Anchor-free SAR target detection method based on supervised attention mechanism

Abstract

The present invention discloses a method for anchor-frame-free SAR target detection based on a supervised attention mechanism, comprising: obtaining an original training set and an original test set based on the original SAR image; obtaining gradient information and CFAR information from the original training set and the original test set, respectively, and constructing a new training set and a new test set based on the obtained information; constructing an anchor-frame-free SAR target detection network based on the fusion of gradient information and CFAR information and a supervised attention mechanism; training the target detection network using the new training set; inputting the new test set into the trained target detection network to obtain preliminary target detection results; mapping the preliminary target detection results to the test image, and performing a NMS operation to remove overlapping target detection frames to obtain the final target detection results. This method can effectively alleviate the problems of computational complexity and positive and negative sample imbalance existing in the anchor-frame target detection network itself, thereby improving the SAR target detection performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and in particular relates to an anchor-free SAR target detection method based on a supervised attention mechanism. Background Art

[0002] Synthetic Aperture Radar (SAR), capable of operating day and night in all weather conditions, has been widely used in a variety of fields, including military reconnaissance, resource exploration, environmental protection, disaster prevention, and scientific research. Automatic Target Recognition (ATR) technology for SAR images is dedicated to locating and identifying targets in complex SAR scenes. It is a core area of ​​SAR imagery application and plays a vital role in both military and civilian fields. The Lincoln Laboratory in the United States pioneered this research and proposed the famous three-stage SAR ATR process in the late 1980s: SAR target detection, discrimination, and recognition. This is the most common processing flow for SAR image interpretation. SAR target detection is a crucial component of SAR image interpretation.

[0003] Since the advent of synthetic aperture radar, the technology of SAR target detection has also developed rapidly. In the early stage of technology development, a large number of scholars mainly studied traditional target detection methods. Among them, the most famous method is the dual-parameter constant false alarm rate (CFAR) detection method proposed by the Lincoln Laboratory in the United States in the 1990s. This method successfully extended the CFAR detector to the field of two-dimensional SAR image target detection. Due to the superior detection performance of the dual-parameter CFAR algorithm in some simple scenes, a large number of studies began to focus on the constant false alarm algorithm, and detectors such as minimum selection CFAR, maximum selection CFAR, unit average CFAR, and sequential statistics CFAR were proposed. Although traditional target detection methods can achieve good target detection performance in some simple scenes, they still have many disadvantages: (1) They usually involve a large number of hyperparameters and need to set thresholds to complete the target detection task. Therefore, in practice, they need to manually adjust the parameters according to the usage scenario. The process is cumbersome, it is difficult to achieve adaptive detection, and the target detection accuracy is low; (2) Traditional target detection methods are usually pixel-by-pixel detection, which takes a long time and is difficult to achieve good target detection performance in some complex scenes. Therefore, traditional target detection methods have gradually failed to meet actual usage needs.

[0004] In recent years, with the great success of deep networks in the field of optical image target detection, SAR target detection methods based on deep learning have become a hot topic of research for many scholars. They have been widely used in SAR target detection tasks, achieving SAR target detection performance far superior to traditional target detection methods. For example, patent CN202210269829.6 provides a dual-stream SSD SAR image target detection method based on CFAR guidance. It fully utilizes the strong scattering characteristics of targets in SAR images and enhances target detection performance by fusing SAR amplitude features and CFAR indicator features in feature space. It then uses the CFAR binary indicator map to make the detector pay more attention to the learning of difficult-to-distinguish negative and positive samples. Finally, the AR-NMS algorithm is proposed to improve the traditional NMS algorithm and improve SAR target detection performance.

[0005] However, SAR ground target data is relatively small and the scenes to be detected are complex, often containing a large amount of natural and artificial clutter that closely resembles the target's characteristics. This makes ground target detection a very difficult task. Currently, research on SAR ground target detection is relatively limited. Most mainstream deep network-based SAR ground target detection methods are based on target detection with anchor boxes. These target detection frameworks typically require a large number of anchor boxes to be pre-laid, involving numerous hyperparameter settings and complex IoU calculations. They also suffer from a serious imbalance between positive and negative samples, which is detrimental to SAR target detection. Furthermore, existing methods do not fully consider the inherent geometric distortion, radiometric distortion, and shadow occlusion of SAR images, nor do they consider how these characteristics can aid SAR target detection. This results in low detection performance for current SAR target detection methods in complex scenes. Summary of the Invention

[0006] In order to solve the above-mentioned problems existing in the prior art, the present invention provides an anchor-free SAR target detection method based on a supervised attention mechanism. The technical idea of ​​the present invention is: by using the exponentially weighted average (Ratio of Exponentially Weighted Averages, ROEWA) edge detection algorithm and the dual-parameter CFAR algorithm on the training samples, the gradient information and CFAR information of each training sample are obtained, and sent together to the constructed anchor-free SAR target detection network based on gradient information, CFAR information fusion and supervised attention mechanism for training, and then the test sample is processed in the same way and input into the trained network model to obtain the final target detection result. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] A supervised attention-based anchor-free SAR target detection method includes:

[0008] Step 1: Obtain the original training set and the original test set based on the original SAR image; use the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain gradient information and CFAR information from the original training set and the original test set respectively, and construct new training set and new test set accordingly;

[0009] Step 2: Construct an anchor-free SAR target detection network based on gradient information, CFAR information fusion, and supervised attention mechanism; wherein the target detection network includes a feature extraction module, a feature processing module, and a network prediction module;

[0010] Step 3: Using the new training set to train the target detection network to obtain a trained target detection network;

[0011] Step 4: Input the new test set into the trained object detection network to obtain preliminary object detection results;

[0012] Step 5: Map the preliminary target detection results to the test image and perform NMS operation to remove overlapping target detection frames to obtain the final target detection results.

[0013] Beneficial effects of the present invention:

[0014] 1. The anchor-free SAR target detection method based on the supervised attention mechanism provided by the present invention starts from the perspective of enhancing target features, introduces gradient amplitude information and CFAR information, and constructs an anchor-free target detection network based on the fusion of gradient information and CFAR information. It achieves the ground SAR target detection task in complex scenes, can effectively alleviate the problems of computational complexity and positive and negative sample imbalance existing in the anchor-free target detection network itself, and improves the SAR target detection performance.

[0015] 2. In terms of feature fusion, the target detection network designed by the present invention adopts a Triple-Interactive Channel-Spatial Attention Fusion (T-ICSAF) module based on the interactive attention mechanism. At the same time, a Combining Supervised-Spatial And SE-Channel Attention Mechanism (CSSCAM) module based on Ground Truth (GT) binary label supervision is proposed to effectively suppress background clutter features. This solves the problem of increased false alarm frames caused by the additional background clutter features introduced by traditional features, further enhances target features, and makes the extracted target features more targeted, thereby achieving better SAR target detection performance.

[0016] 3. Based on the characteristics of SAR images such as geometric distortion, radiation distortion, occlusion and shadow, the present invention designs an Attention binary classification branch based on GT binary label supervision to replace the Centerness branch in the original anchor-free target detection network FCOS, making it more suitable for SAR vehicle target detection tasks and can be used in conjunction with the CSSCAM module proposed in the present invention to further improve the SAR target detection performance.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a method for anchor-free SAR target detection based on a supervised attention mechanism provided by an embodiment of the present invention;

[0019] Figure 2 1 is another flowchart of the anchor-free SAR target detection method based on the supervised attention mechanism provided by an embodiment of the present invention;

[0020] Figure 3 This is a diagram of the anchor-free SAR target detection network framework provided by an embodiment of the present invention;

[0021] Figure 4 1 is a network framework diagram of a T-ICSAF feature fusion module provided by an embodiment of the present invention;

[0022] Figure 5 This is a network framework diagram of the CSSCAM module provided by an embodiment of the present invention;

[0023] Figure 6-11It is the MiniSAR data image used in the experiment of this invention. DETAILED DESCRIPTION

[0024] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0025] Example 1

[0026] Please see the joint Figure 1-2 , Figure 1 This is a flow chart of a method for anchor-free SAR target detection based on a supervised attention mechanism provided by an embodiment of the present invention. Figure 2 This is another flow chart of the anchor-free SAR target detection method based on the supervised attention mechanism provided by an embodiment of the present invention. The anchor-free SAR target detection method based on the supervised attention mechanism provided by the present invention specifically includes:

[0027] Step 1: Obtain the original training set and the original test set based on the original SAR image; use the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain the gradient information and CFAR information from the original training set and the original test set respectively, and construct a new training set and a new test set accordingly.

[0028] 1. Construct a new training set

[0029] First, several original SAR images are selected as training images and sliced ​​to obtain several training slices as the original training set ψ;

[0030] Then, the selected training images are respectively subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain the gradient amplitude training image and the CFAR binary training image;

[0031] Finally, the gradient magnitude training image and the CFAR binary training image are sliced ​​to obtain the gradient magnitude training slice corresponding to the original training set ψ and CFAR binary training slices Together with the original training set ψ, a new training set ψ' is formed.

[0032] Optionally, as an implementation method, this embodiment can select five images from the MiniSAR data set as training images, and obtain 3536 training slices with a step size of 45 and a slice size of 800×1333 to form the original training set ψ; then, the selected training images are subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain a gradient amplitude training image and a CFAR binary training image, and the gradient amplitude training slice corresponding to the original training set ψ is obtained with a step size of 45 and a slice size of 800×1333. and CFAR binary training slices Finally, together with the original training set ψ, it forms the final training set ψ'.

[0033] 2. Build a new test set

[0034] First, an original SAR image is selected as a test image and sliced ​​to obtain several test slices as the original test set T;

[0035] Then, the selected test image is subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain the gradient amplitude test image and the CFAR binary test image;

[0036] Finally, the gradient magnitude test image and the CFAR binary test image are sliced ​​to obtain the gradient magnitude test slice corresponding to the original test set T and CFAR binary test slices Together with the original test set T, a new test set T' is formed.

[0037] Optionally, as an implementation method, an image is selected from the MiniSAR data set as a test image, and 63 test slices are obtained with a step size of 150 and a slice size of 800×1333 to form the original test set T, and the positional relationship Loc corresponding to each test slice and the test image is recorded. Then, the selected test image is subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm respectively to obtain a gradient amplitude test image and a CFAR binary test image, and the gradient amplitude test slice corresponding to the original test set T is obtained with a step size of 150 and a slice size of 800×1333. and CFAR binary test slices Finally, together with the original test set T, the final test set T' is formed.

[0038] Furthermore, in the process of constructing the training set and the test set, it is necessary to use the ROEWA edge detection algorithm to obtain the gradient amplitude image of the SAR image. The operation process is the same, and the specific steps are as follows:

[0039] 1) Input the original SAR image;

[0040] 2) For any pixel point (i, j) in the input SAR image, the ROEWA algorithm is first used to calculate the horizontal gradient of the pixel point in the horizontal direction. and the vertical gradient in the vertical direction Then use the following formula to calculate the gradient magnitude G of pixel (i, j) i,j :

[0041]

[0042] Specifically, in the embodiment, the ROEWA algorithm is used to calculate the horizontal gradient of any pixel point (i, j) and vertical gradient The specific process is as follows:

[0043] a. Calculate horizontal gradient

[0044] For any pixel point (i, j), first calculate the exponential weighted mean M of the pixel amplitude values ​​within the range of (4σ+1)×2σ on the left and right sides of the pixel point L and M R , then M L and M R Take the quotient and logarithm to get the horizontal gradient Where σ is the exponential weighting factor. The calculation formula is as follows:

[0045]

[0046]

[0047]

[0048] Where I(·) represents the amplitude value of the pixel in the SAR image.

[0049] b. Calculate vertical gradient

[0050] For any pixel point (i, j), first calculate the exponentially weighted mean M of the pixel amplitude values ​​within the range of (4σ+1)×2σ on both sides of the pixel point T and M B , then M T and M B Take the quotient and logarithm to get the vertical gradient Where σ is the exponential weighting factor. Vertical gradient The calculation formula is as follows:

[0051]

[0052]

[0053]

[0054] Where I(·) represents the amplitude value of the pixel in the SAR image.

[0055] 3) Repeat the operation 2) to obtain the gradient amplitude of each pixel in the input SAR image, thereby obtaining the gradient amplitude image corresponding to the original SAR image.

[0056] Furthermore, in the process of constructing the training set and test set, it is also necessary to use the two-parameter CFAR algorithm to obtain the CFAR binary image of the SAR image. The operation steps are as follows:

[0057] 1) Input the original SAR image;

[0058] 2) For any pixel (i, j), use its amplitude I i,j Define the CFAR detection statistic:

[0059]

[0060] in and is the maximum likelihood estimate of the mean and standard deviation of the Gaussian distribution, which uses the amplitude value x of the background clutter pixel near the pixel point (i, j) i Calculated:

[0061]

[0062]

[0063] 3) The CFAR detection statistic D calculated in 2) is i,j Compared with the CFAR detection threshold T, if D i,j >T, then the pixel is judged as a target, and the CFAR binary detection result is 1, that is, the corresponding position in the CFAR binary image is set to 1; if D i,j <T, the pixel is judged as background, and the CFAR binary detection result is 0, that is, the corresponding position in the CFAR binary image is set to 0;

[0064] 4) Repeat the above 2) and 3) operations to obtain the CFAR binary detection result corresponding to each pixel in the input SAR image, thereby obtaining the CFAR binary image corresponding to the original SAR image.

[0065] Step 2: Construct an anchor-free SAR target detection network based on gradient information, CFAR information fusion and supervised attention mechanism.

[0066] See Figure 3 , Figure 3 This is a framework diagram of an anchor-free SAR target detection network provided by an embodiment of the present invention. The target detection network includes a feature extraction module, a feature processing module, and a network prediction module.

[0067] The following is a detailed introduction to the above three modules.

[0068] (1) Feature extraction module

[0069] like Figure 3 As shown in Figure 2, the feature extraction module includes three feature extraction sub-networks A, B, and C with exactly the same structure but no shared parameters, namely, amplitude feature extraction network A, gradient feature extraction network B, and CFAR feature extraction network C;

[0070] The three feature extraction sub-networks A, B, and C each include a feature extraction module with ResNet-18 as the backbone and an FPN multi-scale feature fusion module;

[0071] The three feature extraction modules are used to extract the training slices ψ, and Perform feature extraction to obtain the corresponding output feature layer and i=A,B,C;

[0072] The three FPN multi-scale feature fusion modules are used to output feature layers and Perform multi-scale feature fusion to obtain the output feature layer of the entire feature extraction module and

[0073] In this embodiment, the network structure, parameter settings, and corresponding relationships of ResNet-18 used in the feature extraction networks of the three feature extraction subnetworks A, B, and C are as follows:

[0074] The network structure of ResNet-18 used mainly includes five layers: conv1, conv2_x, conv3_x, conv4_x and conv5_x. The image size of the input network is H×W×3:

[0075] Conv1: includes a convolution layer L1, whose convolution kernel window size is 7×7, sliding step size is 2, padding attribute Padding is 3, used to output 64 Feature map of size j represents the jth feature map. The output of conv1 is the input of conv2_x;

[0076] conv2_x: includes a pooling layer P1, two convolution modules M1, M2 and two residual modules R1, R2. Specifically, P1 adopts the maximum pooling method, and its input is the feature layer X output by conv1 1 , the convolution kernel window size is 3×3, the sliding step is 2, and the padding attribute is 1, which is used to output 64 Feature map of size j represents the jth feature map. The output of P1 is the input of M1 and R1; M1 contains two convolutional layers L2 and L3, which are used to output 64 Feature map of size j represents the jth feature map. The convolution kernel window size of L2 and L3 is 3×3, the sliding step is 1, the padding attribute is 1, and the output of M1 is another input of R1; R1 contains an operation that adds the corresponding elements of the output feature layer of P1 and the output feature layer of M1 to output 64 Feature map of size j represents the jth feature map. The output of R1 is the input of M2 and R2; M2 also contains two convolutional layers L4 and L5, which are used to output 64 Feature map of size j represents the jth feature map. The convolution kernel window size of L4 and L5 is 3×3, the sliding step is 1, the padding attribute is 1, and the output of M2 is another input of R2; R2 contains an operation that adds the corresponding elements of the output feature layer of R1 and the output feature layer of M2 to output 64 Feature map of size j represents the jth feature map. The output of R2 is the input of conv3_x;

[0077] conv3_x: includes two convolution modules M3 and M4 and two residual modules R3 and R4. Specifically, M3 contains two convolution layers L6 and L7, which are used to output 128 Feature map of size j represents the jth feature map. The convolution kernel window size of L6 is 3×3, the sliding step size is 2, and the padding attribute is 1. The convolution kernel window size of L7 is 3×3, the sliding step size is 1, and the padding attribute is 1. The output of M3 is one input of R3; the other input of R3 is the feature layer X output of R2. 6 , R3 first sets X 6 Through a convolution layer L8 with a convolution kernel window size of 1×1, a sliding step size of 2, and a padding attribute of 0, it is used to output 128 Feature map of size j represents the jth feature map, and then the feature layer X output by L8 8 And the output feature layer X of M3 7 Perform the sum operation of the corresponding elements to output 128 Feature map of size The output of R3 is the input of R4 and M4; M4 contains two convolutional layers L9 and L 10 , used to output 128 Feature map of size j represents the jth feature map. 10 The convolution kernel window size is 3×3, the sliding step is 1, the padding attribute is 1, and the output of M4 is another input of R4; R4 contains an operation that adds the corresponding elements of the output feature layer of R3 and the output feature layer of M4 to output 128 Feature map of size j represents the jth feature map. The output of R4 is the input of conv4_x. In addition, the output of R4 is the output of the feature extraction network in A, B, and C. Feature layer (i = A, B, C), which is also an input to the corresponding FPN multi-scale feature fusion module;

[0078] conv4_x: includes two convolution modules M5, M6 and two residual modules R5, R6. Specifically, M5 includes two convolution layers L 11 and L 12 , used to output 256 Feature map of size j represents the jth feature map. 11 The convolution kernel window size is 3×3, the sliding step is 2, the padding attribute is 1, and L 12 The convolution kernel window size is 3×3, the sliding step is 1, the padding attribute is 1, the output of M5 is one input of R5; the other input of R5 is the feature layer X output of R4 11 , R5 first sets X 11 Through a convolution layer L with a convolution kernel window size of 1×1, a sliding step of 2, and a padding attribute of 0 13 , used to output 256 Feature map of size j represents the jth feature map, and then L 13 Output feature layer X 13 And the output feature layer X of M5 12 Perform the sum operation of the corresponding elements to output 256 Feature map of size j represents the jth feature map. The output of R5 is the input of R6 and M6; M6 contains two convolutional layers L 14 and L 15 , used to output 256 Feature map of size j represents the jth feature map. 14 and L 15The convolution kernel window size is 3×3, the sliding step is 1, and the padding attribute is 1. The output of M6 is another input of R6; R6 contains an operation to add the corresponding elements of the output feature layer of R5 and the output feature layer of M6 to output 256 Feature map of size j represents the jth feature map. The output of R6 is the input of conv5_x. In addition, the output of R6 is the output of the feature extraction network in A, B, and C. Feature layer (i = A, B, C), which is also an input to the corresponding FPN multi-scale feature fusion module;

[0079] conv5_x: includes two convolution modules M7, M8 and two residual modules R7 and R8. M7 includes two convolution layers L 16 and L 17 , used to output 512 Feature map of size j represents the jth feature map. 16 The convolution kernel window size is 3×3, the sliding step is 2, the padding attribute is 1, and L 17 The convolution kernel window size is 3×3, the sliding step is 1, the padding attribute is 1, the output of M7 is the input of R7; the other input of R7 is the output feature layer X of R6 16 , R7 first sets X 16 Through a convolution layer L with a convolution kernel window size of 1×1, a sliding step of 2, and a padding attribute of 0 18 , used to output 512 Feature map of size j represents the jth feature map, and then L 18 Output feature layer X 18 And the feature layer X output by M7 17 Perform the sum operation of the corresponding elements to output 512 Feature map of size The output of R7 is the input of R8 and M8; M8 contains two convolutional layers L 19 and L 20 , used to output 512 Feature map of size j represents the jth feature map. 19 and L 20 The convolution kernel window size is 3×3, the sliding step is 1, the padding attribute is 1, and the output of M8 is another input of R8; R8 contains an operation that adds the corresponding elements of the output feature layer of R7 and the output feature layer of M8 to output 512 Feature map of size j represents the jth feature map. The output of R8 is the output of the feature extraction network in A, B, and C. Feature layer (i = A, B, C), and it is also an input to the corresponding FPN multi-scale feature fusion module.

[0080] Furthermore, the three feature extraction sub-networks A, B, and C use the FPN module for multi-scale feature fusion. The implementation method, parameter settings, and corresponding relationships of the FPN multi-scale feature fusion module are as follows:

[0081] One input of the FPN multi-scale feature fusion module in A, B, and C is the output feature layer of the corresponding feature extraction network and Taking the feature extraction subnetwork A as an example, the input of its corresponding FPN module is and

[0082] The FPN multi-scale feature fusion module used mainly consists of three parts: FPN1 module, FPN2 module and FPN3 module:

[0083] FPN1 module: The input of FPN1 module is Feature layer, the feature layer size is The specific implementation of this module is as follows: First, Perform a convolution operation with a convolution kernel window size of 1×1, a sliding step size of 1, and a padding attribute of 0, and get 256 Feature map of size Then the obtained feature layer X 22 Perform an upsampling operation to obtain 256 Feature map of size This is the output of the FPN1 module. In addition, the output of the FPN1 module is an input to the FPN2 module;

[0084] FPN2 module: The input of FPN2 module is Feature layer and the output feature layer X of the FPN1 module 23 , and X 23 The feature layer sizes are The specific implementation of this module is as follows: First, Perform a convolution operation with a convolution kernel window size of 1×1, a sliding step size of 1, and a padding attribute of 0, and get 256 Feature map of size Then the obtained feature layer X 24 and X23 Directly add and fuse to get 256 Feature map of size Finally, the feature layer X 25 Perform an upsampling operation to obtain 256 Feature map of size This is the output of the FPN2 module. In addition, the output of the FPN2 module is an input to the FPN3 module;

[0085] FPN3 module: The input feature layer of the FPN3 module is Feature layer and the output feature layer X of the FPN2 module 26 , and X 26 The sizes are and The specific implementation of this module is as follows: First, Perform a convolution operation with a convolution kernel window size of 1×1, a sliding step size of 1, and a padding attribute of 0, and get 256 Feature map of size Then the obtained feature layer X 27 and X 26 Directly add and fuse to get 256 Feature map of size Finally, the feature layer X 28 Perform another convolution operation with a convolution kernel window size of 3×3, a sliding step size of 1, and a padding attribute of 1, and obtain 256 Feature map of size This is the output of the FPN3 module. In addition, the output of the FPN3 module is the output feature layer of the entire FPN module.

[0086] (2) Feature processing module

[0087] like Figure 3 As shown, in this embodiment, the feature processing module includes a T-ICSAF feature fusion module based on an interactive attention mechanism and a CSSCAM module based on a combination of a spatial attention mechanism and a channel attention mechanism supervised by GT binary labels; wherein,

[0088] The T-ICSAF feature fusion module is used to integrate the output feature layer of the feature extraction module. and Perform fusion to obtain fusion feature F3;

[0089] The CSSCAM module is used to process the fusion feature F3 to obtain a feature layer A3 in which the target feature is enhanced and the background clutter feature is suppressed, which is used as the output of the feature processing module.

[0090] See Figure 4 , Figure 4 This is a network framework diagram of the T-ICSAF feature fusion module provided by an embodiment of the present invention, wherein the T-ICSAF feature fusion module mainly includes four sub-modules, namely, a feature preprocessing sub-module, an interactive channel attention sub-module, an interactive spatial attention sub-module, and an attention fusion sub-module.

[0091] In this embodiment, the feature preprocessing submodule is used to convert the output feature layer of the feature extraction part into and Convolution operation and BN batch normalization operation are performed respectively, and three feature layers X are obtained accordingly. 30 、X 31 and X 32 .

[0092] Specifically, the feature preprocessing submodule is mainly implemented in the following ways:

[0093] The output feature layer of the feature extraction part and Perform a convolution operation with a window size of 3×3, a sliding step size of 1, a padding attribute of 1, and a BN batch normalization operation to obtain 3 Feature layer of size X 30 、X 31 and X 32 , which is the output of the feature preprocessing submodule. In addition, the output of the feature preprocessing submodule is X 30 、X 31 and X 32 The feature layer is the input of the interactive channel attention submodule and the interactive spatial attention submodule, and the feature preprocessing submodule outputs X 30 The feature layer is an input to the attention fusion submodule.

[0094] Furthermore, the interactive channel attention submodule is used to 30 、X 31 and X 32 Perform global average pooling, summation, and Sigmoid normalization operations in sequence to obtain the channel attention weight F c .

[0095] In this embodiment, the interactive channel attention submodule is mainly implemented in the following manner:

[0096] First, the output of the feature preprocessing submodule X30 、X 31 and X 32 The feature layer performs a global average pooling operation on the spatial dimension to obtain three feature vectors X of size 1×1×256. 33 、X 34 and X 35 ; Then the obtained X 33 、X 34 and X 35 Perform element-by-element summation and a Sigmoid normalization operation to obtain a channel attention weight F of size 1×1×256 c , which is the output of the interactive channel attention submodule. In addition, the output of the interactive channel attention submodule is an input of the attention fusion submodule.

[0097] Furthermore, the interactive spatial attention submodule is used to 30 、X 31 and X 32 Perform global average pooling, stacking (Concat), convolution and Sigmoid normalization operations on the channel dimension in sequence to obtain the spatial attention weight F s .

[0098] In this embodiment, the interactive spatial attention submodule is mainly implemented in the following manner:

[0099] First, the output of the feature preprocessing submodule X 30 、X 31 and X 32 The feature layer performs a global average pooling operation on the channel dimension, and obtains 3 Size characteristics Figure X 36 、X 37 and X 38 ; Then the obtained X 36 、X 37 and X 38 Perform a Concat operation and get 1 Size characteristics Figure X 39 ; Then get X 39 Perform a convolution operation with a window size of 3×3, a sliding step size of 1, a padding attribute of 1, and a Sigmoid normalization operation to obtain a The spatial attention weight F of size s , which is the output of the interactive spatial attention submodule. In addition, the output of the interactive spatial attention submodule is an input of the attention fusion submodule.

[0100] In addition, the attention fusion submodule is used to integrate the channel attention weights F c The feature layer X output by the feature preprocessing submodule 30 Each channel of the corresponding multiplication is obtained to obtain the feature layer F'3; then the spatial attention weight F s Multiplying the pixels of each channel of the feature layer F'3 one by one, we get the fused feature layer F3, which is the output of the T-ICSAF feature fusion module.

[0101] In this embodiment, the attention fusion submodule is mainly implemented in the following ways:

[0102] First, the channel attention weight F output by the interactive channel attention submodule is c and the output of the feature preprocessing submodule X 30 Each channel of The feature layer F'3 of size; then the spatial attention weight F output by the interactive spatial attention submodule s Multiply the pixels of each channel of the feature layer F'3 one by one, and finally get a The fused feature layer F3 of large and small sizes is the output of the T-ICSAF feature fusion module. In addition, the output of the T-ICSAF feature fusion module is the input of the feature processing module CSSCAM module.

[0103] For further information, see Figure 5 , Figure 5 : This is a network framework diagram of the CSSCAM module provided by an embodiment of the present invention, wherein the CSSCAM module includes three submodules, namely, the spatial attention mechanism submodule for GT binary label supervision, the SE channel attention mechanism submodule, and the attention fusion submodule; wherein,

[0104] The spatial attention mechanism submodule of GT binary label supervision is used to perform global average pooling operation on the channel dimension, global maximum pooling operation on the channel dimension, Concat operation, convolution and Sigmoid normalization operation on the fusion feature layer F3 output by the T-ICSAF feature fusion module to obtain the spatial attention weight A. s .

[0105] Specifically, the spatial attention mechanism submodule of the GT binary label supervision first performs a global average pooling operation on the channel dimension and a global maximum pooling operation on the channel dimension on the fusion feature F3 output by the T-ICSAF feature fusion module, and obtains two Size characteristics Figure X 40 、X 41 ; Then the obtained X 40 、X 41Perform a Concat operation and get 1 Size characteristics Figure X 42 ; Then get X 42 Perform a convolution operation with a convolution kernel window size of 7×7, a sliding step size of 1, a padding attribute of 3, and a Sigmoid normalization operation to obtain a The spatial attention weight of size A s , which is the output of the spatial attention mechanism submodule supervised by the GT binary label.

[0106] The SE channel attention mechanism submodule is used to perform global average pooling and global maximum pooling operations in the spatial dimension, feature vector compression, mapping and decompression operations, summation, and Sigmoid normalization operations on the fusion feature layer F3 output by the T-ICSAF feature fusion module to obtain the channel attention weight A. c .

[0107] Specifically, the SE channel attention mechanism module first performs a global average pooling operation in the spatial dimension and a global maximum pooling operation in the spatial dimension on the fusion feature F3 output by the T-ICSAF feature fusion module, and obtains two feature vectors X of size 1×1×256. 43 、X 44 ; Then the obtained X 43 、X 44 Pass through a fully connected layer L with 16 neurons 21 , a relu activation function layer and a fully connected layer L with 256 neurons 22 Perform compression, mapping, and decompression operations on the feature vector to obtain two 1×1×256 feature vectors X after channel attention. 45 、X 46 , where the fully connected layer L 21 The input feature vector is compressed into a feature vector of size 1×1×16, and the fully connected layer L 22 Decompress the input feature vector back to a feature vector of size 1×1×256; finally, the obtained X 45 、X 46 Perform element-by-element summation and a Sigmoid normalization operation to obtain a channel attention weight A of size 1×1×256 c , which is the output of the SE channel attention mechanism sub-module.

[0108] The attention fusion submodule is used to integrate the spatial attention weight A sMultiply the pixels of each channel of the fusion feature layer F3 output by the T-ICSAF feature fusion module one by one to obtain the feature layer A'3, and then add the channel attention weight A c Multiplying each channel of the feature layer A'3 accordingly, we get a feature layer A3 in which the target feature is enhanced and the background clutter feature is suppressed, which is the output of the CSSCAM module.

[0109] Specifically, first, the spatial attention weight A output by the spatial attention mechanism submodule of the GT binary label supervision is s Multiply the pixels of each channel of the fusion feature F3 output by the T-ICSAF feature fusion module one by one to obtain a The feature layer A'3 of size; then the channel attention weight A output by the SE channel attention mechanism submodule c Multiplying each channel of the feature layer A3' accordingly, we get a target feature that is enhanced and background clutter features that are suppressed. The feature layer A3 of size is the output of the CSSCAM module. In addition, the output of the CSSCAM module is the input of the network prediction part.

[0110] In addition, it should be noted that the target detection network constructed by the present invention also includes a supervised label generation module for generating GT binary labels.

[0111] In this embodiment, the method for constructing the GT binary label is as follows:

[0112] a) Perform true annotation on the training image, labeling the target pixels as 1 and the background pixels as 0, to obtain the labeled image corresponding to the training image;

[0113] b) Using a step size of 45 and a size of 800×1333, we obtain 3536 binary labeled slices GT corresponding to the original training set ψ;

[0114] c) Downsample the 3536 binary labeled slices GT obtained in b) three times to obtain 3536 binary labeled slices GT' of the same size as the feature layer fed into the CSSCAM module and the network prediction part, and save the .mat file corresponding to the binary labeled slices GT' as the final GT binary labels.

[0115] (3) Network prediction module

[0116] Please continue to see Figure 3 , where the network prediction module includes a classification branch submodule for target detection tasks, a regression branch submodule, and an Attention binary classification branch submodule based on GT binary label supervision;

[0117] The classification branch submodule and the regression branch submodule are used to predict the output feature layer A3 of the feature processing module, and obtain the classification score and bounding box regression parameters respectively;

[0118] The Attention two-class branch submodule is used to predict the feature layer C of the classification score. P and a feature layer R for predicting bounding box regression parameters P Processing is performed to obtain the binary classification score.

[0119] In this embodiment, the classification branch submodule is implemented in the following manner:

[0120] First, the feature layer A3 output by the CSSCAM module in the feature processing part is subjected to four convolution operations to obtain a feature layer C for predicting the classification score. P ; Then the feature layer C P Perform convolution and Sigmoid normalization operations in sequence to obtain the classification score predicted by the classification branch submodule Figure X 47 .

[0121] Specifically, the feature layer A3 output by the CSSCAM module of the feature processing part is first subjected to four convolution operations with a window size of 3×3, a sliding step of 1, and a padding attribute of 1, to obtain a prediction score. Feature layer C of size P ; Then get the feature layer C P Perform another convolution operation with a window size of 3×3, a sliding step size of 1, a padding attribute of 1, and a Sigmoid normalization operation to obtain a Classification scores predicted by the classification branch submodule of size Figure X 47 .

[0122] In this embodiment, the regression branch submodule is implemented in the following manner:

[0123] First, the feature layer A3 output by the CSSCAM module in the feature processing part is subjected to four convolution operations to obtain a feature layer R for predicting the bounding box regression parameters. P ; Then the feature layer R P Perform a convolution operation to obtain the bounding box regression parameter X predicted by the regression branch submodule 48 .

[0124] Specifically, the feature layer A3 output by the CSSCAM module of the feature processing part is first subjected to four convolution operations with a window size of 3×3, a sliding step of 1, and a padding attribute of 1, to obtain a convolution operation for predicting the bounding box regression parameters. The feature layer R of size P ; Then get the feature layer R P Perform another convolution operation with a convolution kernel window size of 3×3, a sliding step size of 1, and a padding attribute of 1 to obtain a The bounding box regression parameters X predicted by the regression branch submodule of size 48 .

[0125] In this embodiment, the Attention binary classification branch submodule is implemented as follows:

[0126] First, the feature layer C used to predict the classification score in the classification branch submodule is P And the feature layer R used to predict the bounding box regression parameters in the regression branch submodule P Perform a Concat operation and get the features Figure X 49 Perform convolution operation to obtain feature layer A P ; Finally, feature layer A P Perform convolution and Sigmoid normalization operations to obtain the binary classification scores predicted by the Attention binary classification branch submodule Figure X 50 .

[0127] Specifically, first, the feature layer C used to predict the classification score in the classification branch submodule P And the feature layer R used to predict the bounding box regression parameters in the regression branch submodule P Perform a Concat operation and get 1 Size characteristics Figure X 49 ; Then get X 49 Perform a convolution operation with a convolution kernel window size of 3×3, a sliding step size of 1, and a padding attribute of 1 to obtain a Size feature layer A P , and finally the feature layer A P Perform a convolution operation with a convolution kernel window size of 3×3, a sliding step size of 1, and a padding attribute of 1, and perform a Sigmoid normalization operation to obtain a The binary classification score predicted by the Attention binary classification branch submodule of size Figure X 50 .

[0128] In terms of feature fusion, the target detection network designed in the present invention adopts a feature fusion module T-ICSAF based on a three-branch interactive attention mechanism. At the same time, an attention mechanism module CSSCAM based on GT binary label supervision is proposed to effectively suppress background clutter features, solving the problem of increased false alarm frames caused by additional background clutter features introduced by traditional features, further enhancing target features, and making the target features after feature extraction more targeted, thereby achieving better SAR target detection performance.

[0129] In addition, based on the characteristics of SAR images such as geometric distortion, radiation distortion, and occlusion and shadow, the present invention designs an Attention binary classification branch based on GT binary label supervision to replace the Centerness branch in the original anchor-free target detection network FCOS, making it more suitable for SAR vehicle target detection tasks and can be used in conjunction with the CSSCAM module proposed in the present invention, thereby further improving SAR target detection performance.

[0130] Step 3: Use the new training set to train the target detection network to obtain the trained target detection network ψ'.

[0131] Specifically, during the network training process, the CSSCAM module uses the GT binary label GT' to supervise the spatial attention weight A of the spatial attention mechanism submodule of the GT binary label. s For supervised learning, the loss function uses FocalLoss;

[0132] The classification branch submodule uses the positive and negative sample labels obtained according to the anchor-free positive and negative sample selection strategy for supervised learning, and the loss function uses Focal Loss;

[0133] The regression branch submodule uses the positive sample labels obtained according to the anchor-free positive and negative sample selection strategy for supervised learning, and the loss function uses GIOU Loss;

[0134] The Attention binary classification branch submodule uses the GT binary label GT' at the same position as the positive sample obtained according to the anchor-free positive and negative sample selection strategy for supervised learning, and its loss function uses BCE Loss.

[0135] Specifically, the expression of the Focal Loss loss function is as follows:

[0136] FL(p t )=-(1-p t ) γ log(p t )

[0137]

[0138] Among them, p t Indicates that the network prediction is the corresponding classification label c * The probability of , γ represents the modulation factor, which is generally 2;

[0139] The GIoU Loss loss function is for two bounding boxes A and B, and its expression is as follows:

[0140]

[0141]

[0142] Among them, C represents the minimum circumscribed bounding box that can enclose bounding boxes A and B, and IoU represents the intersection-over-union ratio of bounding boxes A and B;

[0143] The expression of BCE Loss loss function is as follows:

[0144]

[0145] Among them, p represents the network prediction target (c * =1).

[0146] Furthermore, in this embodiment, the strategy for obtaining positive and negative sample labels based on the positive and negative sample selection strategy without anchor boxes is as follows:

[0147] a) For any position (x, y) on the predicted feature layer, map it back to the original image according to the downsampling multiple to obtain the corresponding position on the original image

[0148] b) Judgment Whether it falls inside the target horizontal box, first calculate The minimum distance to each of the four sides of the target horizontal box M is the number of objects in the original image. Assume that any target horizontal box but To B t The distance l between the four sides t 、r t , t t 、b t The calculation formula is as follows:

[0149]

[0150]

[0151] Then, according to the following three situations The corresponding position (x, y) on the prediction feature layer is divided into positive and negative samples and the label is defined:

[0152] b1) If Then The corresponding position (x, y) on the predicted feature layer is divided into negative samples, and the category label Defined as 0, regression label Defined as -1.

[0153] b2) If there is a unique The location It only falls inside a target horizontal frame. Then it is necessary to further judge based on the scale regression range of the predicted feature layer. The position (x, y) on the corresponding prediction feature layer predicts the target horizontal box. Assume that the target horizontal box is B i , first calculate the position (x, y) to B i The maximum distance between the four sides Then it is compared with the size regression range [s min ,s max ] to compare, if Then The corresponding predicted position (x, y) on the feature layer is divided into positive samples and category labels Defined as B i Category, regression label Defined as To B i The distances of the four sides Otherwise The position (x, y) on the corresponding prediction feature layer is classified as a negative sample, that is, the position (x, y) is considered not to predict B i , and the category labels Defined as 0, regression label Defined as -1.

[0154] b3) If there are N The location Falls inside the N target horizontal boxes. Then, similar to case b2), it is necessary to further judge based on the scale regression range of the predicted feature layer The corresponding predicted feature layer position (x, y) predicts whether these target horizontal boxes are predicted. First calculate the position The maximum distance to the four sides of these N target horizontal boxes Then they are regressed with the size range [s min ,s max ] to compare, if there are t (t>2) satisfy:

[0155]

[0156] Then it means The corresponding position (x, y) on the predicted feature layer corresponds to t target horizontal boxes. Since the Fully Convolution One-Stage Object Detection (FCOS) algorithm believes that target horizontal boxes with larger areas can be predicted by positions on deeper scale predicted feature layers, The corresponding position (x, y) on the predicted feature layer will eventually match the target horizontal box B with the smallest area among the t target horizontal boxes. j Corresponding, about to The corresponding predicted position (x, y) on the feature layer is divided into B j Corresponding positive sample, category label Defined as B j Category, regression label Defined as To B j The distances of the four sides If exists satisfy Then it means The corresponding predicted feature layer position (x, y) is only related to one target horizontal box B k Corresponding, about to The corresponding predicted position (x, y) on the feature layer is divided into B k Corresponding positive sample, category label Defined as B k Category, regression label Defined as To B k The distances of the four sides Otherwise The position (x, y) on the corresponding prediction feature layer is divided into negative samples, that is, it is considered that the position (x, y) does not predict any target horizontal box, and the category label Defined as 0, regression label Defined as -1.

[0157] c) Repeat operations a) and b) to define positive and negative samples and their labels at all positions on any feature layer;

[0158] d) Repeat operations a), b), and c) to define positive and negative samples and their labels at all positions on all feature layers;

[0159] It should be noted that this embodiment only uses A3 output by the CSSCAM fusion module in the feature processing part for network prediction, so operation d) is not required. In addition, this embodiment sets the size regression range of the A3 prediction feature layer to [-1,128].

[0160] Step 4: Input the new test set into the trained object detection network ψ' to obtain preliminary object detection results.

[0161] 41) An amplitude test slice and its corresponding gradient amplitude test slice in the new test set CFAR binary test slice The three feature extraction subnetworks in the trained target network are sent to the network for testing respectively, and the classification score bounding box regression parameters and binary classification scores of each position on the feature layer are obtained; among them, the amplitude test slice is a test slice in the original test set T;

[0162] 42) Multiply the classification score and the binary classification score of each position and take the square root to obtain the final target detection score for that position, and compare it with the preset score threshold:

[0163] If the target detection score at any position in the feature layer is less than the preset score threshold, the predicted detection box at that position is discarded;

[0164] Otherwise, the position on the feature layer is mapped back to the original image of the test slice according to the downsampling multiple as the center of the final detection box, and the bounding box regression parameters are integrated to obtain the target detection result at this position;

[0165] 43) Repeat step 42) for each position in the feature layer to obtain preliminary target detection results.

[0166] Step 5: Map the preliminary target detection results to the test image and perform NMS operation to remove overlapping target detection frames to obtain the final target detection results.

[0167] Specifically, based on the Loc obtained when constructing the test set in step 1, the object detection results of each test slice are mapped back to the test image, and an NMS operation is performed to remove overlapping object detection boxes. Repeat steps 4 and 5 to obtain the final object detection results.

[0168] The anchor-free SAR target detection method based on the supervised attention mechanism provided by the present invention introduces gradient amplitude information and CFAR information from the perspective of enhancing target features, constructs an anchor-free target detection network based on the fusion of gradient information and CFAR information, realizes the ground SAR target detection task in complex scenes, can effectively alleviate the problems of computational complexity and imbalance of positive and negative samples existing in the anchor-free target detection network itself, and improves the SAR target detection performance.

[0169] In order to further verify the effectiveness of the anchor-free SAR target detection method based on the supervised attention mechanism proposed in this invention, this embodiment also performs detection on MiniSAR data images.

[0170] See Figure 6-11 , Figure 6-11 It is the MiniSAR data image used in the experiment of the present invention; Table 1 shows the performance of the method proposed in the present invention and the CFAR-Guided-EfficientDet SAR image target detection method with better target detection performance at this stage (abbreviated as CFAR-Guided-EfficientDet, from the paper "Research on SAR Image Target Detection and Identification Methods", doctoral dissertation of Xidian University, Wang Ning, 2021) and the dual-stream SSD SAR image target detection method based on CFAR guidance (abbreviated as ICSAF-CFAR-SSD, from the paper "Research on SAR Target Detection Combining Constant False Alarm Detection and Deep Network", master's thesis of Xidian University, Tang Tiangu, 2022) in Figure 6-11 Vehicle target detection performance indicators on the MiniSAR data image shown.

[0171] Table 1 Different detection networks Figure 6-11 Comparison of target detection performance on MiniSAR data images shown

[0172]

[0173] In Table 1, Pre represents precision, that is, the percentage of true targets in the detected target box; Rec represents recall, that is, the percentage of targets correctly detected; F1-Score represents harmonic mean, and AP represents average accuracy, which are systematic indicators that unify precision Pre and recall Rec.

[0174] It can be seen from Table 1 that the method proposed by the present invention Figure 6-11The SAR target detection performance on the six MiniSAR data images shown is better than that of the CFAR-Guided-EfficientDet target detection network. The SAR target detection performance of the proposed method on images 1, 2, 3 and 4 is better than that of the ICSAF-CFAR-SSD target detection network. The SAR target detection performance on image 5 is significantly higher than that of the ICSAF-CFAR-SSD target detection network except for AP, which is comparable to that of the ICSAF-CFAR-SSD target detection network. In addition, the SAR target detection performance on image 6 is higher than that of the ICSAF-CFAR-SSD target detection network except for Pre and F1-Score, which are lower than those of the ICSAF-CFAR-SSD target detection network.

[0175] Overall, the proposed method significantly outperforms two currently available SAR target detection methods for vehicle detection in complex scenarios in terms of recall rate. It detects more targets while also ensuring significantly better target detection accuracy, F1-Score, and AP. Therefore, through the above experimental analysis, the proposed method achieves superior SAR target detection performance compared to currently available SAR target detection methods, fully validating the effectiveness and superiority of the proposed method.

[0176] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for anchor-free SAR target detection based on a supervised attention mechanism, characterized in that: include: Step 1: Obtain the original training set and the original test set based on the original SAR image; The ROEWA edge detection algorithm and the dual-parameter CFAR algorithm are used to obtain gradient information and CFAR information from the original training set and the original test set respectively, and a new training set and a new test set are constructed accordingly; Step 2: Construct an anchor-free SAR target detection network based on gradient information, CFAR information fusion, and supervised attention mechanism; wherein the target detection network includes a feature extraction module, a feature processing module, and a network prediction module; The feature extraction module includes three feature extraction sub-networks A, B, and C with identical structures but no shared parameters, namely, amplitude feature extraction network A, gradient feature extraction network B, and CFAR feature extraction network C; wherein, The three feature extraction sub-networks A, B, and C each include a feature extraction module with ResNet-18 as the backbone and an FPN multi-scale feature fusion module; The three feature extraction modules are used to extract the original training set , and the original training set Corresponding gradient magnitude training slice and CFAR binary training slices Perform feature extraction to obtain the corresponding output feature layer 、 and , ; The three FPN multi-scale feature fusion modules are used to output feature layers 、 and Perform multi-scale feature fusion to obtain the output feature layer of the entire feature extraction module 、 and ; The feature processing module includes a T-ICSAF feature fusion module based on an interactive attention mechanism and a CSSCAM module based on a combination of a spatial attention mechanism and a channel attention mechanism supervised by GT binary labels; wherein, The T-ICSAF feature fusion module is used to integrate the output feature layer of the feature extraction module. 、 and Fusion is performed to obtain fusion features ; The CSSCAM module is used to fusion the features Processing is performed to obtain a feature layer in which the target features are enhanced and the background clutter features are suppressed , as the output of the feature processing module; Step 3: Using the new training set to train the target detection network to obtain a trained target detection network; Step 4: Input the new test set into the trained object detection network to obtain preliminary object detection results; Step 5: Map the preliminary target detection results to the test image and perform NMS operation to remove overlapping target detection frames to obtain the final target detection results.

2. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 1 is characterized in that Step 1 includes:

1. Build a new training set; Select several original SAR images as training images and slice them to obtain several training slices as the original training set ; The selected training images are respectively subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain the gradient amplitude training image and the CFAR binary training image; Slice the gradient magnitude training image and the CFAR binary training image to obtain the original training set Corresponding gradient magnitude training slice and CFAR binary training slices , and with the original training set Together to form a new training set ; 2. Build a new test set; Select an original SAR image as a test image and slice it to obtain several test slices as the original test set ; The selected test images are respectively subjected to the ROEWA edge detection algorithm and the dual-parameter CFAR algorithm to obtain the gradient amplitude test image and the CFAR binary test image; Slice the gradient amplitude test image and the CFAR binary test image to obtain the original test set Corresponding gradient amplitude test slice and CFAR binary test slices and with the original test set Together form a new test set .

3. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 1 is characterized in that The T-ICSAF feature fusion module mainly includes four submodules, namely feature preprocessing submodule, interactive channel attention submodule, interactive spatial attention submodule and attention fusion submodule; wherein, The feature preprocessing submodule is used to convert the output feature layer of the feature extraction part into 、 and Convolution operation and BN batch normalization operation are performed respectively, and three feature layers are obtained accordingly. 、 and ; The interactive channel attention submodule is used to 、 and Perform global average pooling, summing, and Normalization operation to obtain channel attention weight ; The interactive spatial attention submodule is used to 、 and Perform global average pooling, Concat, convolution and Normalization operation to obtain spatial attention weight ; The attention fusion submodule is used to convert the channel attention weights Feature layer output by feature preprocessing submodule Each channel of the corresponding multiplication is obtained to obtain the feature layer ; Then the spatial attention weight With feature layer The pixels of each channel are multiplied one by one to obtain the fusion feature layer , which is the output of the T-ICSAF feature fusion module.

4. According to the anchor-free SAR target detection method based on the supervised attention mechanism in claim 3, the CSSCAM module includes three submodules: a spatial attention mechanism submodule for GT binary label supervision, a SE channel attention mechanism submodule, and an attention fusion submodule; wherein, The GT binary label supervised spatial attention mechanism submodule is used to fusion feature layer output by T-ICSAF feature fusion module Perform global average pooling operation on the channel dimension and global maximum pooling operation on the channel dimension, Concat operation, convolution and Normalization operation to obtain spatial attention weight ; The SE channel attention mechanism submodule is used to fusion feature layer output by T-ICSAF feature fusion module Perform global average pooling and global maximum pooling operations in spatial dimensions, feature vector compression, mapping and decompression operations, summation, and Normalization operation to obtain channel attention weight ; The attention fusion submodule is used to convert the spatial attention weights The fused feature layer output by the T-ICSAF feature fusion module The pixels of each channel are multiplied one by one to obtain the feature layer , and then the channel attention weight With feature layer Multiply each channel of the corresponding layer to obtain a feature layer in which the target feature is enhanced and the background clutter feature is suppressed , which is the output of the CSSCAM module.

5. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 1 is characterized in that In the anchor-free SAR target detection network constructed in step 2, the network prediction module includes a classification branch submodule for target detection tasks, a regression branch submodule, and an Attention binary classification branch submodule based on GT binary label supervision; The classification branch submodule and the regression branch submodule are respectively used to process the output feature layer of the feature processing module Make predictions and obtain the corresponding classification scores and bounding box regression parameters; The Attention two-class branch submodule is used to predict the feature layer of the classification score and a feature layer for predicting bounding box regression parameters Processing is performed to obtain the binary classification score.

6. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 5 is characterized in that The classification branch submodule is implemented in the following way: First, the feature layer output by the CSSCAM module in the feature processing part Perform four convolution operations to obtain a feature layer for predicting classification scores ; Then the feature layer obtained Convolution and Normalization operation to obtain the classification score map predicted by the classification branch submodule ; The regression branch submodule is implemented in the following way: First, the feature layer output by the CSSCAM module in the feature processing part Perform four convolution operations to obtain a feature layer for predicting bounding box regression parameters ; Then the feature layer obtained Perform a convolution operation to obtain the bounding box regression parameters predicted by the regression branch submodule ; The Attention binary classification branch submodule is implemented as follows: First, the feature layer used to predict the classification score in the classification branch submodule and the feature layer in the regression branch submodule for predicting bounding box regression parameters Perform a Concat operation and get the feature map Perform convolution operation to obtain the feature layer ; Finally, the feature layer Perform convolution operations and Normalization operation to obtain the binary classification score map predicted by the Attention binary classification branch submodule .

7. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 6 is characterized in that: In step 3, during the training of the target detection network, The CSSCAM module uses GT binary labels Spatial attention weights of the spatial attention mechanism submodule for GT label supervision For supervised learning, the loss function uses Focal Loss; The classification branch submodule uses the positive and negative sample labels obtained according to the anchor-free positive and negative sample selection strategy for supervised learning, and the loss function uses Focal Loss; The regression branch submodule uses the positive sample labels obtained according to the anchor-free positive and negative sample selection strategy for supervised learning, and the loss function uses GIOU Loss; The Attention two-classification branch submodule uses the GT binary label at the same position as the positive sample obtained according to the anchor-free positive and negative sample selection strategy For supervised learning, the loss function uses BCE Loss.

8. The anchor-free SAR target detection method based on the supervised attention mechanism according to claim 1 is characterized in that: Step 4 includes: 41) An amplitude test slice and a corresponding gradient amplitude test slice in the new test set are combined , CFAR binary test slice The three feature extraction subnetworks in the trained target network are respectively sent to the test to obtain the classification score bounding box regression parameters and binary classification scores at each position on the feature layer; wherein the amplitude test slice is the original test set A test slice in 42) Multiply the classification score and the binary classification score of each position and take the square root to obtain the final target detection score for that position, and compare it with the preset score threshold: If the target detection score at any position in the feature layer is less than the preset score threshold, the predicted detection box at that position is discarded; Otherwise, the position on the feature layer is mapped back to the original image of the test slice according to the downsampling multiple as the center of the final detection box, and the bounding box regression parameters are integrated to obtain the target detection result at this position; 43) Repeat step 42) for each position in the feature layer to obtain preliminary target detection results.

Images