SAR (Synthetic Aperture Radar) target identification method based on unsupervised domain adaptation

Abstract

The invention discloses an SAR target recognition method based on unsupervised domain adaptation, and the method comprises the steps: inputting a to-be-recognized SAR image into a trained target recognition model, and outputting a corresponding target prediction result; the method comprises the following training steps: augmenting a source domain sample through a progressive multistage wavelet transform data augmentation module, and calculating the loss of the wavelet transform augmented source sample; the classification probability of the source domain samples is calculated through an asymptotic instance-prototype alignment module according to the distance between each source domain sample and each target prototype, and instance-prototype loss is calculated; calculating a consistency alignment loss based on single sample consistency alignment and multi-sample relationship consistency alignment of the weak augmented target sample and the strong augmented target sample; performing parameter optimization on the target identification model based on wavelet transform augmented source sample loss, instance-prototype loss and consistency alignment loss; and training is repeated until the model converges. According to the method, the accuracy of SAR target recognition can be improved under the condition that no labeled sample exists in the target domain.

Description

Technical Field

[0001] The present invention relates to the fields of unsupervised domain adaptation and SAR target recognition, and in particular to a SAR target recognition method based on unsupervised domain adaptation. Background Art

[0002] Compared with existing optical imaging systems, synthetic aperture radar (SAR) can capture high-resolution images under any weather conditions. Therefore, SAR has been widely used in both the military and private sectors. It has shown great potential in the fields of topographic surveying, environmental monitoring, battlefield reconnaissance, etc. Among them, automatic target recognition (ATR) occupies an important position in SAR image processing and application, and is of great significance in the field of military intelligence.

[0003] The emergence of deep learning (DL) has brought significant progress to SAR automatic target recognition, because deep learning can effectively extract high-level semantic features compared to traditional hand-crafted features. SAR automatic target recognition methods based on deep learning have achieved excellent recognition performance. However, a successful deep learning model requires a large number of labeled samples for training. Since SAR targets are usually non-cooperative, it is challenging to collect enough labeled data. In addition, obtaining experimental target data through real flight tests is usually economically costly and time-consuming, which further increases the difficulty of obtaining labeled data. Therefore, how to solve the problem of limited training samples faced by SAR automatic target recognition is one of the main challenges of current research.

[0004] In recent years, a striking trend has emerged in the research of SAR automatic target recognition, which is to use simulated SAR data instead of training deep learning models to solve the problem of insufficient measured data. However, although this solution is attractive in theory, modeling errors in simulation often lead to a significant domain gap between simulated and measured data. In practical applications, this gap may lead to severe performance degradation, making it difficult to apply deep learning-based ATR models trained on simulated data to measured data. Therefore, a solution is urgently needed to bridge the domain gap between simulated and measured data. To address this issue, some researchers have studied techniques based on unsupervised domain adaptation (UDA). In the absence of labeled data in the target domain, UDA aims to effectively align the distribution between simulated and measured domains. In this way, more accurate and reliable ATR models can be trained 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 studied a class-level alignment method based on prototype networks to perform fine class structure alignment, ultimately achieving domain alignment across granularities (domain level and class level).

[0005] In UDA, no data is labeled in the target domain, which poses a great challenge to the ATR model. It is worth noting that the current SSDA methods are built on real optical datasets, while the characteristics of SAR images are very different from those of optical images. Therefore, domain adaptation strategies for SAR images should be considered. In particular, the domain offset between measured and simulated SAR images is unique. This is because simulated SAR images are usually generated by physics-based electromagnetic (EM) modeling. EM modeling is based on a fine CAD model of the observed target, and then the image generation algorithm is performed using the same radar parameters as the measured data. Therefore, although there is a domain offset between the measured and simulated SAR data, there is rich prior information in the measured SAR data that is closely related to the simulated SAR data. Moreover, the two domains are different from the general SAR image domain in terms of domain differences. Therefore, it is necessary to consider a SAR target recognition method suitable for unsupervised domain adaptation assisted by simulated data. Summary of the invention

[0006] In view of the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is: how to provide a SAR target recognition method based on unsupervised domain adaptation, by fusing the high-frequency subband components of the source domain and target domain images on the basis of wavelet transform through PMWTDA, so as to realize domain alignment of data; then by gradually promoting the source domain samples to approach the corresponding category prototypes of the target domain through AIPA, so as to realize category-level alignment of data; finally, by utilizing the strength enhancement consistency of a single sample and the relationship consistency of multiple samples, the consistency alignment of data is realized, so as to improve the accuracy of SAR target recognition in the absence of labeled samples in the target domain.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

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

[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 object recognition model are as follows:

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

[0013] S202: Augmenting the source domain samples through a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculating the wavelet transform augmented source sample loss according to the wavelet transform augmented source samples;

[0014] S203: weakly augmenting and strongly augmenting the target domain samples to obtain weakly augmented target samples and strongly augmented target samples;

[0015] S204: extracting features of the weakly augmented target samples and the strongly augmented target samples respectively through a feature extractor, and constructing feature spaces of the source domain and the target domain;

[0016] S205: in the feature space, calculating the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype through the asymptotic instance-prototype alignment module; calculating the instance-prototype loss according to the classification probability of the source domain sample;

[0017] S206: Calculating 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;

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

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

[0020] S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

[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 samples in the target domain with the highest confidence exceeding the confidence threshold;

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

[0024] S2023: Mix the high frequency sub-bands of the source domain sample and the target domain sample to obtain a mixed high frequency sub-band;

[0025] S2024: The low-frequency subband and the mixed high-frequency subband of the source domain sample are merged and reconstructed by two-dimensional discrete wavelet to obtain a wavelet transform augmented source sample.

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

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

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

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

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

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

[0032] in:

[0033]

[0034] Where: represents 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.

[0035] Preferably, in step S2023, the mixed high frequency sub-band is calculated by 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] Where: Hm 、V m , D m represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

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

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

[0042] Where: s′ i represents the wavelet transform augmented source sample; IDWT2(·) represents the two-dimensional discrete wavelet reconstruction;

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

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

[0045] in:

[0046]

[0047] Where: H F Represents a high-pass filter; L F Represents a low-pass filter.

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

[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 i );

[0051] S2052: Embed f(t according to the feature of the target sample i ) Calculate the target prototype h for each target class (the item whose prediction vector exceeds the confidence threshold is used as the pseudo label to determine the corresponding target class) k ;

[0052] The formula is:

[0053]

[0054] Where: Indicates the number of samples of the kth class involved in calculating the target prototype; Represents the label of the target sample;

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

[0056] The formula is:

[0057]

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

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

[0060]

[0061] Where: N s Represents the number of source domain samples.

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

[0063] S2061: Calculate single sample consistency loss;

[0064] The specific steps include:

[0065] 1) For unlabeled target domain samples Get their weakly augmented target samples respectively And strongly augmented target samples

[0066] 2) Extract weakly augmented target samples through feature extractor And strongly augmented target samples The features of the classifier are then used to output the corresponding classification probabilities p based on the extracted features. w and p s ;

[0067] 3) Based on the classification probability p through the following formula w Calculating pseudo labels

[0068] The formula is:

[0069]

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

[0071]

[0072] Where: H(:,:) represents and p s The cross entropy of 1(max(p w )≥σ) means that only the samples with the maximum probability exceeding the threshold σ are calculated;

[0073] S2062: Calculate multi-sample relationship consistency loss;

[0074] The specific steps include:

[0075] 1) Determine the feature set F of strong augmented samples and weak augmented samples s and F w ;

[0076] 2) Calculate the feature set F of the strongly augmented sample through the Gaussian radial basis function s The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all strongly augmented sample features s ;

[0077] The calculation formula of Gaussian radial basis function is expressed as:

[0078]

[0079] Where: H ij represents f(t i ) and f(t j ) similarity; β represents the hyperparameter;

[0080] The calculation formula of the similarity matrix is ​​expressed as:

[0081]

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

[0083] 3) Calculate the feature set F of the weakly augmented sample through the Gaussian radial basis function w The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all weakly augmented sample features w ;

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

[0085]

[0086] S2063: Based on single sample consistency loss L pl and multi-sample relation consistency loss L msr Calculate the consistency alignment loss L:

[0087] cona

[0088] The formula is:

[0089] L cona =L pl +λ msr L msr ;

[0090] Where: msr is a weight parameter used to balance the pseudo label loss and multi-sample relationship loss.

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

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

[0093] Where: L wte represents the loss of wavelet transform augmented source samples; L pta represents the instance-prototype alignment loss; L cona represents the consistency alignment loss; λ pta and λ cona are the weight parameters of the instance-prototype alignment and consistency alignment loss terms, respectively.

[0094] Compared with the prior art, the SAR target recognition method based on unsupervised domain adaptation in the present invention has the following beneficial effects:

[0095] The present invention proposes a novel unsupervised domain adaptation (UDA) framework, which can realize high-precision analog data-assisted synthetic aperture radar automatic target recognition (ATR) without labeled samples in the target domain. First, the present invention augments the source domain samples through progressive multi-level wavelet transform data augmentation (PMWTDA). PMWTDA fuses the high-frequency subband components of the source domain and target domain images on the basis of wavelet transform, and PMWTDA is gradually realized by adding high-confidence pseudo-labeled samples of the target domain in batches, which can effectively reduce the global domain distribution gap and realize domain alignment of data, thereby improving the accuracy of SAR target recognition. Secondly, the present invention calculates the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype through asymptotic instance-prototype alignment (AIPA). The design purpose of AIPA is to realize the category-level alignment of data by gradually promoting the source domain samples to approach the corresponding category prototype of the target domain, and AIPA is also progressive, because the class prototype of the measurement domain is gradually updated, thereby further improving the accuracy of SAR target recognition. Finally, the present invention calculates the consistency alignment loss based on the consistency learning of weakly augmented target samples and strongly augmented target samples, and realizes the consistency alignment of data by utilizing the strong and weak enhancement consistency of a single sample and the consistency of multi-sample relations, thereby improving the accuracy of SAR target recognition and enhancing the robustness and generalization of the target recognition model. In summary, the unsupervised domain adaptation SAR target recognition framework of the present invention realizes domain alignment, category-level alignment and consistency alignment of data, thereby improving the accuracy of SAR target recognition in the absence of labeled samples in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

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

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

[0099] Figure 3 Schematic diagram for wavelet transform data augmentation;

[0100] Figure 4 It is the schematic diagram of the Asymptotic Instance-Prototype Alignment Module (AIPA);

[0101] Figure 5 This is an example of an image pair from the SAMPLE dataset (simulated-measured);

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

[0103] Figure 7 Visualize the confusion matrix results of various methods on the SAMPLE dataset: In each confusion matrix, the columns represent the predicted categories and the rows represent the true categories;

[0104] Figure 8 t-SNE visualization results of various methods on the SAMPLE dataset: the same color represents the same category, and the number represents the center of the category. DETAILED DESCRIPTION

[0105] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0106] The following is a further detailed description through 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 object recognition model are as follows:

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

[0114] S202: Augmenting the source domain samples through a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculating the wavelet transform augmented source sample loss according to the wavelet transform augmented source samples;

[0115] In this embodiment, the wavelet transform augmented source sample passes through the feature extractor and the classifier to obtain the classification probability, and then the wavelet transform augmented source sample loss is calculated in combination with the true label of the source sample. The feature extractor is a tool for extracting deep features from samples, using the resnet34 network; the classifier is a tool for further final classification of sample features, using a single fully connected layer FC.

[0116] S203: weakly augmenting and strongly augmenting the target domain samples to obtain weakly augmented target samples and strongly augmented target samples;

[0117] S204: extracting features from the weakly augmented target samples and the strongly augmented target samples respectively through a feature extractor, and constructing feature spaces of the source domain and the target domain; the feature space is a vector space composed of feature vectors output by all samples after passing through the feature extractor;

[0118] S205: in the feature space, calculating the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype through the asymptotic instance-prototype alignment module; calculating the instance-prototype loss according to the classification probability of the source domain sample;

[0119] S206: Calculating 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;

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

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

[0122] S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

[0123] The present invention proposes a novel unsupervised domain adaptation (UDA) framework, which can realize high-precision analog data-assisted synthetic aperture radar automatic target recognition (ATR) without labeled samples in the target domain. First, the present invention augments the source domain samples through progressive multi-level wavelet transform data augmentation (PMWTDA). PMWTDA fuses the high-frequency subband components of the source domain and target domain images on the basis of wavelet transform, and PMWTDA is gradually realized by adding high-confidence pseudo-labeled samples of the target domain in batches, which can effectively reduce the global domain distribution gap and realize domain alignment of data, thereby improving the accuracy of SAR target recognition. Secondly, the present invention calculates the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype through asymptotic instance-prototype alignment (AIPA). The design purpose of AIPA is to realize the category-level alignment of data by gradually promoting the source domain samples to approach the corresponding category prototype of the target domain, and AIPA is also progressive, because the class prototype of the measurement domain is gradually updated, thereby further improving the accuracy of SAR target recognition. Finally, the present invention calculates the consistency alignment loss based on the consistency learning of weakly augmented target samples and strongly augmented target samples, and realizes the consistency alignment of data by utilizing the strong and weak enhancement consistency of a single sample and the consistency of multi-sample relations, thereby improving the accuracy of SAR target recognition and enhancing the robustness and generalization of the target recognition model. In summary, the unsupervised domain adaptation SAR target recognition framework of the present invention realizes domain alignment, category-level alignment and consistency alignment of data, thereby improving the accuracy of SAR target recognition in the absence of labeled samples in the target domain.

[0124] Unsupervised domain adaptation (UDA) studies the ( is the source sample s i The mark, N s is the number of source samples) transferred to the unlabeled target domain (N t is the target image t i ). and The labels come from the label space This example uses the SAMPLE dataset as the experimental dataset, where the simulated dataset is The measured data set is Obviously, and The data distribution of is different, resulting in domain shift. and study

[0125] The domain adaptation model is expressed as The parameter is θ, where f represents the feature extractor and g represents the classifier. Figure 1As shown in the figure, the UDA framework proposed in the present invention takes into account domain alignment, class alignment and consistency alignment, among which the key modules include PMWTDA, AIPA and consistency learning. PMWTDA encourages the simulated domain distribution to be close to the measured domain distribution, aiming to achieve domain-level alignment. AIPA can make the simulated domain samples close to 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] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0127] 1. Progressive Multilevel Wavelet Transform Data Augmentation Module (PMWTDA)

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

[0129]

[0130] in, represents 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. Wavelet transform decomposes the SAR image into a low-frequency subband (denoted as A) and three high-frequency subbands (denoted as H, V, D) through equations (1), (2), (3), and (4), where the A component usually represents the contour of the image. In addition, the horizontal, vertical, and diagonal directions of the image detail information are contained in H, V, and D, respectively. For image I, the symbol "DWT2(·)" is used to represent the two-dimensional discrete wavelet decomposition in equation (5):

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

[0132] Through the inverse wavelet transform in equation (6), the original image can be reconstructed using the four sub-bands A, H, V, and D. F Represents a high-pass filter, L F 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 the simulated and measured image pairs. The results show that the difference between the simulated image and the measured image mainly exists in the high-frequency part, not the low-frequency part. This is because the simulated image is calculated based on the target CAD model and the computational EM method. Due to the simulation errors of CAD modeling and EM, there are certain limitations in the high-frequency detail information. In order to reduce the domain difference of high-frequency sub-band information and obtain good cross-domain generalization ability. The present invention designs a progressive wavelet transform data augmentation method to confuse the high-frequency sub-bands of the source image and the high-confidence pseudo-label target image of the corresponding category. The high-frequency part is mixed and then combined with the original low-frequency part of the source image to generate an augmented image through 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: Obtain source domain samples and target domain samples with high confidence pseudo labels t i (max(h(t i ))≥σ); The target domain samples with high confidence pseudo labels are the unlabeled samples in the target domain whose highest confidence exceeds the confidence threshold σ;

[0138] S2022: performing two-dimensional discrete wavelet decomposition on the source domain samples and the target domain samples to obtain low-frequency subbands and high-frequency subbands 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] Where: A s , A t Represent the source domain samples s i and target domain sample t i The low-frequency subband of H s 、V s , D s Represents the source domain sample s i High frequency subband; H t 、V t , D trepresents the target domain sample t i The high frequency subband of ; DWT2(·) represents two-dimensional discrete wavelet decomposition;

[0143] S2023: Mix the high frequency sub-bands of the source domain sample and the target domain sample to obtain a mixed high frequency sub-band;

[0144] The mixed high frequency sub-band is calculated by 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] Where: H m 、V m , D m represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

[0149] S2024: The low-frequency subband and the mixed high-frequency subband of the source domain sample are merged and reconstructed by two-dimensional discrete wavelet to obtain a wavelet transform augmented source sample.

[0150] The wavelet transform augmented source samples are generated by the following formula:

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

[0152] Where: s′ i represents the wavelet transform augmented source sample; IDWT2(·) represents the two-dimensional discrete wavelet reconstruction;

[0153] The advantage of the present invention using the progressive multi-level wavelet transform data augmentation strategy is that, as the model training proceeds, more and more high-confidence pseudo-label target images are generated, and more pseudo-label target samples participate in the augmentation process, which enhances the generalization and sample diversity of the model. The progressive data augmentation in the case of primary wavelet transform decomposition was explored above. In order to extend this method to multi-level decomposition, a 2-level wavelet transform progressive data augmentation was designed. The primary low-frequency components of the source image and the target image are further decomposed into 2-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 through the feature extractor; the corresponding target prediction results are generated based on the extracted features through the classifier; the cross entropy loss is calculated through the target prediction results and the corresponding true labels (the true labels here are the true labels of the source domain samples) as the wavelet transform augmented source sample loss.

[0155] 2. Asymptotic Instance-Prototype Alignment Module (AIPA)

[0156] In order to achieve more refined class-level alignment, an instance-prototype alignment method is developed. Assuming that 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 is worth mentioning that this is instance-prototype alignment, not prototype-prototype alignment. When training the model, the instance-prototype loss is calculated using the source samples and the target prototypes, with the goal of making the former closer to the 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 target samples t with high confidence pseudo labels through feature extractor i (max(h(t i ))≥σ)(unlabeled samples with the highest confidence exceeding the possibility threshold σ) feature embedding f(t i );

[0159] S2052: Embed f(t according to the feature of the target sample i ) Calculate the target prototype h for each target class (the item whose prediction vector exceeds the confidence threshold is used as the pseudo label to determine the corresponding target class) k ;

[0160] The formula is:

[0161]

[0162] Where: Indicates the number of samples of the kth class involved in calculating the target prototype; Represents the label of the target sample;

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

[0164] The formula is:

[0165]

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

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

[0168]

[0169] Where: N s Represents the number of source domain samples.

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

[0171] 3. Consistency Alignment

[0172] In order to further improve the generalization ability of the model, strong and weak augmentation consistency learning is incorporated into the SAR ATR method. Consistency regularization is a technique for effectively utilizing unlabeled data. In essence, weak augmentation is used to create pseudo-labels for unlabeled data, and these pseudo-labels are then combined with strong augmentation of the unlabeled data to determine the cross entropy loss. The robustness of the model to perturbations of the input samples is enhanced, and the regularization performs consistency alignment between the strong and weak augmented versions of the unlabeled target samples. In the present invention, the consistency learning process involves all unlabeled target samples. Two types of consistency are specifically considered, namely single-sample consistency and multi-sample relationship consistency.

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

[0174] S2061: Calculate single sample consistency loss;

[0175] The specific steps include:

[0176] 1) For unlabeled target domain samples Get their weakly augmented target samples respectively And strongly augmented target samples RandAugment is used to obtain strongly augmented target samples.

[0177] 2) Extract weakly augmented target samples through feature extractor And strongly augmented target samples The features of the classifier are then used to output the corresponding classification probabilities p based on the extracted features. w and p s ;

[0178] 3) Based on the classification probability p through the following formula w Calculating pseudo labels

[0179] The formula is:

[0180]

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

[0182]

[0183] Where: H(:,:) represents and p s The cross entropy of 1(max(p w )≥σ) means that only the samples with the maximum probability exceeding the threshold σ are calculated;

[0184] S2062: Calculate multi-sample relationship consistency loss;

[0185] In order to utilize multi-sample relation consistency, multi-sample relation consistency loss is introduced in model training.

[0186] The specific steps include:

[0187] 1) Determine the feature set F of strong augmented samples and weak augmented samples s and F w ;

[0188] 2) Calculate the feature set F of the strongly augmented sample through the Gaussian radial basis function s The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all strongly augmented sample features s ;

[0189] The calculation formula of Gaussian radial basis function is expressed as:

[0190]

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

[0192] The calculation formula of the similarity matrix is ​​expressed as:

[0193]

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

[0195] Among them, H s represents the relationship between the strongly augmented versions of all unlabeled target data, H w Represents the relationship between weakly augmented versions.

[0196] 3) Calculate the feature set F of the weakly augmented sample through the Gaussian radial basis function w The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all weakly augmented sample features w ;

[0197] 4) Calculate the multi-sample relationship consistency loss L by 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 and multi-sample relation consistency loss L msr Calculate the consistency alignment loss L cona :

[0201] The formula is:

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

[0203] Where: msr is a weight parameter used to balance the pseudo label loss and multi-sample relationship loss.

[0204] 4. Volume loss function

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

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

[0207] Where: L wte represents the loss of wavelet transform augmented source samples; L pta represents the instance-prototype alignment loss; L cona represents the consistency alignment loss; λ pta and λ cona are the weight parameters of the instance-prototype alignment and consistency alignment loss terms, respectively.

[0208] 5. Experimental Description

[0209] In this embodiment, experiments are 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 types of targets in the SAMPLE dataset are given.

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

[0212] Figure 6 The comparative experimental results of the method of the present invention and other methods on the SAMPLE dataset are given in Figure . It can be seen from the figure that the classification performance of the method proposed in the present invention is better than other unsupervised domain adaptation methods. This shows that the present invention is obviously better in domain-level and class-level alignment. However, the comparative results obtained by exploring the progressive data augmentation based on level 1 and level 2 wavelets (Ours-1 and Ours-2) respectively show that the method based on level 2 wavelet transform does not show any advantage.

[0213] Figure 7 and Figure 8 The confusion matrix and t-SNE visualization results of the method of the present invention and other methods on the SAMPLE data set are respectively given. As can be seen from the figure, the present invention performs significantly better in reducing domain shift and class confusion. The present invention has better intra-class compactness 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 solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A 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 object recognition model are as follows: S201: Obtain source domain samples and target domain samples for model training as model input; S202: Augmenting the source domain samples through a progressive multi-level wavelet transform data augmentation module to obtain wavelet transform augmented source samples; calculating the wavelet transform augmented source sample loss according to the wavelet transform augmented source samples; S203: weakly augmenting and strongly augmenting the target domain samples to obtain weakly augmented target samples and strongly augmented target samples; S204: extracting features of the weakly augmented target samples and the strongly augmented target samples respectively through a feature extractor, and constructing feature spaces of the source domain and the target domain; S205: in the feature space, calculating the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype through the asymptotic instance-prototype alignment module; calculating the instance-prototype loss according to the classification probability of the source domain sample; S206: Calculating 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; S207: Calculate the overall loss function based on wavelet transform augmented source sample loss, instance-prototype loss and consistency alignment loss to optimize the parameters of the target recognition model; S208: Repeat steps S201 to S207 until the target recognition model converges; S3: The target prediction result output by the target recognition model is used as the target category of the SAR image to be identified.

2. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 1, characterized in that: In step S202, 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 samples in the target domain with the highest confidence exceeding the confidence threshold; S2022: performing two-dimensional discrete wavelet decomposition on the source domain samples and the target domain samples to obtain low-frequency subbands and high-frequency subbands of the source domain samples and the target domain samples, respectively; S2023: Mix the high frequency sub-bands of the source domain sample and the target domain sample to obtain a mixed high frequency sub-band; S2024: The low-frequency subband and the mixed high-frequency subband of the source domain sample are merged and reconstructed by two-dimensional discrete wavelet to obtain a wavelet transform augmented source sample.

3. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 2, characterized in that: In step S2022, the formula for two-dimensional discrete wavelet decomposition is expressed as: A s ,H s ,V s ,D s =DWT 2(s i ); A t ,H t ,V t ,D t =DWT2(t i ); Where: A s , A t Represent the source domain samples s i and target domain sample t i The low-frequency subband of H s 、V s , D s Represents the source domain sample s i The high frequency subband; H t 、V t , D t represents the target domain sample t i The high frequency subband of ; DWT2(·) represents two-dimensional discrete wavelet decomposition; The decomposition logic of two-dimensional discrete wavelet decomposition DWT2(·) is as follows: Perform two-dimensional discrete wavelet decomposition on the image sample I(i,j) to obtain A,H,V,D=DWT2(I); in: Where: represents 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.

4. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 3, characterized in that: In step S2023, the mixed high frequency sub-band is calculated by 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 ; Where: H m 、V m , D m represents the mixed high-frequency subband; α represents the hyperparameter of the mixing ratio.

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

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

7. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 1, characterized in that: In step S205, the processing steps of the asymptotic instance-prototype alignment module are as follows: S2051: Calculate the feature embedding f(t i ); S2052: Embed f(t according to the feature of the target sample i ) Calculate the target prototype h for each target class (the item whose prediction vector exceeds the confidence threshold is used as the pseudo label to determine the corresponding target class) k ; The formula is: Where: Indicates the number of samples of the kth class involved in calculating the target prototype; Represents the label of the target sample; S2053: Calculate the classification probability of the source domain sample according to the distance between each source domain sample and each target prototype; The formula is: Where: y s , C represent the label variable, label constant and number of categories of source domain samples respectively.

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

9. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 1, characterized in that: In step S206, the consistency alignment loss is calculated by the following steps: S2061: Calculate single sample consistency loss; The specific steps include: 1) For unlabeled target domain samples Get their weakly augmented target samples respectively And strongly augmented target samples 2) Extract weakly augmented target samples through feature extractor And strongly augmented target samples The features of the classifier are then used to output the corresponding classification probabilities p based on the extracted features. w and p s ; 3) Based on the classification probability p through the following formula w Calculating pseudo labels The formula is: 4) Calculate the single sample consistency loss L by the following formula pl : Where: H(:,:) represents and p s The cross entropy of 1(max(p w )≥σ) means that only the samples with the maximum probability exceeding the threshold σ are calculated; S2062: Calculate multi-sample relationship consistency loss; The specific steps include: 1) Determine the feature set F of strong augmented samples and weak augmented samples s and F w ; 2) Calculate the feature set F of the strongly augmented sample through the Gaussian radial basis function s The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all strongly augmented sample features s ; The calculation formula of Gaussian radial basis function is expressed as: Where: H ij represents f(t i ) and f(t j ) similarity; β represents the hyperparameter; The calculation formula of the similarity matrix is ​​expressed as: Where: N U Represents the number of unlabeled samples in the target domain; 3) Calculate the feature set F of the weakly augmented sample through the Gaussian radial basis function w The similarity of any two feature vectors in the same sample; construct the similarity matrix H according to the similarity of all weakly augmented sample features w ; 4) Calculate the multi-sample relationship consistency loss L by the following formula: msr : S2063: Based on single sample consistency loss L pl and multi-sample relation consistency loss L msr Calculate the consistency alignment loss L cona : The formula is: L cona =L pl +λ msr L msr ; Where: msr is a weight parameter used to balance the pseudo label loss and multi-sample relationship loss.

10. The SAR target recognition method based on unsupervised domain adaptation as claimed in claim 1, characterized in that: In step S206, the overall loss function L is calculated by the following formula: L=L wte +λ pta L pta +λ cona L cona ; Where: L wte represents the loss of wavelet transform augmented source samples; L pta represents the instance-prototype alignment loss; L cona represents the consistency alignment loss; λ pta and λ cona are the weight parameters of the instance-prototype alignment and consistency alignment loss terms, respectively.

Images