Research On Image Retrieval Based On Interest Points

With the development of digital media technology, more and more digital images are produced. How to organize, store, represent, query and retrieve these images is an urgent problem. The content-based image retrieval (CBIR) provides a helpful method for this goal. In content-based image retrieval, global features related to color or texture and shape are commonly used to describe the image content. The problem with this approach is that these global features cannot capture all parts of the image having different characteristics and can't contain the spatial features. On the other hand, users could be interested in only similar parts or object of images. In this context, global image descriptors are unusable. Therefore, local computation of image information is necessary to improve retrieval result. However, interest points are just used to describe local information of the image.Image presentation and retrieval method based on interest points are studied further. A new method which is based on convex hulls of interest points is presented for overcoming the drawback of the existing efficient method called concentric cirques based on interest points which does not work well when there are some isolated points in the set of interest points. Firstly, we detected interest points by wavelet transform. Secondly, convex hulls of interest points were computed recursively, which were assigned to the buckets. As a result, color histogram feature and Gabor wavelet texture feature based on convex hulls of interest points were presented as a part of image features. With spatial distribution feature, whole color histogram feature and shape feature using wavelet and moment in addition, the system of image retrieval is built. Moreover, a new method which is improving the query point movement relevant feedback by weighted using SVM is presented. It overcomes the drawback of query point movement which does not consider the weight of relevant and non-relevant images which are returned by users.A content-based image retrieval system is built on which presented methods are tested. Experiment shows that image retrieval method and improved relevant feedback method presented by this paper provides more accurate and efficient retrieval performance; combining image retrieval method based on convex hulls of interest points and weighted query point movement using SVM relevant feedback method, the precision and recall can be improved about 20% and 10% respectively.