Multi-information Fused Image Saliency Detection

Based on the visual attention model of human visual system,an image processing technology called saliency detection is proposed.It can find the interference information to act as an information filter and is a useful way to solving some problems of computer vision,such as adaptive compression of images and content based image retrieval.Through saliency detection,every pixel can get a saliency value which could represent the possibility that it is a part of the region of an image which contains the main information.And then to get the goal of removing the interference,some methods can be used to process the saliency value,such as the binarization.In recent years,saliency detection has been developing rapidly and has been a research hotspot.Now,various kinds of saliency detection method are proposed and on the simple image,saliency detection has achieved good effect.In this paper,the background and development of saliency detection are firstly reviewed.And then we study a large number of existing saliency detection methods and analyze their strengths and weaknesses to summarize its developing trends and the basic characteristics of the salient region.By the intensive study on the saliency detection,a new method of fusing two-layer information is proposed.The main characteristics of this method are as follows.First,on the basis of comprehensive analysis of uniform region representation and non-uniform representation of the image,a two-layer region segmentation method including two kinds of region representation is constructed,which is characterized by its simple calculation and reducing the number of region effectively.Second,according to the diverse properties on the constructed two-layer representations,multiple perspectives saliency computing strategy from spatial-structure and color respectively is made,which can make the results more accurate.In addition,“soft” background prior is added to the saliency computing.This prior does not set the region containing the boundary pixels as background region,but uses the percentage of the boundary pixels in each region to adjust the saliency values.At last,we compared the new method with other excellent algorithm through the experiments.And the results prove the method outperforms the state-of-the-arts methods in precision,recall,F-measure and Mean Absolute Error(MAE).