Repetitive Pattern Recognition Based on Representative Descriptors

In this paper we study the recognition of repetitive patterns. While SIFT (Scale-Invariant Feature Transform) is shown to be very distinctive in feature representation and matching, it is hindered from producing better performance in building recognition as repetitive structures in building images make distance ratio criterion not suitable in matching. Instead of using threshold as an alternative to distance ratio, we use a similarity based method to cluster SIFT descriptors and then match only the representative descriptors in all the clusters. This is shown to improve recognition rate and also save computation expenses in experiments with several building database. In order to further speed up the recognition, we propose to apply SIFT matching only to the top N candidates selected by vocabulary tree based matching. This helps to reduce computation load and retain recognition performance maximally. Experiments with building and non-building images verify the validity of our proposed method.