Research On Saliency Detection Based On Opponent Color And Flow Feature

Saliency detection is a conscious activity which selects and remains vital information from the massive information of the real world while ignoring the non-essential information. It is also the efficient and reliable guarantee of human visual perception. However, the introduction of human visual perception to computing vision is still significantly difficult. On the one hand, traditional parallel processing mechanism of visual feature does not match the visual perception process, so it is unable to cope with visual saliency detection problem in the complex environment. On the other hand, the role of motion information to saliency is still in the exploratory stage, and thus restricts the development of video saliency detection theory. In recent years, as the neurophysiology and computer vision have developed, some new cognition has facilitated the developments of color vision and flow feature. This research work has great significance to solve the difficulties of visual saliency detection.Driven by the task of image/video saliency detection, based on opponent inhibition mechanism of color and flow motion feature, we studied visual saliency detection from two aspects:feature coding and significant measurement. The main work is as follows:(1) We summarize the research status of saliency detection and the related computing theory from the perspective of neurophysiology and computer vision. And we emphatically analyzed the data-driven calculation model and the evaluation indicators of saliency;(2) To highlight saliency targets or regions, we have studied the spatial and color selectivity of color neurons to improve the color opponent inhibition coding, and introduced enhancement-inhibiting factor for the surface and edges of color, while suppressing the surrounding background;(3) We studied the influence of spatial position to saliency perception, and defined the criteria for saliency measure based on the difference of color feature and spatial distance of pixels. Eventually, we fused the saliency of color surface and color edge to realize the color saliency calculation. This model shows great advantages of color and direction joint encoding for saliency detection, and achieves better result than the state-of-the-art detection algorithm;(4) We researched and analyzed the difference between video saliency and image saliency, and then proposed a method for motion saliency detection based on flow features. Finally, the video saliency is exacted by fusing the time and spatial information. Experiment shows the reasonability and effectiveness of the proposed algorithm for motion saliency detection, and achieves comparable results with the classical algorithm at present.