Similarity Analysis And Image Recommendation Based On Visual Regions Of Interest

The man-machine interactive system includes two methods:explicit feedback, e.g. mouse, keyboard and other types, and implicit feedback, e.g. the information of eyes’ movement. At present, there are more using of the application of explicit feedback, while the implicit feedback has not been widely used because of technical problems. Having based on the implicit feedback has very high application value, this paper focuses on the research of the similarity-analysis of the image and image recommendation aspects of eye-movement information. This paper use people as the starting point, which basing on the principle of "people-oriented", researching the images’similarity-analysis of image and retrieval.Due to a single or small amount of observations, not enough to explain the interest tendency of the observer, this paper adopted a method to cluster attention regions based on low-level feature vector, and use the clustering results to be the region feature which can be used to reflect the regional characteristics of the interest tendency, that is, the visual region of interest. In the subsequent image retrieval, proposing squared random point selection method for testing image region segmentation, conducting similarity analysis based on image region, and storing the test image which is meet to certain similarity to the similar concentration which is correspond to the visual region of interest. In the clustering process, the average value of the degree of the image which is classified as a class is used to be the degree of interest in this area of interest, which is used to measure the degree of interest of the viewer on such features. In the foundation of the level of interest, conduct the similarity of image for the recommendation in a certain order; in the process of recommendation, this paper propose a more humanization algorithm of recommendation. And on the implementation details of the recommendation algorithm, this article propose a method of self-adjustment of similarity threshold, which can better reconcile the contradictory relationship between the degree and the recommended amount.Through the design and analysis of the whole process, this paper put forward a better solution in a series of aspects:fixation area extraction, the interest tendency under clustering, the similarity analysis followed regional test image, and novel image recommendation. Finally, the recommendation of image get a good effect.