Nearest neighbor rule or k-nearest neighbor rule is a technique of nonparametric pattern recognition. Its algorithm is simple and error is smaller than twice the Bayes error if there are enough training samples. However, it requires enormous computational quantities that is proportional to the number of samples and the number of dimensions of feature vector. In this paper, a fast algorithm for k-nearest neighbor rule based on branch and bound method is proposed. Moreover, a new training algorithm for constructing a search tree that can reduce the computational quantity is proposed. Experimental results show the effectiveness of the proposed algorithms.

pattern recognition, nearest neighbor rule, k-nearest neighbor rule, branch and bound method, character recognition