Image salient region detection by fusing clustering and ranking

Salient region detection is an extremely challenging problem in computer vision. Most salient region detection algorithms determine the saliency of pixels in the image by directly computing the contrast between a pixel or a patch and its neighborhood within a certain range. When the image background is cluttered or the background and the salient objects in the image have some of the same image features, the detection capabilities of these traditional methods are decreased. A salient region detection framework based on re-clustering is used here. First, a clustering algorithm is used to group superpixels into a number of superpixel clusters by automatically computing the scale parameter and the number of clusters. The algorithm automatically selects possible background clusters from the superpixel clusters with the selected clusters used as queries in a ranking algorithm to obtain the final saliency map. Tests on two public salient region detection datasets show that the algorithm gives stable salient region detection results with better performance metrics than five other algorithms. © 2016, Tsinghua University Press. All right reserved.