The Design of a Nearest-Neighbor Classi er and Its Use for Japanese Character Recognition

The nearest neighbor (NN) approach is a powerful nonparametric technique for pattern classi cation tasks. Although the brute-force NN algorithm is simple and has high accuracy, its computation cost is usually very expensive, especially for applications such as Japanese character recognition in which the number of categories is large. Many methods have been proposed to improve the efciency of NN classi ers by reducing the number of prototypes and speeding up NN search. In this paper, algorithms for prototype reduction, hierarchical prototype organization and fast NN search are described. To remove redundant category prototypes and to avoid redundant comparisons, the algorithms exploit geometrical information of a given prototype set which is represented approximately by computing k-nearest/farthest neighbors of each prototype. The performance of a NN classi er using those algorithms for Japanese character recognition is reported. Given a large Japanese character training set, only a small portion of samples in the set are selected as prototypes. The fast NN search algorithm works as accurately as the straightforward algorithm while the average number of comparisons is about two third of that in the straightforward algorithm. The average number of comparisons is further reduced to less than one third of total number of prototypes if prototypes are organized hierarchically.