A method for the off-line recognition of handwritten characters with hidden Markov models (HMMs) is described. Performance of our HMM is compared to a baseline Naive Bayes classifier. Experimental results are given for different variants of our HMM algorithm, with optimizations for ignoring the effects of inter-word transitions when applying the Viterbi algorithm to return the most likely character sequence. Finally, our HMM algorithms are compared to a dictionary creation and lookup algorithm that is less generally applicable but for our particular dataset achieves almost perfect classification results. 1

Abstract—Handwriting recognition is a main topic of Optical Character Recognition (OCR), which has a very wide application area. Hidden Markov model is a popular model for handwriting recognition because of its effectiveness and robustness. In this paper, we propose a hidden Markov model with parameteroptimized k-means clustering for handwriting recognition. We explore two deep features of the images of characters, thus significantly boosting the effectiveness of k-means clustering. The experiments show that our model largely increases the average accuracy of HMM with k-means clustering to 83.5 % when the number of clusters is 3000. Index Terms—OCR; HMM; k-means; clustering; I.