Hierarchical-biased random walk for urban remote sensing image segmentation

Random walk (RW) technique, with benefit of handling complicated boundaries, has recently drawn increasing attention in image segmentation. In this paper, RW is employed for urban remote sensing image segmentation. To deal with the complex spatial distribution with heterogeneous structures, a novel hierarchical-biased RW (HBRW) method is proposed. Firstly, edge regions extracted by fractional differential are combined with histograms to obtain plentiful features. Then, Dirichlet process mixture model is used to generate hierarchical global prior distribution and local seeds, which substitute manual scribbles. Moreover, the proposed model can adapt to different resolution segmentation tasks through adjusting the concentration parameter. Final segmentation output is obtained by biased RW. Experimental results on urban high-resolution remote sensing images demonstrate that the proposed algorithm achieves better performance than other state-of-the-art algorithms.