Saliency Detection Based On Bayesian Framework

Image saliency detection is a very challenging and important research area in computer vision. The saliency detection researchers study on the image parts which can attract human’s brain and visual system attention, and obtain a probability map which indicates how much it is to be the salient object. Saliency detection can be utilized into many computer vision areas, such as image classification and retrieval, image segmentation, video tracking and so on.In this paper, we propose two different bottom-up saliency detection methods based on Bayesian framework. Background Global Framework, which is called BGF, utilizes image background and global information to detect salient object. We adopt the image border to estimate background distribution and get the standard background set. Part of prior map is calculated by color and spatial difference and combined with global cue to derive the prior map. Then we compute the observation likelihood of foreground and background by convex hull which is obtained by Harris point detection. Boundary Soft-segmentation Framework, which is called BSF, utilizes image boundary and soft-segmentation to compute the saliency. We first get a boundary weight map and calculate the prior map by color and spatial distance weighted automatically by the boundary weight map. Then according to the coarsely location of the object, we adopt the soft-segmentation method to improve the convex hull and get the revised observation likelihood. At last, we combine two proposed prior maps with the corresponding observation likelihood via Bayesian framework for final detection result.In this paper, we do the experimental validation on a publicly available data set which contains the human labeled ground truth. We do many experiments, including the validation of two proposed prior map, the validation of the background assumptions, the effectiveness of soft-segmentation and the comparison with several state-of-the-art algorithms. Besides, we apply our detection maps into image segmentation application. All the experiments achieve good performance which indicates that the two proposed methods can effectively detect salient object. Our methods can precisely highlight the object and suppress the background noise effectively. Besides, our saliency detection methods can be expanded to other image processing applications.