Research On Detection And Recognition Of Maneuvering Targets In SAR Images Based On Convolutional Neural Network

The interpretation of synthetic aperture radar images has always been a very important research topic and difficulty,and it has great practical application significance for military and civilian purposes.In recent years,with the rapid development of deep learning,target detection and recognition algorithms based on convolutional neural networks have also been applied to many scenes and achieved excellent results.This paper studies the detection and classification of SAR image vehicles and ship targets in different scenarios based on deep learning.The main tasks are as follows:1.In terms of vehicle target classification in SAR images,the classic VGG-16 and Res Net network models have been studied and analyzed,and the residual module has been improved,resulting in an improved Res Net network,which can better utilize and disseminate feature information at different locations in CNN.Using the expanded MSTAR data set,VGG-16,Res Net-18 and improved Res Net-18 networks are used for classification training and testing,and the network convergence and recognition accuracy are analyzed and compared.2.In terms of vehicle target recognition in SAR images,an improved Faster-RCNN detection algorithm is proposed.First,a feature fusion method is added to the CNN network to extract features to fuse shallow features and deep features;secondly,Soft-NMS is introduced to replace the traditional NMS method,remove redundant frames more reasonably,and reduce the missed detection rate.Using the self-made SAR image target detection data set,the traditional Faster-RCNN and the improved Faster-RCNN are used to train and test the target detection respectively,and the convergence effect,missed and false detection situtation,recognition situation and other aspects are analyzed.3.In the aspect of ship target detection in SAR images,an improved YOLOv3 algorithm is proposed.First,the k-means clustering method is used to optimize the detection prior frame,and then the similarity measure based on GIOU is introduced to further improve the frame regression loss function and confidence loss function,thereby optimizing the overall loss function.Using the AIR-SARShip-2.0 data set,the effectiveness of the improved YOLOv3 algorithm is verified,the recall rate,precision rate and m AP value are improved,and the detection of targets in complex environments and small targets is more accurate.In summary,this article has conducted a lot of research on the detection and recognition of SAR images,and put forward practical and feasible improvement plans.Finally,it analyzes the shortcomings and looks forward to the future work,which has important theoretical research and practical significance.