Research Of Ship Detection In Spaceborne SAR Images With Sea-land Junction

Synthetic Aperture Radar(SAR)is a full-time,all-weather,penetrating,high-resolution radar that is widely used in marine monitoring.Ship detection is extremely important in marine monitoring.Therefore,it is of great value and significance for the research of ship target detection technology in SAR image.The research content of this thesis is mainly divided into three parts: sea-land segmentation,ship detection and false target elimination.The specific work arrangements are as follows:First of all,the sea-land segmentation technology is studied.Firstly,for the general SAR image scene,the sea-land segmentation technology based on adaptive threshold is introduced.In this method,the image filtering method,the binary conversion method,the topological analysis method for binarized image,the morphology and the method for getting the land are mainly introduced.The spaceborne SAR with sea-land junction is used for experiments,and the segmentation method is verified and analyzed.Then,for the complex SAR image scene,a sea-land segmentation technology based on U-net neural network is proposed.U-net is an end-to-end neural network,which was first applied to medical image segmentation.In this technology,the feasibility that the U-net neural network is applied to the sea-land segmentation of SAR images is analyzed.And the structure of U-net is adjusted to fit the sea-land segmentation of SAR images.The U-net network structure and principle,the pre-processing method for large-scale SAR images,the training set and the training method of U-net are introduced in detail.Then the training process is analyzed through visualizing the process of feature extraction.The complex SAR image scene is used to conduct experiments,and the technology is verified and analyzed by comparing with the former method.Finally,because of the insufficiency of U-net on the land edge details,an ensemble segmentation method combining U-net and adaptive threshold is proposed,which combine the advantages of U-net and adaptive threshold.This method get better segmentation results through experimental comparison.Secondly,ship target detection in SAR image algorithm is studied.This thesis chooses the Constant False Alarm Rate(CFAR)detection algorithm as the ship target detection method,and introduces various sea clutter distribution models.The CFAR detection of various statistical model is analyzed and its parameter estimation method is introduced.Then the existing CFAR detectors are introduced,and their advantages,disadvantages and applicability are analyzed.Then,to reduce computational complexity and get pure background window,based on the lognormal distribution model and CA-CFAR detector,the iterative CA-CFAR ship target detection method by fast convolution is deeply studied.Then the complexity analysis is carried out.Compared with the traditional parameter estimation method,the complexity is greatly reduced.Finally,the SAR image data is used for experiments,and the experimental results are compared and analyzed.In the study of false target elimination algorithm,the false target is removed by the size of target initially.This step can eliminate the false target with small size.Then CNN model is designed to remove false target with similar size with real ship target.The principle of convolutional neural network is introduced,including forward propagation and back propagation.Then the CNN network structure,data set,training method and parameter setting is introduced.Finally the training set data and the real SAR image data are used for the comparison of SVM and CNN,which validates the advantages of CNN for false target elimination.Finally,the whole system of this thesis is introduced.The development environment of this thesis is introduced,including hardware environment and development languages.The use of Keras deep learning framework including some APIs is mainly indroduced.Then the system interface is shown.