A small sample SAR target recognition method based on key feature capture and discrimination

Abstract

The present invention discloses a small-sample SAR target recognition method based on key feature capture and discrimination. The method receives the entire SAR target image through a global auxiliary branch structure for initial recognition and provides initial recognition loss. The SAR target image is then segmented using a feature capture module to obtain key image region slices for current recognition. A local enhancement branch structure is then constructed to recognize the segmented key image region slices again and calculate the loss. Finally, the feature recognition module collaborates with the global and local branches to improve the separability of local features of similar pairs of samples between classes and the compactness of local features of dissimilar pairs of samples within a class in the global features, thereby completing the final target recognition. The method of the present invention searches and captures the key areas of the target image through the collaborative global and local branch structures, allowing the recognition model to focus on the effective discrimination area for learning and training, thereby improving the effectiveness and discrimination ability of target image extraction features and the accuracy and robustness of SAR target recognition under small sample conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and in particular relates to a small sample SAR target recognition method based on key feature capture and discrimination. Background Art

[0002] As a widely used active remote sensing technology, synthetic aperture radar (SAR) provides high-resolution, weather-resistant imagery for a variety of civilian applications. SAR target recognition is a key application, and numerous methods have been proposed. Influenced by the rapid development of deep learning, many researchers are leveraging it to address the bottlenecks of SAR target recognition.

[0003] The paper “S.Chen, H.Wang, Sar target recognition based on deep learning, in: 2014 International Conference on Data Science and Advanced Analytics (DSAA), IEEE, 2014, pp.541–547.” introduces the application of convolutional neural networks (CNN) in SAR target recognition and evaluates its performance on the benchmark Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset. The paper “S.A.Wagner, Sar atr by a combination of convolutional neural network and support vector machines, IEEE transactions on Aerospace and Electronic Systems 52(6)(2016)2861–2872.” proposes using CNN to extract feature vectors, which are then input into a support vector machine for classification.

[0004] Currently, most methods focus on improving recognition performance by training the network, assuming sufficient available data. However, when data is severely limited, that is, under small sample conditions, existing methods usually use the entire target image as a whole input to train the recognition network model. The deep learning model may not be able to capture the key areas of the image from the whole, but instead focus on useless or even harmful areas for recognition. Limited label information means that the model cannot obtain sufficient information to determine the focus position of the entire SAR image and extract discriminative features, resulting in poor recognition performance. Therefore, under limited training SAR samples, target recognition methods based on the whole image are not optimal because the extracted whole features contain a large number of useless or even harmful local image area features, rather than discriminative features of key image areas. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a small sample SAR target recognition method based on key feature capture and discrimination.

[0006] The technical solution of the present invention is a small sample SAR target recognition method based on key feature capture and discrimination, and the specific steps are as follows:

[0007] Step 1: Construct a global auxiliary branch structure to receive the entire SAR target image as input for initial recognition and calculate the initial recognition loss;

[0008] Step 2: Use the feature capture module to segment the SAR target image and obtain the key image area slices for current recognition;

[0009] Step 3: Construct a local enhancement branch structure, identify the key image area slices after segmentation again and calculate the loss;

[0010] Step 4: Use the feature recognition module to collaborate with the global and local branches to improve the separability of local features of similar sample pairs between classes and the compactness of local features of dissimilar sample pairs within a class in the global features, and complete the final target recognition.

[0011] Beneficial effects of the present invention: The method of the present invention receives the entire SAR target image through a global auxiliary branch structure for initial recognition and provides initial recognition loss, then uses a feature capture module to segment the SAR target image, obtains key image area slices for current recognition, and then constructs a local enhancement branch structure to recognize the segmented key image area slices again and calculate the loss. Finally, the feature recognition module collaborates with the global and local branches to improve the separability of local features of similar sample pairs between classes and the compactness of local features of dissimilar sample pairs within classes in the global features, thereby completing the final target recognition. The method of the present invention searches and captures the key areas of the target image through the collaborative global and local branch structures, so that the recognition model focuses on the effective discrimination area for learning and training, thereby improving the effectiveness of target image feature extraction, discrimination ability, and SAR target recognition accuracy and robustness under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flowchart of a small sample SAR target recognition method based on key feature capture and discrimination.

[0013] Figure 2 It is an overall framework diagram in an embodiment of the present invention.

[0014] Figure 3 Schematic diagram of a feature recognition module in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described below with reference to the accompanying drawings and examples.

[0016] like Figure 1 As shown in FIG, a flow chart of a small sample SAR target recognition method based on key feature capture and discrimination of the present invention is shown, and the specific steps are as follows:

[0017] Step 1: Construct a global auxiliary branch structure, receive the entire SAR target image as input, perform initial recognition, and calculate the initial recognition loss L whole ;

[0018] like Figure 2 As shown in Step 1, the global auxiliary branch structure is similar to the conventional image recognition framework, which mainly consists of two parts: feature extractor and classifier. i SAR target image sample x i , which is fed into the global auxiliary branch.

[0019] First, the output feature map of the global auxiliary branch is obtained through the feature extractor, denoted as M whole (x i )∈R h ×w×c .

[0020] Among them, R represents the number field, h represents the height of the input image, w represents the width of the input image, and c represents the number of channels of the input image.

[0021] Then, the classifier generates the sample x i The global auxiliary branch category prediction result p whole (y j |x i ), and calculate the initial recognition loss L whole , the expression is as follows:

[0022]

[0023] Among them, p whole (y j |x i ) represents the sample x i The probability of being classified as the jth category, K represents the number of categories contained in the current SAR target image sample set.

[0024] Step 2: Use the feature capture module to segment the SAR target image and obtain the key image area slices for current recognition

[0025] like Figure 2 As shown in Step 2, the feature capture module aims to search and capture the most critical image areas in the input samples for recognition.

[0026] In the work of the global auxiliary branch, by inputting a complete SAR target image into the deep learning model, the model is able to classify the image into a specific target class based on the extracted feature maps. These feature maps are obtained from different spatial local regions of the image sample and play different roles in recognition. The feature capture module uses the above recognition results, including the global auxiliary branch prediction results p whole (y j |x i ) and the output feature map M of the global auxiliary branch whole (x i ), search and locate key image regions that contribute to correct prediction and obtain corresponding weights.

[0027] The feature capture module is divided into two stages and introduced below.

[0028] The first stage: Search for key features that are helpful for final recognition in the global auxiliary branch, and use the final category prediction results of the global auxiliary branch to weight and integrate feature maps to obtain key features.

[0029] The process of weighted integration is as follows:

[0030] Based on step 1, input image sample x i The output feature map of the global auxiliary branch is It is integrated through the global average pooling operation to obtain F(x i )∈R 1×1×c , and then the classifier is processed by SoftMax to obtain the input image sample x i The category prediction result pred={p whole (y1|x i ),...,p whole (y K |x i )}∈R 1×K , and then use the category prediction result pred to adjust the weight W corresponding to the fully connected layer FC , the expression is as follows:

[0031]

[0032] in, Corresponding to M whole (x i ) is the feature map of the jth channel in the vector, arg max(·) represents the position index corresponding to the maximum value in the vector, w1,w2,...,w c represents the parameters of the fully connected layer, Indicates the weights corresponding to the fully connected layer obtained by calculating the update.

[0033] Then calculate the captured key features M fea-capture ∈R h×w as follows:

[0034]

[0035] Among them, resize(·) means adjusting the size of these local features to the input image x by bilinear interpolation i size;

[0036] Phase 2: Based on the key features obtained in the first phase, capture and segment the input image x i The corresponding key image region slice in and will Input into the local enhancement branch.

[0037] Captured slices of key image areas The calculation is as follows:

[0038]

[0039] in, Represents the input image x iThe global average value of B(·) represents the threshold value. Binarization, μ represents the weight of the feature capture module, ⊙ represents the Hadamard product calculation, and the captured key image area slice As the input of the local enhancement branch.

[0040] Step 3: Construct a local enhancement branch structure and slice the key image areas after segmentation Re-identify and calculate losses;

[0041] like Figure 2 As shown in Step 3, the key image area slices obtained based on the segmentation in Step 2 It is input into the local enhancement branch, and then the output result of the local enhancement branch is obtained through the feature extractor and classifier, including the local enhancement branch category prediction result p local (y j |x i ) and the output feature map M of the local enhancement branch local (x i ); recognition loss L of the local enhancement branch local The calculation is as follows:

[0042]

[0043] Step 4: Use the feature recognition module to collaborate with the global and local branches to improve the separability of local features of similar pairs of samples between classes and the compactness of local features of dissimilar pairs of samples within a class in the global features, thus completing the final target recognition;

[0044] like Figure 3 As shown, the feature recognition module is designed to collaborate with the global branch and the local branch, capturing more key image regions in the global branch and extracting more discriminative features in the local branch.

[0045] The feature recognition module calculates a feature recognition loss L disc To update the model, it guides the model to search and capture stronger inter-class separability and intra-class compactness of key image regions based on three constraints, as follows:

[0046] 1) If Figure 3 As shown in the right part, the basic global recognition constraints are divided into two stages, as follows:

[0047] The first stage: input SAR image x i The output feature map M of the global auxiliary branch whole (x i ) and the output feature map M of the local enhancement branch local (x i ), in all input images In , the cosine similarity of global features is calculated as:

[0048]

[0049] in, Indicates that it belongs to the target category label y i where N is the number of training samples for each class, ||·||2 is the L2 norm, and f(·) is the flatten operation of the feature map.

[0050] Then get the similarity ranking table Score of the global features global , for sample x i , find the global features of all different classes that match x i Most similar features and the most dissimilar features within the same class

[0051] The second stage: Calculate the basic global recognition loss L disc-BGD , the calculation formula is as follows:

[0052]

[0053]

[0054]

[0055] in, Represents the sample x in the global feature i The similarity between the most similar inter-class samples, Represents the sample x in the global feature i The similarity between the most dissimilar samples in the same class, the constant ψ represents and The similarity boundary between them, ∑· represents the function operation of adding all vector elements.

[0056] Then the recognition loss L disc-BGD Make the similarity of features of the same category greater than ψ, and the similarity of features of different categories less than ψ.

[0057] 2) If Figure 3 The global-local consistency constraints shown on the left are as follows:

[0058] The second constraint is global-local consistency, which aims to improve the separability of local features for inter-class samples with similar global features, and the compactness of local features for intra-class samples with dissimilar global features.

[0059] For the sample x i,Based on the above basic global recognition constraints, we find global feature samples that are difficult to distinguish between inter-class pairs and intra-class pairs.

[0060] This embodiment uses and Represent the global and local features of the most similar samples between classes, and use and They represent the global and local features of the most dissimilar samples within the class, respectively.

[0061] Then the basic strength (similarity) of inter-class and intra-class pairs of local features can be calculated as:

[0062]

[0063]

[0064] in, Represents the sample x in the local feature i The basic similarity between the most similar inter-class samples, Represents the sample x in the local feature i The basic similarity between the sample and the least similar sample in the class.

[0065] Then, the dynamic strength of pushing apart inter-class pairs and pulling together intra-class pairs is used to adjust the basic similarity in Equations (10) and (11). The calculation operation is as follows:

[0066]

[0067]

[0068] Where γ represents the adjustable consistency parameter, and Represents the sample x in the adjusted local features respectively i Adaptive similarity between the most similar inter-class samples and the least similar inter-class samples.

[0069] Then the global-local consistency loss L disc-GLC Calculated as:

[0070]

[0071] From formula (12), formula (13) and formula (14), we can see that for a sample x i, if it has similar global features to inter-class samples but dissimilar global features to intra-class samples, then Equation (14) will calculate a larger loss to cluster intra-class sample pairs and separate local features of inter-class sample pairs. In this way, the feature recognition module promotes the intra-class compactness and inter-class separability of local features of these dissimilar intra-class pairs or similar inter-class pairs.

[0072] 3) Adaptive local recognition constraints, as follows:

[0073] The third constraint is adaptive local recognition, which aims to ensure that the effectiveness of local features is better than the corresponding global features.

[0074] According to equations (10), (11), (12) and (13), the adaptive boundary ψ ALD (x i ) is calculated as follows:

[0075]

[0076] Where ρ≥0 means adjusting ψ ALD (x i ) value, and

[0077] Therefore, facing Close to -1 and Difficult to identify samples close to 1, ψ ALD (x i ) value can maintain its original high value; Close to 1 and Easy-to-identify samples close to -1, ψ ALD (x i ) value can be attenuated according to the parameter ρ. The larger the ρ value, the more difficult it is to identify the sample’s ψ ALD The value of ψ and other samples ALD The more obvious the difference between the values.

[0078] Then the adaptive local recognition loss L disc-ALD Expressed as:

[0079]

[0080] Finally, the feature recognition module combines the above three constraints to provide a hybrid loss L disc , calculated as follows:

[0081] L disc =L disc-BGD +L disc-GLC +L disc-ALD (17)

[0082] The three constraints of the feature recognition module can guide the model to search and capture key image regions with stronger inter-class separability and intra-class compactness. Even if the global branch performs poorly at the beginning of training, the feature recognition module can guide the global branch to acquire more important image regions and enable the local branch to extract more informative features.

[0083] Integration step 1 obtained global auxiliary branch recognition loss L whole , the recognition loss L of the local enhancement branch obtained in step 3 local , and the mixed loss L obtained in step 4 disc , we get a final loss L, and complete the final accurate target recognition. The calculation formula is as follows:

[0084] L=λ1L whole +λ2L local +λ3L disc (18)

[0085] Here, λ1, λ2, and λ3 represent weighting coefficients.

[0086] In this embodiment, by calculating the gradient based on the loss L and updating the model using backpropagation, this embodiment can enhance the effectiveness of key features in SAR target images and improve recognition performance. Tables 1 and 2 respectively show the comparative experimental results of the method of the present invention and other existing popular SAR target recognition methods on two SAR image measurement benchmark datasets, including the moving and stationary target acquisition and recognition (MSTAR) dataset (as shown in Table 1) and the OpenSARShip dataset (as shown in Table 2, where the numbers in brackets in Table 2 represent the specific number of supervised samples for each type of training corresponding to the recognition rate), verifying the excellent performance of the method of the present invention in SAR target recognition.

[0087] Table 1

[0088]

[0089] Table 2

[0090]

[0091] In summary, the method of the present invention searches for and captures the key areas of the target image through the collaborative global and local branch structures, so that the recognition model is concentrated on the effective discrimination area for learning and training. It has demonstrated excellent performance in both the MSTAR and OpenSARShip datasets, improving the effectiveness of target image feature extraction, the discrimination ability, and the accuracy and robustness of SAR target recognition under small sample sizes.

[0092] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A small sample SAR target recognition method based on key feature capture and discrimination, the specific steps are as follows: Step 1: Construct a global auxiliary branch structure to receive the entire SAR target image as input for initial recognition and calculate the initial recognition loss; Step 2: Use the feature capture module to segment the SAR target image and obtain the key image area slices for current recognition; Step 3: Construct a local enhancement branch structure, identify the key image area slices after segmentation again and calculate the loss; Step 4: Use the feature recognition module to collaborate with the global and local branches to improve the separability of local features of similar pairs of samples between classes and the compactness of local features of dissimilar pairs of samples within a class in the global features, thus completing the final target recognition; The step 2 is specifically as follows: The feature capture module is divided into two stages: The first stage: Search for key features that contribute to the final recognition in the global auxiliary branch, and use the final category prediction results of the global auxiliary branch to weight and integrate feature maps to obtain key features; The process of weighted integration is as follows: Based on step 1, input image sample x i The output feature map of the global auxiliary branch is It is integrated through the global average pooling operation to obtain F(x i )∈R 1×1×c , and then the classifier is processed by SoftMax to obtain the input image sample x i The category prediction result pred={p whole (y1|x i ),...,p whole (y K |x i )}∈R 1×K , and then use the category prediction result pred to adjust the weight W corresponding to the fully connected layer FC , the expression is as follows: in, Corresponding to M whole (x i ) is the feature map of the jth channel in the vector, argmax(·) represents the position index corresponding to the maximum value in the vector, w1,w2,...,w c represents the parameters of the fully connected layer, Indicates the weight corresponding to the fully connected layer obtained by calculating the update; Then calculate the captured key features M fea-capture ∈R h×w as follows: Among them, resize(·) means adjusting the size of these local features to the input image x by bilinear interpolation i size; Phase 2: Based on the key features obtained in the first phase, capture and segment the input image x i The corresponding key image region slice in and will Input into the local enhancement branch; Captured slices of key image areas The calculation is as follows: in, Represents the input image x i The global average value of B(·) represents the threshold value. Binarization, μ represents the weight of the feature capture module, ⊙ represents the Hadamard product calculation, and the captured key image area slice As the input of the local enhancement branch.

2. The small sample SAR target recognition method based on key feature capture and discrimination according to claim 1 is characterized in that: The step 1 is specifically as follows: The global auxiliary branch structure mainly consists of two parts: feature extractor and classifier; given a target category label y i SAR target image sample x i , which is fed into the global auxiliary branch; First, the output feature map of the global auxiliary branch is obtained through the feature extractor, denoted as M whole (x i )∈R h×w×c ; Where R represents the number domain, h represents the height of the input image, w represents the width of the input image, and c represents the number of channels of the input image; Then, the classifier generates the sample x i The global auxiliary branch category prediction result p whole (y j |x i ), and calculate the initial recognition loss L whole , the expression is as follows: Among them, p whole (y j |x i ) represents the sample x i The probability of being classified as the jth category, K represents the number of categories contained in the current SAR target image sample set.

3. The small sample SAR target recognition method based on key feature capture and discrimination according to claim 1 is characterized in that: The step three is as follows: Slices of key image regions obtained based on the segmentation in step 2 It is input into the local enhancement branch, and then the output result of the local enhancement branch is obtained through the feature extractor and classifier, including the local enhancement branch category prediction result p local (y j |x i ) and the output feature map M of the local enhancement branch local (x i ); recognition loss L of the local enhancement branch local The calculation is as follows:

4. The small sample SAR target recognition method based on key feature capture and discrimination according to claim 1 is characterized in that: The step 4 is specifically as follows: The feature recognition module calculates a feature recognition loss L disc To update the model, it guides the model to search and capture stronger inter-class separability and intra-class compactness of key image regions based on three constraints, as follows: 1) Basic global recognition constraints are divided into two stages, as follows: The first stage: input SAR image x i The output feature map M of the global auxiliary branch whole (x i ) and the output feature map M of the local enhancement branch local (x i ), in all input images In , the cosine similarity of global features is calculated as: in, Indicates that it belongs to the target category label y i The j-th SAR image sample in , N represents the number of training samples for each class, ||·||2 represents the L2 norm, and f(·) represents the flatten operation of the feature map; Then get the similarity ranking table Score of the global features global , for sample x i , find the global features of all different classes that match x i Most similar features and the most dissimilar features within the same class The second stage: Calculate the basic global recognition loss L disc-BGD , the calculation formula is as follows: in, Represents the sample x in the global feature i The similarity between the most similar inter-class samples, Represents the sample x in the global feature i The similarity between the most dissimilar samples in the same class, the constant ψ represents and The similarity boundary between them, ∑· represents the function operation of adding all vector elements; Then the recognition loss L disc-BGD Make the similarity of features of the same category greater than ψ, and the similarity of features of different categories less than ψ; 2) Global-local consistency constraints, as follows: For the sample x i ,Based on the above basic global recognition constraints, we find the global feature samples that are difficult to distinguish between inter-class pairs and intra-class pairs; use and Represent the global and local features of the most similar samples between classes, and use and Represent the global and local features of the most dissimilar samples within the class; Then the basic similarity between and within classes of local features can be calculated as: in, Represents the sample x in the local feature i The basic similarity between the most similar inter-class samples, Represents the sample x in the local feature i The basic similarity between the most dissimilar samples in the class; Then, the dynamic strength of pushing apart inter-class pairs and pulling together intra-class pairs is used to adjust the basic similarity in Equations (10) and (11). The calculation operation is as follows: Where γ represents the adjustable consistency parameter, and Represents the sample x in the adjusted local features respectively i Adaptive similarity between the most similar inter-class samples and the least similar inter-class samples; Then the global-local consistency loss L disc-GLC Calculated as: 3) Adaptive local recognition constraints, as follows: According to equations (10), (11), (12) and (13), the adaptive boundary ψ ALD (x i ) is calculated as follows: Where ρ≥0 means adjusting ψ ALD (x i ) value, and Then the adaptive local recognition loss L disc-ALD Expressed as: Finally, the feature recognition module combines the above three constraints to provide a hybrid loss L disc , calculated as follows: L disc =L disc-BGD +L disc-GLC +L disc-ALD (17) Integration step 1 obtained global auxiliary branch recognition loss L whole , the recognition loss L of the local enhancement branch obtained in step 3 local , and the mixed loss L obtained in step 4 disc , we get a final loss L, and complete the final accurate target recognition. The calculation formula is as follows: L=λ1L whole +λ2L local +λ3L disc (18) Here, λ1, λ2, and λ3 represent weighting coefficients.

Images