Comparison of Distance Computation Methods in Trained HMM Clustering for Huge-Scale Online Handwritten Chinese Character Recognition

In this paper, we propose a clustering method to solve computing resource problem that may arise in very large scale on-line Chinese character recognition using HMM and structure code. The basic concept of our method is to collect HMMs that have same number of parameters, then to cluster those HMMs. In our system, the number of classes is 98,639, which makes it almost impossible to load all the models in main memory. We load only cluster centers in main memory while the individual HMM is loaded only when it is needed. We got 0.92 sec/char recognition speed and 96.03% 30-candidate recognition accuracy for Kanji, using less than 250 MB RAM for the recognition system.