Acceleration of Optical Character Recognition (OCR) algorithms is quite important for developing real-time applications on mobile devices with limited computational performances. Multilingual scene text recognition is becoming more important for mobile and wearable devices. Since Japanese and Chinese have thousands of characters, a fast and accurate character recognition algorithm is required. We developed and proposed a tree-based clustering technique combined with Linear Discriminant Analysis (LDA), and it worked fine with ETL9B dataset consisting of Japanese handwritten characters. However, a significant performance degradation with HCL2000 Chinese handwritten character dataset was found. In this paper, we formalize the candidate reduction technique for the Nearest Neighbor (NN) problems, and propose an improved method that works fine with both Japanese and Chinese character sets. Experimental results show that our method is faster and more accurate than the existing acceleration techniques such as Approximate Nearest Neighbor (ANN) search and Locality Sensitive Hashing (LSH).