SAR Image Segmentation Based On Character Of Sketch And Feature Learning By Mean Field Variational Bayesian

SAR image segmentation is the foundation of SAR image interpretation,the results of segmentation directly affect the following processing such as positioning,identification and so on.High-resolution SAR images has the characters of high-dimension,heterogeneity and mixture.All these characteristics make segmentation hard for SAR image.Our team proposed concept of SAR image hierarchical visual semantic model and semantic space,so we can divide SAR image into gathering pixel subspace,structural pixel subspace and homogeneous pixel subspace.For the gathering pixel subspace including a number of extremely inhomogeneous regions,because of its rich and complex structure,the artificial features is difficult to reflect the features of its essential structure.For this reason,this paper presents a method of leaning features based on information of primal sketch and variational bayesian of mean field,then we divide it by clustering the features.What's more,for the line target in structural pixel subspace,we propose a segmentation scheme based on information of primal sketch and superpixel.The innovation points of this paper are as follows:(1)The gathering pixel subspace of the SAR image contains a number of highly inhomogeneous regions.The classification of these regions is the key point to divide the gathering pixel subspace.Due to the largely change of brightness and the scattering characteristics of the ground objects in this pixel subspace,the structural features is complex.Finding the correct features to describe the structure of each region can provide a unsupervided segmentation method.However,the artificial features can hardly reflect the nature of SAR image.So,we proposed the mean field Bayes Network with information of primal sketch,then we initialize the network with G~0 statistic model for SAR image.At the same time,we train the network under the guidance of sketch.We use the weight of network as the features of each region.At last,we use the features coding and clustering method to cluster the features,achieving the segmentation of gathering pixel subspace.(2)Structure pixel subspace include the details information of the SAR image,so finding the line target and the positioning of the boundary is the main task of it's segmentation.In this paper,we propose a segmentation scheme based on information of primal sketch and superpixel.The line target and boundaries are some details of the information,so we need to remain them from the smaller scale.Superpixels is an effective method by combining the adjacent pixels presenting similar features to synthesize the image blocks,so we select superpixels as the basis of segmentation for line target.Finally,combine the local results of three pixel subspaces with the intergrating of details as the final SAR segmentation.