Research Of Video Saliency Detection Algorithm Based On Motion Feature

With the development of science and technology,the level of daily lives continue to increase,which brings more diverse sources of information to the human eye.Videos contain the maximum amount of information among these information carriers.In recent years,how to extract useful saliency region in the complex video scene has become one of the hot topics in video processing.The pixel domain information are usually used to calculate saliency values in traditional video saliency detection technologies.However,the optical flow method is used to obtain motion information in uncompressed domain methods,resulting in poor saliency detection performance.The information which are obtained from the compressed video stream can be used to detect saliency regions more accurately.Therefore,the compressed domain approach is the trend of the video saliency detection.Currently,video saliency detection in compressed domain is in early research stage,and it needs to be improved significantly in terms of saliency detection accuracy.Through analyze and summarize of existing video saliency detection technologies,this paper has proposed a saliency detection model based on the motion feature.Compared with conventional video saliency detection methods,the main research and innovations are listed as follows.1.The human eye is more sensitive to the relative motion according to the human visual mechanism.Optical flow or differential methods are used in existing methods to find motion information.The accuracy of such information is low.The global motion estimation is used to classify motion vectors.It is useful to reduce the noise block in the temporal saliency map by removing background motion vectors.The macro-block coding information can be used to optimize the temporal saliency map.2.Video saliency detection methods are mostly based on pixel domain.There are few saliency detection methods in compressed domain and the detection quality is difficult to ensure in a variety of complex scenes.The information which are obtained from the compressed domain can ensure the accuracy of detection results.The saliency detection method in this paper is based on the compressed domain.3.Saliency regions should be intensive and have great differences with non-saliency regions according to the image saliency feature.Clustering algorithm can be used to optimize the saliency map.This paper proposes an improved clustering algorithm to tag the cluster of each saliency value.The method enhances the performance of the clustering algorithm and avoids setting number of clusters and center coordinates in advance.Based on the experimental results obtained in our study,the performance of the proposed approach is better than those of the other compared approaches.