Research On Key Techniques For Video Saliency Detection

Saliency detection,which detects important regions in visual scenes,plays an important role in computer vision.Researchers in computer vision make much effort on visual attention modeling.They use visual attention model as a key component in machine vision systems for saliency detection and computing resource distribution to improve the performance of computer vision systems to process massive digital media.However,the accuracy of the visual saliency detection algorithms is not high enough at present,which limits their applications.Therefore,constructing effective visual attention models is a key problem that needs to be solved urgently in the field of computer vision.In this paper,we investigate the research on stereoscopic visual attention modeling.The main contributions include the following aspects.We propose a new saliency detection model for stereoscopic video based on Gestalt Theory to combine the spatial and temporal saliency maps adaptively: the law of proximity,which means the closer the distance in the video frame from the salient center point,the more likely it is to be a salient pixel;the law of continuity,which means the stronger the closeness of the image pixel,the more likely it is to be a salient pixel;the law of common fate,which means that a spatial location has the similar movement way to the most concentrated saliency region in an image is more likely to be salient location.Extensive experiment results show that the proposed Gestalt theory based method can get promising performance for stereoscopic video saliency detection compared with other existing stare-of-the-art algorithms.With the wide applications of deep learning in computer vision,we use multi-module full convolutional neural networks to design a stereoscopic video saliency detection model.This method can efficiently and accurately detect salient regions of 3D video.Compared with traditional methods of 3D video saliency detection,the running time of the method reduces significantly.The proposed pixel-level 3D video saliency detection method of end-to-end training can not only detect significant regions of 3D video,but also apply to saliency detection problems of images and 2D video.The large-scale training dataset which includes semantic information training in the first two fully convolutional networks(FCNs)can enhance the performance of the saliency map calculated by the last TD-FCN.Extensive experiment has been conducted to show that the proposed Gestalt theory based method can get promising performance for stereoscopic video saliency detection compared with other existing stare-of-the-art algorithm.