A Two-Step Descriptor-Based Keypoint Filtering Algorithm for Robust Image Matching

Finding robust and correct keypoints in images remains a challenge, especially when repetitive patterns are present. In this article, we propose a universal two-step filtering method to solve the mismatch problem in repetitive patterns. Having applied a mean-shift clustering algorithm to remove obvious mismatches, the proposed confusion reduction (CR) method uses a novel confusion index (CI) in a gridding schema to identify and filter out the remaining confusing keypoints. In both steps, the descriptors’ statistical properties are evaluated using kernel density estimation. Various synthetic and real stereo pairs, along with multiview image blocks, were used to assess the performance of the presented algorithm. The results were also compared with those obtained by several state-of-the-art mismatch removal methods. The experiments showed that, on average, the proposed strategy improves the accuracy of matching by 10% and the accuracy of photogrammetric blocks by 20%–30%.