Building Change Detection Based on Markov Random Field: Exploiting Both Pixel and Corner Features

Map databases usually suffer from obsolete scene details due to frequently occurring changes, therefore automatic change detection has become vital. Previous research has demonstrated that Markov random field (MRF) is an effective method for image classification, in which both per pixel features and contextual relations between neighbouring points are incorporated into one framework, and the problem solved by means of maximum a posteriori (MAP) criterion. However, with the advent of high resolution images, other types of spatial information (e.g. corners and edges) can also be extracted and treated as clues for detecting changes, which is usually ignored in the previous work. In this paper, we propose a framework for building change detection from high resolution images based on Markov random field that exploits all spectral, spatial and contextual features. The initial detection results are obtained based on pixel level classification and MRF. Following that, corners are extracted and building corner candidates are determined via classification. All candidates are then refined based on previous MRF results and connected by a weighted edge map. Hereafter, building changes are initialized by the area included in the connected corners (refined) and the MRF is optimized again to improve previous outputs. Final results are achieved after some suitable post processing steps. Experimental results demonstrate the capability of the proposed method for building change detection and the usefulness of spatial features.