Zone and rule assisted recognition of Meitei-Mayek handwritten characters

This paper presents a two-stage approach for an improved recognition accuracy of handwritten Meitei Mayek characters. The first-stage is a CNN based recognition. The second-stage recognition is based on the zonal information and script-specific orthographic rules. The work is carried out on the finding that low recognition accuracy of Meitei Mayek handwritten characters is due to the presence of highly shape-similar confusing character pairs. First, the confusing character pairs are identified from the confusion matrix obtained while training and testing the CNN. For these confusing characters pairs, the zone information and certain script-specific orthographic rules are incorporated in the second recognition stage to distinguish between them. The second-stage recognition is carried out only on those characters which belong to one of the confusing character pairs. This makes the technique computationally more efficient than those where it is carried out for every output of the recognition stage. For zone identification, a novel method based on the row co-ordinates of characters present in the word is proposed. An improvement in the recognition accuracy is achieved with the two-stage approach compared to the single-stage recognition using the CNN. An improve ment in recognition accuracy of 3.36% is shown by our approach with a recognition accuracy of 91.86%.