Graph-based Video Saliency Detection Algorithm

The human visual system(HVS)has a great ability to quickly perceive most salient regions in static as well as dynamic scenes.Saliency detection based on this ability seeks the attractive regions and movements in video sequences.Saliency detection mostly applying various computer vision,such as object detection and segmentation due to its can reduce the search effort and the computation complexity.Most existing bottom-up saliency detection methods includes spatial saliency detection using statistical characteristics and temporal saliency detection using motion information.However,most of them neglected the connection between salient region and surrounding region.Graph-based video saliency detection method can improve the accuracy of saliency detection considering the connection between salient and surrounding region.Under such a premise,the paper carries out graph-based video saliency detection algorithms.Firstly,the background and significance of studying visual saliency detection is presented together with advanced saliency research status.This paper introduces the study on the theoretical basis of visual attention mechanism and video saliency detection algorithms.Among these algorithms,the graph-based techniques provide relatively good saliency-detection results.Secondly,we propose a novel algorithm to detect visual saliency from video sequence by combining both spatial and temporal information.To obtain spatial saliency map,video frame was represented by a fully connected graph,on which a Markov chain was defined to describe a random walk.Then assigned a weight to each edge,which was proportional to the color,intensity,compactness feature dissimilarity and the spatial proximity between the two connected nodes.The saliency of each node was determined by the visiting frequency of the random walker to the node.Motion probability distribution was obtained based on a recent psychological study of human visual speed perception.Then,probability distribution of motion orientation was computed using motion direction histogram.Temporal saliency maps are generated by combining these probability distribution.In order to take full account of the spatiotemporal information of the video,we proposed an adaptive weight fusion method to compute the final saliency map.Then,the proposed algorithm is on the public video database for simulation and makes a performance comparison with the existing those,demonstrating the performance of the improved algorithm has been further improved.Finally,we proposed a novel video saliency algorithm via graph-based manifold ranking in both spatial and temporal area.Our temporal saliency evaluation algorithm starts with motion consistency and contrast estimation.The saliency probability of each node was calculate via manifold ranking,which setting nodes with larger value of the motion consistency as the query object.Temporal salient region generated by computing saliency probability distribution.Spatial saliency map calculated using existing algorithm named graph-based via manifold ranking,which defined boundary nodes as the query object.In last stage,multiplicative fusion algorithm was utilized to get video saliency map.Experiments on the video database demonstrate that the improved algorithm is superior to the existing those.