Research On High Resolution SAR Image Target Detection Based On Deep Learning Convolutional Neural Network

SAR(Synthetic Aperture Radar)is an active ground-detecting radar system used mainly for spacecraft such as aircraft and satellites.It can observe the ground at 24 o'clock.In recent years,with the continuous advancement of SAR imaging research,the resolution of SAR images is getting higher and higher.Therefore,how to quickly and accurately detect the target types in SAR images and locate them is a current research hotspot.Important research value.The thesis studies the high-resolution SAR image target detection based on the deep learning method.The main work is as follows:The traditional SAR image detection method,due to its cumbersome steps,makes it difficult to further optimize to adapt to the increasingly high resolution SAR image detection.In recent years,the convolutional neural network in deep learning has made great strides in the field of target detection by virtue of its powerful feature extraction ability.Firstly,in the deep learning,the YOLOv3(You Only Look Once v3)algorithm has many missed detections and false detections on the target detection of high resolution SAR images,and the accuracy is low.A SAR based on YOLOv3 algorithm is proposed.Image target detection algorithm SAR-YOLO-960.Aiming at the characteristics of SAR images,this algorithm proposes a Darknet-SAR feature extraction network,and improves the loss function of YOLOv3,which makes the model more suitable for targets that are difficult to classify and have fewer categories.Experiments show that the algorithm has higher accuracy in SAR image target detection tasks and improves the phenomenon of missed detection and false detection.When the SAR-YOLO-960 algorithm detects super-resolution SAR images,there is a problem of large fluctuations in missed detection and loss function curves.Based on the algorithm,a target detection algorithm SAR-YOLT(You Only Look Twice)-960 with integrated sliding window is proposed.Firstly,for the problem of loss curve fluctuation,the migration learning strategy is adopted,and the pre-training is performed on the MSTAR(Moving and Stationary Target Acquisition and Recognition)public data set.Then the K-means algorithm is used to cluster the initial candidate frame.Finally,for the missed detection problem,it is found that the cutting is super When the high-resolution image is input,the target feature of the image cutting edge is destroyed,resulting in missed detection,so the detection model of the integrated sliding window is adopted.The effectiveness of the proposed method is proved by the detection task on the ultra-high resolution SAR image and the loss curve during training.Based on the characteristics of high-resolution SAR images,this paper improves the mainstream YOLOv3 network based on the current deep learning optical image target detection,and proposes the SAR-YOLO-960 and SAR-YOLT-960 algorithms.The experimental results show that the proposed algorithm can effectively solve the problems in high-resolution SAR image detection.