A fast and effective dichotomy-based hash (DBH) search algorithm for image matching

The introduction of the full paper reviews past research^[1-12] and then proposes a new DBH search algorithm. Section 1 explains the DBH search algorithm with the help of Fig. 1; its core consists of: (1) we calculate the distribution of the high-dimensional data set to each dimension; (2) we randomly choose the specified dimensions as the key dimensions; (3) we choose different hash functions to hash the high-dimensional features so that the similarity features of the images to be matched can be hashed into the same bucket, using high probability; (4) we present the procedural steps of the DBH search algorithm that hashes and queries for several times. Section 2 did experiments on image matching with the standard data set from Ref. 13 and compared the image matching performance of our DBH search algorithm with that of BBF (best bin first) search algorithm and LSH (local sensitive hash) search algorithm. The experimental results, presented in Figs. 2 and 4 and Tables 3 and 4, show preliminarily that our DBH search algorithm performs better in both accuracy and speed, and has higher recall vs (1-precision) ratios in different transformations of image pairs with rotation, scale, noise and weak affine change than the famous BBF search algorithm and the classical LSH search algorithm.