Winner-update algorithm for nearest neighbor search

Abstract

This paper presents an algorithm, called the winner-update algorithm, for accelerating the nearest neighbor search. By constructing a hierarchical structure for each feature point in the l/sub p/ metric space, this algorithm can save a large amount of computation at the expense of moderate preprocessing and twice the memory storage. Given a query point, the cost for computing the distances from this point to all the sample points can be reduced by using a lower bound list of the distance established from Minkowski's inequality. Our experiments have shown that the proposed algorithm can save a large amount of computation, especially when the distance between the query point and its nearest neighbor is relatively small. With slight modification, the winner-update algorithm can also speed up the search for k nearest neighbors, neighbors within a specified distance threshold, and neighbors close to the nearest neighbor.