Semi-supervised sar target recognition method and system based on contour enhancement and co-training

Abstract

The present application relates to a kind of semi-supervised SAR target identification method and system based on profile enhancement and collaborative training.The specific method inputs the SAR image to be identified into the trained target identification model to generate corresponding target prediction result.The training process is divided into two stages: the first stage is, profile extraction and enhancement processing are carried out to the unlabeled SAR image sample, then semi-supervised learning is carried out using FixMatch framework.The second stage is, the weak augmented image of label sample in the first stage is input into collaborative training network for training, the decision result of different rounds is summarized, and the reliability of each decision result is judged by entropy value.The common decision result of collaborative training network constitutes a sampling pool, and a certain number of high-confidence samples are extracted from the sampling pool as pseudo-label samples to expand the label training set of the first stage, and then the expanded label training set is used to continue training FixMatch.The present application significantly improves the accuracy, reliability and robustness of SAR target identification.

Description

Technical Field

[0001] This invention relates to the field of radar target detection technology, and specifically to a semi-supervised SAR target recognition method and system based on contour enhancement and collaborative training. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an imaging radar system widely deployed on mobile platforms such as UAVs, aircraft, and satellites. By transmitting microwave pulses towards the observed target and receiving the echo signals, SAR achieves precise imaging observation of the land and sea surfaces. Compared to optical imaging systems, SAR possesses the capabilities of all-weather, all-time data acquisition, high resolution, and long-range operation, thus finding widespread application in various fields such as battlefield reconnaissance, topographic mapping, geological exploration, and marine surveillance. Automatic Target Recognition (ATR) of SAR images is one of the key technologies for SAR data application, significantly improving the efficiency and accuracy of surveillance and reconnaissance missions.

[0003] In recent years, artificial intelligence technologies, represented by deep learning, have empowered SAR target recognition, greatly improving target recognition performance. However, deep learning models require a large number of labeled training samples, while 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. Data augmentation and pseudo-label generation are two important methods in SSL. Data augmentation increases the diversity of training samples, thereby improving the generalization and robustness of deep learning models. Generating pseudo-labels for unlabeled samples effectively expands the labeled training samples, thus improving the recognition accuracy of deep learning models. However, existing SAR target recognition methods based on semi-supervised learning still have the following problems: 1) SAR image data often has small regions of interest and a lot of background noise, which may cause the model to spend a lot of energy learning meaningless features, resulting in poor accuracy in SAR target recognition. 2) In semi-supervised deep learning techniques, when labeled samples are scarce, existing pseudo-label generation methods struggle to obtain highly reliable pseudo-labels, preventing the model from effectively learning reliable features and thus limiting the recognition performance and generalization ability of SAR target recognition models. Therefore, improving the accuracy, reliability, and robustness of SAR target recognition is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a semi-supervised SAR target recognition method based on contour enhancement and co-training, which effectively processes the region of interest and highlights key features by performing contour extraction and enhancement processing on SAR images, while enabling the model to learn a wider range of more reliable SAR target features through co-training, thereby improving the accuracy, reliability and robustness of SAR target recognition.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] Semi-supervised SAR target recognition methods based on contour enhancement and co-training include:

[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 target recognition model are as follows:

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

[0012] S202: Perform contour extraction and enhancement processing on unlabeled SAR image samples to obtain unlabeled contour-enhanced images;

[0013] S203: Perform strong and weak augmentation on the unlabeled contour-enhanced images 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.

[0014] S204: Using unlabeled strong augmented samples, unlabeled weak augmented samples, and labeled weak augmented samples, semi-supervised learning is performed using the FixMatch framework to obtain a target recognition model based on FixMatch.

[0015] S205: Input labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results for each sample at different training stages; make a decision after a specified number of rounds, and take the class with the highest probability value as the prediction label for this round; after multiple decisions, obtain the distribution of the decision results of each sub-model in the collaborative training network for any sample;

[0016] S206: Based on the distribution of the decision results of each sub-model for all samples, a preliminary screening is performed using the characteristics of entropy; high-confidence pseudo-label samples are determined through collaborative decision-making by multiple sub-models; for each type, a certain number of pseudo-label samples are randomly selected.

[0017] S207: Supervised learning of the FixMatch-based target recognition model is performed using extracted pseudo-label samples, that is, the label training set of FixMatch is expanded and training is continued to obtain a well-trained target recognition model.

[0018] S3: Use the target prediction results output by the target recognition model as the target category of the SAR image to be recognized.

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

[0020] S2011: Extract contours from the original SAR image to identify the region of interest;

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

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

[0023]

[0024] In the formula: X mask represents the mask image; p represents the pixel value at the corresponding location in the original SAR image; contour represents the region of interest obtained through contour extraction.

[0025] S2013: The mask image is superimposed on the original SAR image to obtain a contour-enhanced image;

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

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

[0028] In the formula: X new1 X1 represents the contour-enhanced image; X2 represents the original SAR image; X3 represents the contour-enhanced image. mean1 This 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 translating; strong augmentation of the SAR image is achieved by using two strategies: RandAugment and CTAugment.

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

[0031] The formula is expressed as follows:

[0032]

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

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

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

[0036] S2052: Input 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 highest probability value among all target prediction results is selected as its prediction label for that round.

[0038] The formula is expressed as:

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

[0040] In the formula: label_X 1_n y represents the category with the highest probability in the decision result of the first sub-model in round n, i.e., the predicted label of the sample in round n; 1_n This represents the probability distribution of the class among all predicted values ​​of the sample in the nth round of the first sub-model;

[0041] The first sub-model's multi-round decision results for the same sample are represented 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 decision results of each sub-model for the same sample in multiple rounds;

[0045] The formula is expressed as:

[0046]

[0047] In the formula: H(y) represents the entropy value of the sample; y i This represents the multi-round decision results of the sub-model on the same sample; c is the number of categories in the decision results.

[0048] S2062: Delete samples whose entropy value is greater than the preset entropy value;

[0049] S2063: Retain samples with consistent decision results from each sub-model to construct a sampling pool, and use the category in the decision result as the pseudo-label of the sample;

[0050] The formula is expressed as:

[0051]

[0052] In the formula: label_X represents the pseudo-label assigned to 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 continued to be trained by expanding the pseudo-label samples into 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 collaborative supervised loss L. x2 ;

[0056] The formula is expressed as:

[0057]

[0058] In the formula: X sa N represents pseudo-labeled samples; sa This indicates the number of pseudo-labeled samples.

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

[0060] Image input module, used to acquire SAR images 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] The result output module is used to output the target prediction results output by the target recognition model as the target recognition results of the SAR image to be recognized.

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

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

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

[0066] 3) Perform strong and weak augmentation on the unlabeled contour-enhanced images 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, semi-supervised learning is performed using the FixMatch framework to obtain a target recognition model based on FixMatch.

[0068] 5) Input labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results for each sample at different training stages; make a decision after a specified number of rounds, and take the class with the highest probability value as the prediction label for this round; after multiple decisions, obtain the distribution of the decision results of each sub-model in the collaborative training network for any sample.

[0069] 6) Based on the distribution of decision results for all samples by each sub-model, preliminary screening is performed using the characteristics of entropy; high-confidence pseudo-label samples are determined through collaborative decision-making by multiple sub-models; for each type, a certain number of pseudo-label samples are randomly selected.

[0070] 7) Supervised learning of the FixMatch-based target recognition model is performed using extracted pseudo-label samples, that is, the label training set of FixMatch is expanded and training is continued to obtain a well-trained target recognition model.

[0071] Compared with existing technologies, the semi-supervised SAR target recognition method based on contour enhancement and collaborative training in this invention has the following advantages:

[0072] This invention addresses the problem that SAR image data often contains small regions of interest (ROIs) and significant background noise, which can lead to models expending considerable effort learning meaningless features. It proposes a contour extraction-based data augmentation method. Based on the characteristics of SAR images, it is found that ROIs are typically concentrated in a bright area, while the SAR image background contains considerable noise. Therefore, this invention uses contour extraction to highlight the key features of the SAR image target region, helping the model to better learn crucial information. This method expands upon data augmentation techniques in SAR image preprocessing, enabling more effective handling of ROIs. By highlighting key features through contour extraction, it further guides the model to learn crucial information, thereby improving the accuracy of SAR target recognition.

[0073] This invention, based on semi-supervised learning using the FixMatch framework, addresses the significant performance degradation caused by scarce labeled samples by proposing a pseudo-label augmentation mechanism based on a collaboratively trained network. This mechanism aims to introduce discriminative information beyond the original FixMatch framework. The invention employs a collaboratively trained network to generate high-confidence pseudo-labels and selects a certain number of high-reliability pseudo-label samples as labeled data, incorporating them into the target recognition model's learning process to alleviate the label sample scarcity problem. The collaboratively trained network consists of four WideResNets with varying depths and widths, trained using weakly augmented labeled samples generated by the FixMatch framework to more comprehensively learn SAR image features. After each sub-model training, the collaboratively trained network performs collaborative decision-making on the samples and summarizes the decision results from different rounds, using entropy values ​​to evaluate the reliability of each sub-model's decision results. A certain number of pseudo-label samples are extracted from each class of high-confidence decision results, and the common decision results of the four models are combined into a sampling pool. From this pool, a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for further training. This invention introduces a collaborative training network, enabling the target recognition model to learn a wider range of more reliable SAR target features, thereby improving the reliability and robustness of SAR target recognition. Attached Figure Description

[0074] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0075] Figure 1 This is an overall block diagram of a semi-supervised SAR target recognition method based on contour enhancement and co-training;

[0076] Figure 2 A flowchart of the contour enhancement algorithm;

[0077] Figure 3 The diagram shows the pseudo-label selection algorithm based on a collaboratively trained network.

[0078] Figure 4 The impact of different confidence thresholds on model training. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0080] The following detailed explanation illustrates the specific implementation methods:

[0081] Example:

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

[0083] like Figure 1 As shown, a semi-supervised SAR target recognition method based on contour enhancement and co-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; wherein the target recognition model can output all categories and their corresponding prediction probabilities, and take the category with the highest 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 target recognition model are as follows:

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

[0089] S202: Perform contour extraction and enhancement processing on unlabeled SAR image samples to obtain unlabeled contour-enhanced images;

[0090] S203: Perform strong and weak augmentation on the unlabeled contour-enhanced images 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.

[0091] S204: Using unlabeled strong augmented samples, unlabeled weak augmented samples, and labeled weak augmented samples, semi-supervised learning is performed using the FixMatch framework to obtain a target recognition model based on FixMatch.

[0092] S205: Input labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results for each sample at different training stages; make a decision after a specified number of rounds, and take the class with the highest probability value as the prediction label for this round; after multiple decisions, obtain the distribution of the decision results of each sub-model in the collaborative training network for any sample;

[0093] S206: First, based on the distribution of the decision results of each sub-model for all samples, preliminary screening is performed using the characteristics of entropy (screening out "simple" samples); Second, collaborative screening and extraction are performed, and high-confidence pseudo-label samples are determined through collaborative decision-making by multiple sub-models; For each type, a certain number of pseudo-label samples are randomly extracted for training of the FixMatch framework in the first stage.

[0094] S207: Supervised learning of the FixMatch-based target recognition model is performed using extracted pseudo-label samples, that is, the label training set of FixMatch is expanded and training is continued to obtain a well-trained (better performance and more robust) target recognition model.

[0095] S3: Use the target prediction results output by the target recognition model as the target category of the SAR image to be recognized.

[0096] This invention addresses the problem that SAR image data often contains small regions of interest (ROIs) and significant background noise, which can lead to models expending considerable effort learning meaningless features. It proposes a contour extraction-based data augmentation method. Based on the characteristics of SAR images, it is found that ROIs are typically concentrated in a bright area, while the SAR image background contains considerable noise. Therefore, this invention uses contour extraction to highlight the key features of the SAR image target region, helping the model to better learn crucial information. This method expands upon data augmentation techniques in SAR image preprocessing, enabling more effective handling of ROIs. By highlighting key features through contour extraction, it further guides the model to learn crucial information, thereby improving the accuracy of SAR target recognition.

[0097] This invention, based on semi-supervised learning using the FixMatch framework, addresses the significant performance degradation caused by scarce labeled samples by proposing a pseudo-label augmentation mechanism based on a collaboratively trained network. This mechanism aims to introduce discriminative information beyond the original FixMatch framework. The invention employs a collaboratively trained network to generate high-confidence pseudo-labels and selects a certain number of high-reliability pseudo-label samples as labeled data, incorporating them into the target recognition model's learning process to alleviate the label sample scarcity problem. The collaboratively trained network consists of four WideResNets with varying depths and widths, trained using weakly augmented labeled samples generated by the FixMatch framework to more comprehensively learn SAR image features. After each sub-model training, the collaboratively trained network performs collaborative decision-making on the samples and summarizes the decision results from different rounds, using entropy values ​​to evaluate the reliability of each sub-model's decision results. A certain number of pseudo-label samples are extracted from each class of high-confidence decision results, and the common decision results of the four models are combined into a sampling pool. From this pool, a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for further training. This invention introduces a collaborative training network, enabling the target recognition model to learn a wider range of more reliable SAR features, thereby improving the reliability and robustness of SAR target recognition.

[0098] To better illustrate the technical solution of the present invention, this embodiment is described in the following parts.

[0099] I. Contour Enhancement

[0100] The applicant found that the regions of interest (ROIs) in SAR images are typically small, and the background contains a lot of noise, which easily leads to the model wasting a lot of effort on meaningless features. Therefore, this embodiment performs targeted contour extraction on the SAR image to process the main ROIs. For the extracted ROIs, a mask image, referred to as X, is generated. maskThe mask image augments the pixel grayscale values ​​of the region of interest, while the rest of the region remains black.

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

[0102] S2011: Extract contours from the original SAR image to identify the region of interest;

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

[0104] 1) Image preprocessing: Preprocessing SAR images includes steps such as noise reduction (e.g., using filters) and contrast enhancement (e.g., histogram equalization) to improve image quality and make the target area stand out more.

[0105] 2) Binarization: Converting the preprocessed image into a binary image. Binarization divides 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 separated.

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

[0107] 4) Contour Filtering: Since contour detection may identify multiple contours, including some uninteresting areas (such as noise or background clutter), the detected contours need to be filtered. This can be achieved by setting constraints on features such as the area, perimeter, and shape of the contours, retaining only contours that meet specific conditions.

[0108] 5) Extracting the target region of interest: After filtering, the remaining outlines represent the target region of interest. Extract information about these regions as needed, such as coordinates, size, and shape.

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

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

[0111]

[0112] In the formula: X mask represents the mask image; p represents the pixel value at the corresponding location in the original SAR image; contour represents the region of interest obtained through contour extraction.

[0113] S2013: The mask image is superimposed on the original SAR image to obtain a contour-enhanced image;

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

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

[0116] In the formula: X new1 X1 represents the contour-enhanced image; X2 represents the original SAR image; X3 represents the contour-enhanced image. mean1 This 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, while key feature points are highlighted.

[0117] In this embodiment, the mask image X obtained by contour extraction mask This will be overlaid with the original SAR image, corresponding to the mask image X. mask At locations where the pixel value is 0, the original SAR image will also be suppressed, while the corresponding mask image X... mask The location with a pixel value of 1. By overlaying the original SAR image and the mask image, contour enhancement can be achieved. This invention effectively highlights the contour features of key regions in SAR images and reduces noise interference with model training through this data augmentation method.

[0118] II. Weak augmentation and strong augmentation

[0119] In this embodiment, weak augmentation of SAR images is achieved through flipping and translation; strong augmentation of SAR images is achieved through two strategies: RandAugment and CTAugment. The strong and weak augmentation methods are from the FixMatch framework. In the experiment, "flipping and translation" were used for weak augmentation; RandAugment and CTAugment were used for strong augmentation. The main strategy is random sampling, selecting a subset from dozens of augmentation categories and a series of augmentation magnitudes for image enhancement. The core objective is to make the strongly augmented image difficult to identify, while still retaining sufficient semantic information.

[0120] III. Semi-supervised learning and training

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

[0122] In the semi-supervised learning process, FixMatch's loss function includes the supervised loss L. x And unsupervised loss L u ;

[0123] The formula is expressed as follows:

[0124]

[0125] Where: cross-entropy loss L x This is the supervised model training loss for labeled samples, where the input data is labeled weakly augmented samples X1; N x This represents the number of weakly augmented labeled samples; H represents the cross-entropy between the two probability distributions. This represents the target prediction result output by the target recognition model based on the input SAR image x under image enhancement method θ; Represents all categories and their corresponding predicted probabilities; the category with the highest predicted probability is the target prediction result; cross-entropy loss L u This refers to the unsupervised model prediction training loss for unlabeled samples, with the input data being unlabeled strongly augmented samples U. s And unlabeled weakly augmented samples U w N u Indicates the number of unlabeled samples; This indicates that the target recognition model has strong augmentation effect on unlabeled samples U. s The predicted probability for each category; y represents the category label corresponding to the label image; This represents the class label with the highest probability in the target prediction result of the unlabeled strongly augmented sample. Only if the highest probability is greater than the confidence threshold τ will the sample participate in the loss function calculation of the unlabeled sample.

[0126] IV. Pseudo-label selection based on collaborative training networks

[0127] The applicant discovered that FixMatch's performance significantly degrades when labeled samples are extremely scarce. Unlike FixMatch's single pseudo-label selection method, this embodiment proposes a pseudo-label augmentation mechanism based on a co-training network, aiming to introduce discriminative information beyond the original FixMatch framework. Specifically, the proposed model framework consists of two stages. The first stage involves contour augmentation of the SAR dataset before training in FixMatch. The second stage involves the training and selection of the co-training network. First, a weak augmentation method is used to expand the labeled training set from the first stage by a factor of two, which then serves as the training set for the second-stage co-training network. Since simple models can more easily and correctly identify samples with certain obvious features, the co-training network uses four WideResNet sub-models with different depths and widths for joint training. These four sub-models can make judgments on test samples from different perspectives, and their combined judgment results are relatively reliable. After each sub-model training is completed, the judgment results from different rounds are summarized, and the reliability of each judgment result is determined based on the entropy value. Finally, the common decision 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. FixMatch is then trained using the expanded label training set.

[0128] The collaborative training network consists of two parts: batch aggregation of data and data filtering. Batch aggregation aims to collect the decision results of each sub-model on the test samples at different training stages, and by aggregating, the decision distribution of each sub-model on each test sample can be obtained. Subsequently, the properties of entropy are used to filter the data and extract samples with high confidence. The following is a detailed analysis of these two parts.

[0129] 1. Data aggregation in multiple batches

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

[0131] S2051: Construct a co-trained network containing four sub-models, each with a different depth and width.

[0132] In this embodiment, the collaboratively trained network uses four WideResNet sub-models with different depths and widths for joint training. These four sub-models can make decisions on test samples from different perspectives, and their joint decision results are relatively reliable. For certain salient features of SAR images, the simpler model is easier to distinguish, and these decision results should tend to be consistent throughout the training process. Here, WideResNet uses a smaller depth and width and is trained using only weakly augmented data of the labeled samples.

[0133] S2052: Input 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 highest probability value among all target prediction results is selected as its prediction label for that judgment round;

[0135] The formula is expressed as:

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

[0137] In the formula: label_X 1_n y represents the category with the highest probability in the decision result of the first sub-model in round n, i.e., the predicted label of the sample in round n; 1_n This represents the probability distribution of the class among all predicted values ​​of the sample in the nth round of the first sub-model;

[0138] The first sub-model's multi-round decision results for the same sample are represented as follows:

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

[0140] Figure 3 The diagram shows the pseudo-label selection algorithm based on a co-trained network. Four WideResNets with different depths and widths and relatively small size are used as sub-models of the co-trained network. These four models can make simple decisions on test samples from different perspectives, and their common decision results are relatively reliable.

[0141] 2. Data Filtering

[0142] The first step in data screening is entropy-based screening. For multiple rounds of decisions by a sub-model on the same sample, several different predictions may occur. When these results are scattered, it indicates that the sub-model's decision on that sample is not sufficiently certain; only when the predictions are concentrated in one category can they be considered high-confidence labels. 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 decision results of each sub-model for the same sample in multiple rounds;

[0144] The formula is expressed as:

[0145]

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

[0147] S2062: Delete samples whose entropy value is greater than the preset entropy value;

[0148] In this embodiment, a larger entropy value indicates a more uniform distribution of decision results, meaning higher uncertainty in the prediction result for that sample and lower model confidence in that sample. Conversely, if the entropy value of H(y) is very small, or even 0, it indicates that the model always considers the sample to belong to a certain category, meaning the prediction result has a high degree of confidence. In this embodiment, samples with entropy values ​​greater than 0.55 are removed from the prediction results of each sub-model.

[0149] S2063: Retain samples with consistent decision results from each sub-model to construct a sampling pool, and use the category in the decision result as the pseudo-label of the sample;

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

[0151] The formula is expressed as:

[0152]

[0153] In the formula: label_X represents the pseudo-label assigned to 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 decision results of the four sub-models are not completely consistent, the label is not valid. After collaborative screening, the number of qualified samples is significantly reduced, but the quality is greatly improved. Finally, from these retained high-quality pseudo-label samples, a certain number are randomly selected from each class as pseudo-label samples to expand the label training set of the first stage and continue training FixMatch.

[0157] Since these high-quality pseudo-samples may still contain some erroneous samples, they should not be treated as labeled samples in the early stages of training. This could mislead the model into training in the wrong direction, thus affecting the model's accuracy. In this embodiment, when the FixMatch framework reaches its training limit—that is, when a high accuracy has been achieved and further improvement is difficult—the highly reliable pseudo-samples extracted by the co-training network can be treated as labeled samples and added to the FixMatch framework for continued training. This combines the features learned by the original FixMatch model with the broader features provided by the new samples, thereby improving the model's generalization and robustness.

[0158] The proposed co-training network method adds a new loss function term to the two loss functions included in the original FixMatch. This new term is obtained by using pseudo-labeled samples selected by the co-training network to expand the labeled training set in the first stage and continue training FixMatch. During co-training, the loss function of the object recognition model includes the co-supervised loss L. x2 ;

[0159] The formula is expressed as:

[0160]

[0161] In the formula: the third term, collaborative supervised loss, also uses cross-entropy to supervise the samples, and the input data is weakly augmented pseudo-labeled samples X. sa X sa N represents pseudo-labeled samples; sa This indicates the number of pseudo-labeled samples.

[0162] Furthermore, for the training loss function of the sub-models in the co-trained network, only a simple fully supervised loss is used. This is because the number of input samples is small, and the loss of fully supervised labeled samples does not decrease significantly during the training of the sub-models, so the choice of loss function will not have a significant impact.

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

[0164] Loss=λ1L x +λ2L u +λ3L x2 ;

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

[0166] V. Experimental Instructions

[0167] Table 1 presents the comparative experimental results of the proposed method and FixMatch, using the MSTAR dataset. Five, 10, 20, 40, and 80 labeled images per class were selected as the original training data to compare the accuracy differences between the original FixMatch framework and the proposed method. The results are the average of multiple training iterations. It can be seen that the proposed method significantly outperforms FixMatch in terms of accuracy. With only 5 labeled data points, FixMatch's performance drops drastically to 30.65%, while the proposed method maintains an accuracy of 70.62%. When the number of labeled samples increases to 10 or more, the accuracy of the proposed method consistently surpasses that of FixMatch. This indicates that the proposed method outperforms FixMatch and exhibits excellent performance even with very few labeled samples.

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

[0169]

[0170] Table 2 presents the comparative experimental results of various methods. With only 5 and 10 labeled samples per class, the accuracy of the proposed method reaches 70.62% and 99.02%, respectively. When the number of labeled samples per class is large, the proposed method still outperforms SRC, K-SVD, and DCNN by approximately 6% to 20%. Triple-GANs and DCGAN are both generative adversarial network models, showing an advantage only when the number of labeled samples is large; however, their accuracy is still about 3% lower than the accuracy of the proposed method when the number of labeled samples per class is 10, even with 220 labeled samples per class. Notably, for the recent methods SCA and HDLM, no experiments were conducted on SCA with 5 labeled samples per class, while the proposed method's accuracy is only 1.75% lower than HDLM; however, when the number of labeled samples per class is 10, the proposed method's accuracy is 4.2% higher than SCA and 10.86% higher than HDLM. Since the FixMatch main framework experiences a sharp performance drop when labeled data is extremely scarce, the contour augmentation data augmentation method and collaborative training network employed in this invention can effectively compensate for this deficiency. The comparative results of the above methods further demonstrate the superiority of the method proposed in this invention.

[0171] Table 2 Comparison of Experimental Results Using Multiple Methods

[0172]

[0173] Table 3 presents the ablation experiment results to evaluate the impact of each component of the proposed method on model performance. On the MSTAR dataset, the experiment consisted of three parts: the first part was contour extraction; the second part involved adding a co-training network without using entropy to filter the decision results, referred to as the original co-training network; and the third part involved adding a co-training network and using entropy to filter the decision results, referred to as the complete co-training network. According to the ablation experiment results, contour extraction had a negative effect on the model despite its low accuracy. When only contour enhancement was used with 5 samples per label class, its accuracy was 2.03% lower than without contour extraction. However, when the number of samples per label class was 10 and 20, contour enhancement played a significant role, improving the accuracy of the original FixMatch by 3.71% and 2.75%, respectively. If the co-training network does not perform entropy filtering to remove low-confidence samples, it may mislead the model and lead to a decrease in accuracy, even when the accuracy is already high. For example, when the number of samples per label class was 20, the accuracy of the co-training network without entropy filtering was 0.77% lower than without the co-training network. Overall, the co-training network contributed slightly less to the improvement in model accuracy than contour extraction, achieving optimal results of 3.43% and 1.57% for 10 and 20 samples per labeled class, respectively. However, the co-training network played a significant role when there were only 5 samples per labeled class, achieving a 32.28% improvement compared to the original FixMatch. This ablation experiment reveals the independent contributions of the co-training network and contour extraction to the improvement in model accuracy, and the necessity of screening pseudo-labeled samples generated by the co-training network.

[0174] Table 3 Ablation Experiment Results

[0175]

[0176] Figure 4This study demonstrates the impact of different confidence thresholds on the training of the object recognition model, with accuracy representing the average value achievable by the model upon completion of training. To compare the impact of different confidence levels on the model's pseudo-label decision results, the model's decision results on unlabeled data during the third round of training were compared with the actual correct labels. Experimental results show that when the confidence threshold τ is low, the model can involve almost all data in the unlabeled cross-entropy calculation in the initial stage. At a confidence threshold of 0.3, 2402 weakly augmented samples out of 2747 images had a highest predicted probability exceeding 0.3, but only 823 of these were actually correctly labeled, while 1579 were incorrectly labeled. This resulted in the model failing to be effectively trained, with a final accuracy of only 20.41%. When the confidence thresholds were set to 0.6, 0.9, and 0.95, respectively, the model's ability to identify sample categories significantly improved. In the third round of training, the ratio of correct labels to incorrect labels among the pseudo-labels meeting the confidence threshold increased linearly, reaching 2.836, 4.299, and 5.152, respectively. Correspondingly, the model accuracy increased from 88.33% to 94.10%, peaking at 99.02% with a confidence threshold of 0.95. However, further increasing the confidence threshold led to a sharp decrease in the number of samples meeting the criteria. When the confidence threshold was raised to 0.98, only 71 samples met the threshold in the third training round, of which 42 were correctly labeled and 29 were incorrectly labeled, a close ratio. Furthermore, in subsequent rounds, the number of samples meeting the criteria remained around 100, ultimately resulting in an accuracy of only 24.58%. In conclusion, the model performs best with a confidence threshold of 0.95, while thresholds that are too high or too low significantly reduce the model's capabilities.

[0177] Tables 4 and 5 respectively give N x =10 and N x When N = 20, sa The impact on accuracy. Experimental results show that N x Adding 5, 10, and 20 images from the co-training network to the labeled samples for training all improved the original model's performance. Adding 5 images achieved 96.84% accuracy, but accuracy decreased with 10 and 15 images because the probability of drawing incorrect pseudo-labeled samples increased. Adding 20 images significantly increased the number of correctly labeled samples, further improving accuracy, but it was still less accurate than the accuracy obtained with only 5 images. x The overall trend and N when =20 x When the accuracy is 10, the results are generally consistent, but the number of pseudo-label samples extracted from the co-training network is consistently less than the number of samples in the original labeled training set, so there is no sharp drop in accuracy. In summary, the model achieves optimal accuracy by extracting 5 pseudo-label samples from the co-training network as labeled samples to expand the original labeled training set for further training.

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

[0179]

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

[0181]

[0182] Tables 6 and 7 respectively give N sa =5 and N sa The impact of λ3 on accuracy when the loss weight is 20. According to experimental results, the model achieves its maximum accuracy when the collaborative loss weight is set to 0.75. Too small a loss weight makes it difficult for the model to learn new feature information from the pseudo-labeled samples selected by the collaborative training network. For example, with a loss weight of 0.25, the accuracy of extracting 5 and 20 pseudo-labeled samples is not significantly different from the 93.41% accuracy obtained by using FixMatch alone. However, when the loss weight is 1, the model performance decreases slightly compared to 0.75. Therefore, the optimal loss weight value can be determined to be 0.75.

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

[0184]

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

[0186]

[0187] Example 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 Embodiment 1.

[0189] A semi-supervised SAR target recognition system based on contour enhancement and co-training includes:

[0190] Image input module, used to acquire SAR images 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] The result output module is used to output the target prediction results output by the target recognition model as the target recognition results of the SAR image to be recognized.

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

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

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

[0196] 3) Perform strong and weak augmentation on the unlabeled contour-enhanced images 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, semi-supervised learning is performed using the FixMatch framework to obtain a target recognition model based on FixMatch.

[0198] 5) Input labeled weakly augmented samples into a collaborative training network including several sub-models for training, and obtain the target prediction results for each sample at different training stages; make a decision after a specified number of rounds, and take the class with the highest probability value as the prediction label for this round; after multiple decisions, obtain the distribution of the decision results of each sub-model in the collaborative training network for any sample.

[0199] 6) Based on the distribution of the decision results of each sub-model for all samples, the entropy characteristics are used for preliminary screening; high-confidence pseudo-label samples are determined by the collaborative decision of multiple sub-models; for each type, a certain number of pseudo-label samples are randomly selected for the training of the FixMatch framework in the first stage.

[0200] 7) Supervised learning of the FixMatch-based target recognition model is performed using extracted pseudo-label samples. That is, the label training set of FixMatch is expanded and training is continued to obtain a well-trained (better performance and more robust) target recognition model.

[0201] This invention addresses the problem that SAR image data often contains small regions of interest (ROIs) and significant background noise, which can lead to models expending considerable effort learning meaningless features. It proposes a contour extraction-based data augmentation method. Based on the characteristics of SAR images, it is found that ROIs are typically concentrated in a bright area, while the SAR image background contains considerable noise. Therefore, this invention uses contour extraction to highlight the key features of the SAR image target region, helping the model to better learn crucial information. This method expands upon data augmentation techniques in SAR image preprocessing, enabling more effective handling of ROIs. By highlighting key features through contour extraction, it further guides the model to learn crucial information, thereby improving the accuracy of SAR target recognition.

[0202] This invention, based on semi-supervised learning using the FixMatch framework, addresses the significant performance degradation caused by scarce labeled samples by proposing a pseudo-label augmentation mechanism based on a collaboratively trained network. This mechanism aims to introduce discriminative information beyond the original FixMatch framework. The invention employs a collaboratively trained network to generate high-confidence pseudo-labels and selects a certain number of high-reliability pseudo-label samples as labeled data, incorporating them into the target recognition model's learning process to alleviate the label sample scarcity problem. The collaboratively trained network consists of four WideResNets with varying depths and widths, trained using weakly augmented labeled samples generated by the FixMatch framework to more comprehensively learn SAR image features. After each sub-model training, the collaboratively trained network performs collaborative decision-making on the samples and summarizes the decision results from different rounds, using entropy values ​​to evaluate the reliability of each sub-model's decision results. A certain number of pseudo-label samples are extracted from each class of high-confidence decision results, and the common decision results of the four models are combined into a sampling pool. From this pool, a certain number of high-confidence images are extracted as pseudo-label samples and added to the FixMatch framework for further training. This invention introduces a collaborative training network, enabling the target recognition model to learn a wider range of 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 solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A semi-supervised SAR target recognition method based on contour enhancement and co-training, characterized in that, The method comprises the following steps: S1: obtaining a SAR image to be identified; S2: inputting the SAR image to be identified into a trained target identification model to output a corresponding target prediction result; The processing steps during training of the target identification model are as follows: S201: obtaining labeled SAR image samples and unlabeled SAR image samples for model training; S202: performing contour extraction and enhancement processing on the unlabeled SAR image samples to obtain unlabeled contour-enhanced images; S203: respectively performing strong and weak augmentation on the unlabeled contour-enhanced images to obtain unlabeled strong augmented samples and unlabeled weak augmented samples; and performing weak augmentation on the labeled SAR image samples to obtain labeled weak augmented samples; S204: using the unlabeled strong augmented samples, the unlabeled weak augmented samples and the labeled weak augmented samples, and adopting a FixMatch framework to perform semi-supervised learning to obtain a target identification model based on FixMatch; S205: inputting the labeled weak augmented samples into a collaborative training network comprising a plurality of sub-models to train and obtain target prediction results corresponding to each sample at different training stages; after a specified number of rounds, a decision is made, and each time, the class with the maximum probability value is taken as the prediction label of the current round of decision; after multiple decisions, the decision result distribution of each sub-model in the collaborative training network for each sample is obtained; In step S205, the following steps are included: S2051: constructing a collaborative training network comprising four sub-models; the depths and widths of the four sub-models are different; S2052: inputting the labeled weak augmented samples into each sub-model to train and obtain target prediction results output by each sub-model for each sample; S2053: after each predetermined number of rounds, the target prediction results of all samples are decided: for each sample, the class with the maximum probability value in all target prediction results is selected as the prediction label of the current decision round; The formula is as follows: label_X1_n = argmax(y 1_n ) ; In the formula: label_X1_n represents the category of the maximum probability in the decision result of the first sub-model in the nth round, that is, the prediction label of the sample in the nth round; y 1_n represents the probability distribution of the category of all prediction values of the sample in the nth round of the first sub-model; The multi-round decision results of the first sub-model for the same sample are as follows: label_X1={label_X1_1,label_X1_2,…,label_X1_n}; S206: based on the decision result distribution of each sub-model for all samples, a preliminary screening is performed based on the entropy characteristics; high-confidence pseudo-label samples are determined through collaborative decision of multiple sub-models; for each type, a certain number of pseudo-label samples are randomly extracted; S207: performing supervised learning on the target identification model based on FixMatch by using the extracted pseudo-label samples, i.e., expanding the label training set of FixMatch to continue training, to obtain a trained target identification model; S3: taking the target prediction result output by the target identification model as the target category of the SAR image to be identified.

2. The semi-supervised SAR target recognition method based on contour enhancement and co-training of claim 1, wherein: In step S202, contour extraction and enhancement processing are realized through the following steps: S2011: performing contour extraction on the original SAR image to extract a target region of interest; S2012: generating a corresponding mask image based on the target region of interest; The formula of the mask image is as follows: wherein: X mask represents a mask image; p represents a pixel value at a corresponding position of the original SAR image; contour represents a target region of interest obtained by contour extraction; S2013: superimpose the mask image with the original SAR image to obtain a contour-enhanced image; The contour-enhanced image is expressed by a formula as follows: X new1 = (X mean1 X mask + X1) * 0.6; wherein: X new1 represents the contour-enhanced image; X1represents the original SAR image; X mean1 represents the image after smoothing and averaging of the original SAR image X1.

3. The semi-supervised SAR target recognition method based on contour enhancement and co-training of claim 1, wherein: In step S203, weak augmentation of the SAR image is realized by flipping and translation, and strong augmentation of the SAR image is realized by RandAugment and CTAugment strategies.

4. The method of claim 1, wherein the method is based on contour enhancement and co-training for semi-supervised SAR target recognition. In step S204, in the semi-supervised learning process, the loss function of the target recognition model contains a supervised loss L x and an unsupervised loss L u ; The formula is expressed as follows: wherein: N x represents the number of labeled weak 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; N u represents the number of unlabeled samples; represents the prediction probability of each category of the target recognition model for the unlabeled strong augmented sample U s ; y represents the category label corresponding to the labeled image; represents the category label with the maximum probability in the target prediction result of the unlabeled weak augmented sample.

5. The method of claim 1, wherein the method is based on contour enhancement and co-training for semi-supervised SAR target recognition. In step S206, the following steps are specifically included: S2061: calculate an entropy value of the sample based on the decision results of multiple rounds of the same sample by each sub-model; The formula is expressed as follows: In the formula, H(y) represents the entropy value of the sample; y i represents the multi-round decision results of the same sample by the sub-models; c is the number of categories in the decision results; S2062: delete the sample with an entropy value greater than a preset entropy value; S2063: retain the sample with consistent decision results of each sub-model to construct a sampling pool, and take the class in the decision result as a pseudo-label of the sample; The formula is expressed as follows: In the formula, label_X represents the pseudo-label given to the sample after the four sub-models are cooperatively screened. S2064: for each class, a certain number of samples are randomly extracted from the sampling pool as pseudo-label samples.

6. The method of claim 1, wherein the method is based on contour enhancement and co-training for semi-supervised SAR target recognition. In step S207, when the accuracy of the target recognition model adopting FixMatch for semi-supervised learning reaches a preset value and cannot be further improved, the FixMatch-based target recognition model is further trained by expanding the pseudo-label samples to the label training set, that is, cooperative training.

7. The semi-supervised SAR target recognition method based on contour enhancement and co-training of claim 6, wherein: In step S207, in the co-training process, the loss function of the target recognition model includes a co-supervised loss L x2 ; The formula is expressed as follows: In the formula, X sa represents a pseudo-label sample; N sa represents the number of pseudo-label samples.

8. A semi-supervised SAR target recognition system based on contour enhancement and co-training, characterized in that: The semi-supervised SAR target recognition method based on contour enhancement and cooperative training according to claim 1 is implemented, comprising: An image input module configured to obtain a SAR image to be identified; A target recognition module configured to input the SAR image to be identified into a trained target recognition model, and output a corresponding target prediction result; A result output module configured to output the target prediction result output by the target recognition model as a target recognition result of the SAR image to be identified; In the target recognition module, the processing steps of training the target recognition model are as follows: 1) obtaining labeled SAR image samples and unlabeled SAR image samples for model training; 2) performing contour extraction and enhancement processing on the unlabeled SAR image samples to obtain unlabeled contour-enhanced images; 3) respectively performing strong and weak augmentation on the unlabeled contour-enhanced images to obtain unlabeled strong augmented samples and unlabeled weak augmented samples, and performing weak augmentation on the labeled SAR image samples to obtain labeled weak augmented samples; 4) using the unlabeled strong augmented samples, the unlabeled weak augmented samples, and the labeled weak augmented samples, performing semi-supervised learning by using a FixMatch framework to obtain a FixMatch-based target recognition model; 5) inputting the labeled weak augmented samples into a cooperative training network comprising a plurality of sub-models to train and obtain target prediction results corresponding to each sample in different training stages; after a specified number of rounds, a decision is made, and each time the class with the largest probability value is taken as the prediction label of the current round of decision; after multiple decisions, the decision result distribution of each sub-model in the cooperative training network for each sample is obtained; 6) The distribution of the judgment results of all samples based on each sub-model is preliminarily screened through the characteristics of entropy; high-confidence pseudo-label samples are determined through collaborative judgment of multiple sub-models; for each type, a certain number of pseudo-label samples are randomly extracted; 7) The FixMatch-based target recognition model is supervised learning through the extracted pseudo-label samples, that is, the label training set of FixMatch is expanded to continue training, and a trained target recognition model is obtained.

Images