Hierarchical Visual Attention Model Based On Bayesian Belief Propagation

Human visual system helps people to select salient or interesting regions from complex scene and massive information while ignoring and filtering objects unconcerned under the condition of limited resources. The research of visual attention mechanism in the field of neuroscience can not only reveal the information process of human visual system, but also lay a theoretical foundation of biological plausible models, making it better applied in computer vision such as image analysis and object detection.This article studies deeply from the perspective of theory model, selective unit, bias competition and hierarchical mechanism of visual attention, and also visual perception and information transfer in visual cortex based on Bayesian theory. Inspired by the hierarchical processing mechanism of human visual system which can predict interesting regions and quickly detect targets, a hierarchical visual attention model based on Bayesian belief propagation is proposed. The model simulates the information processing function and transfer of visual cortex to extract color feature through multi-channel separation and uses it as bottom-up inference cue in ventral pathway, also extracts orientation feature by Gabor filters and regards it as top-down location bias in dorsal pathway. Then the bottom-up and top-down visual information is integrated in a Bayesian framework and belief propagation algorithm is applied to calculate the posterior of location to obtain the saliency map of visual attention.In order to analyze the effectiveness of the hierarchical model proposed in this paper, experiments are carried out firstly using the Bruce image dataset. Not only algorithm comparison is made with many other visual attention models, but also some evaluation indexes are utilized to evaluate the performance both qualitatively and quantitatively. Results show the saliency maps created by the proposed model are much closer to the eye fixation maps, and the proposed model can better predict human fixations. Meanwhile, the experimental results on the remote sensing dataset using WTA also demonstrate the proposed model can better detect targets then the GB and CA models.