Image Saliency Detection Based On Sparse Construction With Multi-layer Dictionary

A major challenge in computer vision is how to analyze the surrounding scene based on one single image.In order to achieve this specific goal,learning the visual salient information of the image is the most fundamental step.Therefore,in recent years,image saliency detection has gradually become an important research direction in the field of computer vision.The goal of image saliency detection is to identify the most important and informative part of a scene,which is usually the part that attracted the most attention from people who are observing.Image saliency detection later will represent the result on a nearly binarized gray scale map.Inspired by the sparse characteristic of both image signals and the semantics they reflect,this paper propose an image saliency detection algorithm.We first break up the images into superpixels through simple linear iterative clustering(SLIC).Secondly,based on centroid priority we choose some background superpixels to train a background dictionary.Under the help of the dictionary and orthogonal matching pursuit(OMP)algorithm,we can calculate the reconstruction errors,which is set as saliency values of superpixels.Finally,we combine the reconstruction error with K-Nearest Neighbor and Multi-scale Integration to get the final salient map.In the aspect of dictionary learning,we substitute the background dictionary of each scale into those of the next scale,so that the background dictionary can be multilayered,except for the first scale.After obtaining the salient map,we combine it with objectness algorithm to select some most significant superpixels to train a foreground dictionary.With this dictionary,we repeat the above steps.Once compared to the background dictionary based on the centroid prior,the foreground dictionary turns out to be more representative.Therefore,after this process,we will get a better salient map.We summarize this step as a promoting framework for enhancing results of image salient detection.Additionally,the framework is also quite effective for most of the classic algorithms and even some good algorithms proposed recently.In the experiments,the proposed model is compared with some state-of-the-art saliency detection algorithms on five most representative databases.Experimental results have shown that the method we proposed can detect the saliency object accurately and effectively.