COMBINE MARKOV RANDOM FIELDS AND MARKED POINT PROCESSES TO EXTRACT BUILDING FROM REMOTELY SENSED IMAGES
- 1Institute of Spatial Information Technique, Department of Earth Science, Zhejiang University No. 38, Zheda Road, Hangzhou, Zhejiang, 310027, China
- 2Department of Photogrammetry, Institute of Geodesy and Geoinformation, University of Bonn Nussallee 15, 53115 Bonn, Germany
- 3Institute for Information Processing (TNT), Leibniz University Hannover Appelstr. 9A, 30167 Hannover, Germany
Keywords: Building, extraction, segmentation, combination
Abstract. Automatic building extraction from remotely sensed images is a research topic much more significant than ever. One of the key issues is object and image representation. Markov random fields usually referring to the pixel level can not represent high-level knowledge well. On the contrary, marked point processes can not represent low-level information well even though they are a powerful model at object level. We propose to combine Markov random fields and marked point processes to represent both low-level information and high-level knowledge, and present a combined framework of modelling and estimation for building extraction from single remotely sensed image. At high level, rectangles are used to represent buildings, and a marked point process is constructed to represent the buildings on ground scene. Interactions between buildings are introduced into the the model to represent their relationships. At the low level, a MRF is used to represent the statistics of the image appearance. Histograms of colours are adopted to represent the building's appearance. The high-level model and the low-level model are combined by establishing correspondences between marked points and nodes of the MRF. We adopt reversible jump Markov Chain Monte Carlo (RJMCMC) techniques to explore the configuration space at the high level, and adopt a Graph Cut algorithm to optimize configuration at the low level. We propose a top-down schema to use results from high level to guide the optimization at low level, and propose a bottom-up schema to use results from low level to drive the sampling at high level. Experimental results demonstrate that better results can be achieved by adopting such hybrid representation.