Image Saliency Detection Based On Random Walk With Restart And Multi-layer Graph

Saliency object detection is a fundamental and important problem in computer vision and pattern recognition area.The aim of saliency detection is to identify/locate the most salient or informative region of an image that attracts human attention.The salient regions of the input image usually indicate the main objects or discriminative features contained in this image.Thus,saliency detection techniques can be widely used in many computer vision and image processing tasks such as image segmentation,image retrieval and visual object tracking.The study of the image saliency detection method is helpful to the higher level image processing.However,the current methods of image saliency detection are usually not very good in the case of the image with complex background.Aiming at this problem,we propose the algorithms of image saliency based on random walk with restart and multi-layer graph from the perspective of optimizing prior and using multilevel models.The main works of this thesis are as follows:(1)We review the previous works on the development of image saliency detection,and focus on the graph based saliency detection method and image saliency detection methods based on the multi scale.Based on this analysis,several classic image saliency detection methods are introduced,which provide the basic related works for our works.(2)Graph based image saliency detection algorithms usually dependent on the accurate of prior information.When the background of image is complex,existing image saliency detection methods usually cannot obtain the desired saliency result.To overcome this problem,we propose a new graph diffusion model based on restart random walk to optimize prior information,and thus achieve the improvement of image saliency detection results.Our method begins with computing background and foreground priors respectively for the input image.Based on these priors,we then adopt RWR method to obtain more reasonable and accurate background and foreground measurements by further considering the local structure of image.Then,we combine both background and foreground measurements together to obtain a more accurate saliency estimation.Finally,we compare the proposed algorithm with the existing some recent state-of-art methods on four common benchmark datasets to evaluate the effectiveness of the proposed algorithm.(3)Multi-scale analysis of image is the most commonly used technique to obtain more accurate saliency maps.By multi-scale resolution analysis,different important features of the image can be reached on different resolution images,such as larger objects can be analyzed on lower resolution images,and image details can be analyzed on high-resolution images.In view of the advantages of multi-scale analysis in image processing,we propose a new multi-layer graph based diffusion(MLD)model for image saliency detection by adopting random walk with restart(RWR)model.Firstly,we compute background and foreground priors/cues respectively for the input image on different scales.Then,we adopt the proposed diffusion model to obtain more reasonable and accurate background and foreground measurement.Finally,we combine both background and foreground measurements together to obtain a more accurate saliency estimation.One important aspect of the proposed multi-layer diffusion model is that it can conduct diffusion of saliency cues across different layers simultaneously and cooperatively and thus can share and communicate the saliency cues across different image scales.Experimental evaluations on four benchmark datasets demonstrate the benefits and effectiveness of the proposed method.