A high-speed pattern matching algorithm is required for developing real-time character recognition applications especially for mobile devices with limited computational performances. Multilingual scene text recognition has recently become more important for mobile and wearable devices. Since Japanese and Chinese have thousands of characters, not only the accuracy but also the speed of classifiers are crucial. We formalized the candidate reduction technique for the Nearest Neighbor (NN) search with high-dimensional feature vectors, and proposed a tree-based clustering method to realize a fast handwritten character recognition. It works fine with ETL9B dataset consisting of Japanese handwritten characters and HCL2000 Chinese handwritten character dataset. In this paper, we propose an improved candidate reduction method based on our former one. The experimental results show that our method is 60.48% faster and more accurate than the former method.