Image Saliency Detection Based On The Bayesian Model

Image saliency detection is very useful in many computer vision tasks while it still remains a challenging problem. Salience describes the ability that an item outstanding from its neighborhood. The saliency detection algorithm aims to estimate the probable location of the salient object, and output a saliency map in the form of an intensity map. A larger intensity value at a pixel indicates a higher salience value, which means the pixel much probably belongs to the salient object.In this paper, we propose a new computational saliency detection model under the Bayesian framework. In contrast to most previous methods, we firstly get a coarse saliency region provided by the convex hull of saliency points. Then, based on the rough region, we proposed two methods to compute the prior probability at each pixel and combine the prior probability with the observation likelihood to achieve the final saliency map.Different from previous methods which simply use a constant to stand for the prior probability of each pixel in the image, we propose two methods to use the computational method to assign a different probability value for each pixel based on the mid level visual cues via superpixel. In the first method, we compute the Euclidean distance between the color and special difference between superpixels to obtain the prior map. In another method, we propose a new clustering method Laplacian sparse subspace clustering (LSSC) method to computer the prior map which is employed to clustering the superpixels to different groups, and then we combine the clustering results with the coarse salient region (i.e. convex hull) to compute the prior saliency map. In the Bayesian framework, we integrate the two kinds of prior saliency map with the observation likelihood probability to obtain two kinds saliency maps.We implement our experiments on a public available data set with manually labeled ground truth. The power of the proposed clustering method is carefully verified. Experimental results show the effectiveness of the proposed prior map and the strength of our saliency map compared with several previous methods. Compared with existing saliency detection method, our Bayesian saliency model can more properly detect the salient object.