Target recognition classification and application based on center coordinate attention model

Abstract

The application discloses a target recognition classification and application based on a center coordinate attention model, and comprises the following steps: 1, selecting different kinds of SAR images as a sample set; 2, adopting a Lenet-5 network as a backbone network, embedding the center coordinate attention model between the last convolution layer and the full connection layer, and obtaining a fusion weight coefficient matrix; 3, performing optimal feature fusion on the fusion weight coefficient matrix and the output of the last convolution layer through pixel-level horizontal fusion, and obtaining final features; 4, inputting the final features into the full connection layer and a softmax classifier to obtain a prediction result, comparing the prediction result with a real result, and completing a network training process; and 5, inputting a SAR image to be classified into the trained network model to obtain a corresponding category. The application can improve the integrity of target feature representation located in the center of the SAR image, obtain higher classification precision and classification efficiency, and has good engineering application value.

Description

Technical Field

[0001] This invention relates to the field of SAR image target classification technology, specifically to a target recognition and classification method and its application based on a center coordinate attention model. Background Technology

[0002] Synthetic Aperture Radar (SAR) is a modern, all-weather, all-time high-resolution microwave imaging radar capable of acquiring high-resolution radar images of targets under any weather conditions. Compared to optical images, SAR images are easier and more flexible to acquire, and provide richer scattering information. Therefore, SAR images are widely used in military reconnaissance classification and have achieved good classification performance.

[0003] SAR image target classification methods have evolved through four stages: template-based methods, model-based methods, deep learning-based methods, and attention-based methods. Template-based methods classify SAR image targets by designing matching templates, while model-based methods classify SAR image targets by simulating a series of hypothetical models. However, both of these methods heavily rely on manually designed matching templates and hypothetical models, making them not only time-consuming and labor-intensive but also lacking in automation and flexibility. With the development of SAR target classification methods, deep learning-based methods have gained widespread attention. They accomplish the task of SAR image target classification through automatic feature extraction and classification, but they are highly dependent on feature representation capabilities, which limits their application to some extent. With the development of attention mechanisms, attention-based SAR target classification methods have gradually been proposed. These methods supplement the feature representation capabilities of SAR image target classification methods from aspects such as channel attention, spatial attention, and frequency attention, seeking features that best represent SAR image targets from various perspectives.

[0004] Existing SAR image target classification methods complete the SAR image target classification task by deepening network layers, extracting multi-faceted feature information and selecting different convolution fusion methods. However, they cannot simultaneously achieve classification accuracy and efficiency. Furthermore, while extracting a large amount of feature information, they ignore the fact that different feature information has different effects on classification accuracy. The messy multi-faceted feature information will affect classification accuracy and efficiency to a certain extent, resulting in unsatisfactory classification results. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies by proposing a target recognition, classification, and application based on a center coordinate attention model. This approach aims to effectively increase the attention given to the center coordinate information of SAR images, ensure the full utilization of target feature information in SAR images, and thereby improve the accuracy and efficiency of image classification.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The present invention provides a target recognition and classification method based on a center coordinate attention model, characterized by the following steps:

[0008] Step 1: Select different types of SAR images as a sample set for target recognition and classification based on the center coordinate attention model. After unifying the size of the SAR images in the sample set, a preprocessed SAR image set {S1, S2, ..., S} is obtained. i ,…,S n}, where S i Let represent the i-th SAR image sample in the preprocessed SAR image set, and n represent the capacity of the SAR image set; i∈[1,n];

[0009] Step 2: Construct a target recognition and classification network based on the center coordinate attention model, including: backbone network, center coordinate attention model, fusion layer and classification network;

[0010] Step 2.1: The backbone network is based on the Lenet-5 network and includes: a first convolution-pooling block, a second convolution-pooling block, and a convolution block;

[0011] The first convolutional-pooling block has a convolutional kernel K1 with a size of n1×n1 and a stride of c1, and a pooling kernel P1 with a size of m1×m1 and a stride of p1; the second convolutional-pooling block has a convolutional kernel K2 with a size of n2×n2 and a stride of c2, and a pooling kernel P2 with a size of m2×m2 and a stride of p2; the convolutional kernel K3 of the convolutional block has a size of n3×n3 and a stride of c3.

[0012] The i-th SAR image sample S i The data is fed into the backbone network for processing, and a convolutional feature map F is obtained using equation (1). i :

[0013]

[0014] In equation (1), F i,1 F i,2 F i,3 F i,4These represent the outputs of the first convolution-pooling block after passing through the convolution kernel, the outputs of the first convolution-pooling block after passing through the pooling kernel, the outputs of the second convolution-pooling block after passing through the convolution kernel, and the outputs of the second convolution-pooling block after passing through the pooling kernel, respectively. * indicates the convolution operation, and σ(·) is the ReLU function.

[0015] Step 2.2: The center coordinate attention model is composed of a pooling block, a center weighting block, and an output block in sequence;

[0016] The pooling block uses one-dimensional pooling kernels of size H×1 and 1×W to respectively process the convolutional feature map F. i Max pooling is performed to obtain a bar feature map F with dimensions 1×W and H×1 using equation (1). i v F i h :

[0017]

[0018] In equation (2), x and y represent the row and column, respectively, and F i,x,y Represents the convolutional feature map F i The element in the x-th row and y-th column, Represents vertical bar feature map F i v The element in the y-th column, Represents the horizontal bar feature map F i h The element in the xth row;

[0019] The central weighted block uses a central importance weighting kernel to weight the bar feature map F respectively. i v F i h By applying weights, the weighted kernel f of the vertical center importance can be obtained using equation (3). v And horizontal center importance weighted kernel f h :

[0020]

[0021] In equation (3), The weighted kernel f represents the importance of the vertical center. v The element in the x-th row and y-th column, The horizontal center importance weighted kernel f represents the value of the horizontal center. h The element in the x-th row and y-th column, exp(·) represents the exponential function, x * and y * These are bar feature maps F i v Fi h The middle coordinates of the row and column, where ε is the adjustment parameter;

[0022] Using equation (4) to analyze the bar feature map F i v F i h Weighted kernel f with the center importance v f h Weighting is performed to obtain a vertical center importance-weighted bar feature map. Weighted bar chart of horizontal center importance

[0023]

[0024] In equation (4), Represents a weighted bar feature map of vertical center importance. The element in the x-th row and y-th column, A weighted bar feature map representing the importance of the horizontal center. The element in the x-th row and y-th column, MUL(·) represents the multiplication of elements at corresponding positions in the matrix;

[0025] The output block first weights the bar feature map by center importance. and Perform row expansion and column expansion respectively to obtain the convolutional feature map F. i Vertical center importance weighted feature map with consistent size Weighted feature map of horizontal center importance Then, the center importance weighted feature map is applied using equation (5). and After element-wise addition, the final center importance weighted feature map G is obtained. i Finally, the fusion weight coefficient matrix W is obtained using equation (6). i :

[0026]

[0027] W i =Sigmoid(G i (6)

[0028] In equation (5), G i,x,y G represents the center importance weighted feature map. i The element in the x-th row and y-th column is given by Add(·), which means adding the elements at the corresponding positions in the matrix. Represents the weighted feature map of vertical center importance The element in the x-th row and y-th column, Represents the weighted feature map of horizontal center importance. The element in the x-th row and y-th column;

[0029] In equation (6), Sigmoid(·) represents the sigmoid function;

[0030] Step 2.3: The fusion layer uses equation (6) to process the convolutional feature map F. i and the fusion weight coefficient matrix W i Weighted fusion is performed to obtain the fused feature map Y. i :

[0031]

[0032] In equation (6), Y i,x,y Represents the fused feature map Y i The element in the x-th row and y-th column, W i,x,y W represents the fusion weight coefficient matrix. i The element in the x-th row and y-th column;

[0033] Step 2.4: The classification network is composed of a first fully connected layer, a second fully connected layer, and a softmax classifier in sequence;

[0034] Fuse feature map Y i The features are sequentially input into the first fully connected layer and the second fully connected layer for feature integration to obtain the integrated vector; wherein, the dimension of the first fully connected layer is 'a', and the dimension of the second fully connected layer is the same as the number of categories in the SAR image set;

[0035] The integrated vector is fed into the softmax classifier for classification, thereby obtaining the i-th SAR image sample S. i For each category, a posterior probability is assigned, and the category corresponding to the highest posterior probability is selected as the i-th SAR image sample S. i The prediction category;

[0036] The predicted category is compared with the i-th SAR image sample S. i The corresponding real categories are compared, and the target recognition and classification network based on the center coordinate attention model is trained by the backpropagation algorithm to obtain the trained network model.

[0037] Step 3: Input the SAR image to be classified into the trained network model and obtain the classification result.

[0038] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the target recognition and classification method, and the processor is configured to execute the program stored in the memory.

[0039] The present invention provides a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the target recognition and classification method.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. The target recognition and classification based on the center coordinate attention model proposed in this invention makes full use of the center position information of SAR images, effectively improves the integrity of feature information representation, and achieves better performance in SAR image target classification.

[0042] 2. This invention proposes a center coordinate attention model, which uses a center importance weighted kernel to enhance the center position information of the image, effectively highlighting the center coordinate information of the SAR image and reducing the interference of strong clutter at the edge of the SAR image, thereby obtaining more distinct SAR target feature information.

[0043] 3. This invention cleverly utilizes one-dimensional max pooling, applying it to both rows and columns of the feature map, thereby capturing the long-range dependencies of the feature map along both horizontal and vertical spatial dimensions, and obtaining more complete feature information of the SAR image, thus improving the feature representation capability of the SAR image target.

[0044] 4. This invention uses a pixel-level optimal feature fusion method to perform weighted fusion of the fusion weight coefficient matrix calculated from the center coordinate attention model and the feature map, which effectively highlights the importance of each pixel in the SAR image and can better locate the position of the SAR target in the SAR image, thereby improving the SAR target classification accuracy.

[0045] 5. This invention embeds the center coordinate attention model into a shallow convolutional neural network, highlighting the effectiveness of shallow networks in classifying targets in simple SAR images, and greatly improving classification efficiency while ensuring classification accuracy. Attached Figure Description

[0046] Figure 1 This is the central coordinate attention model of the present invention;

[0047] Figure 2 This is a flowchart of the SAR image target classification method in this invention;

[0048] Figure 3 This is the MSTAR experimental dataset of the present invention;

[0049] Figure 4a This is a confusion matrix diagram of the SAR image target classification method of the present invention on the SOC experimental dataset;

[0050] Figure 4b This is a confusion matrix diagram of the SAR image target classification method of the present invention on the EOC-D experimental dataset. Detailed Implementation

[0051] The present invention will be further described below with reference to embodiments, and the present invention includes, but is not limited to, the following embodiments. In this embodiment, a traditional convolutional neural network is added with, such as Figure 2 The center coordinate attention model shown proposes, as follows: Figure 1 The example shown is a target recognition and classification system based on a center coordinate attention model. This network can highlight the center coordinate features that contribute significantly to the final classification while reducing interference from strong edge clutter, thus achieving better classification results. Specifically, a target recognition and classification system based on a center coordinate attention model includes the following steps:

[0052] Step 1: Select different types of SAR images as a sample set for target recognition and classification based on the center coordinate attention model. After downsampling the SAR images in the sample set, a preprocessed SAR image set {S1, S2, ..., S...} with a uniform size of b×b is obtained. i ,…,S n}, where S i Let represent the i-th SAR image sample in the preprocessed SAR image set, and n represent the capacity of the SAR image set; i∈[1,n];

[0053] Step 2: Construct a target recognition and classification network based on the center coordinate attention model, including: backbone network, center coordinate attention model, fusion layer and classification network;

[0054] Step 2.1: The backbone network is based on the Lenet-5 network and includes: a first convolution-pooling block, a second convolution-pooling block, and a convolution block;

[0055] The first convolutional-pooling block has a convolutional kernel K1 with a size of n1×n1 and a stride of c1, and a pooling kernel P1 with a size of m1×m1 and a stride of p1; the second convolutional-pooling block has a convolutional kernel K2 with a size of n2×n2 and a stride of c2, and a pooling kernel P2 with a size of m2×m2 and a stride of p2; the convolutional kernel K3 of the convolutional block has a size of n3×n3 and a stride of c3.

[0056] The i-th SAR image sample S i The data is fed into the backbone network for processing, and a convolutional feature map F of size H×W is obtained using equation (1). i :

[0057]

[0058] In equation (1), F i,1 F i,2 F i,3 F i,4 These represent the outputs of the first convolution-pooling block after passing through the convolution kernel, the outputs of the first convolution-pooling block after passing through the pooling kernel, the outputs of the second convolution-pooling block after passing through the convolution kernel, and the outputs of the second convolution-pooling block after passing through the pooling kernel, respectively. * indicates the convolution operation, and σ(·) is the ReLU function.

[0059] Step 2.2: The center coordinate attention model is composed of a pooling block, a center weighting block, and an output block in sequence;

[0060] The pooling block uses one-dimensional pooling kernels of size H×1 and 1×W to respectively process the convolutional feature map F. i Max pooling is performed to obtain a bar feature map F with dimensions 1×W and H×1 using equation (1). i v F i h :

[0061]

[0062] In equation (2), x and y represent the row and column, respectively, and F i,x,y Represents the convolutional feature map F i The xth row and yth column, Represents bar feature map F i v The y-th column, Represents bar feature map F i h The xth row;

[0063] The central weighted block uses a central importance weighting kernel to weight the bar feature map F respectively. i v F i h By performing weighting, we obtain the vertical center importance weighting kernel f shown in equation (3). v And horizontal center importance weighted kernel f h :

[0064]

[0065] In equation (3), This represents the x-th row and y-th column of the weighted kernel for vertical center importance. The x-th row and y-th column of the horizontal center importance weighting kernel is represented by exp(·), where exp(·) represents the exponential function, x * and y * These are the midpoint coordinates of the row and column of the bar feature map, respectively, and ε is an adjustment parameter;

[0066] Then, the bar feature map F is completed using equation (4). i v F i h Weighted kernel f with the center importance v f h The weighting process yields a vertical center importance-weighted bar feature map. Weighted bar chart of horizontal center importance

[0067] In equation (4), Represents a weighted bar feature map of vertical center importance. The xth row and yth column, A weighted bar feature map representing the importance of the horizontal center. In the x-th row and y-th column, MUL(·) represents the element-wise multiplication of corresponding positions in the matrix;

[0068] The output block first weights the bar feature map by center importance. and Perform row expansion and column expansion respectively to obtain the convolutional feature map F. i Vertical center importance weighted feature map with consistent size Weighted feature map of horizontal center importance Then the center importance weighted feature map and The final center importance weighted feature map G is obtained by adding elements one by one using equation (5). i Finally, the fusion weight coefficient matrix W is obtained using equation (6). i :

[0069]

[0070] W i =Sigmoid(G i (6)

[0071] In equation (5), G i,x,y G represents the center importance weighted feature map. iIn the x-th row and y-th column, Add(·) represents adding the elements at corresponding positions in the matrix. Represents the weighted feature map of vertical center importance The xth row and yth column, Represents the weighted feature map of horizontal center importance. The x-th row and y-th column;

[0072] In equation (6), Sigmoid(·) represents the sigmoid function;

[0073] Step 2.3: The fusion layer performs pixel-level optimal feature fusion on the convolutional feature map F i and the fusion weight coefficient matrix W i We perform weighted fusion to obtain the fused feature map Y using equation (6). i :

[0074]

[0075] In equation (6), Y i,x,y Represents the fused feature map Y i In the x-th row and y-th column, W i,x,y W represents the fusion weight coefficient matrix. i The xth row and yth column;

[0076] Step 2.4: The classification network is composed of a first fully connected layer, a second fully connected layer, and a softmax classifier in sequence;

[0077] Fuse feature map Y i After flattening, the data is sequentially input into the first fully connected layer and the second fully connected layer for feature integration, resulting in an integrated vector. This process also increases the nonlinearity of the network. The first fully connected layer has a dimension of 'a', and the second fully connected layer has the same dimension as the number of categories in the SAR image set. To prevent overfitting due to the large number of parameters in the fully connected layers, this invention uses a Dropout layer in both the first and second fully connected layers, setting the random probability to 0.5. During each training process, half of the nodes are randomly selected and set to 0, thus excluding them from training. This further increases the nonlinearity of the network and effectively avoids the gradient explosion and gradient vanishing problems commonly encountered during convolutional neural network training.

[0078] The integrated vector is fed into the softmax classifier for classification, thereby obtaining the i-th SAR image sample S. i For each category, a posterior probability is assigned, and the category corresponding to the highest posterior probability is selected as the i-th SAR image sample S. i The prediction category;

[0079] The predicted category is compared with the i-th SAR image sample S. i The corresponding real categories are compared, and the target recognition and classification network based on the center coordinate attention model is trained by the backpropagation algorithm to obtain the trained network model.

[0080] Step 3: Input the SAR image to be classified into the trained network model and obtain the classification result.

[0081] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the target recognition and classification method. The processor is configured to execute the program stored in the memory.

[0082] In this embodiment, a computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the target recognition and classification method.

[0083] At this point, the target recognition and classification based on the center coordinate attention model is basically complete.

[0084] The advantages of this invention are further illustrated below through experiments on the MSTAR dataset.

[0085] MSTAR dataset SAR image target classification experiment:

[0086] 1. Experimental setup:

[0087] The MSTAR dataset used in this invention was acquired by the Sandia National Laboratories SAR sensor platform. The dataset was acquired by a high-resolution spot-scan synthetic aperture radar with a resolution of 0.3m × 0.3m. It operates in the X-band and uses a horizontal-horizontal (HH) polarization mode. It includes images of 10 types of military targets under different conditions: 2S1, BRDM-2, BTR-60, D7, T-72, BMP-2, BTR-70, T-62, ZIL-131, and ZSU-234. Among these, T-62 and T-72 are tanks, BMP-2, BRDM-2, BTR-60, and BTR-70 are armored vehicles, 2S1 is a rocket launcher, ZSU-234 is an air defense unit, ZIL-131 is a military truck, and D7 is a bulldozer. Optical images and corresponding SAR images of the ten military targets are shown below. Figure 3As shown. To fully illustrate the advantages of the present invention, this embodiment uses the Standard Operating Condition (SOC) dataset and the Depression-variant Extended Operating Condition (EOC-D) dataset from the MSTAR dataset for experiments.

[0088] 2. Results Analysis:

[0089] In this embodiment, the performance of the proposed method is quantitatively analyzed using accuracy, precision, recall, F1 score, training time, and testing time. To illustrate the superiority of the proposed method, several common object recognition and classification methods based on convolutional neural networks are compared, including Le-Net, M_VDCNN, and VGG-Net. For the training process of all methods, the batch size is set to 20, the number of iterations to 100, the learning rate to 0.001, and the optimization algorithm to be Adam. The comparison results are shown in Table 1, where:

[0090]

[0091]

[0092]

[0093] In equations (7) to (9), TP is the number of positive samples that are judged as positive, FP is the number of positive samples that are judged as negative, and FN is the number of negative samples that are judged as positive.

[0094] The confusion matrix is ​​a visualization of the evaluation results of SAR image target classification methods. The numbers on its diagonal represent the number of correctly classified targets; a larger diagonal number indicates better performance. From... Figure 4a and Figure 4b As can be seen, regardless of whether it is the SOC experimental dataset or the EOC-D experimental dataset, the diagonal numbers of the confusion matrix of this invention are much larger than the other numbers, and some categories even reach 100% classification, which further illustrates the superiority of this invention.

[0095] Table 1 Comparison of target recognition and classification performance based on the central coordinate attention model.

[0096]

[0097] Table 1 compares the performance of various SAR image target classification methods with that of the present invention. It can be seen that the classification performance of the present invention is superior to other methods. Among the aforementioned SAR image target classification methods, Le-Net uses a relatively simple convolutional neural network, resulting in shorter training and testing times, but its classification results are not ideal compared to other methods. M-VDCNN generates SAR data from multiple views, extracts feature information from the input SAR data from multiple views, and then gradually fuses the feature information extracted from different views at different levels, thus improving the network's classification performance to some extent. However, M-VDCNN is highly dependent on the image quality of the original SAR image, and generating multi-view SAR data is very time-consuming. VGG-Net uses a stack of small-scale convolutional kernels instead of large-scale convolutional kernels, thereby improving the network's nonlinear characteristics and increasing its depth. However, it does not show superior performance in simple image classification problems, and it significantly increases network complexity, directly leading to a multiple increase in training and testing times. This invention proposes a target recognition and classification model based on a center coordinate attention model. This model utilizes vertical and horizontal coordinate attention to capture the long-range dependencies of the input feature map, and center coordinate attention to enhance the center coordinate information that contributes most to classification performance, while mitigating interference from strong clutter at image boundaries. Furthermore, this invention employs pixel-level optimal feature fusion to highlight important pixels in the feature map. Finally, this invention uses the simplest traditional CNN structure as its main framework, embedding the center coordinate attention model between the last convolutional layer and the fully connected layer, improving classification accuracy while maintaining algorithm efficiency. In summary, the target recognition and classification model based on the center coordinate attention model proposed in this invention features simple structure, accurate classification, and high performance, making it highly valuable for practical classification projects where targets are distributed at the center of the image.

Claims

1. A target recognition and classification method based on a center coordinate attention model, characterized in that, Includes the following steps: Step 1: Select different types of SAR images as a sample set for target recognition and classification based on the center coordinate attention model. After unifying the size of the SAR images in the sample set, a preprocessed SAR image set {S1, S2, ..., S} is obtained. i ,…,S n }, where S i Let represent the i-th SAR image sample in the preprocessed SAR image set, and n represent the capacity of the SAR image set; i∈[1,n]; Step 2: Construct a target recognition and classification network based on the center coordinate attention model, including: backbone network, center coordinate attention model, fusion layer and classification network; Step 2.1: The backbone network is based on the Lenet-5 network and includes: a first convolution-pooling block, a second convolution-pooling block, and a convolution block; The first convolutional-pooling block has a convolutional kernel K1 with a size of n1×n1 and a stride of c1, and a pooling kernel P1 with a size of m1×m1 and a stride of p1; the second convolutional-pooling block has a convolutional kernel K2 with a size of n2×n2 and a stride of c2, and a pooling kernel P2 with a size of m2×m2 and a stride of p2; the convolutional kernel K3 of the convolutional block has a size of n3×n3 and a stride of c3. The i-th SAR image sample S i The data is fed into the backbone network for processing, and a convolutional feature map F is obtained using equation (1). i : In equation (1), F i,1 F i,2 F i,3 F i,4 These represent the outputs of the first convolution-pooling block after passing through the convolution kernel, the outputs of the first convolution-pooling block after passing through the pooling kernel, the outputs of the second convolution-pooling block after passing through the convolution kernel, and the outputs of the second convolution-pooling block after passing through the pooling kernel, respectively. * indicates the convolution operation, and σ(·) is the ReLU function. Step 2.2: The center coordinate attention model is composed of a pooling block, a center weighting block, and an output block in sequence; The pooling block uses one-dimensional pooling kernels of size H×1 and 1×W to respectively process the convolutional feature map F. i Max pooling is performed to obtain a bar feature map F with dimensions 1×W and H×1 using equation (1). i v F i h : In equation (2), x and y represent the row and column, respectively, and F i,x,y Represents the convolutional feature map F i The element in the x-th row and y-th column, Represents vertical bar feature map F i v The element in the y-th column, Represents the horizontal bar feature map F i h The element in the xth row; The central weighted block uses a central importance weighting kernel to weight the bar feature map F respectively. i v F i h By applying weights, the weighted kernel f of the vertical center importance can be obtained using equation (3). v And horizontal center importance weighted kernel f h : In equation (3), The weighted kernel f represents the importance of the vertical center. v The element in the x-th row and y-th column, The horizontal center importance weighted kernel f represents the value of the horizontal center. h The element in the x-th row and y-th column, exp(·) represents the exponential function, x * and y * These are bar feature maps F i v F i h The middle coordinates of the row and column, where ε is the adjustment parameter; Using equation (4) to analyze the bar feature map F i v F i h Weighted kernel f with the center importance v f h Weighting is performed to obtain a vertical center importance-weighted bar feature map. Weighted bar chart of horizontal center importance In equation (4), Represents a weighted bar feature map of vertical center importance. The element in the x-th row and y-th column, A weighted bar feature map representing the importance of the horizontal center. The element in the x-th row and y-th column, MUL(·) represents the multiplication of elements at corresponding positions in the matrix; The output block first weights the bar feature map by center importance. and Perform row expansion and column expansion respectively to obtain the convolutional feature map F. i Vertical center importance weighted feature map with consistent size Weighted feature map of horizontal center importance Then, the center importance weighted feature map is applied using equation (5). and After element-wise addition, the final center importance weighted feature map G is obtained. i Finally, the fusion weight coefficient matrix W is obtained using equation (6). i : W i =Sigmoid(G i ) (6) In equation (5), G i,x,y G represents the center importance weighted feature map. i The element in the x-th row and y-th column is given by Add(·), which means adding the elements at the corresponding positions in the matrix. Represents the weighted feature map of vertical center importance The element in the x-th row and y-th column, Represents the weighted feature map of horizontal center importance. The element in the x-th row and y-th column; In equation (6), Sigmoid(·) represents the sigmoid function; Step 2.3: The fusion layer uses equation (6) to process the convolutional feature map F. i and the fusion weight coefficient matrix W i Weighted fusion is performed to obtain the fused feature map Y. i : In equation (6), Y i,x,y Represents the fused feature map Y i The element in the x-th row and y-th column, W i,x,y W represents the fusion weight coefficient matrix. i The element in the x-th row and y-th column; Step 2.4: The classification network is composed of a first fully connected layer, a second fully connected layer, and a softmax classifier in sequence; Fuse feature map Y i The features are sequentially input into the first fully connected layer and the second fully connected layer for feature integration to obtain the integrated vector; wherein, the dimension of the first fully connected layer is 'a', and the dimension of the second fully connected layer is the same as the number of categories in the SAR image set; The integrated vector is fed into the softmax classifier for classification, thereby obtaining the i-th SAR image sample S. i For each category, a posterior probability is assigned, and the category corresponding to the highest posterior probability is selected as the i-th SAR image sample S. i The prediction category; The predicted category is compared with the i-th SAR image sample S. i The corresponding real categories are compared, and the target recognition and classification network based on the center coordinate attention model is trained by the backpropagation algorithm to obtain the trained network model. Step 3: Input the SAR image to be classified into the trained network model and obtain the classification result.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the target recognition and classification method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the target recognition and classification method of claim 1.

Images