Image Retrieval Based On Visual Attention Mechanism

In recent years, the content-based image retrieval (CBIR) system is a hot research topic. Traditional image retrieval systems are usually searching with the index which is constructed under the image features such as color, texture and shape etc, but the global approach has some limitations in the expression of the image content, it ignores the fact that the attractive degree is not the same in different regions of the images. In the region-based retrieval methods, most methods of regional division are based on the image segmentation, but at present, the precise technology of the image segmentation is still a problem which is difficult to solve, therefore the search results is not good.Related studies show that human always focus on interesting parts of the image when they observe objects, so the search for the regions of interest is an effective way to express the search intention of user. Based on the summary and analysis of content-based image retrieval, according to the mechanism of selective attention in the psychology of the human eye in recent years, we combine Itti-Koch with Stentiford attention model, proposing a new retrieval method based on the significant areas in an image, which are interesting to users. First, the existing attention mechanism model is improved to obtain salience regions that accord more with human observations; Second, we combination both the overall and local features, considering the stability features of the salience regions, taking full advantage of interregional relations of mutual location of the overall composition of the image, then combine the two to do retrieval. This approach has overcome the shortcomings of traditional method which can not solve the problem of image rotation, translation, brightness change, and it also reflects the human eye's perception of things process. The presented method can automatically extract the salience regions, rejecting the approach chosen by hand to mark a salience area, thus the extracted regions match the target well; In addition, using salience regions as clues to do retrieval can help remove the influence coming from background, and this is closer to the user intention. Experiments show that the performance of the proposed method is better then traditional way based on global features.