A radar image target recognition method based on multi-source cross-angle domain adaptation

Abstract

A radar image target recognition method based on multi-source cross-angle domain adaptation belongs to the field of radar signal processing. In view of the problem of model generalization ability decline caused by domain deviation in synthetic aperture radar target recognition, a novel multi-source consistency self-training framework (MS-CST) is proposed. Through a collaborative optimization strategy, the framework efficiently fuses the supervision knowledge from multiple labeled source domains and the internal structure information from the unlabeled target domain. The MS-CST framework contains two core mechanisms: first, a source domain weighting module based on Wasserstein distance is used to evaluate the similarity between each source domain and the target domain, and the supervision signal of the source domain is adaptively aggregated based on this; second, a consistency self-training module generates pseudo labels for high-confidence target domain samples, and forces the model to produce consistent predictions for different data augmentation views, thereby fully exploiting the value of unlabeled data and ensuring the effectiveness and superiority of the framework in dealing with complex multi-source domain scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of radar signal processing, and particularly relates to a radar image target recognition method based on multi-source cross-angle domain adaptation. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) plays an irreplaceable role in target information acquisition in the fields of reconnaissance and civil observation due to its excellent all-weather and all-time imaging capability. Among them, the automatic target recognition (ATR) technology of radar image aims to automatically interpret the target categories in the radar image, and is the core link of radar image information processing, which has always been widely concerned and promoted a large number of innovative researches.

[0003] In recent years, radar image target recognition methods based on deep learning have made significant progress. However, due to the influence of factors such as azimuth angle and sensor parameters, the data collected under different conditions have significant distribution differences, which will cause serious domain shift problem, resulting in a sharp decline in the generalization ability of the model in unknown scenes, which greatly limits the applicability of the radar image recognition system in real combat environment. Domain adaptation (DA) technology is a key way to cope with this challenge, and its goal is to migrate knowledge from the source domain with rich labels to the target domain without labels.

[0004] Early researches are mostly focused on single-source domain adaptation. However, in actual application, radar image data is often composed of data under multiple imaging conditions, and multi-source domain adaptation (MSDA) can utilize more rich perspectives and information diversity, so it has become a more practical research direction. However, MSDA also faces unique challenges: if all source domain data is simply combined, the source domain with larger distribution difference may interfere with model learning, resulting in negative transfer; at the same time, traditional methods usually focus on feature alignment between source domain and target domain, and fail to fully exploit the valuable structural information contained in the unlabeled data of the target domain. Kihyuk S et al. proposed a consistency self-training mechanism, which generates pseudo labels for high-confidence target samples and forces the model to keep consistent predictions for different enhanced views, providing a powerful idea for utilizing target domain data in a semi-supervised framework. However, existing radar image domain adaptation methods are mostly focused on multi-modal scenarios, and fail to combine the effectiveness of consistency self-training with the framework of multi-source domain adaptation to specifically cope with the domain shift problem caused by the change of pitch angle. SUMMARY

[0005] Based on the above technical background, the application constructs a multi-source consistency self-training (MS-CST) framework. First, an adaptive source domain weighting module is designed, which measures the distribution difference between the source domain and the target domain using the Wasserstein distance, and realizes adaptive source domain weighting. Second, a consistency self-training mechanism is designed, which fully excavates the target domain information through pseudo-labels and consistency regularization. Finally, a joint optimization framework is constructed to collaboratively optimize the weighted multi-source supervision loss and the target domain consistency loss. The experimental results on the public MSTAR dataset show that the method can effectively alleviate the azimuth angle domain bias problem and significantly improve the radar image target recognition performance.

[0006] The application provides a radar image target recognition method based on multi-source cross-angle domain adaptation. The solution of the application is that, under a multi-source domain adaptation framework, an adaptive source domain weighting module is used to assign appropriate weights to different source domains, and a consistency self-training mechanism is used to guide the feature alignment of the unlabeled target domain through pseudo-labels, so that the collaborative use of multi-source knowledge and target domain structure information is realized. For feature learning and domain adaptation, a network structure including a shared backbone network and multiple parallel branches is designed, and the weighted multi-source supervision loss and the target domain consistency loss are jointly optimized, so that the model learns a feature representation with high discriminability and domain invariance, and the recognition performance of the radar image in the cross-domain scene is effectively improved.

[0007] The specific steps of the application are as follows:

[0008] A radar image target recognition method based on multi-source cross-angle domain adaptation includes the following steps:

[0009] Step 1: Obtain a target domain and N s source domain radar image dataset, and perform class prediction on the radar image data based on a target domain adaptive recognition network; the target domain adaptive recognition network includes a pre-trained shared feature extractor, N s parallel Sonnet modules, and a classifier;

[0010] Step 2: Construct a source domain adaptive weighting module, and assign a contribution weight to each source domain based on the quantized distribution difference between the source domain and the target domain;

[0011] Step 3: Construct a consistency self-training module, weakly enhance and strongly enhance the target domain samples respectively, and then perform consistency constraint on the prediction values of the two views through the target domain adaptive recognition network by self-supervised learning;

[0012] Step 4: The trained model is applied to the radar image target recognition task by jointly optimizing the weighted multi-source adaptive loss calculated on the source domain and the consistency self-training loss on the target domain.

[0013] The target domain adaptation recognition network makes a class prediction on the source domain data as follows:

[0014] First, the radar image is input into the pre-trained shared feature extractor for feature transformation and classification, and then the output thereof is input into N s parallel sonnet modules, one for each source domain, for learning a feature transformation for the corresponding source domain data distribution. The output of each Sonnet module is then input into the corresponding classifier to obtain the final class prediction; the Sonnet module is an ADDneck block.

[0015] Step 2 is as follows:

[0016] First, a pre-trained feature extractor is used to extract features of all samples in all domains, and then the distance between each source domain and the target domain is calculated based on the Wasserstein distance. Subsequently, a softmax function controlled by a temperature parameter is used to convert the distance between each source domain and the target domain into a set of normalized weights.

[0017] Step 3 is as follows:

[0018] First, for each target domain sample, two enhanced views are generated: a weak enhanced view obtained by using only simple transformations, and a strong enhanced view obtained by using a combined enhancement method.

[0019] Then, the weak enhanced view is input into the target domain adaptation recognition network to obtain a probability distribution of the prediction result. Then, the class corresponding to the maximum probability is selected to generate a pseudo label, and a confidence threshold is set. Only when the maximum probability value exceeds the threshold, the pseudo label is considered as a reliable supervision signal.

[0020] Finally, a consistency constraint is performed. The consistency loss is calculated by calculating the cross-entropy between the prediction value of the strong enhanced view and the high-confidence pseudo label, so that the model is forced to produce consistent prediction results for the strong enhanced view.

[0021] The joint optimization training is as follows:

[0022] Each source domain defines an adaptive loss, which includes a classification loss and a domain alignment loss. The classification loss is a cross-entropy loss, and the domain alignment loss includes a maximum mean discrepancy loss and a discrepancy loss. The maximum mean discrepancy loss minimizes the distance between the feature distributions of the source domain and the target domain in the Sonnet module, and the discrepancy loss is used to constrain different classifiers to produce consistent prediction results for the same target sample.

[0023] Finally, combined with the pre-calculated contribution weight of each source domain, the total weighted multi-source adaptive loss is defined as the weighted sum of the adaptive loss of all source domains;

[0024] The total loss of joint optimization is as follows:

[0025] ;

[0026] wherein, is the weighted multi-source adaptive loss, is the consistency loss, is a hyperparameter.

[0027] The shared feature extractor is based on the ResNet-50 architecture; specifically, after the convolution layer and the pooling layer, 4 groups of bottleneck modules are stacked in succession, and 4 groups of high-dimensional shared features are output in turn;

[0028] After the high-dimensional shared features pass through the ADDneck block, they pass through the average pooling, the flattening layer and the dropout layer into the fully connected layer as the classifier, and the class prediction is obtained.

[0029] Based on the multi-source consistency self-training (MS-CST) framework, the Wasserstein distance domain weighting mechanism and the consistency self-supervised learning, the core problems of negative transfer and insufficient utilization of target domain information in multi-source domain adaptation are effectively solved, and an innovative solution is provided for radar image target recognition in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is the overall framework diagram of the method of the application.

[0031] Figure 2 It is the shared network structure diagram adopted in the specific embodiment of the application.

[0032] Figure 3 It is a radar image example of four kinds of targets in different observation angles in the MSTAR data set adopted in the specific embodiment of the application.

[0033] Figure 4 It is the confusion matrix diagram of the MS-CST method obtained in the specific embodiment of the application on three different target domains.

[0034] Figure 5 It is the t-SNE visualization comparison diagram of the feature distribution before and after domain adaptation obtained in the specific embodiment of the application. DETAILED DESCRIPTION

[0035] The present application mainly adopts the method of simulation experiment for verification, all steps and conclusions are verified correct on the windows 10 operating system platform through pycharm 2020.1.3. In order to facilitate the technical content of the present application for the personnel in the field to understand, the content of the present application is further described below in combination with the drawings.

[0036] Figure 1 The overall framework of the method is shown, and the collaborative process of the multi-source supervised path and the consistency self-training path is clearly presented. A radar image target recognition method based on multi-source cross-angle domain adaptation includes the following steps:

[0037] Step 1: The data set adopts MSTAR data set, and image data with pitch angles of 15°, 17° and 30° are selected respectively, and the image data of one angle is selected each time as the target domain, and the image data of the remaining two angles is selected as the multi-source domain; the target domain adaptation recognition network is used for class prediction on the source domain data; the target domain adaptation recognition network includes a pre-trained shared feature extractor, a parallel Sonnet module and a classifier;

[0038] Specifically, the radar image is input into the shared feature extractor for feature transformation and classification; the shared feature extractor is based on the widely used ResNet-50 architecture and is pre-trained on the ImageNet data set. The deep backbone network maps the input radar image to a high-dimensional feature space, capturing rich visual pattern hierarchies. After the shared feature extractor , the network is split into parallel branches equal to the number of source domains, and each source domain (i.e. labeled radar target image data set with different pitch angles) corresponds to a branch. Each branch contains a Sonnet module for learning feature transformation specific to the corresponding source domain data distribution. The output of each Sonnet module is then input to the corresponding classifier , which is a fully connected layer; and the Sonnet module is realized by the ADDneck block, which first performs feature transformation adaptation specific to the corresponding source domain on the general high-dimensional features output by the shared feature extractor, i.e. adjusts the distribution and weight of the features according to the label data of the source domain, strengthens the features strongly related to the class in the source domain, and weakens the domain-specific interference information, and the classifier maps the adapted features to the final class prediction under the source domain. Figure 2The adopted target domain adaptive recognition network structure diagram is shown, the hierarchical association and parameter transmission logic of the feature extractor, Sonnet module and classifier are clarified, and the overall hierarchical network structure of the shared feature extractor-multi-source adaptation branch-classifier is determined: taking the radar target image (size Bx128x128x3, B is the batch size) as the input, the high-dimensional shared features are obtained through the shared feature extractor, specifically, the Bx32x32x64 basic features are obtained through the Conv64(7x7) convolution layer and MaxPool(3x3) pooling layer processing, and then 4 groups of Bottleneck modules (the number is 3, 4, 6 and 3 respectively) are stacked in sequence, and the high-dimensional shared features of Bx32x32x256, Bx16x16x512, Bx8x8x1024 and Bx4x4x2048 are output in sequence; then the high-dimensional shared features enter the Sonnet module to complete the source domain specific adaptation, and output the adapted features of Bx4x4x256; then the average pooling and flattening layer are used to convert the Bx256 one-dimensional vector, and then the dropout layer is used to relieve overfitting, and then the classifier composed of the full connection layer is input, and finally the original output value of BxK (K is the number of target categories) is obtained to complete the category prediction, the whole structure embodies the process of “general feature extraction-source domain feature adaptation-category prediction”.

[0039] Step 2: Construct a source domain self-adaptive weighting module, and assign a contribution weight to each source domain based on the quantified distribution difference between the source domain and the target domain;

[0040] To effectively aggregate knowledge from multiple heterogeneous source domains and mitigate the potential risk of negative transfer, an adaptive source domain weighting mechanism is introduced. It is assumed that the source domain that is more similar to the target domain should play a more important role in the training process. To achieve this mechanism, first, the distribution difference between each source domain and the target domain is quantified. The Wasserstein distance (Bulldozer distance) is used as the measurement standard, which is known for its robustness in comparing distributions, especially when the distribution support sets do not overlap. Before the main training loop starts, a fixed, pre-trained ResNet-50 model (denoted as ) is used to extract features of all samples in all domains. Then, the distance of each source domain is calculated, and the formula is as follows:

[0041]

[0042] where, is the sample set in the jth source domain, is the sample set in the target domain. Subsequently, a temperature parameter a softmax function to control the distance of each source domain to a set of normalized weights :

[0043]

[0044] where, is the contribution weight of the jth source domain, k is the source domain index, is the distance of the kth source domain; T is a positive hyper-parameter, called temperature coefficient. When T is small, the weight distribution will be more "sharp", making the model focus more on the source domain with the closest distance; when T is large, the weight distribution is relatively smooth. In this way, each source domain is assigned a reasonable, standardized contribution weight.

[0045] Step 3: Construct a consistency self-training module to perform weak and strong augmentation on the target domain samples, and then perform consistency constraint on the prediction values of the two views through the target domain adaptive recognition network by self-supervised learning;

[0046] To make full use of the rich information contained in the unlabeled target domain data, a consistency self-training module is introduced. This module is based on the basic principle of self-supervised learning: a robust model should produce consistent prediction results for inputs with the same semantics, even under data augmentation. For each target domain sample , two augmented views are generated: weak augmented view , which only uses simple transformations such as random horizontal flipping; strong augmented view , which applies a combination of more aggressive transformations including random cropping. The self-training process includes the following three core steps:

[0047] The first step is weak augmented prediction. The model first performs weak augmentation on a sample in the target domain, such as only doing simple image horizontal flipping, to get the weak augmented view of the sample, and then inputs this weak augmented view into the network to get its corresponding prediction result, which becomes the weak prediction, and its probability distribution is denoted as where is the target domain adaptive recognition network formed by training the network on the source domain;

[0048] The second step is high-confidence pseudo-label generation. The pseudo-label is generated by selecting the class corresponding to the maximum probability: To ensure the reliability of the pseudo-label, a confidence threshold is set. Only when the maximum probability value is greater than the threshold, the pseudo-label is considered as a reliable supervision signal;

[0049] The third step is consistency constraint, the model is forced to require strong augmented views to produce consistent prediction results, consistency loss is achieved by calculating the cross entropy between the prediction value of the strong augmented view and the high-confidence pseudo label :

[0050]

[0051] where, denotes a small batch of samples from the target domain, is an indicator function used to filter out low-confidence samples, and CE is the cross-entropy loss function. This module enables the model to effectively learn from the target domain through a self-supervised learning mechanism.

[0052] Step 4: Jointly optimize the model by combining the weighted multi-source adaptive loss calculated on the source domain and the consistency self-training loss on the target domain, and apply the trained model to the radar image target recognition task.

[0053] The training of the MS-CST framework is driven by a joint optimization objective, which consists of two main components: the weighted multi-source adaptive loss calculated on the source domain, and the consistency self-training loss on the target domain. For each source domain , an adaptive loss is defined, which combines the classification objective and the domain alignment objective. First, the classification loss ensures that the model can correctly classify samples from the source domain. This is a standard cross-entropy loss:

[0054]

[0055] where, is the prediction of the model on the sample in the jth source domain dataset , and is the corresponding true label of the source domain sample . To align the distributions between the source and target domains, two difference losses are adopted: the maximum mean discrepancy (MMD) loss and the difference loss . The MMD loss minimizes the distance between the feature distributions of the source domain and the target domain in the Sonnet-specific space:

[0056]

[0057] where MMD denotes the maximum mean discrepancy calculation, and the difference loss is used to constrain different classifiers to produce consistent prediction results for the same target sample, thereby further promoting domain alignment:

[0058]

[0059] where, is the target sample in the target domain dataset , and is the softmax probability output of the target sample on the i-th branch, is the predicted probability distribution of the k-th source domain branch, is the L1 norm (Manhattan distance) to measure the difference between two predicted probability distributions. Thus, the total adaptation loss of the source domain is defined as:

[0060]

[0061] where, is a trade-off hyper-parameter that gradually increases during training to control the strength of domain alignment. Finally, combining the pre-computed domain weight , the total weighted multi-source adaptation loss is defined as the weighted sum of all source domains:

[0062]

[0063] The ultimate goal of the MS-CST framework is to minimize the joint loss from the source and target domains, the total loss is defined as:

[0064]

[0065] where, is a hyper-parameter to balance the contributions of the supervised multi-source adaptation loss and the unsupervised consistency loss. By jointly optimizing this objective, the model learns features that are both domain-invariant and class discriminative.

[0066] To verify the effectiveness of the method, a comprehensive evaluation of each module is conducted through systematic experiments. Figure 3 Radar image examples of four types of targets in the MSTAR dataset under different observation angles are shown, where Figure 3 (a) in the transmitter, Figure 3 (b) in the ground vehicle target, Figure 3 (c) in the ground target, Figure 3 (d) in the air target, which provides a data basis for the design of the cross-elevation angle domain adaptation task; in terms of performance visualization, Figure 4 ​​The MS-CST method is demonstrated on three different target domains, wherein Figure 4 (a) in FIG. 15 is a classification confusion matrix of the 15° target domain, data shows that the diagonal correct classification ratio is close to 100%, and the non-diagonal error ratio tends to be close to 0, combined with the quantitative results, it can be known that the correct classification rates of the four categories are all higher than 99.4%, there is no obvious category confusion, which indicates that the model fuses the knowledge of the 17° and 30° source domains, and accurately adapts to the feature distribution of the 15° target domain, the classification precision is high, and the category discrimination is balanced, Figure 4 (b) in FIG. 17 is a classification confusion matrix of the 17° target domain, the diagonal element ratio in the data is close to full score, and the row / column error ratio corresponding to the air defense forces tends to be close to 0, combined with the results that the precision, recall rate and F1 score are all 99.83%, it can be indicated that even if the samples are unbalanced, the classification accuracy of the minority class air defense forces does not decrease, the model has no category imbalance error, and the robustness to the minority class is excellent, Figure 4 (c) in FIG. 30 is a classification confusion matrix of the 30° target domain, the correct classification ratio in the data is higher than 99%, the non-diagonal error ratio is uniform and extremely low, combined with the classification indicators, it can be known that the correct classification rate is 99.13%, which is much higher than the baseline method, and there is no high-frequency confusion of two categories of targets, which indicates that the model captures the essential features of the target categories, and effectively avoids the negative transfer under the large domain shift; Figure 5 The t-SNE (t-distributed Stochastic Neighbour Embedding) visualization comparison of the feature distribution before and after the domain adaptation is shown, wherein Figure 5 (a) in FIG. 15 is the feature distribution of the source domain and the target domain before the domain adaptation, Figure 5 (b) in FIG. 17 is the feature distribution of the source domain and the target domain after the domain adaptation, which clearly reflects the effect of the method on aligning the cross-domain features and improving the category discrimination. The detailed statistical information table of the MSTAR dataset is shown in Table 1, which clearly shows the number of training and test samples of each target under different observation angles;

[0067] Table 1: Detailed statistical information table of the MSTAR dataset

[0068]

[0069] Table 2 shows the classification accuracy comparison of different methods on three cross-elevation angle domain adaptation tasks, the first one is that the source domain is 17°+30°, and the target domain is 15°, the second one is that the source domain is 15°+30°, and the target domain is 17°, and the third one is that the source domain is 15°+17°, and the target domain is 30°, which quantitatively verifies the performance advantage of the method;

[0070] Table 2: Classification accuracy comparison table of different methods on three cross-elevation angle domain adaptation tasks

[0071]

[0072] Table 3 gives the detailed performance index table of the MS-CST method on three cross- elevation domain adaptation tasks, further verifies the stability and reliability of the method from the precision, recall and F1 score dimensions, and fully verifies the effectiveness of the proposed method in the multi-source cross-elevation domain adaptive radar image target recognition task.

[0073] Table 3: Detailed performance index table of the proposed MS-CST method on three cross-elevation domain adaptation tasks

[0074]

[0075] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

Claims

1. A radar image target recognition method based on multi-source cross-angle domain adaptation, characterized in that, The method comprises the following steps: Step 1: Obtain a target domain and N s radar image dataset, construct a target domain adaptive recognition network to make category prediction on the radar image data; the target domain adaptive recognition network comprises a pre-trained shared feature extractor, N s parallel Sonnet modules, and a classifier; Step 2: constructing a source domain self-adaptive weighting module, based on the quantized distribution difference between the source domain and the target domain, assigning a contribution weight to each source domain; Step 3: constructing a consistency self-training module, respectively performing weak enhancement and strong enhancement on the target domain samples, and then performing consistency constraint on the prediction values of the two views through the target domain adaptive recognition network through self-supervised learning; Step 4: jointly optimizing and training through the weighted multi-source adaptive loss based on the contribution weight calculated on the source domain and the consistency self-training loss on the target domain, and applying the trained model to a radar image target recognition task; The step 2 is specifically as follows: First, a pre-trained feature extractor is used to extract the features of all samples in all domains, then the distance between each source domain and the target domain is calculated based on the Wasserstein distance, and then a softmax function controlled by a temperature parameter is used to convert the distance between each source domain and the target domain into a set of normalized weights; The step 3 is specifically as follows: First, for each target domain sample, two enhanced views are generated: a weak enhanced view obtained by only using simple transformation, and a strong enhanced view obtained by using a combined enhancement method; Then the weak enhanced view is input into the target domain adaptive recognition network to obtain the probability distribution of the prediction result; then the class corresponding to the maximum probability is selected to generate a pseudo label, and a confidence threshold is set, and only when the maximum probability value exceeds the threshold, the pseudo label is regarded as a reliable supervision signal; Finally, consistency constraint is performed; the consistency loss is calculated by calculating the cross entropy between the prediction value of the strong enhanced view and the high-confidence pseudo label, so that the model is forced to produce consistent prediction results for the strong enhanced view.

2. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 1, characterized in that, The target domain adaptive recognition network performs class prediction on the radar image data, specifically as follows: First, the radar image is input into a pre-trained shared feature extractor for feature transformation and classification, and then its output enters N s parallel sonnet modules, one for each source domain, for learning a feature transformation specific to the data distribution of the corresponding source domain; the output of each Sonnet module is then input into a classifier to obtain the final class prediction; The Sonnet module is an ADDneck block.

3. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 2, characterized in that, The joint optimization training is specifically as follows: Each source domain defines an adaptive loss, which includes a classification loss and a domain alignment loss; the classification loss is a cross-entropy loss, and the domain alignment loss includes a maximum mean difference loss and a difference loss; the maximum mean difference loss minimizes the feature distribution distance between the source domain and the target domain in the Sonnet module, and the difference loss is used to constrain the consistent prediction results of different classifiers for the same target sample; Finally, combined with the pre-calculated contribution weight of each source domain, the total weighted multi-source adaptive loss is defined as the weighted sum of the adaptive losses of all source domains; The total loss of joint optimization is as follows: ； wherein, is a weighted multi-source adaptive loss, is a consistency loss, is a hyper-parameter.

4. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 3, characterized in that, The shared feature extractor is based on a ResNet-50 architecture; specifically, it passes through a convolutional layer and a pooling layer, and then stacks 4 groups of Bottleneck modules in succession to output 4 groups of high-dimensional shared features; The high-dimensional shared features pass through the ADDneck block, and then pass through the average pooling layer, the flattening layer and the dropout layer to enter the fully connected layer as the classifier to obtain the class prediction.

5. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 4, characterized in that, The calculation formula of the consistency loss is specifically as follows: ； wherein, is a target domain sample, denotes a batch of target domain samples, is an indicator function, and CE is a cross-entropy loss function; is a probability distribution of weak augmented views, is a probability distribution of strong augmented views, 1.A radar image target recognition method based on multi-source cross-angle domain adaptation, characterized in that, The method comprises the following steps: is a pseudo label, is a confidence threshold.

Images