Saliency Detection Based On Convex Hull Clustering And DS Evidence Theory

Image saliency detection is an important and challenging research branch in thecomputer vision community.Salience measures the degrees of importance of different parts in an image.The saliency detection result is typically shown as an intensity image,where a larger intensity value indicates that the pixel is more salient.By effectively separating the image into regions of different importance,saliency detection can ignore non-essential regions and apply limited computing resources to interested regions.It functions as an important pre-processing step in many high-level computer vision systems,and attracts consistent attentions from both the academic and industrial societies.This thesis proposes a novel saliency detection algorithm based on convex hull clustering and DS evidence theory.Unlike most previous saliency detection methods,which usually assume the image boundary as the background or the center as the foreground region,the proposed algorithm makes full use of the local contrast and the boundary connectivity of the image.Firstly,the color enhancement Harris corner detection are used to construct the convex hull surrounding the foreground region.Secondly,performing the clustering within the convex hull to remove the background part of the convex hull,so that we can obtain a compact foreground region.Thirdly,a superpixel based two-hop map model is constructed and the aforementioned compact foreground region is set to the absorption nodes of the map model.The initial saliency value of each superpixel is computed based on the random walk model.Finally,taking advantage of three optimization techniques,i.e.,"diffusion of saliency values within clusters","fusion of the boundary connectivity characteristics of each superpixel" and "suppression of background superpixels' salient values" to obtain the optimized saliency map.Experimental results show the effectiveness of these three optimizations.In addition,a fusion mechanism based on the DS evidence theory is also proposed to exploit the results of multiple saliency detection methods at the pixel level and achieve effective results.We compare the effect of each step of our algorithm on three public standard salient object detection datasets,and ablation studies demonstrate the effectiveness and necessity of each step of the proposed algorithm.At the same time,extensive evaluations on the threepublic salient object detection benchmarks show that the proposed method perform favorably against eleven state-of-the-art methods in several assessment metrics.