Co-saliency Detection Via Graph Constract And Saliency Propagation

In the past few years,people have been exploring the human visual attention mechanism and have done a lot of outstanding work in this field.One of the saliency detection is that the machine vision system has the ability to automatically detect significant areas in each individual image.However,with the rapid development of modern technology,the arrival of large-scale image data and the ubiquitous Internet era has shifted people's attention from a single image to a group of images.For researchers,a new one is followed.The field of research is the Co-saliency detection.As a new branch of visual saliency,Co-saliency detection refers to the discovery of a common significant foreground region from two or more related images and suppression of non-co-saliency regions in the image group.Co-saliency detection is widely used in many computer vision tasks,such as image Co-segmentation,video foreground extraction,image retrieval and object detection.The main work of this paper is focused on the following points:1.A Co-saliency detection method based on global compactness prior knowledge and global similarity significant propagation is proposed.Firstly,the SLIC super-pixel segmentation method is used to super-pixel segmentation of a group of images and to construct the local graph structure and global graph structure of the map.The boundary graph connectivity,global center prior and global compactness prior are utilized respectively through local graph construction and global graph construction.A priori knowledge and an initial salience map is obtained through an optimized fusion framework.Thirdly,cooperative salience propagation is performed global similarity-based saliency propagation between the GPB-OWT-UCM segmentation regions in which the super-pixel are located,to obtain an initial Co-saliency map.Finally,we merge the initial salience map with the initial Co-saliency map to get the final Co-saliency map.The proposed synergistic significance model was validated in Sub-iCoseg and iCoseg datasets,and it was always superior in accuracy,algorithm time complexity,F-Measure,and PR curves in subjective image quality comparison and objective evaluation.Advanced synergistic detection model.2.Then,a method of Co-saliency detection via intersaliency propagation and intrasaliency constraint is proposed.Specifically,the algorithm framework is mainly composed of cross-salience propagation of images and salience constraints in images.The salience cross-propagation of images is mainly based on the salient seeds of single images,and the cross-propagation mechanism across images is performed.The co-salient objects in the image group are merged and the non-cosaliency objects are suppressed to obtain a preliminary co-saliency map.Then,the a mechanism of the convex hull prior of the image is introduced to obtain the spatial distribution information of the salient objects in the image,and the common background noise is constrained.And suppress similar background areas to obtain a more accurate co-saliency map.Finally,we conducted comparative test analysis on two standard data sets Cp and iCoseg,and performed well on the evaluation criteria of MAE value,PR curve,ROC,F-Measure curve,etc.The experimental results show that it is also unsupervised.Model,this algorithm model achieves better performance than existing unsupervised collaborative significance detection methods.In summary,our proposed method for co-saliency detection based on global compactness and global similarity saliency propagation and the method of co-saliency detection between images and intra-image constraints are all synergistic detection of a group of images.Methods,these two methods have been well demonstrated on iCoseg,Sub-iCoseg dataset and Cp,iCoseg dataset experimental verification.