Purifying Sets of Matched Features through RANSAC for Image Retrieval

An image is often represented by a set of local invariant features for many computer vision tasks such as object recognition and content-based image retrieval (CBIR), in which correct and reliable feature matching is an essential and challenging issue. Aiming at the problem of the existence of the false matches in CBIR system, we put forward a post-verification method in this paper where RANSAC algorithm is adopted to verify the primary retrieved images on global geometric consensus with the query image so that the false matches are discarded as outliers and only the correct ones are remained as inliers. Experiments show that RANSAC algorithm used in this context can improve the reliability of CBIR systems efficiently.