SAR image road network extraction with scene context priming

Abstract

In this paper, a novel method of road network extraction with scene context priming for Synthetic Aperture Radar (SAR) images is proposed. In this method, the scene contextual information of each site, that this site is in roadexist scene or road-not-exist scene, is got by an SVM classifier. Then this information is used for guiding perceptual grouping of the lines and curves got by a line extractor. The output of perceptual grouping is further refined by Markov Random Field (MRF) based global optimization method to get the final road network. The performance of the method is tested on a high resolution TerraSAR-X image of urban scene.