Image Matching Based on Binarized SIFT Descriptors
Download as PDF
Hui Huang, Yan Ma
The scale invariant feature transform (SIFT) algorithm is particularly effective in distinctive feature extraction. However, its matching is time consuming. The reason lies in that the Euclidean distance is used to measure the similarity of two SIFT descriptors in the SIFT matching. To improve the matching efficiency, in this paper, we present a novel image matching scheme (BI-SIFT) based on Binarized SIFT descriptors. First, 128-D SIFT descriptor is converted into 256-bit binarized SIFT (BSIFT) descriptor which retains the distinctive power of the original descriptor. Generally, the distance similarity meature between BSIFT descriptors by Hamming distance. However, it can introduce some extra false matches in the matching phase. Therefore, to avoid this problem, we also present a novel distance metric method for BSIFT descriptors. We evaluate our method on the UKBench data set. Experimental results show the superior performance of BI-SIFT method outperforms the state-of-the-art algorithms in image matching.
Image matching, Binarization, Hamming distance, Feature description