Radar image target identification method based on multi-source cross-angular domain adaptation

Abstract

The invention discloses a radar image target recognition method based on multi-source cross-angular domain adaptation, and belongs to the field of radar signal processing. The invention provides a novel multi-source consistency self-training framework (MS-CST) for solving the problem that the generalization ability of a model is reduced due to domain migration in synthetic aperture radar target recognition. According to the framework, supervision knowledge from a plurality of labeled source domains and internal structure information from an unlabeled target domain are efficiently fused through a collaborative optimization strategy. The MS-CST framework comprises two core mechanisms: firstly, a source domain weighting module based on a Wasserstein distance is used for evaluating the similarity between each source domain and a target domain, and on this basis, supervision signals of the source domains are adaptively aggregated; and secondly, a consistency self-training module generates a pseudo tag for a target domain sample with high confidence and forces the model to generate consistent prediction for different data enhancement views, so that the value of non-tag data is fully mined, and the effectiveness and superiority of the framework in processing a complex multi-source domain scene are ensured.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, and specifically relates to a radar image target recognition method based on multi-source cross-angle domain adaptation. Background Technology

[0002] Synthetic Aperture Radar (SAR), with its superior all-weather, all-time imaging capabilities, plays an irreplaceable role in target information acquisition in fields such as reconnaissance and civilian observation. Among these technologies, Automatic Target Recognition (ATR) of radar images aims to automatically interpret target categories in radar images. It is a core component of radar image information processing and has received widespread attention, driving numerous innovative research efforts.

[0003] In recent years, deep learning-based radar image target recognition methods have made significant progress. However, radar imaging is affected by various factors such as azimuth angle and sensor parameters, leading to significant distribution differences in data collected under different conditions. This results in a severe domain shift problem, causing a sharp decline in the model's generalization ability in unknown scenarios, which greatly limits the applicability of radar image recognition systems in real-world environments. Domain adaptation (DA) technology is a key approach to addressing this challenge, aiming to transfer knowledge from a richly labeled source domain to an unlabeled target domain.

[0004] Early research focused primarily on single-source domain adaptation. However, in practical applications, radar image data often consists of data from various imaging conditions. Multi-Source Domain Adaptation (MSDA) can utilize richer perspectives and information diversity, thus becoming a more practical research direction. However, MSDA also faces unique challenges: simply merging all source domain data may lead to interference with model learning due to the large differences in the distribution of source domains, resulting in negative transfer. Simultaneously, traditional methods typically focus on feature alignment between the source and target domains, failing to fully exploit the valuable structural information inherent in the unlabeled data of the target domain itself. The consistency self-training mechanism proposed by Kihyuk S et al., by generating pseudo-labels for high-confidence target samples and forcing the model to maintain consistent predictions across different augmented views, provides a powerful approach for utilizing target domain data within a semi-supervised framework. However, existing radar image domain adaptation methods largely focus on multimodal scenarios, failing to combine the effectiveness of consistency self-training with the framework of multi-source domain adaptation to specifically address the domain shift problem caused by pitch angle variations. Summary of the Invention

[0005] Based on the above technical background, this invention constructs a Multi-Source Consistency Self-Training (MS-CST) framework. First, an adaptive source domain weighting module is designed, utilizing the Wasserstein distance metric to measure the distribution difference between the source and target domains, achieving adaptive source domain weighting. Second, a consistency self-training mechanism is designed, fully mining 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. Experimental results on the publicly available MSTAR dataset demonstrate that this method effectively alleviates the azimuth domain offset problem and significantly improves radar image target recognition performance.

[0006] This invention provides a radar image target recognition method based on multi-source cross-domain adaptation. The solution of this invention, within a multi-source domain adaptation framework, assigns appropriate weights to different source domains through an adaptive source domain weighting module, and combines a consistency self-training mechanism to perform pseudo-label-guided feature alignment on unlabeled target domains, achieving synergistic utilization of multi-source knowledge and target domain structural information. For feature learning and domain adaptation, a network structure including a shared backbone network and multi-source parallel branches is designed. By jointly optimizing the weighted multi-source supervision loss and the target domain consistency loss, the model learns feature representations with high discriminativeness and domain invariance, effectively improving the recognition performance of radar images in cross-domain scenarios.

[0007] The specific steps of this invention 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 the target domain and A dataset of source-domain radar images is used to predict the category of 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, A parallel Sonnet module and a classifier;

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

[0011] Step 3: Construct a consistency self-training module, perform weak and strong enhancement on the target domain samples respectively, and then constrain the consistency of the prediction values ​​of these two views through the target domain adaptation recognition network through self-supervised learning;

[0012] Step 4: Jointly optimize and train the model by using the weighted multi-source adaptive loss calculated in the source domain and the consistent self-training loss calculated in the target domain, and then apply the trained model to the radar image target recognition task.

[0013] The target domain adaptive recognition network performs category prediction on the source domain data as follows:

[0014] First, the radar image is input into a pre-trained shared feature extractor for feature transformation and classification. Then, its output is fed into... There are three parallel Sonnet modules, one for each source domain, used to learn feature transformations 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 described in detail below:

[0016] First, a pre-trained feature extractor is used to extract features from all samples in all domains. 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 transform the distance between each source domain and the target domain into a set of normalized weights.

[0017] Step 3 is described in detail below:

[0018] First, for each target domain sample, two augmented views are generated: a weak augmented view obtained by using only simple transformations, and a strong augmented view obtained by using a combination of augmentation methods.

[0019] The weakly enhanced view is then input into the target domain adaptive recognition network to obtain the probability distribution of the prediction results. Then, the category corresponding to the highest probability is selected to generate pseudo-labels, and a confidence threshold is set. Only when the highest probability value exceeds the threshold is the pseudo-label considered a reliable supervision signal.

[0020] Finally, consistency constraints are applied; the consistency loss is calculated by the cross-entropy between the strongly augmented view predictions and the high-confidence pseudo-labels, which forces the model to produce consistent predictions for the strongly augmented views.

[0021] The joint optimization training is specifically described as follows:

[0022] Each source domain defines an adaptive loss, which includes a classification loss and a domain alignment loss; the classification loss is the cross-entropy loss, and the domain alignment loss includes the maximum mean difference loss and the 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 different classifiers to produce consistent prediction results for the same target sample;

[0023] Finally, combining the pre-calculated contribution weights 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;

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

[0025] ;

[0026] in, For weighted multi-source adaptive loss, For consistency loss, It is a hyperparameter.

[0027] The shared feature extractor is based on the ResNet-50 architecture; specifically, it goes through convolutional layers and pooling layers, and then stacks 4 sets of Bottleneck modules in succession to output 4 sets of high-dimensional shared features in sequence.

[0028] The high-dimensional shared features, after passing through the ADDneck block, then through average pooling, flattening layer, and dropout layer, enter the fully connected layer as a classifier to obtain the class prediction.

[0029] Based on the Multi-Source Consistent Self-Training (MS-CST) framework, Wasserstein distance domain weighting mechanism, and consistent self-supervised learning, this invention effectively solves the core problems of negative transfer and insufficient utilization of target domain information in multi-source domain adaptation, and provides an innovative solution for radar image target recognition in complex scenarios. Attached Figure Description

[0030] Figure 1 This is an overall framework diagram of the method of the present invention.

[0031] Figure 2 This is a diagram of the shared network structure used in a specific embodiment of the present invention.

[0032] Figure 3 Examples of radar images of four targets at different observation angles from the MSTAR dataset used in specific embodiments of the present invention.

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

[0034] Figure 5 This is a t-SNE visualization comparison chart of the feature distribution before and after domain adaptation obtained in a specific embodiment of the present invention. Detailed Implementation

[0035] This invention is primarily verified using simulation experiments. All steps and conclusions were verified correctly on the Windows 10 operating system platform using PyCharm 2020.1.3. To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with accompanying drawings, further elaborates on the invention.

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

[0037] Step 1: The dataset used is the MSTAR dataset. Image data with elevation angles of 15°, 17°, and 30° are selected respectively. One angle is selected as the target domain, and the other two angles are used as the multi-source domain. A target domain adaptation recognition network is used to predict the category of 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 a shared feature extractor for feature transformation and classification; the shared feature extractor... Based on the widely applicable ResNet-50 architecture and pre-trained on the ImageNet dataset, this deep backbone network maps input radar images to a high-dimensional feature space, capturing rich hierarchical visual patterns. A shared feature extractor... Then, the network splits into a number equal to the number of source domains. There are three parallel branches, with one branch corresponding to each source domain (i.e., a labeled dataset of radar target images at different elevation angles). Each branch contains a Sonnet module. This is used to learn feature transformations for the corresponding source domain data distribution. The output of each Sonnet module is then fed into the corresponding classifier. The classifier is a fully connected layer; the Sonnet module is implemented through the ADDneck block, which first performs a feature transformation adaptation on the general high-dimensional features output by the shared feature extractor, that is, it adjusts the distribution and weight of features according to the label data of the source domain, strengthens the features that are strongly correlated with the category in the source domain, and weakens the domain-specific interference information. The classifier then maps the adapted features to the final category prediction under the source domain. Figure 2The target domain adaptation recognition network structure diagram is shown, clarifying the hierarchical relationship and parameter passing logic between the feature extractor, Sonnet module, and classifier. The overall structure is a hierarchical network of shared feature extractor - multi-source adaptation branch - classifier: Taking a radar target image (size B×128×128×3, where B is the batch size) as input, the shared feature extractor obtains high-dimensional shared features. Specifically, these features are first processed by a Conv64 (7×7) convolutional layer and a MaxPool (3×3) pooling layer to obtain basic features of B×32×32×64, and then four sets of Bottleneck modules (3 and 4 respectively) are stacked consecutively. The algorithm sequentially outputs high-dimensional shared features of B×32×32×256, B×16×16×512, B×8×8×1024, and B×4×4×2048. These high-dimensional shared features then enter the Sonnet module to complete source domain-specific adaptation, outputting adapted features of B×4×4×256. After average pooling and flattening, these features are transformed into a one-dimensional vector of B×256. After overfitting is alleviated by a dropout layer, the vector is input into a classifier composed of fully connected layers, finally obtaining the original output value of B×K (K being the number of target categories) to complete category prediction. The entire structure embodies the process of "general feature extraction - source domain feature adaptation - classification prediction".

[0039] Step 2: Construct a source domain adaptive weighting module, and assign contribution weights to each source domain based on the quantized distribution differences between the source and target domains;

[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 source domains more similar to the target domain should play a more significant role during training. To implement this mechanism, each source domain is first quantized. With the target domain The distribution differences between them. The Wasserstein distance (bulldozer distance) is used. As a metric, this distance is known for its robustness in comparing distributions, especially when the support sets of the distributions do not overlap. Before the main training loop begins, a fixed, pre-trained ResNet-50 model (denoted as ) is used. This is used to extract features from all samples across all domains. Then, the distance to each source domain is calculated. The formula is as follows:

[0041]

[0042] in, It is the set of samples in the j-th source domain. It is the sample set in the target domain. Then, it is processed by a parameter of temperature. The controlled softmax function transforms the distance of each source domain into a set of normalized weights. :

[0043]

[0044] in, Let k be the contribution weight of the j-th source domain, and k be the source domain index. Let be the distance to the k-th source domain; T is a positive hyperparameter called the temperature coefficient. When T is small, the weight distribution is more "sharp," making the model more focused on the nearest source domain; when T is large, the weight distribution is relatively smooth. In this way, a reasonable and standardized contribution weight is assigned to each source domain.

[0045] Step 3: Construct a consistency self-training module, perform weak and strong enhancement on the target domain samples respectively, and then constrain the consistency of the prediction values ​​of these two views through the target domain adaptation recognition network through self-supervised learning;

[0046] To fully utilize the rich information contained in unlabeled target domain data, a consistency self-training module is introduced. This module is based on the fundamental principle of self-supervised learning: a robust model should produce consistent predictions for semantically identical inputs, maintaining this consistency even under data augmentation conditions. For each target domain sample... This generates two types of enhanced views: weakly enhanced views. Only simple transformations such as random horizontal flipping are used; enhanced view It applies combined enhancement methods that include more radical transformations such as random pruning. The self-training process consists of the following three core steps:

[0047] The first step is weak augmentation prediction. The model first performs a weak augmentation operation on a sample in the target domain, such as simply flipping the image horizontally, to obtain a weakly augmented view of that sample. Then this weakly enhanced view The input is fed into the network, and the corresponding prediction result is obtained, which is called a weak prediction. Its probability distribution is denoted as... , here This refers to a target domain adaptive recognition network formed by a network that has been trained in the source domain;

[0048] The second step is to generate high-confidence pseudo-labels, which involves selecting the category with the highest probability to generate pseudo-labels. To ensure the reliability of fake labels, a confidence threshold is set. Only when the maximum probability value Only when the threshold is exceeded will the pseudo-label be considered a reliable monitoring signal;

[0049] The third step is consistency constraint, which forces the model to conform to strongly augmented views. To produce consistent prediction results, consistency loss Calculate the predicted value of the enhanced view High-confidence pseudo-labels This is achieved through cross-entropy:

[0050]

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

[0052] Step 4: Jointly optimize and train the model by using the weighted multi-source adaptive loss calculated in the source domain and the consistent self-training loss calculated in the target domain, and then 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 comprises two main components: a weighted multi-source adaptive loss computed on the source domain, and a consistent self-training loss computed on the target domain. For each source domain... Define an adaptive loss This loss combines the classification objective and the domain alignment objective. First, the classification loss... Ensure the model can correctly classify samples from the source domain. This is a standard cross-entropy loss:

[0054]

[0055] in, For the j-th source domain dataset Samples in Take the expected value. Source domain sample The corresponding ground truth labels. To align the distributions between the source and target domains, two difference losses were employed: Maximum Mean Difference (MMD) loss. and difference loss The MMD loss minimizes the source domain. Distance between the target domain and the feature distribution in a specific Sonnet space:

[0056]

[0057] Where MMD stands for Maximum Mean Difference, 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] in, For target domain dataset Samples in Take the expected value. It is the target sample In the The softmax probability output on each branch It is the predicted probability distribution corresponding to the k-th source domain branch. yes The Manhattan distance (norm) measures the difference between two predicted probability distributions. From this, we obtain the source domain. The total adaptation loss is defined as:

[0060]

[0061] in, It is a trade-off hyperparameter that is gradually increased during training to control the strength of domain alignment. Finally, it is combined with pre-computed domain weights. The total weighted multi-source adaptive 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, and the total loss. Defined as:

[0064]

[0065] in, This is a hyperparameter used to balance the contributions of supervised multi-source adaptive loss and 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 proposed in this invention, a comprehensive evaluation of each module was conducted through system experiments. Figure 3 This document presents examples of radar images of four targets from the MSTAR dataset at different observation angles. Figure 3 (a) the transmitter, Figure 3 (b) in the image represents a ground vehicle target. Figure 3 (c) in the diagram represents a ground target. Figure 3 (d) in the diagram represents aerial targets, providing a data foundation for designing missions that adapt across the pitch domain; regarding performance visualization, Figure 4The confusion matrix diagrams of the MS-CST method on three different target domains are shown. Figure 4 (a) in the table is the classification confusion matrix for the 15° target domain. The data shows that the correct classification rate for diagonal targets is nearly 100%, while the incorrect classification rate for off-diagonal targets is close to 0. Combined with the quantitative results, it can be seen that the correct classification rate for all four target classes exceeds 99.4%, with no obvious class confusion. This indicates that after the model integrates knowledge from the 17° and 30° source domains, it accurately adapts to the feature distribution of the 15° target domain, achieving high classification accuracy and balanced class discrimination. Figure 4 (b) in the table is the classification confusion matrix for the 17° target domain. The diagonal elements in the data have a near-perfect percentage, and the row / column error rate for air defense units is still close to 0. Combined with the results that the precision, recall, and F1 score are all 99.83%, it shows that even with imbalanced samples, the classification accuracy of the niche air defense unit category did not decrease, the model has no class imbalance error, and it has excellent robustness to niche categories. Figure 4 (c) is the classification confusion matrix of the 30° target domain. The correct percentage of the diagonal is over 99%, while the error percentage of the off-diagonal is uniform and extremely low. Combined with the classification index, the correct classification rate is 99.13%, which is much higher than the baseline method. Moreover, there is no high-frequency confusion between two types of targets, indicating that the model has captured the essential characteristics of the target category and effectively avoided the negative transfer under large domain shift. Figure 5 A visual comparison is made using t-SNE (t-distributed random neighborhood embedding) of feature distributions before and after domain adaptation, where... Figure 5 In this context, (a) represents the characteristic distribution of the source and target domains before domain adaptation. Figure 5 (b) shows the feature distribution after adaptation of the source and target domains, clearly reflecting the effect of the method in aligning cross-domain features and improving class discriminability. In the dataset and quantitative evaluation section, Table 1 provides detailed statistical information for the MSTAR dataset, clarifying the number of training and testing samples for each type of target under different observation angles.

[0067] Table 1: Detailed Statistical Information of the MSTAR Dataset

[0068] type 2S1 (transmitter) BRDM2 (Ground Vehicle Target) T72a64 (Other ground targets) ZSU234 (Aerial Target) 15° 274 / 137 274 / 137 274 / 137 274 / 137 17° 299 / 150 298 / 149 299 / 151 299 / 80 30° 288 / 144 287 / 144 288 / 144 288 / 144

[0069] Table 2 compares the classification accuracy of different methods on three cross-pitch angle domain adaptation tasks. The first method is with a source domain of 17°+30° and a target domain of 15°. The second method is with a source domain of 15°+30° and a target domain of 17°. The third method is with a source domain of 15°+17° and a target domain of 30°. This quantitatively verifies the performance advantages of the method of the present invention.

[0070] Table 2: Comparison of classification accuracy of different methods on three cross-pitch angle domain adaptation tasks

[0071] method 17，30→15 15，30→17 15，17→30 Average accuracy ERM (Experience Risk Minimization) 97.99 98.47 94.27 96.91 AconvNet (Adaptive Convolutional Network) 94.53 95.21 95.49 95.08 DANN (Domain Adversarial Neural Network) 97.32 97.63 96.87 97.27 MFSAN (Multi-Source Feature Alignment Network) 93.00 97.00 93.00 94.33 This method 99.45 99.83 99.15 99.48

[0072] Table 3 presents detailed performance metrics of the MS-CST method on three cross-elevation domain adaptive tasks. The stability and reliability of the method are further confirmed from the dimensions of precision, recall, and F1 score, fully verifying the effectiveness of the proposed method in multi-source cross-elevation domain adaptive radar image target recognition tasks.

[0073] Table 3: Detailed performance indicators of the proposed MS-CST method on three cross-pitch domain adaptation tasks.

[0074] Source domain → Target domain accuracy Recall rate F1 score 17，30→15 99.46 99.45 99.45 15，30→17 99.83 99.83 99.83 15，17→30 99.15 99.13 99.13

[0075] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

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 a source domain radar image dataset, construct a target domain adaptive recognition network to perform category prediction on the radar image data; the target domain adaptive recognition network includes a pre-trained shared feature extractor, a plurality of parallel Sonnet modules, a classifier; 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; Step 3: Construct a consistency self-training module, and then perform weak enhancement and strong enhancement 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 through self-supervised learning; Step 4: Jointly optimize and train 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 apply the trained model to a radar image target recognition task.

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 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 a parallel sonnet module, each source domain corresponding to a Sonnet module, for learning a feature transformation for the corresponding source domain data distribution; 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, Step 2 is specifically as follows: First, a pre-trained feature extractor is used to extract 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.

4. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 3, characterized in that, Step 3 is specifically as follows: First, for each target domain sample, two enhanced views are generated: a weak enhanced view obtained by using only a 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.

5. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 4, 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 different classifiers to produce consistent prediction results 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.

6. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 5, 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; After the high-dimensional shared features pass through the ADDneck block, they pass through an average pooling layer, a flattening layer and a dropout layer into a fully connected layer as a classifier to obtain class prediction.

7. The radar image target recognition method based on multi-source cross-angle domain adaptation according to claim 6, 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, is a pseudo label, is a confidence threshold.

Images