Keypatches: A New Type of Local Features for Image Matching and Retrieval

Abstract

The paper presents a new approach in defining (and detecting, as an additional option) local features that can be used for matching images and/or visual information retrieval. The method is based on the moment-based pattern detectors presented in previous papers. The proposed local features (preliminarily called keypatches) are obtained by approximating circular windows located at keypoints pre-detected using any typical detector with a selection of geometric patterns. At each keypoint, the optimum approximations (for all available patterns) of the window are computed using moment-based equations. For any approximation, its similarity to the actual window content is estimated. The keypatch is defined if a sufficiently accurate approximation exists. Keypoints where the window cannot be approximated with the sufficient accuracy are ignored. If no pre-detector of keypoints is available, the method itself can find the initial locations of the keypoints. The proposed approach is suitable for both grey-level and colour images (though the latter are only briefly discussed in the paper). Exemplary results explaining the method and illustrating its performances are included and discussed.