Bangla Continuous Handwritten Character and Digit Recognition Using Convolutional Neural Network

Abstract

There are few works are available in Bangla Handwritten Character Recognition. To digitalized the analog normal format Handwritten data, we proposed a new methodology to recognized the Bangla character in continuous form. That means in sentence form. We build a system which can take an input of an image of Bangla Handwritten sentence and automatically output the characters exists in the sentence. This system consists of some components like feature extraction, preprocessing, segmentation of character. In Bangla language in written format there is a rigid possibility that two characters are overlapped. This is the main problem in Bangla handwritten format that two consecutive characters overlapped with each other. Some people wrote like this. This becomes difficult to segment the character which overlapped. So, segmentation of the character is important more than to prediction of character with model. To build an OCR system for Bangla Handwritten text, firstly we detect line and segment the line after that word segmentation is done and then individually segment character by character. In this project we used EkushNet dataset model which has the best accuracy till now. This model trained on 85 basic characters, 10 digits, 52 conjunct character, 10 modifiers. By using our algorithm, we can successfully segment 95% true character and recognized by the model. Overall, this current OCR system can deal the recognition system and segmentation of the characters from handwritten Bangla texts effectively.