Remote sensing image segmentation based on information clustering

A new algorithm based on information clustering is presented for remote sensing (RS) image segmentation, which solves the dependency of clustering centers and the sensitive-to-noise problem in the classical clustering image segmentation methods. The intensities of the homogenous region of RS image were assumed to satisfy identical and independent Gaussian distributions. Combining with the characteristics of Gaussian distribution, the joint distribution of pair-wise pixels was established. The objective function was formed based on the mutual information used as similarity measure in clustering process, and the pixel similarity in and between homogeneous regions. The iterative solution of membership between the pixel and homogeneous regions is equivalent to the maximizing solution of the objective function, so as to achieve RS image segmentation. Experiments on simulated and real images were performed to illustrate the efficiency and effectiveness of the proposed algorithm. Results show that the new method can avoid the initial clustering center selection, reduce the noise sensitivity and enhance the stability of image segmentation, which verifies the feasibility and effectiveness of the proposed algorithm. © 2017, Editorial Board of Journal of CUMT. All right reserved.