A wounded person detection method and electronic equipment based on SAR image

Abstract

This invention discloses a casualty detection method and electronic device based on SAR images. Based on the YOLOv8 model, this method replaces the C2F module in the YOLOv8 model with the C2F-PSConv module. The neck network uses a multi-scale feature fusion module for feature fusion. An adaptive rotational convolution module is added before the detection head of the head network to construct a casualty detection model based on an improved YOLOv8. The casualty detection model is trained using a training set and a validation set. SAR images transmitted by an unmanned aerial vehicle are received by a ground control station or a mobile terminal, and the trained casualty detection model is used for detection, obtaining the casualty's location information and issuing an alarm. This invention can improve the accuracy of casualty detection based on SAR images.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a casualty detection method and electronic equipment based on SAR images. Background Art

[0002] In the field of emergency rescue, quickly and accurately obtaining disaster-site information is key to formulating rescue strategies and reducing casualties and property losses. Traditional methods that rely on optical remote sensing and ground surveys have significant limitations: optical imaging is easily restricted by environmental conditions such as clouds, rain, fog, and darkness at night, making it difficult to obtain effective information during the "golden rescue period" after a disaster; ground surveys are limited by factors such as traffic blockages and the risk of secondary disasters, making timeliness and safety difficult to guarantee. Synthetic aperture radar (SAR) has become an irreplaceable sensing tool in emergency rescue due to its all-day, all-weather imaging capabilities and its ability to penetrate the ground. SAR can penetrate clouds, smoke, and shallow vegetation to obtain key information such as earthquake-collapsed building structures, flood-inundated areas, and landslide deformation, providing high-value data support for disaster assessment and rescue decision-making.

[0003] Current target detection algorithms based on SAR images fall into two main categories. Traditional target detection methods based on synthetic aperture radar (SAR) images primarily rely on the statistical distribution of background clutter and manually designed feature extraction algorithms. The constant false alarm rate (CFAR) detection algorithm is a widely used technique. It calculates the background cells around the target to determine a fixed false alarm probability and sets the detection threshold accordingly. However, when processing wide imagery, the CFAR algorithm uses the same statistical model for all clutter within the sliding window, which can lead to mismatches in inappropriate areas. Another category of methods utilizes advanced deep learning target detection techniques, such as YOLO and Faster R-CNN. These methods leverage neural network models to automatically extract and learn high-level semantic features in images, enabling accurate localization and classification of casualty targets. Compared to traditional image processing techniques, deep learning methods exhibit greater robustness in adapting to complex environmental changes and can address a wider range of scenario challenges.

[0004] However, deep learning methods also present inherent challenges. For example, inherent geometric distortion, masking, and speckle noise in SAR images lead to a sharp drop in the detection accuracy of traditional algorithms in dense urban areas or areas covered by vegetation. Furthermore, deep learning models face the high computational complexity and the insignificant difference between the features of casualties in SAR images and those of the surrounding environment and other non-target objects. Therefore, research on efficient and accurate casualty detection techniques to address the shortcomings of these existing methods is crucial for promoting the application of SAR imagery in disaster relief, helping rescue workers locate casualties more quickly and accurately, and thus effectively reducing casualty rates. Summary of the Invention

[0005] In order to solve the problems of false detection and missed detection caused by complex background interference of SAR images and small size of wounded targets, the present invention provides a wounded detection method and electronic equipment based on SAR images.

[0006] The present invention provides the following technical solution: a method for detecting casualties based on SAR images, comprising the following steps:

[0007] Step S1: Construct a SAR image dataset and divide it into a training set and a validation set;

[0008] Step S2: Based on the YOLOv8 model, PSConv convolution (multi-scale convolution) is introduced to replace the bottleneck layer in the C2F module, forming a C2F-PSConv module. The C2F-PSConv module is used to replace the C2F module in the YOLOv8 model. The neck network uses a multi-scale feature fusion module for feature fusion. The head network uses four detection heads, and an adaptive rotation convolution module (ARConv) is added before each detection head. This constructs a casualty detection model based on the improved YOLOv8, and the casualty detection model is trained using the training set and the validation set.

[0009] Step S3: Use the ground control station or mobile terminal to receive the SAR image transmitted by the UAV, use the trained casualty detection model to detect, obtain the location information of the casualty and issue an alarm.

[0010] Further preferably, the collected SAR background image and the wounded target template image are preprocessed, including image size unification and grayscale normalization, feature enhancement of the wounded target template image, and the preprocessed wounded target template image is superimposed on the SAR background image using an image processing library to form a wounded SAR image.

[0011] Further preferably, a transformation operation is performed on the SAR image of the wounded person to simulate wounded targets in different postures and positions.

[0012] Further preferably, an image annotation tool is used to carefully annotate the generated SAR images of the wounded, marking the location information and category labels of the wounded targets; the annotated SAR images of the wounded are processed through data enhancement operations, and the data enhancement operations include multi-angle rotation, random scaling and asymmetric cropping, and the SAR images of the wounded are collected as a SAR image dataset of the wounded.

[0013] Further preferably, the processing process of the adaptive rotation convolution module is expressed as:

[0014] ;

[0015] in It is the output feature map, which represents the result after rotation convolution, and its dimension is consistent with the input feature map X. is the jth convolution kernel, Indicates the rotation transformation of the j-th convolution kernel, is the rotation angle, is the weight coefficient, which is used to perform weighted fusion on the rotated convolution kernel output.

[0016] Further preferably, the processing process of the C2F-PSConv module is expressed as:

[0017] ;

[0018] in, is the output feature map, express module, and They are input feature maps through The two branches of the split after module processing each occupy half of the number of intermediate channels. i is the serial number of the PSConv convolution operation currently being executed, and Concat means splicing along the channel dimension.

[0019] Further preferably, the processing flow of the multi-scale feature fusion module of this embodiment is as follows: the first feature map is input into the first multi-scale attention, and the second feature map is input into the second multi-scale attention; the first feature map and the second feature map are spliced and input into the first depth-separable convolution for processing, and the obtained outputs are respectively input into the second depth-separable convolution, the third depth-separable convolution, the fourth depth-separable convolution, and the fifth depth-separable convolution, and the outputs of the second depth-separable convolution, the third depth-separable convolution, the fourth depth-separable convolution, and the fifth depth-separable convolution are added, and then input into the third multi-scale attention after 1×1 convolution channel by channel; the output of the third multi-scale attention is multiplied by the output of the first multi-scale attention and the second multi-scale attention respectively, and the obtained output is added with the output of the third multi-scale attention, the first feature map, and the second feature map, and then input into the fourth multi-scale attention through a 1×1 convolution operation, and the output of the fourth multi-scale attention is used as the final output of the multi-scale feature fusion module.

[0020] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the instructions are executed by the processor, the processor implements the above-mentioned SAR image-based casualty detection method.

[0021] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the above-mentioned method for detecting casualties based on SAR images.

[0022] Compared with the existing technology, the present invention has the following beneficial effects:

[0023] (1) By synthesizing images of injured persons and constructing an annotated dataset based on image transformation and superposition, the insufficient number and diversity of datasets are compensated, which can effectively improve the generalization ability of the model and enable it to accurately detect injured persons in various complex and changing emergency rescue scenarios.

[0024] (2) The PSConv convolution is introduced and combined with the C2F module in YOLOv8 to generate a new C2F-PSConv module, which realizes multi-scale feature extraction at a finer granularity and improves the model's adaptability to complex scenes (such as dense occlusion and low contrast).

[0025] (3) A multi-scale feature fusion module is designed to generate clearer feature representations, which can promote the feature extraction of local and global context information and improve the model detection accuracy.

[0026] (4) The detection head branch is added and the head network of the YOLOv8 model is improved. The characteristics of the dynamic rectangular convolution kernel and adaptive sampling density of the ARConv module are used to solve the limitations of traditional convolution in multi-scale target detection in SAR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Flow chart of the method of the present invention.

[0028] Figure 2 This is a framework diagram of the improved wounded target detection model based on YOLOv8 of the present invention.

[0029] Figure 3 This is a diagram of the detection head framework of the present invention.

[0030] Figure 4 This is the framework diagram of the C2F-PSConv module of the present invention.

[0031] Figure 5 This is a framework diagram of the multi-scale feature fusion module of the present invention. DETAILED DESCRIPTION

[0032] The present invention is further described below with reference to the embodiments. It is necessary to point out that the following embodiments are only used to further illustrate the present invention and are not to be construed as limiting the scope of protection of the present invention. Non-essential improvements and adjustments made by persons skilled in the art based on the above-mentioned invention still fall within the scope of protection of the present invention.

[0033] like Figure 1 As shown, the present invention provides a technical solution: a method for detecting casualties based on SAR images, comprising the following steps:

[0034] Step S1: Construct a SAR image dataset of casualties and divide it into a training set and a validation set;

[0035] Step S2: Based on the YOLOv8 model, PSConv convolution (multi-scale convolution) is introduced to replace the bottleneck layer (Bottleneck) in the C2F module, forming the C2F-PSConv module. The C2F-PSConv module is used to replace the C2F module in the YOLOv8 model. The neck network uses a multi-scale feature fusion module for feature fusion. The head network uses four detection heads, and an adaptive rotation convolution module (ARConv module) is added before each detection head. This constructs a casualty detection model based on the improved YOLOv8, and the casualty detection model is trained using the training set and the validation set.

[0036] Step S3: Use the ground control station or mobile terminal to receive the SAR image transmitted by the UAV, use the trained casualty detection model to detect, obtain the location information of the casualty and issue an alarm.

[0037] The specific process of constructing the SAR image dataset of casualties and dividing it into training set and validation set is as follows:

[0038] A sufficient number of SAR images were collected as background scenes, covering scenes with varying terrain, weather conditions, and imaging parameter settings. Clear images of the casualty target were obtained. The collected SAR background images and the casualty target template images were preprocessed, including image resizing and grayscale normalization, to ensure that all images had the same format and scale for subsequent overlay processing. Feature enhancement was performed on the casualty target template image to better match the imaging characteristics of SAR images. Using the image processing library OpenCV, the preprocessed casualty target template image was overlaid onto the SAR background image to form the casualty target SAR image. During the overlay process, appropriate positioning was determined to prevent the target from exceeding the image boundaries or causing unreasonable overlap. A series of transformation operations were performed on the overlaid casualty target SAR image to simulate the casualty target in different poses and positions. Image annotation tools were used to carefully annotate the resulting casualty target SAR image, marking the casualty target's location information and category label. The annotated SAR images of injured persons were processed through data augmentation operations, including multi-angle rotation, random scaling, and asymmetric cropping. Multi-angle rotation rotates the injured person images from 0 to 360 degrees at intervals of 5 degrees. Random scaling scales the injured person SAR images by 0.5 to 2 times to simulate target size changes at different imaging distances and enhance the model's adaptability to scale changes. Asymmetric cropping randomly crops 10% to 30% of the image edge while retaining the target area of the injured person to simulate partial occlusion scenarios. Overly blurry or incorrectly annotated SAR images of injured persons were deleted, and the SAR images were pooled as a dataset of injured person SAR images and divided into a training set and a validation set with a ratio of 8:2. Before the division, the entire SAR image dataset of injured persons was randomly shuffled to ensure uniform data distribution and avoid bias in the division results due to data order issues.

[0039] like Figure 2 As shown, the injured person detection model based on the improved YOLOv8 includes a backbone network, a neck network, and a head network. The backbone network is sequentially connected by a first CBS module, a second CBS module, a first C2F-PSConv module, a third CBS module, a second C2F-PSConv module, a fourth CBS module, a third C2F-PSConv module, a fifth CBS module, a fourth C2F-PSConv module, and an SPPF module.

[0040] The neck network performs multi-scale feature fusion, the feature map output by the SPPF module is up-sampled and then spliced with the feature map output by the third C2F-PSConv module for the first time, the feature map after the first splicing is sent to the fifth C2F-PSConv module for processing, the feature map output by the fifth C2F-PSConv module is up-sampled and then spliced with the feature map output by the second C2F-PSConv module for the second time, the feature map after the second splicing is sent to the sixth C2F-PSConv module for processing, the feature map output by the sixth C2F-PSConv module is up-sampled and then spliced with the feature map output by the first C2F-PSConv module for the third time, and the feature map after the third splicing is sent to the seventh C2F-PSConv module for processing; the feature map output by the seventh C2F-PSConv module is processed by the sixth CBS module, and then multi-scale feature fusion is performed on the feature map output by the first EMAF module and the feature map output by the sixth C2F-PSConv module to obtain a first fused feature map, and the first fused feature map is processed by the eighth C2F- PSConv module processing; the feature map output by the eighth C2F-PSConv module is processed by the seventh CBS module, and then the feature map output by the second EMAF module and the fifth C2F-PSConv module is subjected to multi-scale feature fusion to obtain a second fused feature map; the second fused feature map is processed by the ninth C2F-PSConv module, and the feature map output by the ninth C2F-PSConv module is processed by the eighth CBS module, and then the feature map output by the third EMAF module and the SPPF module is subjected to multi-scale feature fusion to obtain a third fused feature map, and the third fused feature map is sent to the tenth C2F-PSConv module for processing; the head network includes four detection heads, namely the first detection head, the second detection head, the third detection head, and the fourth detection head; the feature maps output by the seventh C2F-PSConv module, the eighth C2F-PSConv module, the ninth C2F-PSConv module, and the tenth C2F-PSConv module are selected as the input of the first detection head, the second detection head, the third detection head, and the fourth detection head, respectively.

[0041] like Figure 3 As shown in the figure, the first detection head, the second detection head, the third detection head, and the fourth detection head have the same structure. The feature map of the input detection head is first processed by the adaptive rotation convolution module (ARConv module) and then input into two branches respectively. Each branch consists of two consecutive CBS modules and a Conv2d module.

[0042] The processing process of the ARConv module is expressed as follows:

[0043] ;

[0044] in It is the output feature map, which represents the result after rotation convolution, and its dimension is consistent with the input feature map X. is the jth convolution kernel, Indicates the rotation transformation of the j-th convolution kernel, is the rotation angle, is the weight coefficient, which is used to perform weighted fusion on the rotated convolution kernel output.

[0045] like Figure 4 As shown in Figure 2, the processing of the C2F-PSConv module is expressed as:

[0046] ;

[0047] in, is the output feature map, express module, and They are input feature maps through The two branches of the split after module processing each occupy half of the number of middle channels. i is the serial number of the PSConv convolution operation currently being executed. The PSConv convolution operation is repeated three times. Concat means splicing along the channel dimension.

[0048] like Figure 5 As shown, the multi-scale feature fusion module of this embodiment is an EMAF module, which consists of a first multi-scale attention layer, a second multi-scale attention layer, a third multi-scale attention layer, a fourth multi-scale attention layer, a first depth-wise separable convolution layer, a second depth-wise separable convolution layer, a third depth-wise separable convolution layer, a fourth depth-wise separable convolution layer, a fifth depth-wise separable convolution layer, and two 1×1 convolutions (Conv). The first multi-scale attention layer, the second multi-scale attention layer, the third multi-scale attention layer, and the fourth multi-scale attention layer have the same structure. The depth-wise separable convolutions have the same structure: the convolution kernel size of the first depth-wise separable convolution is 3×3, the convolution kernel size of the third depth-wise separable convolution is 5×5, the convolution kernel size of the fourth depth-wise separable convolution is 7×7, and the convolution kernel size of the fifth depth-wise separable convolution is 9×9.

[0049] The processing flow of the EMAF module is as follows: the first feature map is input into the first multi-scale attention, and the second feature map is input into the second multi-scale attention; the first feature map and the second feature map are spliced and input into the first depth-wise separable convolution for processing, and the outputs are respectively input into the second depth-wise separable convolution, the third depth-wise separable convolution, the fourth depth-wise separable convolution, and the fifth depth-wise separable convolution. The outputs of the second depth-wise separable convolution, the third depth-wise separable convolution, the fourth depth-wise separable convolution, and the fifth depth-wise separable convolution are added, and then 1×1 convolution is performed channel by channel and then input into the third multi-scale attention; the output of the third multi-scale attention is multiplied with the output of the first multi-scale attention and the second multi-scale attention respectively, and the output is added with the output of the third multi-scale attention, the first feature map, and the second feature map, and then input into the fourth multi-scale attention through a 1×1 convolution operation. The output of the fourth multi-scale attention is used as the final output of the EFAM module.

[0050] Set the training parameters to 200 iterations, 640*640 image size, and 16 batch size. Use the training set and validation set obtained in step S1 to train the improved YOLOv8-based casualty detection model. During training, monitor the loss curve and learning rate changes to select the optimal casualty detection model.

[0051] The YOLOv8 model is compared with the casualty detection model based on the improved YOLOv8. Table 1 is a comparison table of YOLOv8 model indicators before and after improvement. YOLOv8-C2F-PSConv represents the YOLOv8 model improved with the C2F-PSConv module, YOLOv8-EMAF represents the YOLOv8 model improved with the EMAF module, YOLOv8-ARConv represents the YOLOv8 model improved with the ARConv module, YOLOv8-C2F-PSConv-EMAF represents the YOLOv8 model improved with the C2F-PSConv module and the EMAF module, and YOLOv8-C2F-PSConv-EMAF-ARConv represents the YOLOv8 model improved with the C2F-PSConv module, the EMAF module, and the ARConv module. The performance evaluation indicators are P, R, mAP50, and mAP50-95. Where: P represents precision, R represents recall, mAP50 represents the average detection accuracy of the target when the intersection-over-union threshold is 0.5, and mAP50-95 represents the average detection accuracy of the target when the intersection-over-union threshold is 0.5-0.95.

[0052] Table 1

[0053]

[0054] As can be seen from Table 1, the C2F-PSConv module, the EMAF module, and the ARConv module can all improve the accuracy of the original YOLOv8 model. At the same time, by fusing the three improvements, it is found that the improved algorithm is better than the original YOLOv8 model in effect.

[0055] The present invention utilizes an unmanned aerial vehicle (UAV) equipped with a MiniSAR system for SAR image acquisition and imaging. When the airborne MiniSAR system flies over a disaster area, it emits high-frequency radar waves that can penetrate various obstacles, such as bricks and tree branches in ruins, and receive reflected signals from trapped individuals or their surroundings. SAR images are generated by accurately analyzing and processing these reflected signals. Combined with the improved YOLOv8-based casualty detection model, the location of trapped individuals can be identified from complex backgrounds and an alert can be issued. Based on this precise location information, rescue teams can quickly develop a rescue plan and use specialized rescue equipment to extricate the trapped individuals. MiniSAR penetrating imaging technology overcomes environmental limitations, while the improved YOLOv8-based casualty detection model improves search and rescue accuracy. Communication relay ensures command efficiency, thus establishing a comprehensive "detection-analysis-action" emergency rescue system.

[0056] A second embodiment of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the instructions are executed by the processor, the processor implements the above-mentioned SAR image-based casualty detection method.

[0057] A third embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the above-mentioned method for detecting casualties based on SAR images.

[0058] The above description merely represents preferred embodiments of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above disclosure to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for detecting casualties based on SAR images, characterized in that: The following steps are involved: Step S1: Construct a SAR image dataset and divide it into a training set and a validation set; Step S2: Based on the YOLOv8 model, the PSConv convolution is introduced to replace the bottleneck layer in the C2F module to form the C2F-PSConv module. The C2F-PSConv module is used to replace the C2F module in the YOLOv8 model. The neck network uses a multi-scale feature fusion module for feature fusion. The head network uses four detection heads, and an adaptive rotation convolution module is added before each detection head. This constructs a casualty detection model based on the improved YOLOv8, and the casualty detection model is trained using the training set and the validation set. The wounded detection model based on improved YOLOv8 includes a backbone network, a neck network, and a head network. The backbone network is sequentially connected by a first CBS module, a second CBS module, a first C2F-PSConv module, a third CBS module, a second C2F-PSConv module, a fourth CBS module, a third C2F-PSConv module, a fifth CBS module, a fourth C2F-PSConv module, and an SPPF module; the neck network performs multi-scale feature fusion, and the feature map output by the SPPF module is up-sampled and then spliced with the feature map output by the third C2F-PSConv module for the first time. The feature map after the second splicing is sent to the fifth C2F-PSConv module for processing. The feature map output by the fifth C2F-PSConv module is upsampled and then spliced with the feature map output by the second C2F-PSConv module for the second time. The feature map after the second splicing is sent to the sixth C2F-PSConv module for processing. The feature map output by the sixth C2F-PSConv module is upsampled and then spliced with the feature map output by the first C2F-PSConv module for the third time. The feature map after the third splicing is sent to the seventh C2F-PSConv module for processing. The feature map output by the seventh C2F-PSConv module is After the image is processed by the sixth CBS module, the feature maps output by the first EMAF module and the sixth C2F-PSConv module are subjected to multi-scale feature fusion to obtain a first fused feature map, which is processed by the eighth C2F-PSConv module; after the feature map output by the eighth C2F-PSConv module is processed by the seventh CBS module, the feature map output by the second EMAF module and the fifth C2F-PSConv module are subjected to multi-scale feature fusion to obtain a second fused feature map; the second fused feature map is processed by the ninth C2F-PSConv module, and the feature map output by the ninth C2F-PSConv module is subjected to the eighth The CBS module is used for processing, and then multi-scale feature fusion is performed on the feature maps output by the third EMAF module and the SPPF module to obtain a third fused feature map, which is sent to the tenth C2F-PSConv module for processing; the head network includes four detection heads, namely the first detection head, the second detection head, the third detection head, and the fourth detection head; the feature maps output by the seventh C2F-PSConv module, the eighth C2F-PSConv module, the ninth C2F-PSConv module, and the tenth C2F-PSConv module are selected as the inputs of the first detection head, the second detection head, the third detection head, and the fourth detection head, respectively; Step S3: Use the ground control station or mobile terminal to receive the SAR image transmitted by the UAV, use the trained casualty detection model to detect, obtain the location information of the casualty and issue an alarm.

2. The method for detecting casualties based on SAR images according to claim 1, wherein: The collected SAR background image and wounded target template image are preprocessed, including image size unification and grayscale normalization, and feature enhancement of the wounded target template image. The preprocessed wounded target template image is superimposed on the SAR background image using the image processing library to form a wounded SAR image.

3. The method for detecting casualties based on SAR images according to claim 2, wherein: The SAR images of casualties are transformed to simulate casualties with different postures and positions.

4. The method for detecting casualties based on SAR images according to claim 3, wherein: The generated SAR images of the wounded were carefully annotated using image annotation tools to mark the location information and category labels of the wounded targets. The annotated SAR images of the wounded were processed through data augmentation operations, including multi-angle rotation, random scaling and asymmetric cropping. The SAR images of the wounded were collected as the SAR image dataset of the wounded.

5. The method for detecting casualties based on SAR images according to claim 1, wherein: The processing process of the adaptive rotation convolution module is expressed as: ； in It is the output feature map, which represents the result after rotation convolution, and its dimension is consistent with the input feature map X. is the jth convolution kernel, Indicates the rotation transformation of the j-th convolution kernel, is the rotation angle, is the weight coefficient, which is used to perform weighted fusion on the rotated convolution kernel output.

6. The method for detecting casualties based on SAR images according to claim 1, wherein: The processing process of the C2F-PSConv module is expressed as follows: ； in, is the output feature map, express module, and They are input feature maps through The two branches split after module processing each occupy half of the number of intermediate channels. i is the serial number of the PSConv convolution operation currently being executed, and Concat means splicing along the channel dimension.

7. The method for detecting casualties based on SAR images according to claim 1, wherein: The processing flow of the multi-scale feature fusion module is as follows: the first feature map is input into the first multi-scale attention, and the second feature map is input into the second multi-scale attention; the first feature map and the second feature map are spliced and input into the first depth-wise separable convolution for processing, and the outputs are respectively input into the second depth-wise separable convolution, the third depth-wise separable convolution, the fourth depth-wise separable convolution, and the fifth depth-wise separable convolution. The outputs of the second depth-wise separable convolution, the third depth-wise separable convolution, the fourth depth-wise separable convolution, and the fifth depth-wise separable convolution are added, and then 1×1 convolution is performed channel by channel and then input into the third multi-scale attention; the output of the third multi-scale attention is multiplied with the output of the first multi-scale attention and the second multi-scale attention respectively, and the output is added with the output of the third multi-scale attention, the first feature map, and the second feature map, and then input into the fourth multi-scale attention through a 1×1 convolution operation. The output of the fourth multi-scale attention is used as the final output of the multi-scale feature fusion module.

8. An electronic device comprising a memory and a processor, wherein the memory stores computer-readable instructions, wherein: When the instructions are executed by the processor, the processor implements the method for detecting casualties based on SAR images according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting casualties based on SAR images according to any one of claims 1 to 7 is implemented.

Images