An Analysis on Bengali handwritten conjunct character recognition and prediction

Abstract

In the very active field of handwriting recognition, a lot of research can be found in the detection of the handwriting of various languages, especially English. However, for languages like Bengali, while they hold some success in handwritten character recognition, a big roadblock is Bengali conjunct characters or “Juktakkhor”. As Bengali conjunct characters are very complex, even today many institutions in Bangladesh still maintain documents as handwritten copies. In this paper, we will present a model that focuses on conjunct character recognition and conversion to textformat. OurproposedsystemwillbetrainedandtestedusingCNNmodelslike VGG19, ResNet-50, GoogleNet, LSTM, ShuffleNet etc. The results generated from preliminary analysis yield that ShuffleNet gives the most accurate results with an accuracy of 91.2% followed by GoogleNet with 73.3%.