Multiscale detection of salient regions

An algorithm of multiscale salient region detection with precise boundaries is presented. Most of the existing salient region detection algorithms can only get position of salient regions in certain scale while no precise boundaries to distinguish them from background, so they are seldom useful for subsequent tasks such as feature extracting. Based on multiscale segmentation, the proposed algorithm is implemented as a bottom-up processing procedure. Global salient measurement is used to measure the visual stimulation of a region. Salient regions are preserved as seed nodes to the upper layer while the rest are merged together. With region merging, a salient region will be detected in a certain layer according to its scale. The position and precise boundary of a salient region can be detected simultaneously. Experiments demonstrate the validity of the algorithm.