SAR Image Segmentation Method Based On Semantic Space And Curvelet Convolution Structure Learning Model

SAR is high resolution imaging radar.It's widely used in the fields like military and civilian.SAR image segmentation closely combines the understanding and interpretation of SAR images,which plays an mportant role in SAR image engineering.Usually,the better the segmentation results is,the better the quality of understanding and efficiency is.Traditional feature extraction methods have the characteristics of artificial extraction,but also through the existing deep learning model to extract features.The characteristics of artificial extraction depends on work experience and knowledge background,so the quality of artificial extraction determines the quality of image segmentation.While the deep learning algorithm can directly learn the characteristics of the image,the filter used is not suitable for extracting the structure features of SAR images.At the same time,the traditional SAR image segmentation method does not consider the image semantics.This can result in the loss of details.Therefore,inorder to deal with above problems,this thesis presents a SAR Image Segmentation Method Based on Semantic Space and Curvelet Convolution Structure Learning Model.The major innovations are as follows:(1)First,based on the region map in the hierarchical visual semantic space,the SAR image is divided into hybrid pixel subspace,homogeneous pixel subspace and structural pixel subspace.For the inhomogeneous region of the hybird pixel subspace,the curvelet atom can well detect and match the features with irregular topological structure.Thus the curvelet atom is used as the curvelet filter.Curvelet filter set is constructed for inhomogeneous region,in which the filter is used to extract the structure characteristics of the inhomogeneous region.(2)When the original image and the feature map is more closely,the better the effect of the extraction.Therefore,this thesis uses the curvelet filter to extract the characteristics and construct the curvelet convolution structure learning model.The model use the frobenius norm of the original image and the feature map to design the objective function.The objedtive function is for guiding the update of the filter.At the same time,the structure fidelity term is designed with the sketch of the SAR image.The structural fidelity term is the structure consistent condition of the original image and the feature map.The characteristic of each region can be obtained by using the curvelet convolution structure learning model to study the characteristics of the inhomogeneous region in the hybrid pixel subspace.For these features,this thesis propses the hybrid pixel subspace unsupervised segmentation method based on learning feature and maximum convergence coding.The method will code the feature set of each region to obtain the structure feature vector,and hierarchical cluster the structure feature vector,get the segmentation results of the hybrid pixel subspace.(3)For the independent target in the structural pixel subspace,according to the imaging characteristics of the independent target,this thesis proposes an independent target segmentation method in the structural pixel subspace.The method firstly filters the sketch lines in the structure area,and connected these sketch lines.Secondly,use the gather feature to extract the suspect lines.Then use the watershed segmentation algorithm to extract the super-pixel around the suspected independent target,and finally merge the super pixels belonging to the independent target.After segmenting the line target and the homogeneous pixel subspace in the structural pixel subspace,the segmentation results of the three pixel subspaces are integrated to obtain the final segmentation result of the SAR image.