Research On Image Segmentation Method Based On Spectral Clustering

Image segmentation is a classical and crucial problem in the field of computer vision and image understanding. Because of extracting only local information with disconnected boundary of the segmented region, and lack of ability to integrate global and local knowledge about the segmented objects, classical non-model based image segmentation techniques cannot satisfy the requirements of complex image vision applications. The algorithm which is based on spectral clustering flourishes in recent years. Derived from graph cut theory, these algorithms need no prior information to be given except the similarity between each two data, which makes it extremely simple to be implemented and promising in image segmentation.Primarily, the mathematical basis of graph cut is expatiated. Then the relationship between graph cut and spectral clustering is described through the Laplacian matrix point of view, the application to the image segmentation is analyzed as well. Then, the implementation of two classical segmentation rules, Ncut and RatioCut, in the spectral clustering algorithm is deduced and the main steps of spectral clustering are confirmed. By solving the decomposition problem of Laplacian matrix, it is a feasible and powerful tool in image segmentation.According to the weakness of huge calculating complexity in spectral clustering, combined with edge operator and neighborhood information, based on the super-pixel method, a novel scalable spectral clustering is proposed in this paper. In the proposed algorithm, image is firstly divided into several blocks, and then the first-level spectral clustering is implemented combined with information within the neighborhood for each pixel. The second-level spectral clustering is carried on based on the 16-bin histogram of each clustering areas. After that, the morphological operator combined with edge detection is utilized to obtain the final segmentation results. This algorithm maintains high speed and fine local details in segmenting images and need not estimate number of class.In order to segment polarimetric SAR image using scalable spectral clustering, the Wishart distribution and revised Wishart measurement is introduced in this paper. Aiming at characteristics of polarimetric SAR image, the mean Wishart coherence matrices are utilized in second-level clustering to calculate the pairwise revised Wishart similarity of areas. Experimental results show that the proposed method obtain excellent segmentation performance. The randomly preset of the coefficients does little affects on the final segmentation in proposed algorithm. Besides, the contours in the polarimetric SAR image have been reserved and incorrectly partitions within the clusters have been reduced dramatically.