Selective visual attention is a common psychological phenomenon. In cognitive psychological and neurophysiological domain, it is a very hot research area. With the development of active vision research, the important role of visual attention in computer vision area is coming to be recognized by more and more people. The main task of this paper is to setup an efficient computational model of bottom-up object-based selective attention. Numerous psychological experiments prove that "object" is the unit of visual attention but most of the computational models of visual attention in literature are space-based. Therefore, we proposed a novel computational model of selective attention completely from the perspective of"perceptual object". In this way, our model is not only more accordant with the psychological experiments but also more beneficial to high-level visual processes. First, we give out a reasonable evaluation function to evaluate the possibility of a given region to be a perceptual object. The evaluation value of this function is called "homogeneity value". The greater is the homogeneity value of the given region, the higher is the possibility of it to be a perceptual object and the more salient is it. Next, we design a sub-optimal algorithm to find the region that has the greatest homogeneity value all over the give image. This region will be the first focus of attention. This algorithm is fast and in most cases it is successful. Finally, we design a hierarchical mechanism of attention shift. This mechanism is beneficial to the general tasks of image analysis. The advantages of our model include: 1) It is based on "object"; 2) Due to the hierarchical character of our attention shift mechanism, our model is multi-scale in both of physical and feature space; 3) Our model provides multiple interfaces to combine high-level guidance. Measured by the four standards proposed by us, our model is efficient and can be easily used by general image-analysis systems.
修改评论