Structural Information Analysis Based Saliency Detection

With the rapid development of technology,multimedia or smart terminal devices such as mobile phones,cameras,and tablets have become popular.People's lives are filled with a large amount of digital information,which makes the scale of image data to be stored and analyzed greatly increase,and the processing of visual information becomes difficult and challenging.The saliency detection of images makes use of the human attention selective mechanism to enable the computer to extract the prominent objects in the image.As a preprocessing process for other applications,it has deep research significance and application prospect in the field of computer vision,image processing and artificial intelligence.In order to improve the accuracy and stability of the existing saliency detection model,this thesis constructs two saliency detection models from the perspectives of image contrast and boundary structure connectivity based on the topology structural information.First,some existing methods usually use distance metrics to calculate the image contrast,and it is easy to ignore the structural information of an image or describe the topology structure inaccurately,which results in the unsatisfactory detection effect.To tackle this problem,this thesis proposed a manifold learning based saliency detection method.This method adopts the diffusion maps of manifold learning,which diffuses through the iterative process from the local neighborhood to the global image.During this process,the feature difference and the topology structure are combined together to obtain the global image information.Besides,this method proposes an average diffusion distance to measure the image contrast and shows the improved method performance during experiment in terms of stability and robustness against image noises.In order to solve the problem of unclear edges caused by illumination changes or noise in complex background images,this thesis proposes a structural connectivity measurement method based on anisotropic diffusion,and combines structural connectivity with appearance contrast to form the saliency detection model.Anisotropic diffusion has a restrictive and directional perception of image edges,which can effectively retaining important edge structures and removing redundant noise in an image,and can be used as an effective method for measuring the boundary connectivity of image elements.The experimental results show that the model combines structural information and feature comparison,which improves the accuracy and stability of salient object detection and effectively solves the edge blurring problem in the image.