentrated in one category can it be regarded as a high-confidence label. The goal of this embodiment is to screen out reliable high-confidence pseudo-labels based on the uncertainty of multiple rounds of prediction results. In this embodiment, the screening process is quantified by calculating entropy. Specifically, it includes the following steps:

[0143] S2061: Calculate the entropy value of the sample based on the judgment results of multiple rounds of the same sample by each sub-model;

[0144] The formula is:

[0145]

[0146] Where: H(y) represents the entropy value of the sample; yi represents the multiple rounds of judgment results of the sub-model on the same sample; c is the number of categories in the judgment result;

[0147] S2062: Delete samples whose entropy values ​​are greater than a preset entropy value;

[0148] In this embodiment, the larger the entropy value, the more uniform the distribution of the judgment results, that is, the higher the uncertainty of the prediction result of the sample, the lower the confidence of the model in the sample. On the contrary, if the H(y) entropy value is very small, or even 0, it means that the model always believes that the sample belongs to a certain category, which means that the prediction result has a high confidence. In this embodiment, samples with entropy values ​​greater than 0.55 are deleted from the prediction results of each sub-model.

[0149] S2063: retain samples with consistent judgment results from each sub-model to build a sampling pool, and use the categories in the judgment results as pseudo labels for the samples;

[0150] In this embodiment, the second step of data screening is collaborative screening and extraction. After the first step of screening, the judgment results of each sub-model have found "simple" samples based on certain features, and these samples have high confidence in their respective sub-models. Next, collaborative screening is used to merge the simple features extracted by each sub-model, and only the samples with consistent judgment results of the four sub-models are retained.

[0151] The formula is:

[0152]

[0153] Where: label_X represents the pseudo label of the sample after collaborative screening by the four sub-models;

[0154] S2064: For each category, a certain number of samples are randomly selected from the sampling pool as pseudo-label samples.

[0155] 3. Collaborative training

[0156] When the judgment results of the four sub-models are not completely consistent, the label is not valid. After collaborative screening, the number of samples that meet the requirements is significantly reduced, but the quality is greatly improved. Finally, among these retained high-quality pseudo-label samples, a certain number of pseudo-label samples are randomly selected from each category to expand the first stage of label training set and continue to train FixMatch.

[0157] Since these high-quality pseudo samples may still contain some erroneous samples, they should not be regarded as labeled samples in the early stage of training, which may mislead the model to train in the wrong direction, thereby affecting the accuracy of the model. In this embodiment, when the training of the FixMatch framework reaches its limit, a high accuracy rate has been obtained and it is difficult to continue to improve. At this time, the highly reliable pseudo samples extracted by the collaborative training network can be regarded as labeled samples and added to the FixMatch framework for continued training. In this way, the generalization and robustness of the model can be improved by combining the features learned by the original FixMatch model and the broader features provided by the new samples.

[0158] The proposed collaborative training network method adds a new loss function term based on the two loss functions contained in the original FixMatch. This new term is obtained by using pseudo-label samples screened by the collaborative training network to expand the label training set of the first stage and continue to train FixMatch. During the collaborative training process, the loss function of the target recognition model includes the collaborative supervised loss L x2 ;

[0159] The formula is:

[0160]

[0161] Where: The third collaborative supervised loss also uses cross entropy to supervise the sample, and the input data is the weakly augmented pseudo-label sample X sa , X sa represents a pseudo-label sample; N sa Represents the number of pseudo-label samples.

[0162] In addition, for the training loss function of the sub-model in the collaborative training network, only a simple fully supervised loss is used. This is because the number of input samples is small, and during the sub-model training process, the loss of fully supervised label samples will not decrease significantly, so the selection of the loss function will not have much impact.

[0163] In summary, in this embodiment, the specific formula of the total loss function of the target recognition model is as follows:

[0164] Loss = λ1L x +λ2L u +λ3L x2 ;

[0165] Among them, λ1, λ2, λ3 are the weights for a single loss function.

[0166] 5. Experimental Description

[0167] Table 1 shows the comparative experimental results between the method of the present invention and FixMatch, and the experimental data set is MSTAR. 5, 10, 20, 40 and 80 labeled images of each category were selected as the original training data to compare the difference in accuracy between the original framework of FixMatch and the proposed method. The results are the average of multiple training results. It can be seen that the performance of the method proposed in the present invention is significantly better than that of FixMatch in terms of accuracy. When there are only 5 labeled data, the performance of FixMatch seriously declines to 30.65%, while the method proposed in the present invention can still maintain an accuracy of 70.62%. When the number of labeled samples increases to 10 or more, the accuracy of the method proposed in the present invention continues to be better than FixMatch. This shows that the performance of the method of the present invention surpasses FixMatch and can still show excellent performance when there are only very few labeled samples.

[0168] Table 1 Comparison experimental results of the method of the present invention and FixMatch

[0169]

[0170] Table 2 shows the comparative experimental results of various methods. When there are only 5 and 10 labeled samples per class, the accuracy of the method proposed in the present invention reaches 70.62% and 99.02% respectively. When the number of labeled samples per class is large, the method of the present invention still leads the SRC, K-SVD and DCNN methods by about 6% to 20%. Triple-GANs and DCGAN are both generative adversarial network models, which only show advantages when the number of labeled samples is large, but their accuracy is still about 3% lower than the accuracy of the example of the present invention when the number of labeled samples per class is 10 when the number of labeled samples per class is 220. It is worth noting that for the recent methods SCA and HDLM, when the number of labeled samples per class is 5, the SCA method has not been tested, and the accuracy of the method of the present invention is only 1.75% lower than that of HDLM; but when the number of labeled samples per class is 10, the accuracy of the method of the present invention is 4.2% higher than that of SCA and 10.86% higher than that of HDLM. Since the performance of the FixMatch main framework will drop sharply when the label data is extremely scarce, the contour enhancement data augmentation method and the collaborative training network adopted by the present invention can effectively make up for this defect. The comparison results of the above methods further prove the superiority of the method proposed by the present invention.

[0171] Table 2 Comparison experimental results of various methods

[0172]

[0173] Table 3 shows the results of ablation experiments to evaluate the impact of various components of the proposed method on model performance. On the MSTAR dataset, the experiment is divided into three components: the first part is contour extraction; the second part is the addition of the collaborative training network, but entropy is not used to screen the judgment results, which is called the original collaborative training network; the third part is the addition of the collaborative training network and the use of entropy to screen the judgment results, which is called the complete collaborative training network. According to the results of the ablation experiment, contour extraction has a counter-effect on the model when its own accuracy is low. When only contour enhancement is used and the number of labeled samples per class is 5, its accuracy is 2.03% lower than that without contour extraction. However, when the number of labeled samples per class is 10 and 20, contour enhancement plays a significant role, improving the accuracy of the original FixMatch by 3.71% and 2.75%, respectively. If the collaborative training network does not perform entropy screening to eliminate low-confidence samples, then when the accuracy itself is already very high, it may mislead the model and cause a decrease in accuracy. For example, when the number of labeled samples per class is 20, the accuracy of the collaborative training network without entropy screening is 0.77% lower than that without adding the collaborative training network. Overall, the contribution of the collaborative training network to the improvement of model accuracy is slightly lower than that of contour extraction. The best results are 3.43% and 1.57% when the number of labeled samples per class is 10 and 20, respectively. However, when the number of labeled samples per class is only 5, the collaborative training network plays an important role, improving 32.28% compared to the original FixMatch. This ablation experiment reveals the independent contributions of the collaborative training network and contour extraction to the improvement of model accuracy, as well as the necessity of screening the pseudo-label samples generated by the collaborative training network.

[0174] Table 3 Ablation experiment results

[0175]

[0176] Figure 4The influence of different confidence thresholds on the training of the target recognition model is shown, and the accuracy is the average value that can be achieved when the model training is completed. In order to compare the influence of different confidence levels on the model's pseudo-label judgment results, the model's judgment results on unlabeled data in the third round of training were selected and compared with the actual correct labels. The experimental results show that when the confidence threshold τ is low, the model can almost allow all data to participate in the unlabeled cross entropy calculation in the initial stage. When the confidence threshold is 0.3, the highest prediction probability of 2402 weakly augmented samples in 2747 images exceeds 0.3, but only 823 of them are actually correct labels and 1579 are wrong labels, which leads to the failure of the model to be effectively trained, and the final accuracy is only 20.41%. When the confidence threshold is set to 0.6, 0.9 and 0.95 respectively, the model's ability to identify sample categories is significantly improved. In the third round of training, the ratio of the number of correct labels to the number of wrong labels in the number of pseudo-labels that meet the confidence threshold rose sharply, reaching 2.836, 4.299 and 5.152 respectively. Correspondingly, the model accuracy also increased from 88.33% to 94.10%, reaching a peak of 99.02% at a confidence threshold of 0.95. However, continuing to increase the confidence threshold will lead to a sharp decrease in the number of samples that meet the conditions. When the confidence threshold was raised to 0.98, only 71 samples of the model met the threshold in the third round of training, of which 42 were correct labels and 29 were incorrect labels, which were close. In addition, in subsequent rounds, the number of qualified samples remained at about 100, ultimately resulting in an accuracy of only 24.58%. In summary, the model can perform best when the confidence threshold is 0.95, while thresholds that are too high or too low will significantly reduce the model's capabilities.

[0177] Tables 4 and 5 give the N x =10 and N x =20, N sa The impact on accuracy. According to the experimental results, N x = 10, extracting 5, 10, and 20 images from the collaborative training network and adding them to the labeled sample training can improve the performance of the original model. When extracting 5 images, an accuracy of 96.84% can be obtained, but when extracting 10 and 15 images, the accuracy decreases. This is because the probability of extracting wrong pseudo-label samples increases. When extracting 20 images, the accuracy is improved again due to the greatly increased number of expanded correct pseudo-label samples, but it is still not as good as the accuracy obtained by extracting only 5 images. x =20 and the overall trend and N x =10, but the number of pseudo-label samples extracted from the collaborative training network is always less than the number of original label training sets, so there is no sharp drop in accuracy. In summary, the model can obtain the best accuracy by extracting 5 pseudo-label samples from the collaborative training network as label samples to expand the original label training set and continue training.

[0178] Table 4 N x =10 when N sa Impact on accuracy

[0179]

[0180] Table 5 N x =20 hours N sa Impact on accuracy

[0181]

[0182] Tables 6 and 7 give the N sa =5 and N sa =20. According to the experimental results, when the collaborative loss weight is set to 0.75, the accuracy of the model reaches the maximum. Too small a loss weight will make it difficult for the model to learn new feature information from the pseudo-label samples screened by the collaborative training network. For example, when the loss weight is 0.25, the accuracy of extracting 5 and 20 pseudo-label samples is not much different from the accuracy of 93.41% obtained by using FixMatch alone. When the loss weight is 1, the model performance is slightly reduced compared to 0.75. Therefore, it can be determined that the optimal loss weight value is 0.75.

[0183] Table 6 N sa The influence of λ3 on the accuracy when λ=5

[0184]

[0185] Table 7 N sa The influence of λ3 on the accuracy when λ=20

[0186]

[0187] Embodiment 2:

[0188] This embodiment discloses a semi-supervised SAR target recognition system based on contour enhancement and collaborative training, which is implemented based on the semi-supervised SAR target recognition method in the first embodiment.

[0189] A semi-supervised SAR target recognition system based on contour enhancement and collaborative training, including:

[0190] An image input module is used to obtain a SAR image to be identified;

[0191] The target recognition module is used to input the SAR image to be recognized into the trained target recognition model and output the corresponding target prediction result;

[0192] A result output module is used to output the target prediction result output by the target recognition model as the target recognition result of the SAR image to be recognized;

[0193] In the target recognition module, the processing steps for training the target recognition model are as follows:

[0194] 1) Obtain labeled SAR image samples and unlabeled SAR image samples for model training;

[0195] 2) Perform contour extraction and enhancement processing on the unlabeled SAR image samples to obtain unlabeled contour enhanced images;

[0196] 3) Perform strong and weak augmentation on the unlabeled contour enhanced images respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; perform weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples;

[0197] 4) Using unlabeled strong augmented samples, unlabeled weak augmented samples, and labeled weak augmented samples, the FixMatch framework is used for semi-supervised learning to obtain a FixMatch-based object recognition model;

[0198] 5) Input the labeled weakly augmented samples into the collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after the specified rounds, and take the category with the largest probability value as the prediction label for the judgment of this round; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample;

[0199] 6) Based on the distribution of the judgment results of all samples of each sub-model, the entropy characteristics are used to perform a preliminary screening; multiple sub-models are used to collaboratively determine the pseudo-label samples with high confidence; for each type, a certain number of pseudo-label samples are randomly selected for the first stage of FixMatch framework training;

[0200] 7) The extracted pseudo-label samples are used to supervise the FixMatch-based target recognition model, that is, the label training set of FixMatch is expanded to continue training, so as to obtain a well-trained (better performance and better robustness) target recognition model.

[0201] The present invention aims at the problem that the main area of ​​interest in SAR image data is small and there are many background noise points, which may cause the model to spend a lot of energy on learning meaningless features, and proposes a data augmentation method based on contour extraction. Based on the characteristics of SAR images, it is found that the target area of ​​interest is usually concentrated in a bright spot area, and there are many noise points in the background of SAR images. Therefore, the present invention uses contour extraction to further highlight the key features of the target area of ​​SAR images, helping the model to better learn key information. This method expands the data enhancement method in SAR image preprocessing, can process the area of ​​interest more effectively, can highlight key features through contour extraction, and further guide the model to learn key information, thereby improving the accuracy of SAR target recognition.

[0202] Based on the semi-supervised learning based on the FixMatch framework, the present invention proposes a pseudo-label expansion mechanism based on a collaborative training network to address the problem that the performance will be significantly reduced due to the scarcity of label samples, aiming to introduce discriminant information outside the original FixMatch framework. The present invention adopts a collaborative training network to generate high-confidence pseudo-labels, and selects a certain number of high-reliability pseudo-label samples as label data, and adds them to the learning process of the target recognition model, thereby alleviating the problem of scarce label samples. The collaborative training network is composed of four WideResNets with inconsistent depth and width, and is trained using weakly augmented label samples generated by the FixMatch framework to learn SAR image features more comprehensively; after the training of each sub-model is completed, the collaborative training network makes collaborative judgments on the samples and summarizes the judgment results of different rounds, and then uses entropy to evaluate the reliability of the judgment results of each sub-model; a certain number of pseudo-label samples are extracted from each type of high-confidence judgment results, and the common judgment results of the four models form a sampling pool, from which a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for continued training. The present invention introduces a collaborative training network so that the target recognition model can learn more extensive and more reliable SAR target features, thereby improving the reliability and robustness of SAR target recognition.

[0203] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A semi-supervised SAR target recognition method based on contour enhancement and collaborative training, characterized in that: include: S1: Acquire the SAR image to be identified; S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result; The processing steps when training the object recognition model are as follows: S201: Obtain labeled SAR image samples and unlabeled SAR image samples for model training; S202: performing contour extraction and enhancement processing on the unlabeled SAR image sample to obtain an unlabeled contour enhanced image; S203: performing strong and weak augmentation on the unlabeled contour enhanced image respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; performing weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples; S204: using the unlabeled strong augmented samples, the unlabeled weak augmented samples, and the labeled weak augmented samples, and using the FixMatch framework to perform semi-supervised learning, to obtain a FixMatch-based object recognition model; S205: Input the labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after a specified round, and take the category with the largest probability value as the prediction label of the judgment in this round for each judgment; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample; S206: Preliminarily screening the judgment result distribution of all samples based on each sub-model through the characteristics of entropy; determining pseudo-label samples with high confidence through collaborative judgment of multiple sub-models; and randomly extracting a certain number of pseudo-label samples for each type; S207: supervise the object recognition model based on FixMatch through the extracted pseudo-label samples, that is, expand the label training set of FixMatch to continue training, and obtain a trained object recognition model; S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

2. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 1, characterized in that: In step S202, contour extraction and enhancement processing are implemented through the following steps: S2011: Perform contour extraction on the original SAR image to extract the target area of ​​interest; S2012: generating a corresponding mask image based on the target region of interest; The formula of the mask image is expressed as: Where: X mask represents the mask image; p represents the pixel value of the corresponding position of the original SAR image; contour represents the target area of ​​interest obtained by contour extraction; S2013: superimposing the mask image with the original SAR image to obtain a contour enhanced image; The formula for contour enhancement image is expressed as: X new1 =(X mean1 X mask +X1)*0.6; Where: X new1 represents the contour enhanced image; X1 represents the original SAR image; X mean1 Represents the image after smoothing and averaging the original SAR image X1.

3. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 1, characterized in that: In step S203, weak augmentation of the SAR image is achieved by flipping and translating; strong augmentation of the SAR image is achieved by RandAugment and CTAugment strategies.

4. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 1, characterized in that: In step S204, during the semi-supervised learning process, the loss function of the target recognition model includes the supervised loss L x and the unsupervised loss L u ; The formula is as follows: Where: N x represents the number of labeled weakly augmented samples; H represents the cross entropy between two probability distributions; represents the target prediction result output by the target recognition model based on the input SAR image x under the image enhancement mode θ; Represents all categories and their corresponding predicted probabilities; N u Represents the number of unlabeled samples; Represents the target recognition model for the unlabeled strong augmentation sample U s The predicted probability of each category; y represents the category label corresponding to the label image; Represents the category label with the highest probability in the target prediction results of the unlabeled weakly augmented samples.

5. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 1, characterized in that: Step S205 specifically includes the following steps: S2051: Construct a collaborative training network consisting of four sub-models; the four sub-models have different depths and widths; S2052: Input the labeled weakly augmented samples into each sub-model for training, and obtain the target prediction result output by each sub-model for each sample; S2053: After each preset round, the target prediction results of all samples are judged: for each sample, the category with the largest probability value among all target prediction results is selected as the prediction label of the judgment round; The formula is: label_X 1_n =argmax(y 1_n ); Where: label_X 1_n represents the category with the maximum probability in the judgment result of the first sub-model in the nth round, that is, the predicted label of the sample in the nth round; y 1_n Represents the probability distribution of the categories in all predicted values ​​of the samples of the first sub-model in round n; The first sub-model's multiple rounds of judgment results for the same sample are expressed as follows: label_X1={label_X 1_1 ,label_X 1_2 ,…,label_X 1_n }。 6. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 5, characterized in that: Step S206 specifically includes the following steps: S2061: Calculate the entropy value of the sample based on the judgment results of multiple rounds of the same sample by each sub-model; The formula is: Where: H(y) represents the entropy value of the sample; y i represents the multiple rounds of judgment results of the sub-model on the same sample; c is the number of categories in the judgment results; S2062: Delete samples whose entropy values ​​are greater than a preset entropy value; S2063: retain samples with consistent judgment results from each sub-model to build a sampling pool, and use the categories in the judgment results as pseudo labels for the samples; The formula is: Where: label_X represents the pseudo label of the sample after collaborative screening by the four sub-models; S2064: For each category, a certain number of samples are randomly selected from the sampling pool as pseudo-label samples.

7. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 1, characterized in that: In step S207, when the accuracy of the target recognition model using FixMatch for semi-supervised learning reaches a preset value and cannot be further improved, the target recognition model based on FixMatch is further trained by expanding the pseudo-label samples to the label training set, i.e., collaborative training.

8. The semi-supervised SAR target recognition method based on contour enhancement and collaborative training as claimed in claim 7, characterized in that: In step S207, during the collaborative training process, the loss function of the target recognition model includes the collaborative supervised loss L x2 ; The formula is: Where: X sa represents a pseudo-label sample; N sa Represents the number of pseudo-label samples.

9. A semi-supervised SAR target recognition system based on contour enhancement and collaborative training, characterized by: The implementation of the semi-supervised SAR target recognition method based on contour enhancement and collaborative training described in claim 1 includes: An image input module is used to obtain a SAR image to be identified; The target recognition module is used to input the SAR image to be recognized into the trained target recognition model and output the corresponding target prediction result; A result output module is used to output the target prediction result output by the target recognition model as the target recognition result of the SAR image to be recognized; In the target recognition module, the processing steps for training the target recognition model are as follows: 1) Obtain labeled SAR image samples and unlabeled SAR image samples for model training; 2) Perform contour extraction and enhancement processing on the unlabeled SAR image samples to obtain unlabeled contour enhanced images; 3) Perform strong and weak augmentation on the unlabeled contour enhanced images respectively to obtain unlabeled strongly augmented samples and unlabeled weakly augmented samples; perform weak augmentation on the labeled SAR image samples to obtain labeled weakly augmented samples; 4) Using unlabeled strong augmented samples, unlabeled weak augmented samples, and labeled weak augmented samples, the FixMatch framework is used for semi-supervised learning to obtain a FixMatch-based object recognition model; 5) Input the labeled weakly augmented samples into the collaborative training network including several sub-models for training, and obtain the target prediction results corresponding to each sample at different training stages; make judgments after the specified rounds, and take the category with the largest probability value as the prediction label for the judgment of this round; after multiple judgments, obtain the judgment result distribution of each sub-model in the collaborative training network for any sample; 6) Based on the distribution of the judgment results of all samples of each sub-model, the entropy characteristics are used to perform a preliminary screening; multiple sub-models are used to collaboratively determine the pseudo-label samples with high confidence; for each type, a certain number of pseudo-label samples are randomly selected; 7) The extracted pseudo-label samples are used to supervise the FixMatch-based target recognition model, that is, the label training set of FixMatch is expanded to continue training and a trained target recognition model is obtained.

Images