Saliency Detection By Graph-based Manifold Ranking

With the development of science and technology,digital images are increasing,and information is more and more widely used in various fields of human life.Therefore,the field of saliency detection in computer vision has become more and more important.The saliency detection of images is based on the visual attention mechanism,which aims to recognize the salient objects noticed by the human eye with a computer,and is used in various fields such as image retrieval and object detection.Among the existing saliency detection algorithms,the MR(Saliency detection via graph-based manifold ranking)algorithm applies manifold ranking to saliency detection and has achieved good experimental results.It is a classic and representative saliency detection model.However,the MR algorithm is still insufficient for the processing of some complex images.Based on this,this paper improves it.The specific work is as follows:(1)In the preprocessing stage,SLIC and Pedro are used to segment the input image to construct a multi-scale superpixel segmentation map.Then this paper use the four boundaries as the background seed points to manifold ranking,and get the rough boundary saliency map.In this way,the boundary superpixels can be constrained,and to a certain extent,the situation that a significant object is mistakenly regarded as a background when it appears at the boundary can be avoided.(2)In order to avoid bringing too much boundary information,this paper adopts a new fusion strategy to fuse rough boundary saliency maps to obtain rough background saliency maps.Then the adaptive threshold strategy is used to obtain the foreground seed points,and then obtains a rough foreground saliency map by manifold ranking..(3)In order to optimize the rough foreground saliency map,the MS(Saliency Detection with Multi-Scale Superpixels)algorithm is introduced in this paper.The saliency map obtained by this method is merged with the rough foreground saliency map.The more accurate foreground seed points are selected and then obtain the final saliency map by manifold ranking.For the manifold ranking at this stage,this article will no longer apply the average value of the superpixel block in the color space,but will use the significant value obtained previously to replace it,highlighting the foreground information more accurately and suppressing the background information.This paper compares the four representative data sets with eight classical saliency detection algorithms.The comparison results show the superiority of the proposed algorithm and fully prove the effectiveness of the proposed algorithm.