Extracting Man-made Region(s) based on Attention driven Level-set Evolution

This work proposed an attention driven level set method for extracting man-made regions from aerial or satellite images. Compared with other level-set segmentation, the main re-modification of the novel approach are as following aspects. Firstly, by detecting focuses of attention and compact convex-hull polygons, salient initial contour(s) can be generated for level-set evolution adoptively and quickly, which can be much close to the real boundaries of man-made areas from natural ground. Secondly, by using a novel variational formulation, the zero level set curves can be evolved without costly re-initialization. Thirdly, a saliency map and an improved Mumford-Shah model are combined to drive the level set evolution for better segmentation. Experimental results with real images showed that the approach artfully avoids much redundant computation, and pops out the efficiency perceptually.