A novel segmentation method of high resolution remote sensing image based on object-oriented Markov random fields model

Abstract

A novel methodology base on object-oriented MRF is proposed in order to obtain precise segmentation of high resolution satellite image. Conventional pixel-by-pixel MRF model methods only consider spatial correlation and texture of each pixel fixed square neighborhood. The segmentation method based on pixel-by-pixel MRF model usually suffers from salt and pepper noise. Based on the analysis of problems existing in pixel-by pixel MRF model methods of highresolution remote sensed images, an object-oriented MRF-based segmentation algorithm is proposed. The proposed method is made up of two blocks: (1) Mean-Shift algorithm is employed to obtain the over-segmentation results and the primary processing units are generated based on which the object adjacent graph (OAG) can be constructed. (2) MRF model is easily defined on the OAG, in which special features of pixels are modeled in the feature field model and the neighbor system, potential cliques and energy functions of OAG are exploited in the labeling model. The proposed segmentation method is evaluated on high resolution remote sensed image data-IKONOS. The experimental results show the proposed method can improve the segmentation accuracy while simultaneously obviating "salt and pepper noise" phenomenon and reducing the computational complexity greatly.