Study On The Methods Of SAR Ship Detection And Recognition Based On Deep Learning

As a hot area of marine remote sensing application,SAR(Synthetic Aperture Radar)ship detection and recognition is the key technology of SAR image interpretation.In recent years,the development of SAR technology has brought the massive data for SAR ship monitoring and also challenges the efficiency and precision of SAR image interpretation technology.Deep learning based on big data has broken the limitation of traditional methods in the field of computer vision and made breakthrough progress.Therefore,aiming at the characteristic of SAR ship targets,exploring the SAR ship target detection and recognition based on deep learning technology can take innovation for SAR image interpretation and also bring the broader prospects for SAR application.Based on the the problems of traditional SAR ship target detection and recognition methods,this paper analyzes the deep learning methods and expolores the new methods for SAR ship target detection and recognition.The theory of deep learning is the basis of SAR ship detection and recognition.This paper introduces the neural network model,back propagation algorithm,bounding box regression and Softmax classification.It also describes three common deep networks,that is,Stacked Autoencoder(SAE),Convolution Neural Network(CNN)and Faster R-CNN.The summary of merits and difficulties of SAR ship detection and recognition based on those deep models establishes the foundation of this paper.The task of SAR ship target detection is to quickly locate the ship target from the large scene image,which is the prescreening process of target recognition.Aiming at the problem of missing detection on Faster R-CNN,this paper presents a modified Faster R-CNN based on CFAR(Constant False Alarm Rate)algorithm for SAR ship detection.Utilizing the distribution of clutter in background window of CFAR,it estimates a threshold to be compared with the mean magnitude of pixels in the target window,which is a proposal generated by Faster R-CNN.With the modified CFAR detector,ships are picked up with a slightly increase of false alarm rate.In addition,for the missing detection of the small and weak targets of Faster R-CNN,this paper presents the SAR ship target detection method based on contextual R-CNN with multi-layer fusion.On the basis of Faster R-CNN,it fuses multiple shallow layers of the network for the purpose of improving the network resolution and increasing the classification information of small targets.Then,the contextual information surrounding the proposal is added to the network in order to help the classification of the target detection network and rule out of false alrms.Based on the experiments of Sentinel-1,GF-3 and RS-3,the proposed method of SAR ship target detection based on Faster R-CNN can effectively improve the overall performance of the detector and achieve excellent generalization ability.On the basis of target detection,this paper studies the target recognition method of SAR ships.In order to reduce the network parameters and achieve efficient detection of SAR ship targets,this paper presents a method of SAR ship target recognition with feature fusion based on SAE.It extracts 25 kinds of baseline features and TPLBP(Three-Patch Local Binary Pattern)features.After that,features are connected in series and fed into SAE to obtain more distinctive fusion features.At last,fusion features are imported into Softmax classifier for target recognition.Besides,in order to extract the robustness deep feature and realize the automatic ship target recognition,this paper presents the method of ship target recognition based on CNN(Convolution Neural Network).Using smaller convolution kernels,a deep convolution neural network is designed to prevent the overfitting and increase the depth of the network.Then the fully connection layers are utilized for dimension reduction of features and further to complete the target recognition.The experiments of target recognition based on Sentinel-1 and MSTAR explore the performance of feature fusion.It also shows that the algorithm based on SAE can improve the recognition accuracy and efficiency of the network.Compared with other algorithms,SAR ship target recognition algorithm based on CNN has better performance on recognition.