Saliency Propagation With High-dimensional Feature Space

With the development of information technology,more and more digital devices are applied to record and propagate information in work and life,leading to the explo-sive growth of photos and videos.How to screen effective information from massive visual media information quickly becomes a hot topic in the field of visual information processing.Visual saliency detection attempts to simulate the human visual mechanism,detects and locates regions of human interest in images and videos automatically,which makes it possible to rapidly process the visual data.After in-depth analysis of domestic and foreign saliency research,we find that the performance of existing models are not satisfactory in complex scenarios.For example,repeated small colorful structures in the background are likely to be detected as fore-ground.Colors of salient objects varying widely would lead to that only part of salient object is detected.The background being more bright-coloured than the foreground would be mistaken as foreground and it would lead to unclear boundary when the back-ground and foreground are similar.For the above complex scenes,this thesis focuses on the problem that the simple feature space is insufficient to reflect the differences of pixels and that the propagation algorithm lacks global information and the background prior is unreliable.The main research contents and innovations are as follows:1.A high-dimensional feature space suitable for complex scenes is constructed.A phenomenon that makes it difficult to distinguish the subtle differences of different superpixels results from the simple use of the color space as feature space in the existing algorithm.Consequently,this thesis constructs a high-dimensional feature space and attempts to use more features to describe superpixels,including CIELab color space,color histogram,texture space,object integrity featur-e and deep convolution network based high level semantic feature.Object integrity feature contains not only the edge of a salient object,but also the interior.We spread the pixel-level saliency value on the edge of salient object to the interior of it by means of superpixel segmentation,contour detection and object proposals,forming the superpixel-level object integrity feature.To further confirm the location of salient object,this thesis designs a deep convolutional network to extract more convincing high-level semantic feature.2.The existing propagation method is improved.Current saliency propagation models lack global information and are excessively dependent on prior knowledge.The existing literature deals with prior knowledge simply,primitively taking the image boundary as the background prior or choosing three from four boundaries or screen-ing out the high contrast pixels using color contrast,which fails when salient objects touch the boundary.Hence,the saliency propagation model is improved by boundary screening,boundary connection and introduction of sink points.Boundary screening is to remove the superpixels that may belong to the foreground in the image bound-ary.Boundary connection connects superpixels to filtered boundary,which provides global information for saliency propagation.Introduction of sink points is to set stable background reference points at the boundary of the image.This thesis performs experiments in seven public saliency datasets with ten state-of-the-art algorithms for qualitative and quantitative analysis.Our algorithm has signif-icant improvement on ROC curve,precision-recall ratio,F-measure curve and average absolute error(MAE).The high-dimensional feature space described in this thesis also does good to other algorithms.