SAR image wounded target detection method based on multi-scale feature alignment

Abstract

This invention discloses a SAR image casualty target detection method based on multi-scale feature alignment, belonging to the field of SAR image processing technology. Addressing the problems of difficulty in representing weakly scattering small targets, inconsistent multi-scale features, and insufficient feature fusion capabilities in existing SAR image casualty target detection methods, this invention first constructs a SAR casualty image dataset. Then, based on the Faster R-CNN framework, it employs a scattering-guided adaptive deformable convolution module to enhance the feature representation of weakly scattering targets, a scattering-guided distribution alignment module to constrain the distribution consistency of multi-scale features, and a scale-prior scattering perception feature fusion module to adaptively weight and fuse multi-scale features, thereby constructing a SAR image casualty detection model. This invention can fully utilize the all-weather, all-day advantages of SAR imaging, improving the accuracy and robustness of casualty detection in complex disaster environments, and is applicable to scenarios such as post-disaster search and rescue.

Description

Technical Field

[0001] This invention relates to the field of synthetic aperture radar (SAR) image target detection technology, specifically a method for detecting wounded targets in SAR images based on multi-scale feature alignment. Background Technology

[0002] In the event of sudden disasters such as earthquakes, landslides, mudslides, and building collapses, quickly and accurately locating trapped individuals is a crucial prerequisite for improving emergency rescue efficiency and reducing casualties. Traditional methods of casualty search and rescue primarily rely on manual searches, optical remote sensing, infrared imaging, and life detection equipment. However, in complex disaster environments, the applicability of these methods is significantly limited by factors such as smoke, dust, obstructions, lighting conditions, and weather conditions, making it difficult to meet the practical needs of rapid search and rescue in complex scenarios. Therefore, researching a detection method that is not significantly affected by lighting and weather conditions and can effectively locate casualty targets in complex backgrounds has significant practical implications and application value.

[0003] Synthetic Aperture Radar (SAR) offers advantages such as all-weather, all-day imaging and long-range capability, enabling it to acquire surface scene information under conditions of clouds, low light, and obstruction. It is widely used in disaster monitoring, ground feature identification, and target detection. Unlike optical images, SAR images characterize target and background information through echo scattering characteristics, reflecting the target's structure, scattering intensity, and its geometric relationship with the surrounding environment. Therefore, it has unique advantages in scenarios such as post-disaster search and rescue and target detection in complex environments. Applying SAR imaging technology to casualty target detection is expected to compensate for the shortcomings of optical methods in complex scenarios, providing new technical support for emergency search and rescue.

[0004] Currently, target detection methods based on SAR images mainly include traditional methods based on statistical models such as Constant False Alarm Rate (CFAR) and deep learning-based methods. Traditional methods rely on local background statistical characteristics, extracting anomalous scattering points by setting detection thresholds. They have advantages such as clear principles and low computational cost, but are prone to high false alarm rates and insufficient robustness in complex backgrounds, with significant clutter interference, and with small target scales. In contrast, deep learning-based detection methods can automatically extract more discriminative features, demonstrating stronger advantages in complex background suppression, small target detection, and multi-scale feature representation, and have gradually become an important research direction in SAR target detection.

[0005] Existing deep learning-based SAR target detection methods are mostly applied to the detection of targets such as ships, vehicles, and aircraft. However, for casualty target detection, a series of technical challenges remain. On the one hand, casualty targets in SAR images are usually small in size and have weak scattering characteristics, making them easily submerged in complex background clutter and speckle noise. On the other hand, due to attitude changes, occlusion, and imaging geometry, the appearance of casualty targets in SAR images is not stable, making it difficult to form a uniform appearance description. Furthermore, target information in SAR images is not only reflected in bright scattering areas but is also closely related to the background clutter around the target, local spatial structure, and feature distribution at different scales. If detection is based solely on a single scale or a single scattering feature, it is easy to produce missed or false detections.

[0006] In existing research, conventional convolutional neural networks have certain advantages in local texture extraction. However, the fixed sampling method of traditional convolutional modules is difficult to adapt to the irregular scattering patterns of wounded targets in SAR images, and the response to the main scattering region of weakly scattering small targets is insufficient. At the same time, there are often problems of inconsistent distribution and unstable representation of the main scattering region among features of different scales, which can easily lead to information misalignment or suppression of weak target features during the fusion of multi-scale features. On the other hand, traditional feature fusion methods usually use a unified or implicitly learned weight allocation method for features of different scales, which is difficult to fully highlight the weak wounded target details contained in shallow high-resolution features. Therefore, how to enhance the feature representation ability of the main scattering region of wounded targets, improve the consistency between multi-scale features, and achieve feature fusion that is more favorable to weakly scattering small targets remains a key problem that urgently needs to be solved in current technology.

[0007] Furthermore, due to the difficulty in acquiring SAR casualty target data in real disaster scenarios, the number of existing publicly available data samples is limited, and the data distribution is uneven, making it difficult to directly support sufficient training of deep learning models. How to construct SAR casualty target images with high physical consistency by combining the electromagnetic scattering characteristics of the human body, and generate a dataset that meets the model training requirements under real SAR background conditions, is also a crucial fundamental problem in achieving casualty target detection based on SAR images.

[0008] Therefore, a method for detecting casualties based on SAR images is proposed. By constructing a SAR casualty detection dataset and combining scattering enhancement feature extraction, multi-scale feature alignment and multi-scale feature fusion, a casualty target detection model based on improved Faster R-CNN is built to achieve effective detection of casualty targets under complex background conditions. This is of great significance for improving search and rescue efficiency and detection accuracy in disaster scenarios. Summary of the Invention

[0009] To address the problems in existing SAR image-based casualty target detection technologies, such as difficulty in representing small targets with weak scattering, susceptibility to false alarms and missed detections in complex backgrounds, insufficient multi-scale feature expression and fusion capabilities, and lack of training samples, this invention provides a casualty target detection method based on SAR images to improve the accuracy and reliability of casualty target detection in complex scenarios.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting wounded targets in SAR images based on multi-scale feature alignment, comprising the following steps: Step S1: Construct a SAR wounded soldier image dataset; Step S2: Based on the Faster R-CNN model, A scattering-guided adaptive deformable convolution (SGADConv) module is used to replace the convolution module in the feature extraction stage for extracting multi-scale features. The scattering-guided adaptive deformable convolution (SGADConv) module includes a deformable convolution branch and a spatial attention branch based on Gaussian clustering attention. The Gaussian clustering attention and spatial attention are fused by logarithmic modulation. The output of the deformable convolution branch is weighted and concatenated with the output of the standard convolution branch and then fused by a 1×1 convolution. The Scattering Guided Distribution Alignment (SGDAM) module is used to align features at multiple scales. The SGDAM module weights the aligned features at each scale based on the foreground guided weight map, calculates the foreground weighted mean and foreground weighted variance of the aligned features at each scale, and performs distribution alignment using mean consistency constraints, variance consistency constraints, and symmetric KL divergence constraints. The Scale Prior Scattering Aware Feature Fusion (SSAFM) module is used to fuse the aligned multi-scale features. The SSAFM module performs average pooling and local variance pooling on the aligned features at each scale, generates a spatial attention response after residual dilated depth convolution context aggregation, and introduces scale prior bias for normalized weighted fusion. Thus, a SAR image casualty detection model based on an improved Faster R-CNN was constructed, and the SAR image casualty detection model was trained using a SAR casualty image dataset; Step S3: Input the SAR casualty image to be detected into the trained SAR image casualty detection model for detection processing, and output the location and category detection results of the casualty target.

[0011] Specifically, the workflow of the scattering-guided adaptive deformable convolution module is as follows: The input feature maps are respectively fed into the standard convolution branch, deformable convolution branch, spatial attention branch, and Gaussian clustering attention branch; The output of the spatial attention branch is fused with the output of the Gaussian clustering attention branch to generate a log-modulated Gaussian spatial attention. The output of the deformable convolution branch is weighted using the log-modulated Gaussian spatial attention to obtain attention-enhanced deformable features. The attention-enhanced deformable features are concatenated with the output features of the standard convolutional branch, and then fused through a 1×1 convolution to obtain the output feature map.

[0012] Specifically, the workflow of the scattering-guided distribution alignment module is as follows: The multi-scale features are input into the scattering consistency alignment unit, and the position response of the main scattering region of the human target in each scale feature is adjusted to obtain the aligned multi-scale features. Subsequently, a foreground guiding weight map is generated based on the aligned multi-scale features; Subsequently, the foreground-guided weight map was used to perform foreground-weighted statistics on the aligned multi-scale features, and the foreground-weighted mean and foreground-weighted variance of the aligned features at each scale were obtained respectively. After obtaining the foreground-weighted mean and foreground-weighted variance of the aligned features at each scale, Gaussian distribution modeling is performed. Based on the statistical distribution differences between the aligned features at different scales, mean consistency constraints, variance consistency constraints, and symmetric KL divergence constraints are established to construct the total alignment loss of the scattering-guided distribution alignment module.

[0013] Specifically, the workflow of the scale prior scattering perception feature fusion module is as follows: First, the aligned multi-scale features are used as the input feature map set, and then the input feature map sets at each scale are processed. Joint scattering texture pooling is performed to obtain a stitched pooled feature map; then, the stitched pooled feature map is input into a residual hole depth separable convolutional structure to obtain a context-enhanced feature map; then, a spatial attention map of the corresponding scale is generated by pointwise convolution of PWConv. Subsequently, the spatial attention maps corresponding to each scale are input into a softmax fusion unit with scale prior bias to obtain the fusion weights corresponding to each scale feature. Finally, the fusion weights are weighted and fused with the input feature map group to obtain the final output fused feature map.

[0014] Specifically, the process of inputting the SAR casualty image to be detected into the trained SAR image casualty target detection model for detection processing in step S3 is as follows: After the SAR casualty images to be detected are input into the trained SAR image casualty detection model, multi-scale features are extracted through the backbone feature extraction network with an embedded scattering-guided adaptive deformable convolution module. The multi-scale features are input into the scattering-guided distribution alignment module to obtain aligned multi-scale features. The aligned multi-scale features are input into the scale prior scattering perception feature fusion module to obtain the fused feature map; The fused feature map is input into the region proposal network to generate candidate regions. Then, the detection head performs classification and bounding box regression to obtain the category probability and bounding box regression results, and finally outputs the target location and category detection results of the injured person.

[0015] Specifically, the process of constructing the SAR wounded soldier image dataset in step S1 is as follows: In a simulation environment, a human feature scattering library is established based on the electromagnetic scattering characteristics of the human body. By adjusting the relative spatial position of multiple scattering point targets and the intensity of radar scattering cross section, the scattering echo characteristics of the human body under different trapped postures are simulated to generate a clean foreground image with human target scattering characteristics. The SAR background image obtained from actual measurements by a low-altitude UAV was selected as the background layer. The clean foreground image and the background image are adaptively fused, and the human target is randomly mapped and normalized to generate the original synthetic SAR image sample. The original synthetic SAR image samples are uniformly cropped, filtered, and augmented to obtain a SAR casualty image dataset, which is then divided into a training set, a validation set, and a test set.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned SAR image casualty target detection method based on multi-scale feature alignment.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention introduces a scattering-guided adaptive deformable convolution module in the Faster R-CNN framework. The deformable convolution branch adapts to the irregular scattering pattern of the wounded target, and combines logarithmically modulated Gaussian clustering attention to enhance the response of the weak scattering main region, thereby improving the feature representation ability of weakly scattering small targets and reducing missed detections.

[0018] (2) This invention employs a scattering-guided distribution alignment module, which reduces the positional offset and distribution differences of the main scattering regions between features of different scales and improves the consistency of multi-scale features through foreground-guided weighted statistics and three constraints: mean, variance, and symmetric KL divergence. Simultaneously, a scale-prior scattering perception feature fusion module is used, which strengthens the contribution of shallow high-resolution features to the details of small wounded individuals through scattering texture pooling and scale-prior bias mechanisms, achieving adaptive weighted fusion. These improvements collectively enhance the accuracy and positioning precision of wounded target detection under complex background conditions, and reduce the false alarm rate and false negative rate.

[0019] (3) The present invention is based on SAR images to carry out casualty target detection, which can make full use of the technical advantages of SAR imaging, which is available all day and all weather and has strong anti-obstruction capabilities, and improve the applicability of casualty target detection in complex disaster environments; (4) This invention alleviates the problems of difficulty in obtaining real samples and insufficient sample size by constructing a SAR casualty detection dataset with high physical consistency (based on the simulation of a pure foreground based on the electromagnetic scattering characteristics of the human body and adaptively fused with the measured SAR background), and provides a reliable data foundation for training the detection model. Attached Figure Description

[0020] Figure 1 The flowchart of the wounded target detection method based on SAR images of this invention.

[0021] Figure 2 This is a schematic diagram of the structure of a SAR image casualty detection model based on an improved Faster R-CNN.

[0022] Figure 3 A schematic diagram of the scatter-guided adaptive deformable convolution (SGADConv) module structure.

[0023] Figure 4 This is a schematic diagram of the scatter-guided distribution alignment (SGDAM) module structure.

[0024] Figure 5 This is a schematic diagram of the Scale-Prior Scattering Awareness Feature Fusion (SSAFM) module. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions or modifications made without departing from the concept of the present invention should fall within the scope of protection of the present invention.

[0026] like Figure 1As shown, this invention provides a method for detecting wounded targets in SAR images based on multi-scale feature alignment, comprising the following steps: Step S1: Construct a SAR wounded soldier image dataset; Step S2: Based on the Faster R-CNN model, the convolutional module in the feature extraction stage is replaced by the scattering-guided adaptive deformable convolution (SGADConv) module, the multi-scale features are aligned by the scattering-guided distribution alignment (SGDAM) module, and the multi-scale features are fused by the scale prior scattering-aware feature fusion (SSAFM) module, thereby constructing a SAR image casualty detection model based on the improved Faster R-CNN, and the SAR image casualty detection model is trained using the SAR casualty image dataset; Step S3: Input the SAR casualty image to be detected into the trained SAR image casualty detection model for detection processing, and output the location and category detection results of the casualty target.

[0027] In this embodiment, the specific process of constructing the SAR casualty image dataset in step S1 is as follows: First, a human feature scattering library is established in a simulation environment based on the electromagnetic scattering characteristics of the human body. By adjusting the relative spatial position of multiple scattering point targets and the intensity of the radar cross section, the scattering echo characteristics of the human body under different trapped postures are simulated to generate a clean foreground image with human target scattering characteristics. Second, a SAR background image obtained from actual measurements by a low-altitude UAV is selected as the background layer. Then, the clean foreground image and the background image are adaptively fused, and the human target is randomly mapped and normalized to generate original synthetic SAR image samples. Finally, the original synthetic SAR image samples are uniformly cropped, filtered, and data-enhanced to obtain the SAR casualty image dataset. The SAR casualty image dataset is divided into a training set, a validation set, and a test set for subsequent model training, validation, and detection.

[0028] The specific process of uniformly cropping, filtering, and data augmenting the original synthetic SAR image samples is as follows: the original synthetic SAR image samples are uniformly cropped into fixed-size images, such as 256×256; the cropped images are filtered to delete invalid samples that do not meet the detection requirements; the filtered valid samples are subjected to data augmentation processing, which includes horizontal flipping, vertical flipping, multi-angle rotation, and adding speckle noise.

[0029] like Figure 2As shown, in this embodiment, the SAR image casualty target detection model constructed in step S2 uses Faster R-CNN as the basic detection framework. The SAR casualty image to be detected is first input into the ResNet50 backbone network with the SGADConv module for feature extraction, resulting in multi-scale features (composed of the first scale feature C2 output by the second stage of the ResNet50 backbone network, the second scale feature C3 output by the third stage, the third scale feature C4 output by the fourth stage, and the fourth scale feature C5 output by the fifth stage; the first scale feature C1 output by the first stage of the ResNet50 backbone network is discarded). Subsequently, the multi-scale features are input into the Scattering Guided Distribution Alignment (SGDAM) module to perform distribution consistency alignment of the main scattering regions of the human target in different scale features, resulting in aligned multi-scale features. Then, the aligned multi-scale features are input into the Scale Prior Scattering Awareness Feature Fusion (SSAFM) module for multi-scale feature fusion, resulting in a fused feature map.

[0030] The fused feature map is input into a Region Proposal Network (RPN) to generate candidate regions through 3×3 convolutions, classification branches, and regression branches. Simultaneously, the fused feature map and the candidate regions are input into a Region of Interest Alignment (RoI Align) module for candidate region feature mapping. The RoI-aligned feature map is then further input into a detection head for classification and bounding box regression, yielding class probabilities and bounding box regression results, ultimately outputting the location and class of the injured target.

[0031] Therefore, the SAR image casualty target detection model, while retaining the candidate region generation, region of interest alignment, and classification and regression capabilities of Faster R-CNN, enhances the feature representation capability of weakly scattering casualty targets by introducing the SGADConv module into the ResNet50 backbone network, improves the distribution consistency between multi-scale features by using a multi-scale feature input scattering-guided distribution alignment module, and enhances the fusion and discrimination capability of aligned multi-scale features by using a scale prior scattering perception feature fusion module. This improves the accuracy and localization capability of SAR image casualty target detection under complex background conditions.

[0032] In this embodiment, the scattering-guided adaptive deformable convolution (SGADConv) module described in step S2 is mainly used to enhance the feature representation capability of the main scattering region of the wounded target. Traditional convolution modules use fixed sampling positions, which are difficult to adapt to the irregular scattering patterns of human targets in SAR images. Therefore, this embodiment introduces a deformable convolution branch on the basis of the standard convolution branch to improve the adaptive sampling capability of the local scattering structure of the human target; at the same time, a Gaussian clustering attention branch is introduced to extract the response of the main scattering region of the human body, and a logarithmic modulation mechanism is added in the spatial attention generation process to enhance weakly scattering targets and suppress strong background clutter. Through the above processing, the effective response of the weakly scattering target of the human body can be highlighted while maintaining background stability.

[0033] like Figure 3 As shown, the workflow of the SGADConv module is as follows: The input feature map X is fed into a standard convolutional branch, a deformable convolutional branch, a spatial attention branch, and a Gaussian clustering attention branch. The standard convolutional branch extracts basic local features, the deformable convolutional branch adaptively samples the irregular scattering structure of the human target, the spatial attention branch generates a spatial attention response, and the Gaussian clustering attention branch extracts the clustering response of the main scattering region of the human body. Subsequently, the outputs of the spatial attention branch and the Gaussian clustering attention branch are fused to generate a log-modulated Gaussian spatial attention; then, the output of the deformable convolutional branch is weighted using the log-modulated Gaussian spatial attention to obtain attention-enhanced deformable features; finally, the attention-enhanced deformable features are concatenated with the output features of the standard convolutional branch and fused using a 1×1 convolution to obtain the output feature map Y of the SGADConv module. The SGADConv module improves the adaptability to the irregular scattering structure of human targets by changing the fixed sampling method to a deformable sampling method. At the same time, it extracts the response of the main scattering region of the human body through Gaussian clustering attention and enhances weak scattering targets by combining logarithmically modulated spatial attention, thereby improving the feature representation ability of wounded targets under complex background conditions.

[0034] In this embodiment, the spatial attention branch includes, in sequence, spatial attention convolution, logarithmic transformation, and a sigmoid activation function. The output of the spatial attention branch can be expressed as: ; Where X is the input feature map. Represents spatial attention convolution. This represents the modulation coefficient used to adjust the logarithmic modulation intensity. Represents the natural logarithm function. This represents the Sigmoid activation function. The spatial attention response is represented by [missing information]. This spatial attention response is fused with the Gaussian clustering attention response, and the output of the deformable convolution branch is weighted and then concatenated with the output features of the standard convolution branch. Finally, it is fused using a 1×1 convolution to obtain the output feature map Y of the SGADConv module, which can be represented as: ; in, This indicates a feature concatenation operation. This represents a 1×1 convolution fusion operation. This indicates attention-enhanced deformable features. This represents the output features of the standard convolution branch.

[0035] In this embodiment, the Scatter-Guided Distribution Alignment (SGDAM) module described in step S2 is mainly used to reduce the differences in scattering distribution between features of different scales. For SAR casualty detection tasks, the responses of the human body's main scattering region in features of different scales often exhibit positional offsets and statistical inconsistencies. Direct fusion can easily lead to misalignment of effective features or their being submerged by background statistics. The SGDAM module, by introducing foreground-guided statistical modeling and distribution consistency constraints, reduces the differences in scattering distribution between features of different scales and improves the consistency of the expression of the human body's main scattering region in features of different scales, thereby providing a more stable feature foundation for subsequent multi-scale feature fusion.

[0036] like Figure 4 As shown, the workflow of the SGDAM module is as follows: First, the input multi-scale features (C2, C3, C4, and C5) are input into the scattering consistency alignment unit. The positional response of the main scattering region of the human target in each scale feature is adjusted to make the spatial distribution of the target region in different scale features as consistent as possible, thus obtaining the aligned multi-scale features (aligned features of the first scale). Aligned features at the second scale Aligned features at the third scale Aligned features with the 4th scale ).

[0037] Subsequently, a foreground guiding weight map is generated based on the aligned multi-scale features. Specifically, the aligned multi-scale features are input into the foreground guiding weight map generation module, and the foreground guiding weight map is obtained through feature mapping and nonlinear activation. The foreground guiding weight map Used to highlight the main scattering region of the human target and suppress background clutter regions.

[0038] In this embodiment, for ease of formula expression, the aligned multi-scale features are... , , and Unified representation as the first Aligned features at each scale ,in, For scale indexing. The foreground guided weight map. The generation process can be represented as: ; in, ; In the formula, This represents a scale transformation operation, used to map aligned features of different scales to the same spatial size; This indicates a channel-level concatenation operation; This represents a 1×1 convolution operation; This represents the Sigmoid activation function; This represents the feature obtained by concatenating aligned multi-scale features; This represents a foreground-guided weighting graph.

[0039] Subsequently, the foreground guiding weight map is used. Foreground-weighted statistics were performed on the aligned multi-scale features to obtain the foreground-weighted mean of the aligned features at each scale. Prospect-weighted variance Among them, the first Foreground-weighted mean of aligned features at each scale It can be represented as: ; No. Foreground-weighted variance of features at each scale It can be represented as: ; In the formula, Indicates the first Aligned features at each scale; This represents the foreground guided weight map that matches the aligned features at the k-th scale. Indicates the spatial location index in the feature map; This represents a very small constant used to prevent the denominator from being zero.

[0040] After obtaining the foreground-weighted mean and foreground-weighted variance of the aligned features at each scale, a Gaussian distribution model is performed. Based on the statistical distribution differences between the aligned features at different scales, mean consistency constraints, variance consistency constraints, and symmetric KL divergence constraints are established to construct the total alignment loss of the scattering-guided distribution alignment (SGDAM) module, which can be expressed as: ; In the formula, This represents the mean consistency constraint term; This represents the variance consistency constraint term; Represents the symmetric KL divergence constraint term; , and These are the weighting coefficients for each loss item; This represents the total alignment loss of the scattering-guided distribution alignment module.

[0041] In this embodiment, the Scale Prior Scattering Awareness Feature Fusion (SSAFM) module described in step S2 is mainly used to enhance the multi-scale feature fusion capability after alignment. For SAR casualty detection tasks, casualty targets typically exhibit characteristics such as weak scattering, small scale, and fragmented local structures. Different scale features have varying representation capabilities for target details, contextual semantics, and background clutter. If fusion is performed using simple addition or splicing, the weak target details in shallow high-resolution features are easily submerged by deep semantic features, and background clutter interference may also be introduced, reducing the ability to distinguish casualty targets in complex scenes. The SSAFM module introduces joint scattering texture pooling, residual hole depth separable convolution, PWConv pointwise convolution (1×1 convolution), and softmax fusion with scale prior bias to adaptively weight and fuse the aligned multi-scale features. This highlights the effective responses related to the main scattering region of the human target in different scale features, suppresses redundant background information, and improves the multi-scale feature fusion effect for weakly scattering small targets.

[0042] like Figure 5 As shown, the workflow of the SSAFM module is as follows: First, the aligned multi-scale features output by the SGDAM module are used as the input feature map set. ,in, This represents the scale index. Then, the input feature maps for each scale are processed. Perform joint scattering texture pooling to obtain the stitched pooled feature map. Subsequently, the spliced ​​pooling feature map Inputting a residual-based dilation depth separable convolutional structure yields context-enhanced feature maps. Then, through pointwise convolution using PWConv (i.e., 1×1 convolution), a spatial attention map of the corresponding scale is generated. .

[0043] Subsequently, the spatial attention maps corresponding to each scale were generated. A softmax fusion unit with a scale prior bias is input to obtain the fusion weights corresponding to features at each scale. The scale prior bias is used to ensure that features at different scales have differentiated response capabilities during the fusion process, thereby enhancing the contribution of shallow high-resolution features to the detail information of weakly scattering small targets. Finally, the fusion weights are combined with the input feature map group. Weighted fusion is performed to obtain the final output fused feature map. .

[0044] In this embodiment, the first The concatenated pooled feature map obtained by pooling the aligned features at each scale after joint scattering texture can be represented as: ; in, ( () indicates the average pooling operation, used to extract the overall intensity response information in the feature map; This represents a local variance pooling operation, used to extract local scattering texture information from the feature map; ( () indicates a channel-level concatenation operation; Indicates the first Pooled feature map after splicing at each scale.

[0045] Furthermore, the context-enhanced feature map can be represented as: ; in, This indicates a separable convolution operation with residual hole depth. Indicates the first Context-enhanced feature maps at various scales.

[0046] After obtaining the context-enhanced feature map, a spatial attention map is generated through pointwise convolution using PWConv, which can be represented as: ; in, express Convolution operation; Indicates the first Each scale input feature map group The corresponding spatial attention map.

[0047] In this embodiment, the first The fusion weights corresponding to each scale feature can be expressed as: ; in, Indicates the first input feature map group The corresponding spatial attention map; Indicates the first The scale prior bias corresponding to each scale; Indicates the first The scale prior bias corresponding to each scale; Indicates the number of scales involved in the integration; Indicates the first The fusion weights corresponding to each scale feature.

[0048] Furthermore, the final output fused feature map after multi-scale feature fusion can be represented as: ; in, This represents element-wise multiplication; This represents the final output fused feature map.

[0049] In this embodiment, the specific process of inputting the SAR casualty image to be detected into the trained SAR image casualty target detection model for detection processing is as follows: The SAR casualty image to be detected is first input into the trained SAR image casualty target detection model. After the backbone feature extraction network extracts initial features at different levels, it is input into the scattering-guided adaptive deformable convolution module to enhance the feature representation ability of the main scattering region of the human target, thus obtaining enhanced multi-scale features. Subsequently, the multi-scale features are input into the scattering-guided distribution alignment module to perform distribution consistency alignment of features at different scales, thus obtaining aligned multi-scale features. Then, the aligned multi-scale features are input into the scale prior scattering perception feature fusion module to perform adaptive weighted fusion of the aligned multi-scale features, thus obtaining a fused feature map. Finally, the fused feature map is input into the region proposal network to generate candidate target regions, and combined with the region of interest alignment module and the detection head, the casualty target classification and bounding box position regression are completed, and the casualty target location and category detection results in the SAR casualty image to be detected are output.

[0050] In this embodiment, the specific process of outputting the location and category detection results of the casualty targets in the SAR casualty image to be detected is as follows: After the trained SAR image casualty target detection model completes the generation of candidate target regions, target classification, and bounding box position regression, it filters the detection results to obtain casualty target candidate regions that meet preset conditions; based on the classification results corresponding to the candidate regions, it is determined whether there are casualty targets in the SAR casualty image to be detected, and the location information of the casualty targets in the SAR casualty image is determined based on the bounding box position regression results; finally, the location and category detection results of the casualty targets are output to provide a basis for target location and rescue decision-making in subsequent disaster search and rescue.

[0051] To verify the effectiveness of the SAR image casualty target detection method based on multi-scale feature alignment described in this invention, training parameters were set as follows: number of iterations: 200; image size: 640×640; batch size: 16. The training set and validation set obtained in step S1 were used to train the SAR image casualty detection model based on the improved Faster R-CNN. During training, the optimal SAR image casualty detection model was selected by monitoring the loss curve and learning rate changes.

[0052] The Faster R-CNN model and the SAR image casualty detection model based on the improved Faster R-CNN are compared. Table 1 shows the comparison of the performance indicators of the Faster R-CNN model before and after the improvement. Faster R-CNN-SGADConv represents the Faster R-CNN model improved with the SGADConv module, Faster R-CNN-SGDAM represents the Faster R-CNN model improved with the SGDAM module, Faster R-CNN-SSAFM represents the Faster R-CNN model improved with the SSAFM module, Faster R-CNN-SGADConv-SGDAM represents the Faster R-CNN model improved with the SGADConv and SGDAM modules, Faster R-CNN-SGDAM-SSAFM represents the Faster R-CNN model improved with the SGDAM and SSAFM modules, and Faster R-CNN-SGADConv-SGDAM-SSAFM represents the Faster R-CNN model improved with the SGADConv, SGDAM and SSAFM modules. The performance evaluation indicators are P, R, mAP50 and mAP50-95. Where: P represents precision, R represents recall, mAP50 represents the average detection precision of the target when the cross-union ratio threshold is 0.5, and mAP50-95 represents the average detection precision of the target when the cross-union ratio threshold is between 0.5 and 0.95.

[0053] Table 1

[0054] As shown in Table 1, the SGADConv module, SGDAM module, and SSAFM module can all improve the performance of the original Faster R-CNN model. Furthermore, after fusing the three improvements, the improved Faster R-CNN-SGADConv-SGDAM-SSAFM model outperforms the original Faster R-CNN model in detection performance, indicating that the improved method proposed in this invention can effectively improve the performance of SAR image wounded target detection.

[0055] In summary, this invention constructs a SAR casualty detection dataset with high physical consistency and sequentially introduces the SGADConv, SGDAM, and SSAFM modules onto the Faster R-CNN framework. This enables enhanced feature extraction, multi-scale feature distribution alignment, and adaptive weighted fusion for casualty targets with weak scattering, small scale, and irregular shapes, thereby improving the accuracy, positioning precision, and robustness of SAR casualty target detection under complex background conditions.

[0056] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting a wounded target in a SAR image based on multi-scale feature alignment, characterized in that, Includes the following steps: Step S1: Construct a SAR wounded soldier image dataset; Step S2: Based on the Faster R-CNN model, A scattering-guided adaptive deformable convolution module is used to replace the convolution module in the feature extraction stage for extracting multi-scale features. The scattering-guided adaptive deformable convolution module includes a deformable convolution branch and a spatial attention branch based on Gaussian clustering attention. The Gaussian clustering attention and spatial attention are fused by logarithmic modulation. The output of the deformable convolution branch is weighted and then concatenated with the output of the standard convolution branch and fused by 1×1 convolution. A scattering-guided distribution alignment module is used to align multi-scale features. The scattering-guided distribution alignment module weights the aligned features at each scale based on the foreground guided weight map, calculates the foreground weighted mean and foreground weighted variance of the aligned features at each scale, and performs distribution alignment using mean consistency constraints, variance consistency constraints, and symmetric KL divergence constraints. A scale-prior scattering perception feature fusion module is used to fuse aligned multi-scale features. The scale-prior scattering perception feature fusion module performs average pooling and local variance pooling on the aligned features at each scale, generates a spatial attention response after residual-type dilated depth convolution context aggregation, and introduces scale prior bias for normalized weighted fusion. Thus, a SAR image casualty detection model based on an improved Faster R-CNN was constructed, and the SAR image casualty detection model was trained using a SAR casualty image dataset; Step S3: Input the SAR casualty image to be detected into the trained SAR image casualty detection model for detection processing, and output the location and category detection results of the casualty target.

2. The method according to claim 1, wherein, The workflow of the scattering-guided adaptive deformable convolution module is as follows: The input feature maps are respectively fed into the standard convolution branch, deformable convolution branch, spatial attention branch, and Gaussian clustering attention branch; The output of the spatial attention branch is fused with the output of the Gaussian clustering attention branch to generate a log-modulated Gaussian spatial attention. The output of the deformable convolution branch is weighted using the log-modulated Gaussian spatial attention to obtain attention-enhanced deformable features. The attention-enhanced deformable features are concatenated with the output features of the standard convolutional branch, and then fused through a 1×1 convolution to obtain the output feature map.

3. The method according to claim 2, wherein, The output of the spatial attention branch is represented as follows: ； in, Represents spatial attention convolution. Represents the logarithmic modulation coefficient. This represents the Sigmoid activation function. Indicates spatial attention response, This is the input feature map.

4. The method of claim 1, wherein the method comprises: The workflow of the scattering-guided distribution alignment module is as follows: The multi-scale features are input into the scattering consistency alignment unit, and the position response of the main scattering region of the human target in each scale feature is adjusted to obtain the aligned multi-scale features. Subsequently, a foreground guiding weight map is generated based on the aligned multi-scale features; Subsequently, the foreground-guided weight map was used to perform foreground-weighted statistics on the aligned multi-scale features, and the foreground-weighted mean and foreground-weighted variance of the aligned features at each scale were obtained respectively. After obtaining the foreground-weighted mean and foreground-weighted variance of the aligned features at each scale, Gaussian distribution modeling is performed. Based on the statistical distribution differences between the aligned features at different scales, mean consistency constraints, variance consistency constraints, and symmetric KL divergence constraints are established to construct the total alignment loss of the scattering-guided distribution alignment module.

5. The method of claim 4, wherein the method further comprises: The foreground guidance weight map The generation process is represented as: ； ； In the formula, This represents a scale transformation operation, used to map aligned features of different scales to the same spatial size; This indicates a channel-level concatenation operation; This represents a 1×1 convolution operation; This represents the Sigmoid activation function; This represents the feature obtained by concatenating aligned multi-scale features; For the aligned features of the first scale, For the aligned features of the second scale, This represents the aligned features at the third scale. This is the aligned feature at the fourth scale.

6. The method of claim 1, wherein the method further comprises: The workflow of the scale prior scattering sensing feature fusion module is as follows: First, the aligned multi-scale features are used as the input feature map set, and then the input feature map sets at each scale are processed. Joint scattering texture pooling is performed to obtain a stitched pooled feature map; then, the stitched pooled feature map is input into a residual hole depth separable convolutional structure to obtain a context-enhanced feature map; then, a spatial attention map of the corresponding scale is generated by pointwise convolution of PWConv. Subsequently, the spatial attention maps corresponding to each scale are input into a softmax fusion unit with scale prior bias to obtain the fusion weights corresponding to each scale feature. Finally, the fusion weights are weighted and fused with the input feature map group to obtain the final output fused feature map.

7. The method of claim 6, wherein the method further comprises: No. The fusion weights corresponding to each scale feature are represented as follows: ； in, Indicates the first Each scale input feature map group The corresponding spatial attention map, Indicates the first input feature map group The corresponding spatial attention map; Indicates the first The scale prior bias corresponding to each scale; Indicates the first The scale prior bias corresponding to each scale; Indicates the number of scales involved in the integration; Indicates the first The fusion weights corresponding to each scale feature. 8.The method of claim 1, wherein, The specific process of inputting the SAR casualty image to be detected into the trained SAR image casualty target detection model in step S3 is as follows: After the SAR casualty images to be detected are input into the trained SAR image casualty detection model, multi-scale features are extracted through the backbone feature extraction network with an embedded scattering-guided adaptive deformable convolution module. The multi-scale features are input into the scattering-guided distribution alignment module to obtain aligned multi-scale features. The aligned multi-scale features are input into the scale prior scattering perception feature fusion module to obtain the fused feature map; The fused feature map is input into the region proposal network to generate candidate regions. Then, the detection head performs classification and bounding box regression to obtain the category probability and bounding box regression results, and finally outputs the target location and category detection results of the injured person.

9. The SAR image casualty target detection method based on multi-scale feature alignment according to claim 1, characterized in that, The specific process of constructing the SAR wounded soldier image dataset in step S1 is as follows: In a simulation environment, a human feature scattering library is established based on the electromagnetic scattering characteristics of the human body. By adjusting the relative spatial position of multiple scattering point targets and the intensity of radar scattering cross section, the scattering echo characteristics of the human body under different trapped postures are simulated to generate a clean foreground image with human target scattering characteristics. The SAR background image obtained from actual measurements by a low-altitude UAV was selected as the background layer. The clean foreground image and the background image are adaptively fused, and the human target is randomly mapped and normalized to generate the original synthetic SAR image sample. The original synthetic SAR image samples are uniformly cropped, filtered, and augmented to obtain a SAR casualty image dataset, which is then divided into a training set, a validation set, and a test set.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the SAR image wound target detection method based on multi-scale feature alignment as described in any one of claims 1 to 9.

Images