Beyond Keypoints: Novel Techniques for Content-Based Image Matching and Retrieval

Keypoints are a well established tool for image matching and retrieval problems. The paper reports development of novel techniques that (by exploiting advantages of keypoints and trying to correct their certain inadequacies) provide higher accuracy and reliability of content-based image matching. The area of ultimately intended applications is near-duplicate image fragment retrieval, a difficult problem of detecting visually similar fragments embedded into images of unknown and unpredictable contents. Two supplementary approaches are proposed: (1) image warping for non-linearly distorted images to obtain the best match between related fragments and (2) detection of maximum regions that are related by affine transformations. Other relevant results are also briefly mentioned. The reported work is a part of an ongoing project so that further improvements and modifications of the proposed methods can be expected in the near future.