Astudy On Visual Saliency Based On Contrast Analysis

When we humankind need to systematically analyze complex visual scene, ourvisual system can accurately determine the salient objects according to high-contraststimuli. This selective visual attention mechanism breaks down the bottleneck ofinformation processing, which helps us to quickly and effectively filter the incomingsensory information. By studying on visual saliency, we not only understand someprinciples of the human visual cortex, but also apply them to computer vision andimage processing to solve the problems of computational complexity and resourceconsuming.Firstly, the pros and cons of state-of-the-art saliency detection methods areanalyzed. Then to overcome the limitations or shortcomings of these saliencyapproaches, this paper proposes four saliency models, oriented to pure computation,which integrate the feature combination, scale space, color space and Matting.Furthermore, this paper makes a detailed study on contrast based saliency detectionmethods that focus on feature contrast of texture, color and space distribution. In theexperimental evaluations of precision, recall, and mean absolute error, our fouralgorithms achieve superior results. The major contributions of this paper are listed asfollowed.(1) Chapter3proposes a saliency model based on HOG and color, which dependson the contrast of texture and color. Firstly, the spatial distribution variance of HOG isused to extract edges of salient objects and its inner complicated textured regions.Then, the similarity and space distribution of element color are applied to describecolor saliency of input images. Experiment shows that this saliency model coulddetect edges of salient objects and its inner textured regions, and highlight the wholesalient objects to a large extent.(2) Chapter4proposes a saliency model based on scale space, which borrows thethoughts of color contrast measure in Chapter3. To begin with, scale space theory isincorporated to measure element color contrast on each scale layer, which ensures thecompleteness of salient objects on coarse scales and inhibits the edge blurring ofsalient objects on elaboration scales. Then a simplified and linearly-weightedcenter-surround operator is developed to enhance and uniformly highlight the salient objects. Experiment results indicate that this saliency model could uniformly highlightthe entire salient objects, sharpen edges of salient objects and inhibit backgrounds,while significantly improve the precision and recall at the same time.(3) Chapter5proposes a saliency model based on double color models, which isbased on the awareness that the values of CIE Lab and RGB for saliency detection aredifferent. First of all, saliency model based on scale space is rebuilt in CIE Lab andRGB spaces, respectively. Consequently, a nonlinear combination method based onentropy distribution is proposed to extract the final saliency map. Experiments showthat this saliency model could overcome the drawback of using single color model forsaliency detection.(4) Chapter6proposes a saliency model based on Matting, which adopts theoptimal scale space analysis. Firstly, it rectifies the saliency map in current scale space.Then it guides the subsequent saliency detection in the high scale space. Experimentsdemonstrate that this saliency model could integrate the optimal scale space analysis,protect salient objects, and inhibit backgrounds.Chapter7makes a summary and expectation of the work of this paper, andevaluates the comprehensive performance of our four saliency models. Experimentresults indicate that the saliency models based on scale space, double color models,and Matting all outperform the saliency model based on HOG and color. Moreover,the saliency models based on double color models and Matting both improve theperformance of saliency model based on scale space.