Research On Target Detection And Recognition In SAR Images Based On Human Visual Attention Mechanism

Synthetic Aperture Radar(SAR)is not subject to weather and time constraints,and can be set on a variety of different flight platforms such as unmanned aerial vehicles,aircraft and satellites.It has a wide range of applications in environmental detection,agricultural survey,building mapping,disaster detection and so on.Under complex background,SAR image presents the characteristics of dense targets,mutual occlusion,strong background clutter scattering and high similarity with targets,so it is easy to appear false alarm and missing alarm.To solve the above problems,this thesis introduces the human visual attention mechanism into the research of SAR image target detection and recognition technology.Based on the human visual attention mechanism,target detection and recognition in SAR images are studied.The main research contents are as follows:1.A target detection and recognition method based on feature integration in SAR images is designed.Aiming at solving the problem of the low accuracy of SAR image target detection and recognition under complex background,a target detection method based on feature integration is designed to improve the SAR image target detection performance under complex background.Then,based on the characteristics of information complementarity between different classifiers,a combination classifier is designed to improve the performance of SAR image target recognition.2.A target detection method of SAR image based on object prior information is designed.Aiming at solving the problem of high noise interference and weak semantic information in SAR image under complex background,firstly,a significance detection algorithm based on target prior information is designed,which uses contour features to obtain target prior information and fuses it with the underlying salient map to generate total salient map,which inhibits background clutter,highlights target features and enhances SAR image semantic information.Then the backbone network of YOLOv5 was improved,and the original C3 module was replaced by CNB＿4 module based on ConvNext,which improved the capability of the backbone network to extract the target features.3.A target detection and recognition method of SAR image combining frequency domain and space domain is designed.In order to solve the problem of insufficient feature information in a single dimension,the identification ability of small-size targets and lowscattering targets was improved by combining the frequency domain features of the salient graph and the improved spatial features represented by the YOLOv5 model,and combining the information of different dimensions.At the same time,a super resolution module is designed to improve the ability of the backbone network to extract highresolution information,aiming at solving the problem of the loss pattern information and texture information caused by the up-sampling operation to restore the size of the feature map.The experimental results show that the recognition rate of the combination classifier designed in this thesis reaches 93.1%,which is 5.1% higher than that of SVM classifier.Based on SRSDD data set,the mAP index reaches 83.7%,which is 8.1% higher than that of YOLOv5 model,and effectively improves the SAR image target detection and recognition performance under complex background.