Co-Saliency Detection Based On Multi-Level Combination

With the development of technology,computer vision becomes the new research direction.As a hot topic in computer vision in recent years,more and more scholars come into the study of image saliency detection.With the rapid development of multimedia and the social network,there are groups of photos in our life,and these photos contain the same or similar objects of them.The saliency detection problem for these images in group can be called co-saliency detection.Co-saliency detection aims to detect the co-salient objects in one group of images.It is quite difficult for co-saliency detection,and it has more profound practical significance.Co-saliency detection can be used for many image processing domains,such as image segmentation,co-object location and image retrieval.Nowadays,most of co-saliency detection methods are based on single level.For example,some methods are based on pixel,and others are based on superpixel or object proposal.The algorithms based on different levels have their own strengths,but also have some shortcomings.Therefore,a method based on only one level usually has disadvantages.Different from existing approaches,this paper propose a co-saliency detection algorithm based on multi-level combination by synthesizing the object level and superpixel level.First of all,we make the object proposal selection in the object level.We make a coarse-to-fine selection to select the more accurate object proposal.In the fine selection,we regard the object proposal selection as an outlier detection problem,and make the fine selection by an outlier detection algorithm.At last,we get the correct templates for images.Later on,we use these templates to guide the operations in the surperpixel level.In the superpixel level,we assign label for each superpixel by superpixel classification.In this paper,we improve the traditional classification model by adding a Laplace constraint to it,which plays a good smoothing effect.Then we use the templates from the object level to choose the pseudo ground truth label,to guide the superpixel classification,and get the final saliency map by the classification result.This paper test on the i Coseg and MSRC dataset,and the results show that the proposed method achieves good performance.