Research On Visual Saliency Detection Based On Spatio-temporal Features

With the development of computer information technology,various pictures and videos have become an important means of information expression as a carrier of communication.Neurobiologists have found that the human visual system can quickly and adaptively select the interested objects in the scene and process them centrally in the face of these massive information.This is the selective attention mechanism of the human visual system,and the human target of interest is the salient area.Thus,the ability of human visual system to quickly screen interested objects is applied to the computer,so that the computer can self-adaptively understand and process video.In recent years,spatial-temporal saliency detection has developed rapidly,but it is still in its infancy.The existing spatial-temporal saliency detection methods get more redundant background information,less detailed information,and poor adaptability to different motion patterns.In order to solve these problems,this thesis studies the existing spatial-temporal saliency detection algorithms in depth.The specific work is as follows:Firstly,a video saliency detection method based on low-level saliency fusion and geodesic is proposed.Spatial-temporal edge saliency map and smoothed spatial color saliency map are extracted and fused to obtain low-level spatio-temporal saliency map.High-precision spatio-temporal saliency map is obtained by the geodesic distance from the subsequent frame to the background area.SegTrack V1,SegTrack V2 and ViSal are selected as experimental datasets and compared with other 11 methods in recent years.Subjective results show that the proposed algorithm can eliminate the background redundancy information,and the internal edge information and details information are clearer.The objective evaluation results are generally better than other methods.Secondly,aiming at the narrow application of the proposed method,a saliency detection method based on minimum directional contrast and spatio-temporal background priori is proposed.The method calculates the minimum directional contrast saliency map and spatio-temporal background priori saliency map respectively,and then fuses the two saliency maps to obtain the final spatio-temporal saliency map.This method is suitable for scenes with more complex background and multiple motion modes.Experiments on three open video datasets,FBMS,MCL and DAVIS,are carried out and compared with the saliency models: MCT,MDC,SAG,GF and STBP.Subjective and objective results show that the proposed algorithm can extract the saliency region of the target well and has advantages in applicability and accuracy.