A SAR target recognition method based on unsupervised domain adaptation

Abstract

This invention discloses a SAR target recognition method based on unsupervised domain adaptation, comprising: inputting the SAR image to be recognized into a trained target recognition model, and outputting the corresponding target prediction result; the training steps are as follows: augmenting the source domain samples using a progressive multi-level wavelet transform data augmentation module, and calculating the wavelet transform augmented source sample loss; calculating the classification probability of the source domain samples based on the distance between each source domain sample and each target prototype using a progressive instance-prototype alignment module, and calculating the instance-prototype loss; calculating the consistency alignment loss based on single-sample consistency alignment and multi-sample relationship consistency alignment of weakly augmented target samples and strongly augmented target samples; optimizing the parameters of the target recognition model based on the wavelet transform augmented source sample loss, instance-prototype loss, and consistency alignment loss; and repeating the training until the model converges. This invention can improve the accuracy of SAR target recognition even when there are no labeled samples in the target domain.

Description

Technical Field

[0001] This invention relates to the fields of unsupervised domain adaptation and SAR target recognition, and specifically to a SAR target recognition method based on unsupervised domain adaptation. Background Technology

[0002] Compared to existing optical imaging systems, synthetic aperture radar (SAR) can capture high-resolution images under any weather conditions. Therefore, SAR has found widespread application in both the military and private sectors, demonstrating enormous potential in areas such as topographic surveying, environmental monitoring, and battlefield reconnaissance. Automatic target recognition (ATR) plays a crucial role in SAR image processing and applications, and is of great significance in the field of military intelligence.

[0003] The emergence of deep learning (DL) has brought significant progress to automatic target recognition in SAR because it can effectively extract high-level semantic features compared to traditional hand-crafted features. Deep learning-based automatic target recognition methods for SAR have achieved excellent recognition performance. However, a successful deep learning model requires a large number of labeled samples for training. Since SAR targets are typically non-cooperative, collecting sufficient labeled data is challenging. Furthermore, obtaining experimental target data through real flight tests is usually costly and time-consuming, further increasing the difficulty of acquiring labeled data. Therefore, addressing the problem of limited training samples in automatic target recognition for SAR is one of the main challenges in current research.

[0004] In recent years, a notable trend has emerged in SAR automatic target recognition research: utilizing simulated SAR data to replace training deep learning models to address the problem of insufficient experimental data. However, while this solution is theoretically attractive, modeling errors in simulations often lead to significant domain gaps between simulated and experimental data. In practical applications, this gap can cause severe performance degradation, making it difficult to apply deep learning-based ATR models trained on simulated data to experimental data. Therefore, a solution to bridge the domain gap between simulated and experimental data is urgently needed. To address this issue, some researchers have investigated techniques based on unsupervised domain adaptation (UDA). In the absence of labeled data in the target domain, UDA aims to effectively align the distributions between the simulated and experimental domains. This allows for the training of more accurate and reliable ATR models using only simulated data. For example, Shi et al. proposed a gradient-weighted adversarial learning method to align the global distribution between domains at the domain level, and also investigated a class-level alignment method based on prototype networks for fine-grained class structure alignment, ultimately achieving cross-granularity (domain-level and class-level) domain alignment.

[0005] In UDA, no data is labeled within the target domain, posing a significant challenge to ATR models. It's worth noting that current SSDA methods are based on real optical datasets, while the characteristics of SAR images differ considerably from those of optical images. Therefore, domain adaptation strategies for SAR images should be considered. Specifically, the domain shift between measured and simulated SAR images is unique. This is because simulated SAR images are typically generated by physics-based electromagnetic (EM) modeling. EM modeling is based on a detailed CAD model of the observed target, followed by image generation using the same radar parameters as the measured data. Therefore, although a domain shift exists between measured and simulated SAR data, the measured SAR data contains rich prior information closely related to the simulated SAR data. Furthermore, the domain differences between these two domains differ from those of a typical SAR image domain. Therefore, it is necessary to consider an unsupervised domain adaptation SAR target recognition method suitable for simulation-assisted modeling. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem this invention aims to solve is: how to provide a SAR target recognition method based on unsupervised domain adaptation, which achieves data domain alignment by fusing high-frequency subband components of source and target domain images based on wavelet transform using PMWTDA; then, gradually promoting source domain samples to approach the corresponding category prototypes in the target domain using AIPA, achieving data category-level alignment; finally, achieving data consistency alignment by utilizing the consistency enhancement of single sample strength and the consistency of multi-sample relationships, thereby improving the accuracy of SAR target recognition even when there are no labeled samples in the target domain.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A SAR target recognition method based on unsupervised domain adaptation includes:

[0009] S1: Acquire the SAR image to be identified;

[0010] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result;

[0011] The processing steps when training the target recognition model are as follows:

[0012] S201: Obtain source domain samples and target domain samples for model training as model input;

[0013] S202: Augment the source domain samples using a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculate the wavelet transform augmented source sample loss based on the wavelet transform augmented source samples.

[0014] S203: Perform weak augmentation and strong augmentation on the target domain samples to obtain weakly augmented target samples and strongly augmented target samples;

[0015] S204: Feature extraction is performed on weakly augmented target samples and strongly augmented target samples respectively using a feature extractor to construct feature spaces for the source domain and the target domain;

[0016] S205: Within the feature space, the asymptotic instance-prototype alignment module calculates the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype; and calculates the instance-prototype loss based on the classification probability of the source domain sample.

[0017] S206: Calculate the consistency alignment loss based on single-sample consistency alignment and multi-sample relation consistency alignment of weakly augmented target samples and strongly augmented target samples;

[0018] S207: Optimize the parameters of the target recognition model by calculating the overall loss function based on wavelet transform augmented source sample loss, instance-prototype loss and consistency alignment loss;

[0019] S208: Repeat steps S201 to S207 until the target recognition model converges;

[0020] S3: Use the target prediction results output by the target recognition model as the target category of the SAR image to be recognized.

[0021] Preferably, in step S202, the processing steps of the progressive multi-level wavelet transform data augmentation module are as follows:

[0022] S2021: Obtain source domain samples and target domain samples with high-confidence pseudo-labels; target domain samples with high-confidence pseudo-labels are the samples in the target domain whose highest confidence exceeds the confidence threshold;

[0023] S2022: Perform two-dimensional discrete wavelet decomposition on the source domain samples and the target domain samples to obtain the low-frequency subband and high-frequency subband of the source domain samples and the target domain samples, respectively.

[0024] S2023: Mix the high-frequency subbands of the source domain samples and the target domain samples to obtain a mixed high-frequency subband;

[0025] S2024: Merge the low-frequency subband and mixed high-frequency subband of the source domain sample and reconstruct it using two-dimensional discrete wavelet to obtain the wavelet transform augmented source sample.

[0026] Preferably, in step S2022, the formula for two-dimensional discrete wavelet decomposition is expressed as follows:

[0027] A s H s V s Ds =DWT2(s i );

[0028] A t H t V t D t =DWT2(t i );

[0029] In the formula: A s 、A t Represent the source domain samples s respectively i and target domain sample t i The low-frequency subband; H s V s D s Represents source domain sample s i High-frequency subband; H t V t D t Represents the target domain sample t i The high-frequency subband; DWT2(·) represents two-dimensional discrete wavelet decomposition;

[0030] The decomposition logic of the two-dimensional discrete wavelet decomposition DWT2(·) is as follows:

[0031] Two-dimensional discrete wavelet decomposition is performed on image sample I(i,j) to obtain A,H,V,D=DWT2(I);

[0032] in:

[0033]

[0034] In the formula: Indicates a high-pass filter; This represents a low-pass filter; m and n represent the translation parameters of the wavelet function in the row and column directions, respectively.

[0035] Preferably, in step S2023, the hybrid high-frequency subband is calculated using the following formula:

[0036] H m =αH s +(1-α)H t ;

[0037] V m =αV s +(1-α)V t ;

[0038] D m =αD s +(1-α)D t ;

[0039] In the formula: Hm V m D m α represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

[0040] Preferably, in step S2024, wavelet transform augmented source samples are generated using the following formula:

[0041] s i =IDWT 2(A s H m V m D m );

[0042] In the formula: s′ i IDWT2(·) represents wavelet transform augmented source samples; IDWT2(·) represents two-dimensional discrete wavelet reconstruction.

[0043] The reconstruction logic of two-dimensional discrete wavelet reconstruction is as follows:

[0044] Two-dimensional discrete wavelet reconstruction is performed on the four sub-bands A, H, V, and D to obtain I = IDWT2(A,H,V,D);

[0045] in:

[0046]

[0047] In the formula: H F L represents a high-pass filter. F This indicates a low-pass filter.

[0048] Preferably, in step S204, features of the wavelet transform augmented source samples are extracted by a feature extractor; a corresponding target prediction result is generated by a classifier based on the extracted features; and the cross-entropy loss is calculated as the wavelet transform augmented source sample loss using the target prediction result and the corresponding real label.

[0049] Preferably, in step S205, the processing steps of the asymptotic instance-prototype alignment module are as follows:

[0050] S2051: Calculate the feature embedding f(t) of the target sample with high-confidence pseudo-labels. i );

[0051] S2052: Embed f(t) based on the features of the target sample i The target prototype h is calculated for each target class (the item in the predicted vector that exceeds the confidence threshold is used as the pseudo-label, thus determining the corresponding target class). k ;

[0052] The formula is expressed as:

[0053]

[0054] In the formula: This represents the number of samples from the k-th class used in the calculation of the target prototype; Labels representing the target samples;

[0055] S2053: Calculate the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype;

[0056] The formula is expressed as:

[0057]

[0058] In the formula: y s C and C represent the label variable, label constant, and number of categories of the source domain sample, respectively.

[0059] Preferably, in step S205, the instance-prototype loss L is calculated using the following formula. pta :

[0060]

[0061] Where: N s This indicates the number of samples in the source domain.

[0062] Preferably, in step S206, the consistency alignment loss is calculated through the following steps:

[0063] S2061: Calculate the single-sample consistency loss;

[0064] The specific steps include:

[0065] 1) For unlabeled target domain samples Obtain their weak augmentation target samples respectively and enhanced target sample

[0066] 2) Extract weakly augmented target samples using a feature extractor and enhanced target sample The features are then used to generate a classifier that outputs the corresponding classification probability p based on the extracted features. w and p s ;

[0067] 3) Based on the classification probability p using the following formula w Calculate pseudo-labels

[0068] The formula is expressed as:

[0069]

[0070] 4) Calculate the single-sample consistency loss L using the following formula. pl :

[0071]

[0072] In the formula: H(:,:) represents and p s Cross-entropy; 1(max(p) w )≥σ) means that only samples with the highest probability exceeding the threshold σ are calculated;

[0073] S2062: Calculate the consistency loss of multi-sample relationships;

[0074] The specific steps include:

[0075] 1) Determine the feature sets F of strongly augmented samples and weakly augmented samples. s and F w ;

[0076] 2) Calculate the feature set F of the strongly augmented sample using Gaussian radial basis functions. s The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all strongly augmented samples. s ;

[0077] The formula for calculating the Gaussian radial basis function is expressed as follows:

[0078]

[0079] In the formula: H ij f(t) i ) and f(t) j The similarity between ) and β represents the hyperparameter;

[0080] The formula for calculating similarity matrices is expressed as:

[0081]

[0082] Where: N U This indicates the number of unlabeled samples in the target domain;

[0083] 3) Calculate the feature set F of the weakly augmented sample using Gaussian radial basis functions. w The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all weakly augmented samples. w ;

[0084] 4) Calculate the multi-sample relationship consistency loss L using the following formula. msr :

[0085]

[0086] S2063: Based on Single Sample Consistency Loss L pl Consistency loss of multiple sample relationships L msr Calculate the consistency alignment loss L:

[0087] cona

[0088] The formula is expressed as:

[0089] L cona =L pl +λ msr L msr ;

[0090] In the formula: λ msr is a weighting parameter used to balance the pseudo-label loss and the multi-sample relationship loss.

[0091] Preferably, in step S206, the overall loss function L is calculated using the following formula:

[0092] L = L wte +λ pta L pta +λ cona L cona ;

[0093] In the formula: L wte L represents the wavelet transform augmented source sample loss; pta Represents instance-prototype alignment loss; L cona λ represents the consistency alignment loss; pta and λ cona These are the weight parameters for the instance-prototype alignment and consistency alignment loss terms, respectively.

[0094] Compared with existing technologies, the SAR target recognition method based on unsupervised domain adaptation in this invention has the following advantages:

[0095] This invention proposes a novel unsupervised domain adaptation (UDA) framework that enables high-precision analog data-assisted synthetic aperture radar (ATR) target recognition (ATR) even without labeled samples in the target domain. First, this invention augments source domain samples using Progressive Multi-Level Wavelet Transform Data Augmentation (PMWTDA). PMWTDA, based on wavelet transform, fuses high-frequency sub-band components of the source and target domain images. Furthermore, PMWTDA is implemented incrementally by adding high-confidence pseudo-labeled samples from the target domain in batches, effectively reducing global domain distribution discrepancies and achieving domain alignment of data, thereby improving the accuracy of SAR target recognition. Second, this invention employs Asymptotic Instance-Prototype Alignment (AIPA) to calculate the classification probability of each source domain sample based on the distance between each target prototype and the source domain sample. AIPA is designed to progressively align source domain samples to the corresponding class prototypes in the target domain, achieving class-level alignment of data. AIPA is also progressive because the class prototypes in the measurement domain are updated incrementally, further improving the accuracy of SAR target recognition. Finally, this invention calculates the consistency alignment loss based on consistency learning of weakly augmented target samples and strongly augmented target samples. By utilizing the consistency of strong and weak augmentation of individual samples and the consistency of relationships among multiple samples, it achieves data consistency alignment, thereby improving the accuracy of SAR target recognition and enhancing the robustness and generalization of the target recognition model. In summary, the unsupervised domain-adaptive SAR target recognition framework of this invention achieves domain alignment, category-level alignment, and consistency alignment of data, thereby improving the accuracy of SAR target recognition even when there are no labeled samples in the target domain. Attached Figure Description

[0096] 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:

[0097] Figure 1 This is a logical block diagram of a SAR target recognition method based on unsupervised domain adaptation.

[0098] Figure 2 A schematic diagram of wavelet decomposition for simulated (source domain) and measured (target domain) image pairs;

[0099] Figure 3 A schematic diagram illustrating the principle of wavelet transform data augmentation;

[0100] Figure 4 This is a schematic diagram of the Asymptotic Instance-Prototype Alignment Module (AIPA).

[0101] Figure 5 Example of image pairs for the SAMPLE dataset (simulation-experiment);

[0102] Figure 6 Experimental results of various methods on the SAMPLE dataset;

[0103] Figure 7 Visualization of confusion matrices for various methods on the SAMPLE dataset: In each confusion matrix, columns represent the predicted classes and rows represent the true classes;

[0104] Figure 8 Visualization of t-SNE results for various methods on the SAMPLE dataset: the same color represents the same category, and the number represents the center of the category. Detailed Implementation

[0105] 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.

[0106] The following detailed explanation illustrates the specific implementation methods:

[0107] Example:

[0108] This embodiment discloses a SAR target recognition method based on unsupervised domain adaptation.

[0109] like Figure 1 As shown, the SAR target recognition method based on unsupervised domain adaptation includes:

[0110] S1: Acquire the SAR image to be identified;

[0111] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result;

[0112] The processing steps when training the target recognition model are as follows:

[0113] S201: Obtain source domain samples and target domain samples for model training as model input;

[0114] S202: Augment the source domain samples using a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculate the wavelet transform augmented source sample loss based on the wavelet transform augmented source samples.

[0115] In this embodiment, the wavelet transform augmented source samples are processed by a feature extractor and a classifier to obtain classification probabilities. These probabilities are then combined with the true labels of the source samples to calculate the wavelet transform augmented source sample loss. The feature extractor, using a ResNet34 network, is a tool for extracting deep features from the samples; the classifier, using a single fully connected layer (FC), is a tool for further classifying the sample features.

[0116] S203: Perform weak augmentation and strong augmentation on the target domain samples to obtain weakly augmented target samples and strongly augmented target samples;

[0117] S204: Feature extraction is performed on weakly augmented target samples and strongly augmented target samples respectively through a feature extractor to construct feature spaces of the source domain and target domain; the feature space is a vector space composed of the feature vectors output by all samples after passing through the feature extractor.

[0118] S205: Within the feature space, the asymptotic instance-prototype alignment module calculates the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype; and calculates the instance-prototype loss based on the classification probability of the source domain sample.

[0119] S206: Calculate the consistency alignment loss based on single-sample consistency alignment and multi-sample relation consistency alignment of weakly augmented target samples and strongly augmented target samples;

[0120] S207: Optimize the parameters of the target recognition model by calculating the overall loss function based on wavelet transform augmented source sample loss, instance-prototype loss and consistency alignment loss;

[0121] S208: Repeat steps S201 to S207 until the target recognition model converges;

[0122] S3: Use the target prediction results output by the target recognition model as the target category of the SAR image to be recognized.

[0123] This invention proposes a novel unsupervised domain adaptation (UDA) framework that enables high-precision analog data-assisted synthetic aperture radar (ATR) target recognition (ATR) even without labeled samples in the target domain. First, this invention augments source domain samples using Progressive Multi-Level Wavelet Transform Data Augmentation (PMWTDA). PMWTDA, based on wavelet transform, fuses high-frequency sub-band components of the source and target domain images. Furthermore, PMWTDA is implemented incrementally by adding high-confidence pseudo-labeled samples from the target domain in batches, effectively reducing global domain distribution discrepancies and achieving domain alignment of data, thereby improving the accuracy of SAR target recognition. Second, this invention employs Asymptotic Instance-Prototype Alignment (AIPA) to calculate the classification probability of each source domain sample based on the distance between each target prototype and the source domain sample. AIPA is designed to progressively align source domain samples to the corresponding class prototypes in the target domain, achieving class-level alignment of data. AIPA is also progressive because the class prototypes in the measurement domain are updated incrementally, further improving the accuracy of SAR target recognition. Finally, this invention calculates the consistency alignment loss based on consistency learning of weakly augmented target samples and strongly augmented target samples. By utilizing the consistency of strong and weak augmentation of individual samples and the consistency of relationships among multiple samples, it achieves data consistency alignment, thereby improving the accuracy of SAR target recognition and enhancing the robustness and generalization of the target recognition model. In summary, the unsupervised domain-adaptive SAR target recognition framework of this invention achieves domain alignment, category-level alignment, and consistency alignment of data, thereby improving the accuracy of SAR target recognition even when there are no labeled samples in the target domain.

[0124] Unsupervised domain adaptation (UDA) research starts from labeled source domains ( For source sample s i The mark, N s (Number of source samples) transferred to the unlabeled target domain (N t For the target image t i (Number of items). and The tags all come from the tag space. This embodiment uses the SAMPLE dataset as the experimental dataset, where the simulated dataset is... The actual test dataset is Obviously, and The data distribution is different, leading to domain shift. Utilizing... and study

[0125] The domain adaptation model is represented as The parameter is θ, where f represents the feature extractor and g represents the classifier. For example... Figure 1As shown, the UDA framework proposed in this invention considers domain alignment, class alignment, and consistency alignment, with key modules including PMWTDA, AIPA, and consistency learning. PMWTDA encourages the simulated domain distribution to approximate the measured domain distribution, aiming to achieve domain-level alignment. AIPA can make simulated domain samples approximate the measured domain prototypes of the corresponding classes, thereby achieving class alignment. Consistency learning includes single-sample strong and weak augmentation consistency and multi-sample relationship consistency.

[0126] To better illustrate the technical solution of the present invention, this embodiment is described in the following parts.

[0127] I. Progressive Multi-Level Wavelet Transform Data Augmentation Module (PMWTDA)

[0128] For a single SAR image I(i,j), its two-dimensional discrete wavelet decomposition is obtained using the Mallat algorithm.

[0129]

[0130] in, Indicates a high-pass filter; represents a low-pass filter; m and n represent the translation parameters of the wavelet function in the row and column directions, respectively. The wavelet transform decomposes the SAR image into one low-frequency subband (denoted as A) and three high-frequency subbands (denoted as H, V, and D) using equations (1), (2), (3), and (4), where the A component typically represents the image contour. Furthermore, the horizontal, vertical, and diagonal directions of image detail information are contained in H, V, and D, respectively. For image I, the two-dimensional discrete wavelet decomposition in equation (5) is represented by the symbol “DWT2(·)”:

[0131] A,H,V,D=DWT2(I) (5)

[0132] The original image can be reconstructed using the four subbands A, H, V, and D through the inverse wavelet transform in equation (6). F L represents a high-pass filter. F This represents a low-pass filter. Here, the symbol “IDWT2(·)” represents the two-dimensional discrete wavelet reconstruction in equation (7).

[0133]

[0134] I = IDWT2(A,H,V,D) (7)

[0135] Figure 2The schematic diagram shows the decomposition results of simulated and measured image pairs. The results indicate that the differences between the simulated and measured images mainly exist in the high-frequency components, not the low-frequency components. This is because the simulated image is calculated based on the target CAD model and the EM calculation method. Due to simulation errors in CAD modeling and EM, high-frequency detail information has certain limitations. To reduce the domain differences in high-frequency sub-band information and obtain good cross-domain generalization ability, this invention designs a progressive wavelet transform data augmentation method, which confuses the high-frequency sub-bands of the source image and the corresponding category's high-confidence pseudo-label target image. The high-frequency components are mixed and then combined with the original low-frequency components of the source image to generate an augmented image using IDWT2.

[0136] Combination Figure 3 As shown, the processing steps of the progressive multi-level wavelet transform data augmentation module are as follows:

[0137] S2021: S2021: Obtain source domain samples and target domain samples with high-confidence pseudo-labels t i (max(h(t i ))≥σ); The target domain sample with a high confidence pseudo-label is the unlabeled sample in the target domain whose highest confidence exceeds the confidence threshold σ;

[0138] S2022: Perform two-dimensional discrete wavelet decomposition on the source domain samples and the target domain samples to obtain the low-frequency subband and high-frequency subband of the source domain samples and the target domain samples, respectively.

[0139] The formula for two-dimensional discrete wavelet decomposition is expressed as:

[0140] A s H s V s D s =DWT2(s i (8)

[0141] A t H t V t D t =DWT2(t i (9)

[0142] In the formula: A s 、A t Represent the source domain samples s respectively i and target domain sample t i The low-frequency subband; H s V s D s Represents source domain sample s i High-frequency subband; H t V t D tRepresents the target domain sample t i The high-frequency subband; DWT2(·) represents two-dimensional discrete wavelet decomposition;

[0143] S2023: Mix the high-frequency subbands of the source domain samples and the target domain samples to obtain a mixed high-frequency subband;

[0144] The hybrid high-frequency subband is calculated using the following formula:

[0145] H m =αH s +(1-α)H t (10)

[0146] V m =αV s +(1-α)V t (11)

[0147] D m =αD s +(1-α)D t (12)

[0148] In the formula: H m V m D m α represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

[0149] S2024: Merge the low-frequency subband and mixed high-frequency subband of the source domain sample and reconstruct it using two-dimensional discrete wavelet to obtain the wavelet transform augmented source sample.

[0150] Wavelet transform augmented source samples are generated using the following formula:

[0151] s′ i =IDWT2(A s H m V m D m (13)

[0152] In the formula: s′ i IDWT2(·) represents wavelet transform augmented source samples; IDWT2(·) represents two-dimensional discrete wavelet reconstruction.

[0153] The advantage of using a progressive multi-level wavelet transform data augmentation strategy in this invention is that, as the model training progresses, more and more high-confidence pseudo-labeled target images emerge, and more pseudo-labeled target samples participate in the augmentation process, enhancing the model's generalization ability and sample diversity. Previously, progressive data augmentation under single-level wavelet transform decomposition was explored. To extend this method to multi-level decomposition, a two-level wavelet transform progressive data augmentation was designed. The first-level low-frequency components of the source and target images are further decomposed into two-level H, V, and D components, and operations similar to those in the first-level decomposition are performed.

[0154] Regarding the loss of augmented source samples: the features of the wavelet transform augmented source samples are extracted by the feature extractor; the corresponding target prediction results are generated by the classifier based on the extracted features; the cross-entropy loss is calculated as the wavelet transform augmented source sample loss using the target prediction results and the corresponding ground truth labels (the ground truth labels here are the ground truth labels of the source domain samples).

[0155] II. Asymptotic Instances - Prototype Alignment Module (AIPA)

[0156] To achieve finer-grained class-level alignment, an instance-prototype alignment method was developed. Assuming a prototype in the feature space represents a corresponding target class, the distance between each target prototype and the source sample is calculated to obtain the recognition result. It's worth noting that this is instance-prototype alignment, not prototype-prototype alignment. During model training, the instance-prototype loss is calculated using both source samples and target prototypes, with the aim of making the former closer to their respective target class prototypes.

[0157] Combination Figure 4 As shown, the processing steps of the asymptotic instance-prototype alignment module are as follows:

[0158] S2051: Calculate the target sample t with high-confidence pseudo-label using a feature extractor. i (max(h(t i The feature embedding f(t) of unlabeled samples whose highest confidence exceeds the probability threshold σ is greater than or equal to σ. i );

[0159] S2052: Embed f(t) based on the features of the target sample i The target prototype h is calculated for each target class (the item in the predicted vector that exceeds the confidence threshold is used as the pseudo-label, thus determining the corresponding target class). k ;

[0160] The formula is expressed as:

[0161]

[0162] In the formula: This represents the number of samples from the k-th class used in the calculation of the target prototype; Labels representing the target samples;

[0163] S2053: Calculate the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype;

[0164] The formula is expressed as:

[0165]

[0166] In the formula: y s C represents the label variable, label constant (representing 0 to 9 categories), and number of categories (C = 10) of the source domain sample, respectively.

[0167] Specifically, the instance-prototype loss L is calculated using the following formula. pta :

[0168]

[0169] Where: N s This indicates the number of samples in the source domain.

[0170] As the number of high-confidence pseudo-labeled target samples increases, the updated target prototype gradually approximates the true distribution. Therefore, AIPA plays an important role in class-level alignment.

[0171] III. Consistent Alignment

[0172] To further improve the model's generalization ability, strong and weak augmentation consistency learning is incorporated into the SAR ATR method. Consistency regularization is a technique that effectively utilizes unlabeled data. Essentially, weak augmentation is used to create pseudo-labels for unlabeled data, which are then combined with strong augmentations of unlabeled data to determine the cross-entropy loss. The model's robustness to input sample perturbations is enhanced by this regularization, which performs consistency alignment between the strong and weak augmented versions of unlabeled target samples. In this invention, the consistency learning process involves all unlabeled target samples. Two types of consistency are specifically considered: single-sample consistency and multi-sample relational consistency.

[0173] Specifically, the consistency alignment loss is calculated through the following steps:

[0174] S2061: Calculate the single-sample consistency loss;

[0175] The specific steps include:

[0176] 1) For unlabeled target domain samples Obtain their weak augmentation target samples respectively and enhanced target sample The RandAugment method was used to obtain the strongly augmented target samples.

[0177] 2) Extract weakly augmented target samples using a feature extractor and enhanced target sample The features are then used to generate a classifier that outputs the corresponding classification probability p based on the extracted features. w and p s ;

[0178] 3) Based on the classification probability p using the following formula w Calculate pseudo-labels

[0179] The formula is expressed as:

[0180]

[0181] 4) Calculate the single-sample consistency loss L using the following formula. pl :

[0182]

[0183] In the formula: H(:,:) represents and p s Cross-entropy; 1(max(p) w )≥σ) means that only samples with the highest probability exceeding the threshold σ are calculated;

[0184] S2062: Calculate the consistency loss of multi-sample relationships;

[0185] To leverage multi-sample relationship consistency, a multi-sample relationship consistency loss is introduced during model training.

[0186] The specific steps include:

[0187] 1) Determine the feature sets F of strongly augmented samples and weakly augmented samples. s and F w ;

[0188] 2) Calculate the feature set F of the strongly augmented sample using Gaussian radial basis functions. s The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all strongly augmented samples. s ;

[0189] The formula for calculating the Gaussian radial basis function is expressed as follows:

[0190]

[0191] In the formula: H ij f(t) i ) and f(t)j The similarity of H) ij ∈[0,1]. In the embedding space, H ij The value represents the similarity between two samples. β is a hyperparameter.

[0192] The formula for calculating similarity matrices is expressed as:

[0193]

[0194] Where: N U This indicates the number of unlabeled samples in the target domain;

[0195] Where H s H represents the relationship between strongly augmented versions of all unlabeled target data. w This indicates the relationship between weak augmented versions.

[0196] 3) Calculate the feature set F of the weakly augmented sample using Gaussian radial basis functions. w The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all weakly augmented samples. w ;

[0197] 4) Calculate the multi-sample relationship consistency loss L using the following formula. msr :

[0198]

[0199] By optimizing the loss, the model is encouraged to learn global semantic knowledge.

[0200] S2063: Based on Single Sample Consistency Loss L pl Consistency loss of multiple sample relationships L msr Calculate the consistency alignment loss L cona :

[0201] The formula is expressed as:

[0202] L cona =L pl +λ msr L msr (twenty two)

[0203] In the formula: λ msr is a weighting parameter used to balance the pseudo-label loss and the multi-sample relationship loss.

[0204] IV. Volume Loss Function

[0205] In this embodiment, the overall loss function L is calculated using the following formula:

[0206] L = L wte +λ pta Lpta +λ cona L cona (twenty three)

[0207] In the formula: L wte L represents the wavelet transform augmented source sample loss; pta Represents instance-prototype alignment loss; L cona λ represents the consistency alignment loss; pta and λ cona These are the weight parameters for the instance-prototype alignment and consistency alignment loss terms, respectively.

[0208] V. Experimental Instructions

[0209] In this embodiment, an experiment is conducted on the SAR target recognition method based on unsupervised domain adaptation of the present invention.

[0210] Figure 5 Examples of simulated and measured image pairs for ten target classes from the SAMPLE dataset are given.

[0211] This embodiment uses the SAMPLE dataset to evaluate the effectiveness of the proposed method. This dataset contains pairs of simulated and measured SAR images of ten target classes. The ten target classes (2S1, BMP2, BTR70, M1, M2, M35, M548, M60, T72, and ZSU23) are the same as those in the MSTAR dataset. For each measured image, a corresponding simulated image is generated by calculating the radar echo using EM. Then, a simulated image of the simulated MSTAR dataset is obtained using a SAR imaging algorithm with parameters of 0.3m range resolution, 128×128px image size, HH polarization, and an azimuth range of 10–80°.

[0212] Figure 6 The figure presents comparative experimental results of the proposed method and other methods on the SAMPLE dataset. As can be seen from the figure, the classification performance of the proposed method is superior to other unsupervised domain adaptation methods. This indicates that the proposed method performs significantly better in terms of domain-level and class-level alignment. However, comparative results obtained by exploring progressive data augmentation based on level 1 and level 2 wavelets (Ours-1 and Ours-2) show that the method based on level 2 wavelet transform does not demonstrate any advantage.

[0213] Figure 7 and Figure 8 The confusion matrices and t-SNE visualizations of the proposed method and other methods on the SAMPLE dataset are presented respectively. As can be seen from the figures, the proposed method significantly outperforms other methods in mitigating domain shift and class confusion. The proposed method exhibits better intra-class tightness and inter-class separation than other methods, and the learned features are more discriminative.

[0214] 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 SAR target recognition method based on unsupervised domain adaptation, characterized in that, include: S1: Acquire the SAR image to be identified; S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction result; The processing steps when training the target recognition model are as follows: S201: Obtain source domain samples and target domain samples for model training as model input; S202: Augment the source domain samples using a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculate the wavelet transform augmented source sample loss based on the wavelet transform augmented source samples. The processing steps of the progressive multi-level wavelet transform data augmentation module are as follows: S2021: Obtain source domain samples and target domain samples with high-confidence pseudo-labels; target domain samples with high-confidence pseudo-labels are the samples in the target domain whose highest confidence exceeds the confidence threshold; S2022: Perform two-dimensional discrete wavelet decomposition on the source domain samples and the target domain samples to obtain the low-frequency subband and high-frequency subband of the source domain samples and the target domain samples, respectively. S2023: Mix the high-frequency subbands of the source domain samples and the target domain samples to obtain a mixed high-frequency subband; S2024: Merge the low-frequency subband and mixed high-frequency subband of the source domain sample and reconstruct it using two-dimensional discrete wavelet to obtain wavelet transform augmented source sample; S203: Perform weak augmentation and strong augmentation on the target domain samples to obtain weakly augmented target samples and strongly augmented target samples; S204: Feature extraction is performed on weakly augmented target samples and strongly augmented target samples respectively using a feature extractor to construct feature spaces for the source domain and the target domain; S205: Within the feature space, the asymptotic instance-prototype alignment module calculates the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype; and calculates the instance-prototype loss based on the classification probability of the source domain sample. The processing steps for the asymptotic instance-prototype alignment module are as follows: S2051: Calculate the feature embedding f(t) of the target sample with high-confidence pseudo-labels. i ); S2052: Embed f(t) based on the features of the target sample i The target prototype h of each target class is calculated. k ; The formula is expressed as: In the formula: This represents the number of samples from the k-th class used in the calculation of the target prototype; Labels representing the target samples; S2053: Calculate the classification probability of the source domain sample based on the distance between each source domain sample and each target prototype; The formula is expressed as: In the formula: y s C and C represent the label variable, label constant, and number of categories of the source domain sample, respectively; S206: Calculate the consistency alignment loss based on single-sample consistency alignment and multi-sample relation consistency alignment of weakly augmented target samples and strongly augmented target samples; S207: Optimize the parameters of the target recognition model by calculating the overall loss function based on wavelet transform augmented source sample loss, instance-prototype loss and consistency alignment loss; S208: Repeat steps S201 to S207 until the target recognition model converges; S3: Use the target prediction results output by the target recognition model as the target category of the SAR image to be recognized.

2. The SAR target recognition method based on unsupervised domain adaptation as described in claim 1, characterized in that: In step S2022, the formula for two-dimensional discrete wavelet decomposition is expressed as follows: A s ,H s ,V s ,D s =DWT2(s i ); A t ,H t ,V t ,D t =DWT2(t i ); In the formula: A s 、A t Represent the source domain samples s respectively i and target domain sample t i The low-frequency subband; H s V s D s Represents source domain sample s i High-frequency subband; H t V t D t Represents the target domain sample t i The high-frequency subband; DWT2(·) represents two-dimensional discrete wavelet decomposition; The decomposition logic of the two-dimensional discrete wavelet decomposition DWT2(·) is as follows: Two-dimensional discrete wavelet decomposition is performed on image sample I(i,j) to obtain A,H,V,D=DWT2(I); in: In the formula: Indicates a high-pass filter; This represents a low-pass filter; m and n represent the translation parameters of the wavelet function in the row and column directions, respectively.

3. The SAR target recognition method based on unsupervised domain adaptation as described in claim 2, characterized in that: In step S2023, the hybrid high-frequency subband is calculated using the following formula: H m =αH s +(1-α)H t ; V m =αV s +(1-α)V t ; D m =αD s +(1-α)D t ; In the formula: H m V m D m α represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

4. The SAR target recognition method based on unsupervised domain adaptation as described in claim 3, characterized in that: In step S2024, wavelet transform augmented source samples are generated using the following formula: s′ i =IDWT 2(A s ,H m ,V m ,D m ); In the formula: s i ′ represents wavelet transform augmented source sample; IDWT2(·) represents two-dimensional discrete wavelet reconstruction; The reconstruction logic of two-dimensional discrete wavelet reconstruction is as follows: Two-dimensional discrete wavelet reconstruction is performed on the four sub-bands A, H, V, and D to obtain I = IDWT2(A,H,V,D); in: In the formula: H F L represents a high-pass filter. F This indicates a low-pass filter.

5. The SAR target recognition method based on unsupervised domain adaptation as described in claim 4, characterized in that: In step S204, features of the wavelet transform augmented source samples are extracted by a feature extractor; corresponding target prediction results are generated by a classifier based on the extracted features; and cross-entropy loss is calculated as the wavelet transform augmented source sample loss using the target prediction results and the corresponding real labels.

6. The SAR target recognition method based on unsupervised domain adaptation as described in claim 5, characterized in that: In step S205, the instance-prototype loss L is calculated using the following formula. pta : Where: N s This indicates the number of samples in the source domain.

7. The SAR target recognition method based on unsupervised domain adaptation as described in claim 1, characterized in that: In step S206, the consistency alignment loss is calculated through the following steps: S2061: Calculate the single-sample consistency loss; The specific steps include: 1) For unlabeled target domain samples Obtain their weak augmentation target samples respectively and enhanced target sample 2) Extract weakly augmented target samples using a feature extractor and enhanced target sample The features are then used to generate a classifier that outputs the corresponding classification probability p based on the extracted features. w and p s ; 3) Based on the classification probability p using the following formula w Calculate pseudo-labels The formula is expressed as: 4) Calculate the single-sample consistency loss L using the following formula. pl : In the formula: H(:,:) represents and p s Cross-entropy; 1(max(p) w )≥σ) means that only samples with the highest probability exceeding the threshold σ are calculated; S2062: Calculate the consistency loss of multi-sample relationships; The specific steps include: 1) Determine the feature sets F of strongly augmented samples and weakly augmented samples. s and F w ; 2) Calculate the feature set F of the strongly augmented sample using Gaussian radial basis functions. s The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all strongly augmented samples. s ; The formula for calculating the Gaussian radial basis function is expressed as follows: In the formula: H ij f(t) i ) and f(t) j The similarity between ) and β represents the hyperparameter; The formula for calculating similarity matrices is expressed as: Where: N U This indicates the number of unlabeled samples in the target domain; 3) Calculate the feature set F of the weakly augmented sample using Gaussian radial basis functions. w The similarity between any two feature vectors within the matrix; construct a similarity matrix H based on the similarity of features from all weakly augmented samples. w ; 4) Calculate the multi-sample relationship consistency loss L using the following formula. msr : S2063: Based on Single Sample Consistency Loss L pl Consistency loss of multiple sample relationships L msr Calculate the consistency alignment loss L cona : The formula is expressed as: L cona L pl +λ msr L msr ; In the formula: λ msr is a weighting parameter used to balance the pseudo-label loss and the multi-sample relationship loss.

8. The SAR target recognition method based on unsupervised domain adaptation as described in claim 1, characterized in that: In step S206, the overall loss function L is calculated using the following formula: L=L wte +λ pta L pta +λ cona L cona ; In the formula: L wte L represents the wavelet transform augmented source sample loss; pta Represents instance-prototype alignment loss; L cona This represents the consistency alignment loss; λ pta and λ cona These are the weight parameters for the instance-prototype alignment and consistency alignment loss terms, respectively.

Images