Saliency detection is one of the natural analyses of human visual system. Points or objects which attract human visual attention is named salient. These objects have some features such as movement, intensity contrast and color contrast. Regardless of real function of man’s brain on implementation of visual attention, finding computational models which can detect salient points in an image have extensively found its application in different fields of image processing and machine vision. Several computional models have been proposed for saliency detection. A group of these models use statistical methods to detect the location of salient points. Some of these methods use Baysian framework to define and compute saliency. Typical Baysian frameworks for saliency detection use many simplifying assumptions. These assumptions lead to saliency detection methods which can not covers all the different situations. In this dissertation a novel Baysian framework for saliency detection is proposed, which covers the defecincies in typical Baysian frameworks. In this framework, saliency is computed in each point from three parts, Location-based saliency, Feature-based saliency and Center-soround saliency. In the mentioned framework, location-based saliency is the part of saliency which shows the relationship between location in image and saliency and it shown that this part of saliency has greater values at the center of the image in comparsion to the boundries of the image. If we choose color components at each point as feature vector, the feature-based saliency shows how probable for each color vector to be salient. In this desertion, it shown that the colors with red and orange hue and also the high intensity colors are more probable to be salient. Center-surround saliency models the statistical relationship between saliency in a point and the feature vectors of that point and the surrounding points. As the proposed method for center-surround saliency leads to distrbtion estimation in high dimensional space with numerous training datas, we use Dirchlet process mixture models to estimate the distributions. The proposed method is tested using five different databases. Using the CAT2000 database, we compute AUC-Borji metric for the proposed method as 0.80, where this metric for the “one human” saliency detection is computed as 0.67 and for the “infinite human” group is computed as 0.84. It shown that, the proposed method outperforms all the state-of-the-art saliency detection methods, where using AUC-Borji metric, the proposed method ranks first in three databases and ranks seconds in two other ones.