Saliency Detection Combined Foreground With Background Based On Manifold Ranking

By simulating human visual attention mechanism,image saliency detection can automatically filter the redundant information of the image,and detect the saliency regions,which make it widely applied in many areas,such as image segmentation,image scaling and edition,target relocation,image retrieval and so on.Reliable detection technology can greatly improve the accuracy and efficiency of image processing.On the basis of analysis and comparison of existing saliency detection algorithms,a method which called saliency detection combined foreground with background based on manifold ranking is proposed to improves the accuracy.The main work is as follows:(1)Based on human visual attention model and the biological foundation of saliency detection,we studied the principle and performance of eight salient detection algorithms.After the analysis,it is discovered that good graph structure is beneficial to improve the effect of saliency detection.And instead of using fixed values,the connection weights between graph nodes in manifold ranking algorithm was calculated using adaptive parameter to solve the problem that the image effect is not ideal because of the fixed value.(2)We proposed a method which called saliency detection combined foreground with background based on manifold ranking.Firstly,the SLIC(Simple Linear Iterative Clustering)algorithm was adopted to divide the input images which have been smoothed into superpixels.Through considering comprehensively with the color,position and background constraints,the rarity of super-pixel is defined.After compared with the threshold which calculated according to rarity,the foreground query nodes are selected.Then the foreground saliency map is obtained after these query nodes sorted by manifold ranking algorithm.Secondly,to avoid the problem that we often mistake the salient object as background query node when it is located at the image boundary,in this paper,we used threshold to eliminate boundary super-pixels which are not belonging to background.Then the query nodes are sorted according to manifold ranking again to obtain a background saliency map.Finally,since the foreground method can suppress background noise well,while background method can highlight the foreground object evenly although it is not enough to suppress noise.Thus,we get the final map after integrate foreground and background saliency map which obtained by manifold ranking.(3)Comparing with other seventeen saliency detection algorithms on public datasets such as MSRA-1000,SED2,and ECSSD-1000,and then evaluating them using indicators such as accuracy,recall,F-measure,and mean absolute error(MAE).The experimental results show that our method is more accurate than other algorithms.(4)In this paper,we discuss two application methods which combined with saliency detection algorithms,that is image segmentation and image zooming of object-of-interest.The saliency map obtained by our algorithm was used as a priori constraint to integrate the C-V(Chan-Vese)model for image segmentation experiments.Taking the saliency map as an index of importance,an image scaling experiment was performed using the SC(Seam Carving)method.