A SAR airplane target detection method and system based on strong scattering guidance

Abstract

The application belongs to the technical field of SAR image target detection, and discloses a SAR airplane target detection method and system based on strong scattering guidance, which comprises the following steps: S1. Based on the SAR image to be detected, the key points of the airplane target are extracted; S2. Based on the position information of the key points, a strong scattering guidance network is constructed and trained; S3. The SAR image is input into the strong scattering guidance network for prediction to generate a prediction frame result; and S4. A non-maximum suppression operation is performed on the prediction frame result to obtain a final prediction result. The key point position information is used to guide the network training, so that the discrete feature learning ability and the detection precision are improved; the texture and scattering feature fusion module combines the high and low layer feature to enhance the target saliency and suppress the background interference; and the scattering perception dynamic fusion module uses the key point prior dynamic weighted feature to strengthen the target physical structure representation ability.

Description

Technical Field

[0001] This invention belongs to the field of SAR image target detection technology, specifically relating to a SAR aircraft target detection method and system based on strong scattering guidance. Background Technology

[0002] Detecting aircraft targets in SAR images presents significant challenges due to the imaging mechanism of SAR images and the structural characteristics of aircraft. First, the features are discrete, making complete detection difficult. Aircraft in SAR images consist of bright pixels generated by scattering from various components and dark pixels generated by weak scattering from smooth areas like the fuselage. Aircraft target features are discrete, structural features are incomplete, and the correlation between components is weak. Existing deep learning algorithms tend to generate multiple prediction boxes for aircraft targets, leading to difficulties in complete detection. Second, strong background clutter makes effective differentiation difficult. SAR and optical sensors differ fundamentally in detection wavelength and imaging mechanism, making SAR image interpretation often more challenging. In SAR images, buildings, vehicles, and some metal structures around aircraft in airports easily generate scattering similar to that of aircraft targets; these complex background clutters can cause visual confusion to some extent.

[0003] To overcome the aforementioned difficulties, various aircraft target detection algorithms targeting the characteristics of SAR images have emerged in recent years. The scattering physics information of aircraft targets contains their electromagnetic response characteristics, which can profoundly reflect the target's essential features such as geometric structure and material properties. Constructing a detection framework based on the synergistic integration of aircraft scattering physics mechanisms and deep learning feature extraction capabilities is a key approach to improving the accuracy of SAR aircraft target detection, suppressing false alarms, and enhancing the physical interpretability of the model. This approach can enhance the model's ability to analyze the inherent electromagnetic scattering characteristics of aircraft targets through physical features at the scattering source level, and can also utilize deep learning to uncover complex spatial and intensity correlation patterns between different scattering sources. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0005] A SAR aircraft target detection method based on strong scattering guidance includes the following steps:

[0006] S1. Extract key points of the aircraft target based on the SAR image to be detected;

[0007] S2. Based on the location information of the key points of the aircraft target, construct and train a strong scattering guidance network;

[0008] S3. Input the SAR image into the strong scattering guidance network for prediction and generate prediction box results;

[0009] S4. Perform nonmaximum suppression on the predicted bounding box results to obtain the final prediction results.

[0010] Preferably, step S1 includes the following steps:

[0011] The SIFT algorithm is used to extract key point features of the aircraft target. Each pixel of the aircraft target is traversed and marked as a key point.

[0012] By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

[0013] Preferably, step S2 includes the following steps:

[0014] Use the location information of key points of the aircraft target as labels to generate SAR image training samples;

[0015] An initial strong scattering guidance network is constructed, and SAR image training samples are input into the initial strong scattering guidance network for training to obtain the prediction probability of key points.

[0016] The loss for the initial strong scattering-guided network prediction is calculated using cross-entropy loss:

[0017] Loss (x,y) =-[qlog(p)+(1-q)log(1-p)],

[0018] Where (x,y) represents the pixel position, q is the ground truth label, and p is the keypoint probability score predicted by the network model.

[0019] Preferably, the strong scattering guidance network includes: a texture and scattering feature fusion module and a scattering perception dynamic fusion module;

[0020] The texture and scattering feature fusion module is used to fuse the low-level texture features and high-level semantic features of the feature pyramid.

[0021] The scattering perception dynamic fusion module is used to dynamically weight and classify feature maps and scattering feature maps based on prior information of key points of the aircraft target.

[0022] Preferably, the structure of the texture and scattering feature fusion module includes:

[0023]

[0024] Among them, C i C represents the input of the low-level feature map. i+1 This represents the high-level feature map input, MAP represents local max pooling, and Conv... s2 DConv represents a convolution with a stride of 2 and a kernel size of 3×3.5×5 DSConv represents a dilated convolution with a kernel size of 5×5. 3×3 This indicates a depthwise separable convolution with a kernel size of 3×3. 1×1 This represents a convolution with a kernel size of 1×1. Indicates channel splicing, C′ i+1 P represents the fused feature map. i+1 This represents the final output of the texture and scattering feature fusion module.

[0025] Preferably, the structure of the scattering sensing dynamic fusion module includes:

[0026] Q = W q F c K = W k F s ,

[0027] A = softmax(QK) T ),

[0028] W c =softmax(GAP(F) c )),W s =softmax(GMP(F s )),

[0029]

[0030] Among them, F c F represents the classification feature map. s Represents the scattering characteristic map, W q W k All represent initial weights, A represents the spatial correlation weights between features, GAP represents global average pooling, GMP represents global max pooling, and W represents... c W represents the channel importance weights of categorical features. s The channel importance weights represent the scattering features, and softmax and sigmoid are activation functions. F' represents element-wise multiplication, Q represents the query space, K represents the key space, T represents the transpose matrix, and F' represents the key matrix. c It is the final output of the scattering sensing dynamic fusion module.

[0031] Preferably, step S4 includes the following steps: performing non-maximum suppression on multiple predicted boxes of the same aircraft target to remove redundant boxes; merging the suppressed detection boxes into a final target detection box.

[0032] The present invention also provides a SAR aircraft target detection system based on strong scattering guidance. The system is used to implement the above method and includes: an extraction module, a construction module, a detection module and a post-processing module.

[0033] The extraction module is used to extract key points of the aircraft target based on the SAR image to be detected;

[0034] The construction module is used to construct and train a strong scattering guidance network based on the location information of the key points of the aircraft target;

[0035] The detection module is used to input the SAR image into the strong scattering guidance network for prediction and generate prediction box results;

[0036] The post-processing module is used to perform non-maximum suppression on the prediction box results to obtain the final prediction results.

[0037] Preferably, the workflow of the extraction module includes:

[0038] The SIFT algorithm is used to extract key point features of the aircraft target. Each pixel of the aircraft target is traversed and marked as a key point.

[0039] By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

[0040] Preferably, the workflow of the building module includes:

[0041] Use the location information of key points of the aircraft target as labels to generate SAR image training samples;

[0042] An initial strong scattering guidance network is constructed, and SAR image training samples are input into the initial strong scattering guidance network for training to obtain the prediction probability of key points.

[0043] The loss for the initial strong scattering-guided network prediction is calculated using cross-entropy loss:

[0044] Loss (x,y) =-[qlog(p)+(1-q)log(1-p)],

[0045] Where (x,y) represents the pixel position, q is the ground truth label, and p is the keypoint probability score predicted by the network model.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention extracts key points of aircraft targets from SAR images and constructs a strong scattering-guided network based on the key point location information to improve the network's learning ability of discrete target features and increase detection accuracy. A novel texture and scattering feature fusion module is designed to enhance the saliency of targets under different conditions, reduce background interference, and minimize missing aircraft targets. This module combines high-level semantic information from the feature pyramid with low-level texture details to enhance the identifiable features of the target and minimize the influence of the background. A novel scattering-aware dynamic fusion module is also designed, using key point predictions as priors to guide the fusion of classification feature maps, enabling the network to focus more on the key physical structural parts of the target and improve the representation and discrimination capabilities of the features. Attached Figure Description

[0048] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the detection results in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1:

[0054] In this embodiment, as Figure 1 As shown, a SAR aircraft target detection method based on strong scattering guidance includes the following steps:

[0055] S1. Extract key points of the aircraft target based on the SAR image to be detected.

[0056] The SIFT algorithm is used to extract key point features of the aircraft target. The process involves traversing each pixel of the aircraft target, marking each pixel as a local extremum point, and marking all local extremum points as key points (i.e., local extremum points, extracted by the SIFT algorithm).

[0057] The SIFT algorithm described above generates scale-space images through Gaussian convolution:

[0058] L(x,y,σ)=G(x,y,σ)*I(x,y),

[0059] Where (x,y) represents the pixel position, I(x,y) represents the gray value of the pixel at (x,y), σ is the standard deviation of the Gaussian distribution, which controls the blurring degree of the Gaussian filter, and G(x,y,σ) is the Gaussian kernel function, expressed as:

[0060]

[0061] The SIFT algorithm calculates the difference Gaussian to detect extrema in scale space:

[0062] D(x,y,σ)=L(x,y,kσ)-L(x,y,σ),

[0063] Here, k is a scale factor, which identifies potential keypoints by comparing images at adjacent scales.

[0064] By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

[0065] S2. Based on the location information of key points, construct and train a strong scattering guidance network.

[0066] A strong scattering-guided network is constructed to perform feature fusion processing on the SAR image to be detected, and the key point location information is used as a label to assist network training.

[0067] The location information of key aircraft target points is used as labels, and together with the SAR image and its initial labels, it is input into an initial strong scattering guidance network for training, yielding the prediction probability of aircraft target key points. Cross-entropy loss is used to calculate the network's loss for predicting key points.

[0068] Loss (x,y) =-[qlog(p)+(1-q)log(1-p)],

[0069] Where (x,y) represents the pixel position, q is the true label (0 or 1), when q is 1, it means that the point is a key point obtained according to the method of obtaining the key point position information of the aircraft target, and p is the key point probability score predicted by the network model.

[0070] S3. Input the SAR image into the trained strong scattering guidance network to make prediction decisions and generate prediction box results.

[0071] In this embodiment, the strong scattering guidance network includes: a texture and scattering feature fusion module and a scattering perception dynamic fusion module.

[0072] The texture and scattering feature fusion module enhances the saliency of targets under different conditions, reduces background interference, and minimizes missing aircraft targets. This module downsamples low-level feature maps to match the size of high-level feature maps while preserving as much key local texture information as possible. The downsampled low-level feature maps are concatenated with the high-level feature maps along the channel dimension to achieve preliminary feature fusion. Subsequently, a two-branch parallel convolution is performed to fuse local details with the global context. This combines high-level semantic information from the feature pyramid with low-level texture details to enhance the identifiable features of the target and minimize the influence of the background.

[0073] The structure of the texture and scattering feature fusion module includes:

[0074]

[0075] Among them, C i C represents the input of the low-level feature map. i+1 This represents the high-level feature map input, MAP represents local max pooling, and Conv... s2 DConv represents a convolution with a stride of 2 and a kernel size of 3×3. 5×5 DSConv represents a dilated convolution with a kernel size of 5×5. 3×3 This indicates a depthwise separable convolution with a kernel size of 3×3. 1×1 This represents a convolution with a kernel size of 1×1. Indicates channel splicing, C′ i+1 P represents the fused feature map. i+1 This represents the final output of the texture and scattering feature fusion module.

[0076] The scattering perception dynamic fusion module uses the prediction of key points as a priori, generates attention weights from the scattering feature map and the classification feature map respectively, and fuses the feature maps in a residual manner, so that the network pays more attention to the key physical structure of the target and improves the feature representation and discrimination ability.

[0077] The structure of the scattering sensing dynamic fusion module includes:

[0078] Q = W q F c K = W k F s ,

[0079] A = softmax(QK) T ),

[0080] W c =softmax(GAP(F) c )),W s =softmax(GMP(F s )),

[0081]

[0082] Among them, F c F represents the classification feature map. s Represents the scattering characteristic map, W q W k All represent initial weights, A represents the spatial correlation weights between features, GAP represents global average pooling, GMP represents global max pooling, and W represents... c W represents the channel importance weights of categorical features. s The channel importance weights represent the scattering features, and softmax and sigmoid are activation functions. F' represents element-wise multiplication, Q represents the query space, K represents the key space, T represents the transpose matrix, and F' represents the key matrix. c It is the final output of the scattering sensing dynamic fusion module.

[0083] S4. Perform non-maximum suppression on the detection box results, and merge the suppressed detection boxes to obtain the final prediction result.

[0084] In this embodiment, since the detection boxes are generated based on pixel samples, the same aircraft target may be contained in multiple detection boxes. Therefore, non-maximum suppression is performed on the detection boxes. The steps include: first, sorting all predicted boxes in descending order according to classification confidence, selecting the box with the highest confidence as the benchmark, traversing the remaining boxes and calculating their intersection-union ratio (IU) with the benchmark box. If the IU of a box with the benchmark box exceeds a preset threshold, it is determined to be a redundant box and directly suppressed and removed, retaining only the benchmark box. This process is iterated, sequentially selecting the highest confidence box among the remaining boxes as the new benchmark and repeating the suppression operation until all boxes have been processed. Finally, a set of highly confident and independent predicted boxes is output, generating the final detection result. Figure 2 As shown.

[0085] Example 2:

[0086] In this embodiment, a SAR aircraft target detection system based on strong scattering guidance includes: an extraction module, a construction module, a detection module, and a post-processing module.

[0087] The following will describe in detail, with reference to this embodiment, how the present invention solves the technical problems in practical work.

[0088] First, the extraction module extracts key points of the aircraft target based on the SAR image to be detected. The SIFT algorithm is used to extract the features of the key points of the aircraft target. The process involves traversing each pixel of the aircraft target, marking each pixel as a local extremum, and marking all local extremums as key points (i.e., local extremums, extracted by the SIFT algorithm).

[0089] The SIFT algorithm described above generates scale-space images through Gaussian convolution:

[0090] L(x,y,σ)=G(x,y,σ)*I(x,y),

[0091] Where (x,y) represents the pixel position, I(x,y) represents the gray value of the pixel at (x,y), σ is the standard deviation of the Gaussian distribution, which controls the blurring degree of the Gaussian filter, and G(x,y,σ) is the Gaussian kernel function, expressed as:

[0092]

[0093] The SIFT algorithm calculates the difference Gaussian to detect extrema in scale space:

[0094] D(x,y,σ)=L(x,y,kσ)-L(x,y,σ),

[0095] Here, k is a scale factor, which identifies potential keypoints by comparing images at adjacent scales.

[0096] By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

[0097] Subsequently, the construction module builds and trains a strong scattering guidance network based on the location information of the key points.

[0098] A strong scattering-guided network is constructed to perform feature fusion processing on the SAR image to be detected, and the key point location information is used as a label to assist network training.

[0099] The location information of key aircraft target points is used as labels, and together with the SAR image and its initial labels, it is input into an initial strong scattering guidance network for training, yielding the prediction probability of aircraft target key points. Cross-entropy loss is used to calculate the network's loss for predicting key points.

[0100] Loss (x,y) =-[qlog(p)+(1-q)log(1-p)],

[0101] Where (x,y) represents the pixel position, q is the true label (0 or 1), when q is 1, it means that the point is a key point obtained according to the method of obtaining the key point position information of the aircraft target, and p is the key point probability score predicted by the network model.

[0102] The detection module inputs the SAR image into the trained strong scattering guidance network for prediction and decision-making, and generates prediction box results.

[0103] In this embodiment, the strong scattering guidance network includes: a texture and scattering feature fusion module and a scattering perception dynamic fusion module.

[0104] The texture and scattering feature fusion module enhances the saliency of targets under different conditions, reduces background interference, and minimizes missing aircraft targets. This module downsamples low-level feature maps to match the size of high-level feature maps while preserving as much key local texture information as possible. The downsampled low-level feature maps are concatenated with the high-level feature maps along the channel dimension to achieve preliminary feature fusion. Subsequently, a two-branch parallel convolution is performed to fuse local details with the global context. This combines high-level semantic information from the feature pyramid with low-level texture details to enhance the identifiable features of the target and minimize the influence of the background.

[0105] The structure of the texture and scattering feature fusion module includes:

[0106]

[0107] Among them, C i C represents the input of the low-level feature map. i+1 This represents the high-level feature map input, MAP represents local max pooling, and Conv... s2 DConv represents a convolution with a stride of 2 and a kernel size of 3×3. 5×5 DSConv represents a dilated convolution with a kernel size of 5×5. 3×3 This indicates a depthwise separable convolution with a kernel size of 3×3. 1×1 This represents a convolution with a kernel size of 1×1. Indicates channel splicing, C′ i+1 P represents the fused feature map. i+1 This represents the final output of the texture and scattering feature fusion module.

[0108] The scattering perception dynamic fusion module uses the prediction of key points as a priori, generates attention weights from the scattering feature map and the classification feature map respectively, and fuses the feature maps in a residual manner, so that the network pays more attention to the key physical structure of the target and improves the feature representation and discrimination ability.

[0109] The structure of the scattering sensing dynamic fusion module includes:

[0110] Q = W q F c K = W k F s ,

[0111] A = softmax(QK) T ),

[0112] W c =softmax(GAP(F) c )),W s =softmax(GMP(F s )),

[0113]

[0114] Among them, F c F represents the classification feature map. s Represents the scattering characteristic map, W q W k All represent initial weights, A represents the spatial correlation weights between features, GAP represents global average pooling, GMP represents global max pooling, and W represents... c W represents the channel importance weights of categorical features. s The channel importance weights represent the scattering features, and softmax and sigmoid are activation functions. Let F represent element-wise multiplication, Q represent the query space, K represent the key space, T represent the transpose matrix, and F represent the key space. c It is the final output of the scattering sensing dynamic fusion module.

[0115] Finally, the post-processing module performs non-maximum suppression on the detection box results and merges the suppressed detection boxes to obtain the final prediction result.

[0116] In this embodiment, since the detection boxes are generated based on pixel samples, the same aircraft target may be contained in multiple detection boxes. Therefore, non-maximum suppression is performed on the detection boxes. The process includes: first, sorting all predicted boxes in descending order according to classification confidence, selecting the box with the highest confidence as the benchmark, traversing the remaining boxes and calculating their intersection-union ratio (IU) with the benchmark box. If the IU of a box with the benchmark box exceeds a preset threshold, it is determined to be a redundant box and directly suppressed and removed, retaining only the benchmark box. This process is iterated, sequentially selecting the highest confidence box among the remaining boxes as the new benchmark and repeating the suppression operation until all boxes have been processed. Finally, a set of highly confident and independent predicted boxes is output, generating the final detection result. Figure 2 As shown.

[0117] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A SAR aircraft target detection method based on strong scattering guidance, characterized in that, Includes the following steps: S1. Extract key points of the aircraft target based on the SAR image to be detected; S2. Based on the location information of the key points of the aircraft target, construct and train a strong scattering guidance network; The strong scattering guidance network includes: a texture and scattering feature fusion module and a scattering perception dynamic fusion module; the texture and scattering feature fusion module is used to fuse the low-level texture features and high-level semantic features of the feature pyramid; the scattering perception dynamic fusion module is used to dynamically weight the classification feature map and the scattering feature map based on the prior information of the key points of the aircraft target. The structure of the texture and scattering feature fusion module includes: ， ， in, Ci This represents the input of the low-level feature map. Ci+1 This represents the high-level feature map input. MAP This indicates local max pooling. Convs2 This represents a convolution with a stride of 2 and a kernel size of 3×3. DConv5×5 This represents a dilated convolution with a kernel size of 5×5. DSConv3×3 This indicates a depthwise separable convolution with a kernel size of 3×3. Conv1×1 This represents a convolution with a kernel size of 1×1. Indicates channel splicing. This represents the fused feature map. This represents the final output of the texture and scattering feature fusion module; The structure of the scattering sensing dynamic fusion module includes: ， ， ， ， in, Fc Represents the classification feature map, Fs Represents the scattering feature map. Wq , Wk Both represent the initial weights. A The weights represent the spatial correlation between features. GAP Indicates global average pooling. GMP Indicates global max pooling. Wc Channel importance weights representing categorical features Ws Channel importance weights representing scattering characteristics softmax and sigmoid It is an activation function. This indicates element-wise multiplication. Q Indicates the query space. K Represents the key space. T Represents the transpose matrix. This is the final output of the scattering sensing dynamic fusion module; S3. Input the SAR image into the strong scattering guidance network for prediction and generate prediction box results; S4. Perform nonmaximum suppression on the predicted bounding box results to obtain the final prediction results.

2. The SAR aircraft target detection method based on strong scattering guidance according to claim 1, characterized in that, S1 includes the following steps: The SIFT algorithm is used to extract key point features of the aircraft target. Each pixel of the aircraft target is traversed and marked as a key point. By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

3. The SAR aircraft target detection method based on strong scattering guidance according to claim 1, characterized in that, S2 includes the following steps: Use the location information of key points of the aircraft target as labels to generate SAR image training samples; An initial strong scattering guidance network is constructed, and SAR image training samples are input into the initial strong scattering guidance network for training to obtain the prediction probability of key points. The loss for the initial strong scattering-guided network prediction is calculated using cross-entropy loss: ， in,( x, y ) represents the pixel position. q For real labels, p The probability score of key points predicted by the network model.

4. The SAR aircraft target detection method based on strong scattering guidance according to claim 1, characterized in that, S4 includes the following steps: performing non-maximum suppression on multiple predicted boxes of the same aircraft target to remove redundant boxes; merging the suppressed detection boxes into a final target detection box.

5. A SAR aircraft target detection system based on strong scattering guidance, the system being used to implement the method according to any one of claims 1-4, characterized in that, include: Extraction module, construction module, detection module, and post-processing module; The extraction module is used to extract key points of the aircraft target based on the SAR image to be detected; The construction module is used to construct and train a strong scattering guidance network based on the location information of the key points of the aircraft target; The detection module is used to input the SAR image into the strong scattering guidance network for prediction and generate prediction box results; The post-processing module is used to perform non-maximum suppression on the prediction box results to obtain the final prediction results.

6. The SAR aircraft target detection system based on strong scattering guidance according to claim 5, characterized in that, The workflow of the extraction module includes: The SIFT algorithm is used to extract key point features of the aircraft target. Each pixel of the aircraft target is traversed and marked as a key point. By mapping the relative positions of key points in the aircraft target to their relative positions in the SAR image, the positional information of key points of the aircraft target in the SAR image is obtained.

7. The SAR aircraft target detection system based on strong scattering guidance according to claim 5, characterized in that, The workflow of the building module includes: Use the location information of key points of the aircraft target as labels to generate SAR image training samples; An initial strong scattering guidance network is constructed, and SAR image training samples are input into the initial strong scattering guidance network for training to obtain the prediction probability of key points. The loss for the initial strong scattering-guided network prediction is calculated using cross-entropy loss: ， in,( x, y ) represents the pixel position. q For real labels, p The probability score of key points predicted by the network model.

Images