Summary: Automatic image tagging seeks to assign relevant words to images that describe the actual content found in the images without intermediate manual annotation. One common problem shared by most previous learning approaches for automatic image tagging is that the segmented regions in the image were considered as equally important and were processed equally. The goal of this paper is to develop a novel annotation approach based on regions of interest to take into account the users’ real experience and fix a visual weight for each region according to the degree of interest. To do this, we firstly segmented the image into several regions. And then it calculated the degree of interest for each region according to the experiments of human visual attention and cognitive psychology. Each region will be assigned a visual weight at the third step. We can obtain the prior probability of the region given the concept. At the stage of the automatic annotation, we can calculate posterior probability with the Bayesian Theorem to get the most likely concept to tag the unseen image. The proposed methodology is examined in a well-known benchmark image collection and the results demonstrated its competitiveness.