Remote Sensing Image Segmentation By Combining Region And Feature

In order to solve the problems that there are a large number of false segmented pixels and a single spectral measure cannot express the image information,an image segmentation method combining region tessellation and multiple image statistical features is proposed.Image regionalization is to divide the image domain into several subregions by using the tessellation technique,and each subregion is used as the processing unit.The image statistical features of the center pixel are calculated based on its neighborhood,including mean,variance,energy,entropy,contrast,correlation and homogeneity,and then the pixel feature vector is constructed by combining pixel spectral measure and above refined feathers.Then the global potential energy function is defined by using the potential energy function and neighborhood relation potential functions.Among them,the heterogeneous potential function characterizes the distribution differences between subregions,and the neighborhood relation potential function describes the interaction between neighbor subregions.And the global potential energy function is transformed into the form of probability distribution by using the unconstrained Gibbs expression and used as image segmentation model.Finally,the Metropolis-Hastings(MH)algorithm is designed to sample the probability distribution,so as to obtain the optimal image segmentation results.Experiments on different types of images are performed and the validity of the proposed algorithm is verified by both qualitative and quantitative results.The proposed method can make full use of the image information,and provide a new way for image segmentation with complex scenes.