SERP: SURF Enhancer for Repeated Pattern

This paper proposes an object-matching method for repetitive patterns. Mismatching problems occur when descriptor-based features like SURF or SIFT are applied to repeated image patterns due to the use of the usual distance-ratio test. To overcome this, we first classify SURF descriptors in the image using mean-shift clustering. The repetitive features are grouped into a single cluster, and each non-repetitive feature has its own cluster. We then evaluate the similarity between the converged modes (descriptors) resulting from mean-shift clustering. We thus generate a new descriptor space that has a distinct and reliable descriptor for each cluster, and we use these to find correlations between images. We also calculate the homography between two images using the descriptors to guarantee correctness of the match. Experiments with repeated patterns show that this method improves recognition rates. This paper shows the results of applying this method to building recognition; the technique can be extended to matching various repeated patterns in textiles and geometric patterns.