Salient Object Detection Based On Diffused Model

In recent years,visual saliency detection has been widely used in image segmentation,image retrieval,target tracking,image compression,pedestrian detection and so on,which make it becomes a hot topic in the field of computer vision.And as the preprocessing part of the image,salient object detection can effectively extract useful areas in the image and suppress useless background information,which greatly reduces the complexity of the scene analysis and improves the efficiency of computer.However,the current methods are still limited in detecting salient objects of those images with complex backgrounds,multiple targets and the background is similar to the foreground.To solve these problems,this thesis proposes different saliency detection improvement algorithms based on the diffusion model.The main achievements of research works in this thesis are summarized as follows:(1)Salient object detection via multi-feature manifold ranking:The similarity matrix is reconstructed by using laplacian matrix and spectral clustering principle,and the seed nodes are obtained according to the high-level prior method,thus the method of manifold ranking is constructed.Then the final saliency map is obtained by nonlinearly integrating the corresponding saliency maps,which generated by the manifold ranking method from the different features of the image.The proposed algorithm can detect the salient areas of the image more completely by using multi-feature of the image.(2)Salient object detection via absorbing markov chain:First,the preliminary saliency map of the image is obtained by calculating the absorbed time of each node according to the background node.Then,the final saliency map of the image is obtained by calculating the absorption probability of each node according to the foreground node.The proposed algorithm can not only highlight the salient areas of the image,but also suppress the background areas.(3)Salient object detection via label propagation and background prior:The preliminary saliency map is obtained according to the label propagation method based on background node and background prior method.Then,the preliminary salient value is taken as the middle feature of the image,and the final saliency map of the image is obtained via the improved label propagation method.The proposed algorithm can highlight the salient area of the image more prominently,which effectively improve the accuracy of the saliency detection.