Research On Remote Sensing Image Segmentation Method Based On Spectral Clustering

There are two problems in spectral clustering algorithm for remote sensing image segmentation.The first one is how to construct the weight maltrix by defining connection relation and similarity between two pixels to model the image graph model.The last one is how to utilize the graph model to automatically and accurately identify class number.For this reason,this paper propose a method to construct graph model based on the pixel neighborhood and spectral measure differences.Then based on the analysis of the graph model,the relationship between the number of classes and the data obtained from the graph model is build to identify class number accurately and then guide remote sensing segmentation.Firstly,the proposed algorithm connects neighbor pixels to construct affinity matrix;The calculation method of the scale parameter are defined to calculate scale parameters of all pixels;Relative similarity of two pixels in each pixel's scale parameter is defined,and similarity is taken as the product of two relative similarity;The connection weight can be calculated by multiplying connecting edge and similarity,and then the graph model is constructed.Secondly,the Laplacians matrix is calculated and then eigenvector matrix is taken by computing eigenvectors of Laplacians matrix,and the feature point can be constructed by making each row of eigenvector matrix corresponding to one pixel;The index of clustering degree is defined by exploiting the clustering property of pixel feature points belonging to real target class;Calculate clustering degree corresponding to different class number and discuss the law that clustering degree varies with class number;The number of segmentation classes is selected as the class number when clustering degree is the last one to jump with a greater degree.Finally,the FCM algorithm is used to divide feature point set corresponding to estimated class number to realize remote sensing image segmentation.Simulated,synthesized and real remote sensing images are used for testing the proposed algorithm.The results show that the proposed algorithm can accurately identify the number of image categories and realize the optimal segmentation of region and edge in remote sensing image.The qualitative and quantitative evaluation results show that the proposed algorithm is feasible and effective.