Aiming at the inaccuracy of superpixel segmentation and the value calculation of saliency map in image saliency detection, this paper proposes an alternative form of saliency detection and updated algorithm based on the level set superpixels and Bayesian framework. A division or merger of the superpixels will achieve a level set method to form superpixels which are adaptive to the different regions of different size. The model's original map will be constructed by using color and space contrast between boundary and inner superpixels. The saliency region can be indicated with the level set superpixels and then three updated algorithms are effectively put forward via the Bayesian framework and based on K-means clustering algorithm. As a result of this update to the original mapping system, the final saliency result can be discovered in a manner that is a vast improvement of the existing method while still achieving a better level. Using our algorithm can improve the accuracy, recall rate, F value and reduce the mean absolute error value. Finally, this saliency detection algorithm for the face recognition can be effectively utilized to deal with the pictures containing people. The experimental results on three open databases show that our proposed algorithm is superior to FT, CA, XL, MR, wCO, BSCA and other saliency detection algorithms in accuracy, recall rate, F value and mean absolute error.