Research On Image And Video Saliency Detection Method

With the rapid development of Internet and image acquisition equipment,the number of digital images and videos has rapidly increased.It has already become a key problem to process massive images and to extract effective information quickly and automatically.Inspired by the selective attention mechanism of human visual system,a computing model of visual saliency is established.The visual sa-liency detection model enables to filter redundant information in the scene auto-matically and to provide relevant effective information for subsequent image seg-mentation,image retrieval,scene semantic understanding and other visual infor-mation processing tasks.The research work and contribution of this dissertation mainly include:First of all,aiming at the shortcomings of the bottom-up and top-down methods,a novel image saliency detection algorithm is proposed,which combines the biological low-level visual characteristics and high-level prior knowledge(the Gestalt visual psychology theory and the F-norm minimization constraint).Then,according to the features of the traffic scene,based on the singular value decomposition(SVD),a saliency detection algorithm of traffic scene is proposed.Finally,the proposed algorithm is extended to video scene.Static and motion saliency detection and fusion algorithms are proposed respectively.In view of the multi-frame video image,the motion saliency is detected by the improved optical flow method and the motion amplitude of the light flow point is used to measure the motion saliency.Based on the difference in the visual sensitivity between mo-tion and static information and the difference in visual sensitivity between color information and gray information,the saliency detection results of the former steps are represented by the Munsell color system and the gray color model in same image respectively.Experimental results show that compared with other same type algorithms,the four algorithms proposed above have significantly improved in various per-formance evaluation capacity indices,and have better robustness.The proposed image saliency detection algorithm makes up for the shortcomings of the top-down or bottom-up model,and enhances the effective expression of the underlying fea-tures.The proposed saliency detection algorithm for traffic scene can detect the saliency targets effectively and accurately,such as traffic signs.The proposed sa-liency detection algorithm for video scene can suppress the noise influence effec-tively.The proposed fusion algorithm for static saliency targets and motion sali-ency targets improves the visual result of saliency targets detection in video scene.