Image matching based on representative local descriptors

Abstract

While straightforward image matching with keypoint based local descriptors produces a high matching accuracy, it is usually accompanied by enormous computation load. In this paper we present a representative local descriptors (RLDs) based approach to improve image matching efficiency without sacrificing matching accuracy. Firstly, local descriptors in one image are clustered with a similarity based method where descriptors are clustered into one group if they are similar enough to their mean. Then only the RLD in each group is used in matching and the number of matched RLDs is used to evaluate the similarity of two images. Experiments indicate that the RLDs approach produces better matching accuracy than both straightforward matching with original descriptors and visual words matching. © 2010 Springer-Verlag Berlin Heidelberg.