SAR target detection method based on wavelet decomposition and attention feature fusion

Abstract

The present invention relates to a SAR target detection method based on wavelet decomposition and attention feature fusion, comprising: performing wavelet decomposition on the original SAR image to obtain low-frequency coefficients and high-frequency coefficients, and performing denoising on the high-frequency coefficients using Bayesian threshold shrinkage; utilizing a dual-branch structure of ConvNeXt and Swin Transformer to extract high-frequency features and low-frequency features at different levels, and performing feature adaptive fusion based on FGAM at each level; inputting the fused features of each level into FPN for further feature fusion to obtain multiple feature maps with the same number of channels; and inputting each feature map into RPN to obtain target detection results. The present invention effectively suppresses missed detection and false alarm phenomena in complex scenarios by explicitly utilizing frequency domain prior knowledge.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and in particular to a SAR target detection method based on wavelet decomposition and attention feature fusion. Background Art

[0002] Due to the all-weather observation capabilities of Synthetic Aperture Radar (SAR), target detection methods based on SAR images have become an important means for reconnaissance and surveillance, with significant application value in both military and civilian fields. However, the acquisition and transmission of SAR images are often subject to interference from multiplicative speckle noise and complex background clutter. This interference, which typically originates from various stages of image acquisition, processing, encoding, and transmission, can affect signal quality in visible or invisible ways, severely limiting the accuracy and robustness of target detection. Specifically, multiplicative speckle noise reduces local contrast, while complex background clutter increases false alarm rates. These factors significantly increase the difficulty of detecting targets such as aircraft and ships.

[0003] In recent years, deep learning-based SAR target detection methods have achieved breakthrough progress in the optical field, with the emergence of detectors such as Faster R-CNN, YOLO, and DETR. However, directly applying existing detector frameworks to SAR imagery for target detection often results in limited performance. This is because SAR images exhibit significant frequency-domain separability. High-frequency components (HF) primarily consist of target edges, textures, and multiplicative speckle noise, exhibiting locally sharp energy distributions. Low-frequency components (LF), on the other hand, reflect the scene's global structure, background clutter, and large-scale features, exhibiting slowly varying energy. Deep learning methods achieve target representation through end-to-end feature learning, but their implicit "black-box" feature coupling mechanism faces a fundamental challenge in SAR scenarios. This involves the indiscriminate fusion of high-frequency noise and low-frequency semantics in the feature space, making it difficult for models to balance noise suppression with target feature fidelity. In addition, existing detection frameworks based on convolutional neural networks (CNN) or Transformer generally use a single-branch architecture to directly process raw data. This design will make it difficult for the limited receptive field of the convolution kernel to effectively distinguish high-frequency noise from target edges, causing spectral confusion in the feature learning process. At the same time, the same level of the network needs to model high-frequency details and low-frequency semantics at the same time, resulting in conflicts in the gradient update direction and weakening the model's representation ability. Summary of the Invention

[0004] The purpose of the present invention is to provide a new SAR target detection strategy. By designing a wavelet denoising method based on Bayesian shrinkage threshold, a dual-branch feature extraction structure composed of ConvNeXt and Swin Transformer, and a feature adaptive fusion method based on FGAM, the SAR image feature extraction and fusion process are optimized, and the speckle noise and background clutter interference problems of SAR images in complex scenes are solved.

[0005] To achieve the above object, the present invention provides a SAR target detection method based on wavelet decomposition and attention feature fusion, comprising the following steps:

[0006] S1. Perform wavelet decomposition on the original SAR image to obtain low-frequency coefficients and high-frequency coefficients, and use Bayesian threshold shrinkage to denoise the high-frequency coefficients;

[0007] S2, using the dual-branch structure of ConvNeXt and Swin Transformer to extract high-frequency features and low-frequency features at different levels, and performing feature adaptive fusion based on FGAM at each level;

[0008] S3, input the fusion features of each level into FPN for feature fusion again to obtain multiple feature maps with the same number of channels;

[0009] S4. Input each feature map into RPN to obtain the target detection result.

[0010] Preferably, the step S1 includes:

[0011] The following model is constructed for the original SAR image:

[0012] (1);

[0013] Among them, y, x, and e represent the original SAR image wavelet coefficients, the real SAR image wavelet coefficients, and the noise wavelet coefficients, respectively; m and n represent the coordinates of the corresponding coefficient matrix in the horizontal and vertical directions, respectively;

[0014] Taking the logarithm of formula (1), the expression is as follows:

[0015] (2);

[0016] Among them, Y, X, and N represent the wavelet coefficients of the original SAR image after logarithmic transformation, the wavelet coefficients of the real SAR image, and the noise wavelet coefficients, respectively;

[0017] The model is decomposed by wavelet to obtain the low-frequency coefficient LL and high-frequency coefficient of the SAR image, where the high-frequency coefficient includes the horizontal high-frequency coefficient LH, the vertical high-frequency coefficient HL and the diagonal high-frequency coefficient HH. All high-frequency coefficients are subjected to threshold denoising. Specifically, the part of the high-frequency coefficient that is less than the threshold is set to zero to remove the noise, and the high-frequency coefficients after noise removal are added together to obtain the new high-frequency coefficients.

[0018] Preferably, in step S1, the threshold used in the denoising process is calculated according to the Bayesian shrinkage method, and the expression is as follows:

[0019] (4);

[0020] Where G represents the horizontal high frequency coefficient LH, the vertical high frequency coefficient HL or the diagonal high frequency coefficient HH, Represents the threshold corresponding to the high-frequency coefficient, represents the variance of the real SAR image wavelet coefficients, It represents the variance of the corresponding high-frequency coefficients after wavelet decomposition. The expression is as follows:

[0021] (5);

[0022] in, represents the median function, Represents the matrix corresponding to the high-frequency coefficient, that is, the horizontal high-frequency coefficient matrix , vertical high frequency coefficient matrix or the diagonal high-frequency coefficient matrix .

[0023] Preferably, step S2 includes:

[0024] Construct a dual-branch structure consisting of ConvNeXt and Swin Transformer, extract high-frequency features through ConvNeXt, and extract low-frequency features through Swin Transformer;

[0025] Construct a feature adaptive fusion network based on FGAM;

[0026] The high-frequency features extracted and low-frequency characteristics The spatial attention feature map X is added and input into FGAM. After the feature map X passes through the parallel spatial attention module and channel attention module, the corresponding weights are output respectively. and , the expression is as follows:

[0027] (6);

[0028] (7);

[0029] in, Indicates a channel connection operation, Represents the ReLU function activation operation, Indicates that the convolution kernel size is The convolution operation, 、 Respectively represent the features in the spatial dimension after the global average pooling operation and the global maximum pooling operation, Represents the features in the channel dimension after the global average pooling operation;

[0030] Based on weight and After the addition, the feature map X is subjected to a channel connection operation, and then a channel shuffling operation, a 7×7 convolution operation, and a Sigmoid function activation operation are performed in sequence to obtain a spatial attention feature map U with optimized feature weights.

[0031] Based on the feature weights in the feature map U, high-frequency features are added through weighted summation and residual connection. and low-frequency characteristics The added features are then projected through a 1×1 convolutional layer to obtain the final fusion features. , the expression is as follows:

[0032] (8);

[0033] in, W is the weight assigned to the feature map U, and I is a matrix whose elements are all 1.

[0034] It should be noted that feature fusion can integrate feature information at different levels in high-frequency features and low-frequency features, enrich the description of the target object, and thus enhance the network model's discrimination of target attributes. Generally speaking, the simplest feature fusion method is to implement it through element-by-element addition. However, this method is not flexible enough and cannot effectively capture the relationship between features. Some studies have also adopted the splicing method to achieve feature fusion. This method helps the model learn more complex information by increasing the dimension of the features, but it may also introduce some irrelevant information. More importantly, these two types of feature fusion methods do not take into account the problem of mismatched receptive fields of features with different frequency characteristics. For example, a pixel in a high-frequency feature may come from a pixel area in a low-frequency feature. Therefore, simple addition or splicing cannot solve the receptive field matching problem before feature fusion.

[0035] FGAM fully mixes channel attention weights and spatial attention weights through channel shuffling to promote information interaction, enhances the weight of important information in the feature map, and weakens the weight of unimportant information, which is conducive to improving target detection accuracy. Furthermore, FGAMF further refines the feature map according to the frequency characteristics of the input features, guiding the model to focus more on the key areas of concern, thereby improving the model's anti-clutter interference performance.

[0036] Preferably, in the dual-branch structure, Swin Transformer selects the Base version, and ConvNeXt selects the Base version.

[0037] Preferably, the dual-branch structure has four feature extraction levels, and the corresponding channel numbers are 128, 256, 512 and 1024 respectively.

[0038] Preferably, step S3 includes:

[0039] After multiple fusion features from different levels are subjected to a 1×1 convolution operation, they are interpolated with the convolution operation results of the previous level and added together. Each summation result is then subjected to a 3×3 convolution operation, and the corresponding feature map is output.

[0040] For those layers that do not have the previous level for interpolation and addition operations, directly take the convolution operation result and perform a convolution operation with a convolution kernel of 3×3, output the corresponding feature map, repeat the convolution operation with a convolution kernel of 3×3 on the feature map, and output the corresponding feature map.

[0041] The present invention has at least the following beneficial effects:

[0042] 1. Aiming at the speckle noise characteristics of SAR images, the present invention introduces a Bayesian shrinkage threshold denoising method based on wavelet transform. The SAR image is decomposed into low-frequency coefficients and high-frequency coefficients through wavelet transform. The low-frequency coefficients represent large-scale structural information, and the high-frequency coefficients contain detailed features and residual noise. The speckle noise contained in the high-frequency coefficients is regarded as Gaussian distribution, and the Bayesian shrinkage method is used to calculate the threshold for adaptive denoising. This method can better simulate the characteristics of speckle noise and remove noise.

[0043] 2. To address the clutter interference problem in SAR images, this paper introduces a feature adaptive fusion method based on the FGAM attention mechanism. This method can fully consider the receptive field matching problem of high-frequency features and low-frequency features, and adaptively assign appropriate weights to them according to the frequency characteristics of the input features. This not only reduces the information loss in the feature fusion process and ensures the integrity of high-frequency and low-frequency features, but also focuses on enhancing the feature response of the target-related area, thereby improving the model's anti-clutter interference performance.

[0044] 3. In view of the different information contained in the high-frequency coefficients and low-frequency coefficients obtained by wavelet decomposition, the present invention designs a dual-branch structure of ConvNeXt and Swin Transformer to extract high-frequency features and low-frequency features respectively. Among them, Swin Transformer adopts a window attention mechanism and hierarchical feature processing method to enable the model to obtain global attention while greatly reducing the computational complexity of the model, reducing the computational complexity to a linear relationship with the image size. ConvNeXt combines the advantages of traditional convolutional neural networks and Transformer models, aiming to provide more efficient computing performance. At the same time, through deep convolution, the model can process input features in a finer-grained manner and retain a higher accuracy. With the CNN branch focusing on high-frequency detail learning, the Transformer branch captures low-frequency global features, which can avoid feature competition and maximize the utilization of prior knowledge.

[0045] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements.

[0047] Figure 1 This is an architecture diagram of a SAR target detection method based on wavelet decomposition and attention feature fusion according to Example 1 of the present invention;

[0048] Figure 2 This is a comparison diagram of the results of threshold denoising using the Bayesian shrinkage method in Example 1 of the present invention, where (a) is the original image and (b) is the denoised image;

[0049] Figure 3 for Figure 2 The operation flow chart from (a) to (b);

[0050] Figure 4 4 is an architecture diagram of a feature adaptive fusion network based on FGAM in Example 1 of the present invention;

[0051] Figure 5 for Figure 4 The architecture diagram of the FGAM network;

[0052] Figure 6 This is an architecture diagram of the FPN network in Example 1 of the present invention;

[0053] Figure 71. A comparison of the results of SAR ship target detection in Example 2 of the present invention, where (a) uses the YOLOv8 algorithm and (b) uses the method of the present invention;

[0054] Figure 8 1 is a comparison chart of the results of SAR vehicle target detection in Example 2 of the present invention, where (a) adopts the YOLOv8 algorithm and (b) adopts the method of the present invention. DETAILED DESCRIPTION

[0055] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0056] Example 1

[0057] The idea of ​​the technical solution of the present invention is as follows Figure 1 As shown in the figure, the input SAR image is decomposed into high-frequency coefficients and low-frequency coefficients by wavelet transform method, the noise in the high-frequency coefficients is removed by Bayesian shrinkage method, and the high-frequency coefficients after noise removal are added and recombined into new high-frequency coefficients. For the high-frequency coefficients and low-frequency coefficients obtained after wavelet decomposition and denoising, the present invention designs a dual-branch structure consisting of ConvNeXt branch and Swin Transformer branch, aiming to utilize the powerful feature extraction capabilities of ConvNeXt for high-frequency features and Swin Transformer for low-frequency features to fully extract target information and improve target detection accuracy. The extracted features are adaptively fused based on the Feature-guided Attention Mechanism (FGAM), so that the weight distribution can be optimized more reasonably to reduce information loss in the feature fusion process. The fused features are passed through Feature Pyramid Networks (FPN) and Region Proposal Network (RPN) to output the target detection results.

[0058] The technical details of the present invention will be described in more detail below with reference to derivations and formulas.

[0059] S1. Perform wavelet decomposition on the original SAR image to obtain low-frequency coefficients and high-frequency coefficients, and use Bayesian threshold shrinkage to denoise the high-frequency coefficients. Specifically:

[0060] Since the speckle noise of SAR images is multiplicative noise, the following model can be constructed for the original SAR image:

[0061] (1);

[0062] Among them, y, x, and e represent the original SAR image wavelet coefficients, the real SAR image wavelet coefficients, and the noise wavelet coefficients, respectively; m and n represent the coordinates of the corresponding coefficient matrix in the horizontal and vertical directions, respectively;

[0063] In order to facilitate subsequent processing, we take the logarithm of equation (1) and convert the multiplicative noise into additive noise. The expression is as follows:

[0064] (2);

[0065] Among them, Y, X, and N represent the wavelet coefficients of the original SAR image after logarithmic transformation, the wavelet coefficients of the real SAR image, and the noise wavelet coefficients, respectively;

[0066] Since there is independence between the input original SAR image and the noise, it can be obtained from formula (2):

[0067] (3);

[0068] in, 、 、 They represent the variance of the original SAR image wavelet coefficients, the variance of the real SAR image wavelet coefficients, and the variance of the noise wavelet coefficients respectively;

[0069] After taking the logarithm operation, the multiplicative noise model that obeys the Gamma distribution can be approximated to a Gaussian distribution;

[0070] The model is decomposed by wavelet to obtain the low-frequency coefficient LL and high-frequency coefficient of the SAR image. The high-frequency coefficient includes the horizontal high-frequency coefficient LH, the vertical high-frequency coefficient HL and the diagonal high-frequency coefficient HH. All high-frequency coefficients are subjected to threshold denoising. Specifically, the part of the high-frequency coefficient that is smaller than the threshold is set to zero to remove the noise. The high-frequency coefficients after noise removal are added together to obtain the new high-frequency coefficients.

[0071] In the above denoising process, the threshold is calculated according to the principle of Bayesian shrinkage method, and the expression is as follows:

[0072] (4);

[0073] Where G represents the horizontal high frequency coefficient LH, the vertical high frequency coefficient HL or the diagonal high frequency coefficient HH, Represents the threshold corresponding to the high-frequency coefficient, Represents the variance of the wavelet coefficients of the real SAR image, which is obtained by calculating the matrix variance formula. It represents the variance of the corresponding high-frequency coefficients after wavelet decomposition. The expression is as follows:

[0074] (5);

[0075] in, represents the median function, Represents the matrix corresponding to the high-frequency coefficient, that is, the horizontal high-frequency coefficient matrix , vertical high frequency coefficient matrix or the diagonal high-frequency coefficient matrix .

[0076] In order to verify the effectiveness of the noise removal method in step S1, an original SAR image is selected from the data set for denoising. The low-frequency coefficients and new high-frequency coefficients obtained by wavelet decomposition are recombined and restored by inverse wavelet transform to obtain the denoised SAR image. The results are shown in Figure 2. Figure 2 As shown in the figure, the presence of speckle noise will cause the originally clear outlines or textures in the image to be submerged by noise, resulting in a significant decrease in image contrast. However, the SAR image obtained after wavelet decomposition and denoising has a higher contrast and a significant denoising effect. The above-mentioned decomposition-denoising-recombination-restoration process is as follows: Figure 3 shown.

[0077] S2. Use ConvNeXt and Swin Transformer to extract high-frequency features and low-frequency features at different levels. Each level performs feature adaptive fusion based on FGAM. Specifically:

[0078] Construct a dual-branch structure consisting of ConvNeXt and Swin Transformer; Swin Transformer is a universal backbone network designed based on the Transformer architecture, and includes four versions: Tiny (T), Small (S), Base (B), and Large (L). In this invention, Swin Transformer-B is selected as the backbone network for low-frequency feature extraction to balance the model's performance and computational efficiency; ConvNeXt is an efficient convolutional neural network (CNN) model, which also includes four versions: T, S, B, and L. In order to correspond to Swin Transformer, ConvNeXt-B is selected as the backbone network for high-frequency feature extraction in this invention;

[0079] Build as Figure 4 The feature adaptive fusion network based on FGAM shown in Figure 1 The FGAMF module in the FGAM has the following structure: Figure 5 As shown, the high-frequency features extracted and low-frequency characteristics The spatial attention feature map X is added and input into FGAM. After the feature map X passes through the parallel spatial attention module and channel attention module, the corresponding weights are output respectively. and , the expression is as follows:

[0080] (6);

[0081] (7);

[0082] in, Indicates a channel connection operation, Represents the ReLU function activation operation, Indicates that the convolution kernel size is The convolution operation, 、 Respectively represent the features in the spatial dimension after the global average pooling operation and the global maximum pooling operation, Represents the features in the channel dimension after the global average pooling operation;

[0083] Based on weight and After the addition, the feature map X is subjected to a channel connection operation, and then a channel shuffling operation, a 7×7 convolution operation, and a Sigmoid function activation operation are performed in sequence to obtain a spatial attention feature map U with optimized feature weights.

[0084] Based on the feature weights in the feature map U, high-frequency features are added through weighted summation and residual connection. and low-frequency characteristics The added features are then projected through a 1×1 convolutional layer to obtain the final fusion features. , the expression is as follows:

[0085] (8);

[0086] in, W is the weight assigned to the feature map U, and I is a matrix with all elements equal to 1;

[0087] In the present invention, the extraction of high-frequency features and low-frequency features by the dual-branch structure is set to four different levels, and the corresponding channel numbers are 128, 256, 512 and 1024 respectively, that is, four fusion features C1-C4 are output.

[0088] S3: Input the fused features of each level into FPN for feature fusion again to obtain multiple feature maps with the same number of channels, specifically:

[0089] Since the FPN structure can achieve a mixture of "coarse-grained" features and "fine-grained" features by promoting cross-level feature fusion, FPN is introduced in the present invention. The key to achieving feature fusion in FPN lies in upsampling and downsampling operations, where upsampling refers to interpolation of the input to improve resolution, and downsampling refers to reducing the data dimension from the original data through convolution operations. The original SAR image of the present invention is output as a list of gradually smaller feature maps after the above-mentioned wavelet decomposition, Bayesian threshold denoising, and dual-branch structure, which is equivalent to having completed the downsampling process.

[0090] Therefore, if Figure 6 As shown in the figure, the fused features C1-C4 from four different levels are subjected to a convolution operation with a convolution kernel of 1×1, and then interpolated and added with the convolution operation results of the previous level in sequence to obtain the corresponding addition operation results M1-M4. Among them, M4 has no previous level to perform interpolation and addition operations, and the convolution operation result of C4 is directly taken;

[0091] For M1-M3, a convolution operation with a convolution kernel of 3×3 is performed, and the corresponding feature maps P1-P3 are output. For M4, a convolution operation with a convolution kernel of 3×3 is performed, and the corresponding feature map P4 is output. P4 repeats the convolution operation with a convolution kernel of 3×3 and outputs the corresponding feature map P5. The number of channels of feature maps P1-P5 is the same.

[0092] S4. Input each feature map into RPN to obtain the target detection result, specifically:

[0093] The five feature maps P1-P5 with the same number of channels are input into the detection heads of five identical RPNs. Each branch performs foreground and background classification and bbox regression. The loss value is calculated based on the IoU value between the regressed bbox and the ground truth value in the classification label. The model will iteratively optimize the object detection model weights by returning the gradient based on the loss in multiple epochs (that is, a complete traversal of the model over the entire training dataset) to make the predicted bbox and category close to the ground truth value.

[0094] Example 2

[0095] This example verifies the SAR target detection method in Example 1. A dataset for this experiment is constructed by collecting data on-site in Changde, Hunan. The dataset contains a total of 7,270 SAR images, and the data is divided into a training set: test set: validation set ratio of 8:1:1.

[0096] In this example, a self-made SAR data set is used as the target detection data set, and YOLOv8 is used as the comparative experimental algorithm to perform SAR target detection on the method of the present invention and the YOLOv8 algorithm. The results are shown in the figure. Figure 7 and Figure 8 As shown, specifically: Figure 7 In the example, YOLOv8 mistakenly identified the ship as a bridge, while the proposed method correctly identified it. Figure 8 In the results, the completeness and accuracy of vehicle recognition by YOLOv8 are obviously inferior to those of the method of the present invention. It can be seen that the method of the present invention has significant advantages in detecting small targets in SAR images.

[0097] While various embodiments of the present disclosure have been described above, the foregoing description is intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0098] The above are merely optional embodiments of the present disclosure and are not intended to limit the present disclosure. Those skilled in the art will readily appreciate that the present disclosure may be modified and varied in various ways. Any modifications, equivalent substitutions, improvements, and the like made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

Claims

1. A SAR target detection method based on wavelet decomposition and attention feature fusion, characterized in that: The following steps are involved: S1. Perform wavelet decomposition on the original SAR image to obtain low-frequency coefficients and high-frequency coefficients, and use Bayesian threshold shrinkage to denoise the high-frequency coefficients; S2, using the dual-branch structure of ConvNeXt and Swin Transformer to extract high-frequency features and low-frequency features at different levels, and performing feature adaptive fusion based on FGAM at each level; The feature adaptive fusion based on FGAM includes: The high-frequency features extracted and low-frequency characteristics The spatial attention feature map X is added and input into FGAM. After the feature map X passes through the parallel spatial attention module and channel attention module, the corresponding weights are output respectively. and , the expression is as follows: （6）； （7）； in, Indicates a channel connection operation, Represents the ReLU function activation operation, Indicates that the convolution kernel size is The convolution operation, 、 Respectively represent the features in the spatial dimension after the global average pooling operation and the global maximum pooling operation, Represents the features in the channel dimension after the global average pooling operation; Based on weight and After the addition, the feature map X is subjected to a channel connection operation, and then a channel shuffling operation, a 7×7 convolution operation, and a Sigmoid function activation operation are performed in sequence to obtain a spatial attention feature map U with optimized feature weights. Based on the feature weights in the feature map U, high-frequency features are added through weighted summation and residual connection. and low-frequency characteristics The added features are then projected through a 1×1 convolutional layer to obtain the final fusion features. , the expression is as follows: （8）； in, W is the weight assigned to the feature map U, and I is a matrix with all elements equal to 1; S3, input the fusion features of each level into FPN for feature fusion again to obtain multiple feature maps with the same number of channels; S4. Input each feature map into RPN to obtain the target detection result.

2. The SAR target detection method based on wavelet decomposition and attention feature fusion according to claim 1, characterized in that: The step S1 comprises: The following model is constructed for the original SAR image: （1）； Among them, y, x, and e represent the original SAR image wavelet coefficients, the real SAR image wavelet coefficients, and the noise wavelet coefficients, respectively; m and n represent the coordinates of the corresponding coefficient matrix in the horizontal and vertical directions, respectively; Taking the logarithm of formula (1), the expression is as follows: （2）； Among them, Y, X, and N represent the wavelet coefficients of the original SAR image after logarithmic transformation, the wavelet coefficients of the real SAR image, and the noise wavelet coefficients, respectively; The model is decomposed by wavelet to obtain the low-frequency coefficient LL and high-frequency coefficient of the SAR image, where the high-frequency coefficient includes the horizontal high-frequency coefficient LH, the vertical high-frequency coefficient HL and the diagonal high-frequency coefficient HH. All high-frequency coefficients are subjected to threshold denoising. Specifically, the part of the high-frequency coefficient that is less than the threshold is set to zero to remove the noise, and the high-frequency coefficients after noise removal are added together to obtain the new high-frequency coefficients.

3. The SAR target detection method based on wavelet decomposition and attention feature fusion according to claim 2, characterized in that: In step S1, the threshold used in the denoising process is calculated according to the Bayesian shrinkage method, and the expression is as follows: （4）； Where G represents the horizontal high frequency coefficient LH, the vertical high frequency coefficient HL or the diagonal high frequency coefficient HH, Represents the threshold corresponding to the high-frequency coefficient, represents the variance of the real SAR image wavelet coefficients, It represents the variance of the corresponding high-frequency coefficients after wavelet decomposition. The expression is as follows: （5）； in, represents the median function, Represents the matrix corresponding to the high-frequency coefficient, that is, the horizontal high-frequency coefficient matrix , vertical high frequency coefficient matrix or the diagonal high-frequency coefficient matrix .

4. The SAR target detection method based on wavelet decomposition and attention feature fusion according to claim 3, characterized in that: In the dual-branch structure, Swin Transformer selects the Base version, and ConvNeXt selects the Base version.

5. The SAR target detection method based on wavelet decomposition and attention feature fusion according to claim 3 is characterized in that: The number of feature extraction levels of the dual-branch structure is four, and the corresponding channel numbers are 128, 256, 512 and 1024 respectively.

6. The SAR target detection method based on wavelet decomposition and attention feature fusion according to claim 3, characterized in that: The step S3 comprises: After multiple fusion features from different levels are subjected to a 1×1 convolution operation, they are interpolated with the convolution operation results of the previous level and added together. Each summation result is then subjected to a 3×3 convolution operation, and the corresponding feature map is output. For those layers that do not have the previous level for interpolation and addition operations, directly take the convolution operation result and perform a convolution operation with a convolution kernel of 3×3, output the corresponding feature map, repeat the convolution operation with a convolution kernel of 3×3 on the feature map, and output the corresponding feature map.

Images