Inferring region salience from binary and gray-level images

摘要

We introduce a method that uses contour fragments to highlight regions of interest. Our method obtains as input either a binary image or the gradient map of a gray-level image. It produces a saliency map that reflects for every point in the image our belief that it belongs to a salient region. Saliency is determined by criteria such as closure, convexity, and size. In addition, gaps in the boundaries of regions diminish their saliency. Explicit scale parameter determines the size of interest. The method is implemented by a convolution of the input edge image with a linear filter that specifies the region of influence of a contour point over the image. Experiments demonstrate the utility of the method for saliency and segmentation.