Superpixel Clustering Based Co-saliency Detection

Co-saliency has become a research hotspot in recent years.Due to more information can be obtained from multiple images,it has a very wide range of applications in image retrieval,collaborative segmentation,video compression,target tracking and many other fields.Generally speaking,co-saliency detection involves two parts: saliency detection and common detection.This paper improves the existing saliency detection method to obtain a better saliency,then acquires the global correlation through the superpixel clustering,and further calculates the synergy by different cues.The main works are summarized as follows:(1)Aiming at the problem that existing saliency object detection model based on guided learning deals boundary of saliency object roughly,this paper proposes a content-sensitive based multi-scale saliency detection method.In order to represent the saliency boundary preferably,the image is segmented by the MSLIC method to obtain content-sensitive superpixels.Then,through calculating saliency of each superpixel with low-level feature,a weak saliency map is generated.A strong classifier based on multi-kernel enhanced learning method with 3 features(RGB,CIELab and LBP)and 4 kernel functions(Linear,Polynomial,RBF and Sigmoid)is trained,by which a strong saliency map is predicted.At the same time,superpixel pyramids are constructed to get the saliency of different scales.Lastly,we fuse the weak and strong saliency maps on different scales to generate the final saliency map.The experimental results on the MSRA and PASCAL-S datasets show that the improved saliency detection method achieves better detection results.(2)Contraposing to the deficiency that common co-saliency detection methods are inefficient at calculating co-saliency for each pixel,a superpixel clustering based co-saliency detection method is proposed.This method divides all the images in the group into superpixel blocks by using the content-sensitive superpixel segmentation method.Then,through clustering all the superpixel blocks with two color features(RGB and CIELab)and Gabor texture feature,global correlation is obtained.After that,using three kind of metrics(contrast,spatial and corresponding)to calculate co-saliency of each superpixel,a weak co-saliency map is generated by multi-scale fusion.Lastly,we fuse the saliency map obtained by the saliency detection method based on content-aware multi-scale and the weak co-saliency map to form the final saliency map.The experimental results on the iCoseg dataset and MSRC dataset show that the proposed method of this paper achieves faster computation speed and better performance.(3)Considering the problem that the change of background,which caused by movement of camera along with moving target,leads to failure of background modeling,this paper designs a target tracking method based on co-saliency detection.We use superpixel clustering based co-saliency detection method to detect the moving targets in two frame or more frames.Then,target tracking is achieved by calculating the speed of moving targets in detected frames.Target tracking is achieved by estimating the target position between the detection frames by calculating the moving speed of the target.The feasibility of the proposed method is verified by the Youtube-Objects video dataset.