SAR (Synthetic Aperture Radar) image intelligent target detection method fused with lightweight potential target activation module

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) image intelligent target detection method fused with a lightweight potential target activation module. The SAR image intelligent target detection method comprises the following steps: S1, segmenting SAR large image data: segmenting the SAR large image data acquired by a radar into slices with the size of 1024 * 1024; s2, slice target detection: performing feature extraction on each slice by using Backbone of a target detection algorithm, inputting the extracted features into a PTAM module, judging whether a target exists in the slice or not, performing accurate detection to output a target category and a bounding box if the target exists in the slice, and inputting a next picture if the target does not exist in the slice; and S3, result integration: converting the detection result of each slice into a corresponding large image coordinate, and finally obtaining a target detection result of the large image. According to the method, false alarms can be effectively suppressed, the detection performance and robustness are remarkably improved, and the system can better adapt to complex and changeable application environments.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing detection technology, and in particular to a SAR image intelligent target detection method integrating a lightweight potential target activation module. Background Art

[0002] With the rapid development of remote sensing technology, Synthetic Aperture Radar (SAR), as an important remote sensing detection method, has shown broad application prospects in target detection, military reconnaissance, terrain mapping, environmental monitoring and other fields. SAR radar generates high-resolution radar images by emitting microwave signals and receiving the echoes reflected from the ground. Compared with optical images, SAR radar is not affected by weather conditions and lighting, and can penetrate obstructions such as clouds to obtain clear images of ground targets. Therefore, SAR radar has unique advantages in complex environments such as cities, forests, oceans and mountains. However, SAR images also have their own unique characteristics and challenges. First, the target features in complex scenes are not obvious, and there are many interference factors, such as terrain undulations, vegetation coverage, building shadows, etc. These factors will seriously affect the accuracy and robustness of target detection. Secondly, the speckle noise, signal attenuation and Doppler shift phenomena unique to SAR images will further increase the difficulty of target detection.

[0003] The application of deep learning, especially convolutional neural network (CNN), makes SAR image target detection more accurate and efficient. Its advantage lies in its powerful feature extraction capability. It no longer needs to manually design features, can process complex image information, learn the implicit internal connections in the data, and has strong generalization ability. Driven by deep learning technology and large-scale databases, large-format synthetic aperture radar (SAR) image target detection has made significant progress in accuracy. However, SAR images, especially spaceborne SAR images, have a large scene format and sparse targets. For example, a scene image of GF-3 in beamforming mode is 10000×10000 pixels with a resolution of 1m. According to the conventional 1024×1024 slices, a scene image needs to be divided into about 100 slices for detection. If each slice is fully detected, it will be a great waste of computing power and time, and it is easy to generate more false alarms. Therefore, how to achieve efficient target detection in large-format images is still a very challenging task. Summary of the invention

[0004] In view of the problems existing in the prior art, the purpose of the present invention is to provide a SAR image intelligent target detection method that integrates a target detection algorithm with a lightweight potential target activation module (PTAM). The method embeds the PTAM module into the target detection algorithm, determines whether there is a target in the image through the PTAM module, and only passes the image slice containing the target to the detector for detection. This method can effectively suppress false alarms, improve detection performance and system robustness, and thus better adapt to complex and changing application environments.

[0005] To achieve the above object, the present invention provides a SAR image intelligent target detection method integrating a lightweight potential target activation module, the method comprising the following steps: S1: SAR large image data segmentation: The SAR large image data collected by the radar is segmented into slices of size 1024*1024; S2. Slice target detection: Use the Backbone of the target detection algorithm to extract features from each slice; input the extracted features into the PTAM module to determine whether there is a target in the slice; S3: Result integration: convert the detection result of each slice into the corresponding large image coordinates, and finally obtain the target detection result of the large image.

[0006] Furthermore, the method embeds a PTAM module into the YOLOv5 deep learning algorithm, determines whether there is a target in the image through the PTAM module, and only passes the image slices containing the target to the detector for detection.

[0007] Further, step S1 comprises: S1.1. Input data: First, obtain the SAR image collected by the airborne radar; S1.2. Cropping operation: Use a cropping program to crop the SAR image into multiple slices of a fixed size, each slice is a slice; each slice has the same resolution, and there is an overlapping area between the slices to avoid loss of target information.

[0008] Further, step S2 specifically includes: S2.1 Feature extraction: Use the Backbone of the object detection algorithm to extract features from each slice; S2.2 Target existence judgment: The extracted features are input into the PTAM module to determine whether there is a target in the slice; S2.3 Target detection: The extracted features are input into the detection head of the target detection algorithm to perform target detection; S2.4 Output detection results: After completing the target detection, output the target category and bounding box of the slice.

[0009] After S2.5 completes the detection of the current slice, return to step 2.1 to process the next slice until all slices are detected.

[0010] Further, in step S2.1, each cropped slice is input into a specific method for processing in turn; the Backbone network in the target detection algorithm model is used to extract features of the input slice. The Backbone network extracts multi-scale features of the image through a series of convolutional layers to generate a feature map.

[0011] Further, in step S2.2, if it is determined that the target exists in the current slice, continue to step 2.3; if it is determined that the target does not exist, return to step 2.1 and continue to process the next slice.

[0012] Further, in step S2.2, the PTAM module outputs an activation map of the potential target of the image, and uses the potential target probability value of each grid of the slice corresponding to the activation map to take the maximum value of each grid as the potential target value of the slice, which is compared with the threshold to determine whether there is a target in the slice.

[0013] Furthermore, S2.3 target detection is specifically performed as follows: for a slice determined to contain a target, a detection head of a target detection algorithm is used to perform target detection on the slice; the target detection algorithm detection head classifies and locates the target in the slice and generates a detection result.

[0014] Further, S2.4 outputs the target category and bounding box of the slice including: the output includes the coordinates of the upper left corner of the target box, the width and height of the target box, the target category, and the confidence of the detection.

[0015] Further, step S3 includes: S3.1 Result mapping: convert the detection results of each slice into the corresponding large image coordinates; S3.2 integrates the detection results of all slices and restores these results to the original large image.

[0016] Beneficial Effects In the method proposed in the present invention, a lightweight Potential Targets Activate Module (PTAM) is used to extract image features, which shares the same backbone network with the target detection algorithm. PTAM determines whether an image slice contains a target by analyzing these features, thereby guiding the YOLOv5 detector to focus only on image slices containing targets. This method can effectively suppress false alarms, significantly improve detection performance and robustness, and enable the system to better adapt to complex and changing application environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the architecture of the SAR image intelligent target detection method integrating a lightweight potential target activation module; Figure 2 It is a schematic diagram of the correspondence between the PTAM structure and the output activation map results and the input image slices; Figure 3 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0020] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] The following combination Figure 1-Figure 3 The specific embodiments of the present invention are described in detail. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0022] Targets in large-format synthetic aperture radar (SAR) images are usually sparsely distributed and concentrated in certain areas, such as ports. After these original large images are cut, a large number of image slices that do not contain targets are usually formed, which greatly consumes the computing resources of the detector, restricts detection efficiency, and increases false alarms. Therefore, how to accurately determine whether an image slice contains a target is of great research significance to improve detection efficiency and reduce false alarms.

[0023] According to the fusion lightweight potential target activation module (PTAM) of the present invention ( Figure 2 ) SAR image intelligent target detection method, wherein the target detection algorithm can be a mainstream end-to-end method, and the target detection algorithm adopts the YOLO series method. The technical concept of the present invention is to embed the PTAM module into the target detection algorithm, and use the PTAM module to determine whether there is a target in the image, and only pass the image slice containing the target to the detector for detection. The lightweight mentioned in the present invention means that the parameter scale is less than 100M. The PTAM module structure is as follows Figure 2 As shown in , it consists of 1 layer of 3×3, 2 layers of 1×1 convolutional networks, and the last layer of 1×1 convolutional network. The input features are reduced to 1 channel and the probability of potential targets on the activation map of 16×16 size is calculated. The activation map corresponds to the probability of potential targets in the 16×16 grid vertebrae of the slice, and then the gate node is used to Figure 1 middle , according to the set threshold, it is determined whether there is a target. In this method, the PTAM module and the detection head in the target detection algorithm share the output features of the backbone in the target detection algorithm. First, the PTAM module uses the features to determine whether there is a target in the image. If there is, the detection head implements image target detection according to the input features. If there is no target, the detection is skipped and the next image is directly input. The detection head preferably uses the YOLOv5 detection head. This method can effectively suppress false alarms, improve detection performance and system robustness, and thus better adapt to complex and changing application environments.

[0024] Figure 3 The flowchart of the present invention is as follows. The process of the SAR image intelligent target detection method integrating the lightweight potential target activation module of the present invention can be summarized as follows: the image is input into the target feature extraction module, the features are extracted, and the features are input into the PTAM module. The PTAM module first determines whether there is a potential target. If not, the subsequent detection is not continued and the next image is input. If there is a potential target, the features extracted by the target feature extraction module are input into the detection head to detect the target result.

[0025] The SAR image intelligent target detection method of the present invention integrating a lightweight potential target activation module comprises the following steps: Step S1: SAR large image data segmentation, cutting the SAR image into slices.

[0026] The large SAR image data collected by the radar is divided into slices of size 1024*1024.

[0027] Specifically include: S1.1. Input data: First, obtain SAR images collected by airborne radar. These images are usually large-size and high-resolution data.

[0028] S1.2. Cropping operation: Use the cropping program to crop the SAR image into multiple slices of a fixed size (1024×1024 pixels). Each slice should maintain the same resolution. The slicing technology uses the conventional slicing processing method in opencv to ensure that there is enough overlap between slices to avoid loss of target information.

[0029] Step S2: Slice target detection, specifically including: S2.1 Feature extraction: Use the Backbone of the target detection algorithm to extract features from each slice (1024*1024). Specifically: Slice input: Each slice obtained by cutting is input into Backbone for processing in turn.

[0030] Use the target detection algorithm Backbone for feature extraction: Use the Backbone part of the target detection algorithm model (such as CSPDarknet53 or other lightweight backbone networks with a parameter scale of less than 100M) to extract features from the input slices.

[0031] The Backbone network extracts multi-scale features of the image through a series of convolutional layers to generate a feature map that contains deep semantic information of the background and target in the slice for subsequent processing.

[0032] S2.2 Target existence judgment: PTAM module generates activation map; The features extracted in S2.1 are input into the PTAM module to determine whether there is a target in the slice.

[0033] If the target exists, continue to step 2.3; if not, return to step 2.1 and continue processing the next slice.

[0034] The PTAM module is used to further process the output features in Backbone and determine whether there is a network module for the target in the image. The PTAM module structure is as follows: Figure 2 As shown in , it consists of 1 layer of 3×3 and 2 layers of 1×1 convolution modules. The input features are reduced to 1 channel. At the same time, the probability of potential targets in the feature map is calculated. The feature map size is 16×16. Each pixel corresponds to the grid divided in the original image. Therefore, the probability of potential targets in each grid can be obtained, and then the probability of potential targets in each grid can be calculated through the gate node (i.e. Figure 2 ), and determines whether there is a target based on the preset threshold.

[0035] The specific judgment method is as follows: S2.2.1. Input features: Input the feature map extracted in step S2 into the PTAM module.

[0036] S2.2.2. Generate activation map: The PTAM module outputs the potential target activation map of the image through three convolutional layers. The response values ​​at different positions of the activation map represent the probability that the corresponding grid contains the target.

[0037] S2.2.3. Score calculation: The 16×16 features of the activation map output are organized into features of dimension [1, 256], representing the probability distribution of 256 regions in the image. The maximum value from the activation map is taken as the score of whether the target exists in the current slice.

[0038] S2.2.4.Threshold determination: Compare the score obtained in step 2.2.3 with the pre-set threshold, such as Figure 2 As shown in the figure, the right side of the figure is the potential target probability value of each grid of the slice corresponding to the activation map. The maximum value of each grid is taken as the potential target value of the slice, and compared with the threshold to determine whether there is a target in the slice.

[0039] S2.2.5. Target existence judgment: If the score exceeds the threshold, it is considered that there is a target in the slice and proceed to step S2.3; otherwise, return to step S2.1 to process the next slice.

[0040] S2.3 Target Detection: Input the features extracted in step 2.1 into the detection head of the target detection algorithm to perform target detection. Specifically: For a slice determined to contain an object, the detection head of the object detection algorithm is used to perform object detection on the slice. The detection head of the object detection algorithm classifies and locates the object in the slice and generates a detection result.

[0041] S2.4 Output test results: After completing the target detection, the target category (classes) and bounding box (boxes) of the slice are output. Specifically, the output includes the coordinates of the upper left corner of the target box (x, y), the width and height of the target box (w, h), the target category (class), and the confidence of the detection (confidence).

[0042] S2.5 processes the next slice: After the detection of the current slice is completed, return to step 2.1 to process the next slice until all slices are detected. If all slices have been processed, go to step S3.

[0043] Step S3: Result integration, the integrated results are restored to a large image.

[0044] The detection results of each slice are converted into the corresponding large image coordinates, and finally the target detection results of the large image are obtained. Specifically: S3.1 Result Mapping: The detection result of each slice is converted into the corresponding large image coordinates (the large image is the original image). This conversion process needs to take into account the position offset of the slice in the large image and the possible overlapping area of ​​the slices.

[0045] S3.2 integrates the detection results of all slices and restores them to the original large image. Finally, the target detection results of the large image are output, including the bounding boxes, categories, and confidences of all targets. The output of the end-to-end target detection algorithm itself contains the confidence of the target, which is the confidence of the target's bounding box.

[0046] The invention proposes a lightweight Potential Targets Activate Module (PTAM), which shares the same backbone network with the target detection algorithm to extract image features. PTAM determines whether the image slice contains the target by analyzing these features, thereby guiding the target detection algorithm detection head to focus only on the image slice containing the target. This method can effectively suppress false alarms, significantly improve detection performance and robustness, and enable the system to better adapt to complex and changing application environments.

[0047]

[0048] In the above formula, F represents the feature map output by Backbone, M represents the response value at different positions of the activation map, and PTAM calculates the probability of the existence of the target in the slice through the image features output by Backbone. Used to decide whether to perform detection on the current slice.

[0049] Table 1 shows the test using a self-built dataset, with a total of 1,377 images, of which 470 contain targets and 567 contain only backgrounds. Taking YOLOv5s as an example, Table 1 shows that the proposed YOLOV5s+PTAM detection strategy can filter out most of the background images, improve the detection speed and reduce false alarms, and the overall detection accuracy has been significantly improved (from 35.1% to 50.9%).

[0050] Table 1 Data comparison table after adding PTAM module to YOLOV5s (using Shengjing II MiniSAR to collect data)

[0051] Any process or method description in the flowchart of the present invention or described in other ways herein can be understood as a module, segment or part of a code including one or more executable instructions for implementing the steps of a specific logical function or process, which can be implemented in any computer-readable medium for use by an instruction execution system, device or equipment, and the computer-readable medium can be any medium containing storage, communication, propagation or transmission programs for use by execution systems, devices or equipment, including read-only memories, magnetic disks or optical disks, etc.

[0052] In the description of this specification, the description with reference to the terms "embodiment", "example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. In addition, those skilled in the art can combine or combine different embodiments or examples described in this specification and the features therein without causing any contradiction.

[0053] Although the above content has shown and described the embodiments of the present invention, it can be understood that the above embodiments are exemplary and cannot be understood as limitations of the present invention. Ordinary technicians in the field can change, modify, replace, modify and other update operations on the above embodiments within the scope of the present invention.

Claims

1. A SAR image intelligent target detection method integrating a lightweight potential target activation module is characterized in that: The method comprises the following steps: S1: SAR large image data segmentation: The SAR large image data collected by the radar is segmented into slices of size 1024*1024; S2. Slice object detection: Use the Backbone of the object detection network to extract features from each slice; input the extracted features into the PTAM module to determine whether there is an object in the slice; S3: Result integration: convert the detection result of each slice into the corresponding large image coordinates, and finally obtain the target detection result of the large image.

2. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 1 is characterized in that: The method embeds the PTAM module into the deep learning target detection algorithm, determines whether there is a target in the image through the PTAM module, and only passes the image slices containing the target to the detector for detection. The target detection algorithm adopts the YOLO series method.

3. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 1 is characterized in that: Step S1 includes: S1.

1. Input data: First, obtain the SAR image collected by the airborne radar; S1.

2. Cropping operation: Use a cropping program to crop the SAR image into multiple slices of a fixed size, each slice is a slice; each slice has the same resolution, and there is an overlapping area between the slices to avoid loss of target information.

4. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 1 is characterized in that: Step S2 specifically includes: S2.1 Feature extraction: Use the Backbone of the object detection algorithm to extract features from each slice; S2.2 Target existence judgment: The extracted features are input into the PTAM module to determine whether there is a target in the slice; S2.3 Target detection: The extracted features are input into the detection head of the target detection algorithm to perform target detection; S2.4 Output detection results: After completing the target detection, output the target category and bounding box of the slice; After S2.5 completes the detection of the current slice, return to step 2.1 to process the next slice until all slices are detected.

5. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 3 is characterized in that: In step S2.1, each cropped slice is input into a specific method for processing in turn; the Backbone network in the target detection algorithm is used to extract features of the input slices. The Backbone network extracts multi-scale features of the image through a series of convolutional layers to generate a feature map.

6. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 3 is characterized in that: In step S2.2, if it is determined that the target exists in the current slice, continue to step 2.3; if it is determined that it does not exist, return to step 2.1 and continue to process the next slice.

7. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 3 is characterized in that: In step S2.2, the PTAM module outputs the potential target activation map of the image. The activation map corresponds to the potential target probability value of each grid in the slice, and the maximum value of each grid is taken as the potential target value of the slice. It is compared with the threshold to determine whether there is a target in the slice.

8. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 3 is characterized in that: In step S2.3, the specific method of target detection is: for a slice determined to contain a target, the detection head of the target detection algorithm performs target detection on the slice; the detection head of the target detection algorithm classifies and locates the target in the slice and generates a detection result.

9. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 3 is characterized in that: S2.4 Output slice target category and bounding box includes: The output includes the upper left corner coordinates of the target box, the width and height of the target box, the target category, and the confidence of the detection.

10. The SAR image intelligent target detection method integrating a lightweight potential target activation module according to claim 1, characterized in that: Step S3 includes: S3.1 Result mapping: convert the detection results of each slice into the corresponding large image coordinates; S3.2 integrates the detection results of all slices and restores these results to the original large image.

Images