Myanmar Handwritten Character Recognition from Similar Character Groups Using K-Means and Convolutional Neural Network

The recognition of characters from similar character groups has gained popularity because it considers only corresponding similar character groups instead of all possible characters when one character is to be recognized. In this paper, we present a novel Myanmar optical character recognition (MOCR) system for handwritten text. Firstly, the system uses tensor-based subspace learning (TSL) to reduce dimension space for only relevant features. The system then gathers similar character groups using the K-means clustering algorithm. The convolutional neural network (CNN) is adopted to discriminate the input character string against the groups of characters that are the most similar to it. The system finally classifies and produces the results as the recognized characters in the phrase level. The experimental works are conducted to evaluate the results in terms of precision, recall, processing time, and error rate. The results have shown that they are superior to other contemporary MOCR works.