A Study On Image Co-saliency Detection

Human visual attention mechanism shows that different objects in visual scene have different attraction to human beings.In computer vision,saliency means that the degree of different image regions attracting human attention.Image saliency detection aims to make the computer has the capability of discovering salient regions in images accurately.Image saliency detection has attracted a lot of attention and a wealth of saliency detection approaches have been proposed in recent years.However,existing approaches mainly focus on detecting saliency regions in a single image.In the real word,the interesting targets often have similar attributes or the same category in a image group.Therefore,people need to combine the context information in the image group to determine the target of interest.The essential problem of co-saliency detection is how to detect the common objects from the group of images.Although co-saliency detection has attracted a large amount of research interests,this problem tends to be much more difficult as in practice the relevant images in a group may have a wide range of variations caused by varying illumination conditions,diverse backgrounds,changing viewpoints,and the target itself may also be affected by non-rigid deformation,occlusion,and camouflage.Therefore,existing methods have not achieved satisfactory results,and it is urgent to propose more effective co-saliency detection methods.The dissertation studies the methods of co-saliency detection systematically.The main work and contributions are as follows:1.A hypergraph optimization and salient seed propagation framework is proposed for unsupervised co-saliency detection.The proposed method iteratively optimizes hy-pergraph structure,feature learning and salient seed propagation.In particular,The high-order relationship between superpixels is adaptively modeled by optimizing the hypergraph structure,which provides a basis for the propagation of salient seed label.The experimental results show that the proposed method can effectively deal with the existing noisy image in the image group and the large appearance change of common objects,and improve the performance of co-saliency detection.2.A novel feature-adaptive semi-supervised(FASS)framework is proposed for co-saliency detection,which seamlessly integrates multiview feature learning,graph structure optimization and cosaliency prediction in a unified solution.The multi-view feature learning lead to an effective representation robust to feature noise and redun-dancy as well as adaptive to the task at hand.It predicts co-saliency map by optimizing co-saliency label prorogation over a graph of both labeled and unlabeled image regions.The graph structure is optimized jointly with feature learning and co-saliency predic-tion to precisely characterize underlying correlation among regions.The experimental results show that the proposed method can obtain effective and reasonable features,and effectively improves the accuracy of co-saliency detection on multiple public datasets.3.A multi-task deep learning network is proposed for co-saliency detection,which consists of a co-semantic learning branch and a co-saliency detection branch.The in-teraction relationship between images are effectively learned by the common learning of images.The multi-scale fusion strategy is used to improve the detail expression a-bility of the salient map.In addition,inter-image and intra-image attention mechanisms is proposed to enhance the accuracy of co-semantic representation.The experimental results show that the proposed method can obtain the reliable co-salient map.