Chinese Journal of Computers   Full Text
Title	Image Retrieval by Convex Hulls of Interest Points and SVM-Based Weighted Feedback
Authors	SU Xiao-Hong DING Jin MA Pei-Jun
Address	(School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001)
Year	2009
Issue	No.11(2221—2228)
Abstract &   Background	Abstract To solve the problem of image retrieval method based on annular color histogram, a new image characteristics extraction method based on convex hulls of interest points is presented. Firstly, the interest points on an image are detected by wavelet transform. Then, convex hulls of interest points are calculated recursively and these points are assigned to some buckets by a specific algorithm to form a color histogram for every bucket. The similarity of two images is calculated by the similarity between histograms of two buckets. Combined with spatial distribution feature and Gabor texture feature based on convex hulls of interest points, the system of image retrieval is built. Experiments on image database show that this method works well when isolated points exist in the interest points set and so provide more accurate retrieval performance comparing with other retrieval method based on interest points. Further more, a novel relevance feedback method is presented. It improves the query point movement relevance feedback method by setting weights based on support vector machine cluster results. The experiments show that by using this method combined with the image retrieval method based on convex hulls of interest points, the precision and recall can be improved about 20% and 10% respectively. Keywords image retrieval; wavelet transform; interest points; convex hull; support vector machine; relevance feedback Background In content-based image retrieval, users may be interested in some part of the image but not the global features. To solve this problem, the method for image retrieval based on interest points was proposed. A critical step of this method is how to match feature based on interest points. Till now, many solutions are based on point matching which is from computer vision. But the way point matching needs too much time to get results, and it does not consider the character of image retrieval; in the relevant feedback of content-based image retrieval, classical RF method called query point movement has the drawback obviously. Unluckily, many years passed and it was not paid attention to. This paper attaches importance to the method for how to extract and match feature based on interest points. A new method for extracting and matching features based on convex hull of interest points is proposed which can work well when isolated points exist in the interest points set. Compared with other methods for image retrieval based on interest points, the new method needs less time and provides more accurate retrieval performance; in addition, RF method called weighted query point movement using SVM is proposed which has better retrieval accuracy than classical query point movement method.