SAR Image Segmentation Based On Semantic Space And Stochastic Gradient Variational Bayesian Feature Learning

SAR image segmentation is the premise and foundation of SAR image understanding,which has important influence on the subsequent analysis and interpretation.SAR image segmentation is particularly difficult because of the rich direction and the changeable scale.Based on hierarchical visual semantic model of SAR image proposed by our team,we can divide SAR image into three different pixel subspaces.In hybrid pixel subspace,the feature extraction of extremely uneven regionals is the focal point and difficult point of segmentation.The traditional method of extracting features from SAR image segmentation is complex and difficult to control.Bayesian machine learning has unique advantages in the uncertain representation and data modeling.According to the prior model of SAR image,we can get effective features of SAR image by inference and learning the Bayesian Network.In view of the above situations,we propose a new method for SAR image segmentation,based on semantic space and random gradient variational bayesian feature learning,The main innovations are as follows:(1)For the hybrid pixel subspace of SAR image,we propose a new model for feature learning of the extremely uneven regionals based on sketch constraint and random gradient variational bayesian.Firstly,according to the hierarchical visual semantic model of SAR image proposed by our team,we can obtain region map by the sketch map,and then divide SAR image into hybrid pixel subspace,structural pixel subspace and homogeneous pixel subspace by the gathering region and homogeneous region and no sketch line region in sketch map.For each extremely uneven regional,construct a sketch constraint and stochastic gradient variational bayesian network,and establish a G~0 distribution statistical model.After estimating the probability density function of G~0 distribution,we can obtain a set of random matrices to initialize the weights of network.The information of the sketch map is used to train the network,and the weight of the trained network is taken as the regional feature.Then,the feature vectors of each region are obtained by feature coding of maximum pooling.Finally,we carry out the clustering of extremely inhomogeneous regions by using the hierarchical clustering algorithm,and then obtain the segmentation result of hybrid pixel subspace.(2)For the homogeneous SAR image pixel sub space,in view of it's Simple structure,Gentle gray change,and Coherent speckle noise,we propose a new image segmentation method of homogeneous pixel subspace based on sketch characteristics and super pixel segmentation.The method is based on the position and orientation information of SAR image sketch lines,and combined with watershed super pixel segmentation method.First,select referenced sketch line from the sketch line of boundary and line target in structural pixel subspace,and select seminal super pixels on both sides of the reference sketch line.Design local region segmentation algorithm according to the direction and position information of referenced sketch lines,and then partition the whole region According to the algorithm,obtain Multiple homogeneous regions of homogeneous pixel subspace.Then extract texture features of each homogeneous region by gray co-occurrence matrix,and cluster features by hierarchical clustering method,obtain segmentation result of homogeneous pixel subspace.Finally,the final result of SAR image is obtained by fusing the results of hybrid pixel subspace,structural pixel subspace and homogeneous pixel subspace.