Method and system for detecting artificial small target in SAR (Synthetic Aperture Radar) image

Abstract

The invention discloses a method and system for detecting an artificial small target in an SAR image, and the method comprises the steps: employing a multi-scale geospatial context attention network MGCAN to obtain a prediction result of an input image, wherein the MGCAN comprises a feature extraction backbone network which is used for extracting feature maps P1-P3, a high-efficiency pyramid convolution attention fusion module which is used for performing high-efficiency pyramid convolution attention fusion to enhance multi-scale context information so as to improve the detection precision of different-scale targets and obtain feature maps C1-C3, a parallel residual space attention module which is used for screening effective target space information so as to improve the distinguishing capability of effective features, and a detection layer which is used for predicting and outputting a prediction bounding box with scores. According to the method, different postures of the artificial small target in the SAR image can be efficiently captured, the essential features of the target are fully extracted, and the small target can be rapidly and accurately detected from the high-resolution large-scene SAR image.

Description

technical field [0001] The invention relates to an aircraft detection technology oriented to SAR images, in particular to a method and system for detecting artificial small targets in SAR images. Background technique [0002] Synthetic aperture radar (SAR) is a wide-coverage active microwave imaging radar, which has the capability of penetrating clouds and fog, and all-weather and all-weather earth observation, which makes it an indispensable part of the earth observation system. missing important detection techniques. The automatic detection of small targets (such as aircraft, vehicles, ships, tanks, etc.) based on the complex background of large-scale SAR images has important application value in both military and civilian applications. Before the war, the detection and recognition of typical military targets in the high-resolution SAR images of the opponent's position can provide an in-depth understanding of the deployment of the opponent's position, and provide importan...

AI Technical Summary

Problems solved by technology

The small target itself carries less information, and its representative features are easily submerged as the depth of the network increases, resulting in a low detection rate

[0005] 2. The attitude change of small man-made targets in SAR images

[0006] In SAR images, due to the different attitudes of artificial small targets, the received target scattering characteristics are different, so that the same target presents different characteristics in SAR images, which also brings great difficulty to target detection.

like figure 2 As shown in , the parking direction of the aircraft is different. Due to the SAR side-view imaging, the electromagnetic scattering characteristics of the aircraft obtained are also very different. Therefore, the characteristics of the aircraft with different attitudes in the SAR intensity map are quite different, which also causes the aircraft The difficulty of feature extraction increases, and the problem of missed detection is prone to occur

[0007] 3. The complex background interference of artificial small targets in SAR images is serious

[0009] 4. Interference of coherent speckle noise in SAR images

The main problems of this method are the common problems of deep learning application in SAR data, insufficient samples, weak network generalization ability, etc.

Unlike the aircraft targets in optical remote sensing images that have clearer texture and other features (the fuselage and wings are obvious), the aircraft in SAR images presents a bright spot form composed of multiple discrete scattering centers, and the available texture and outline The detailed features such as very limited, and there is serious interference in the background, so it is very difficult to detect

Images