ABSTRACT
Bangla is one of the most frequently used and spoken language of the world but not much research has been conducted in recognizing Bangla handwritten characters. Almost every government and non-government offices are being digitized now a days and their previous records need to be digitized too. It is a very tough and dismal human job to convert all the records manually. So, there is a keen need to develop a system that will automatically detect and recognize all the Bangla characters in a document and generate a digital copy of them. This research proposes a Convolution Neural Network based approach in recognizing handwritten
Bangla characters. Primary goal of this research is not just to create a recognition system; rather to obtain a CNN based systematic approach to recognize Bangla handwritten characters. This research focuses on improving the accuracy of character recognition and draws a comparative analysis of validation accuracy between different character classes. For this experiment, Banglalekha-Isolated dataset was being used which is an 84-class dataset and contains 1,66,105 handwritten Bangla characters of vowels, consonants, numbers and special joint letters. ICT division, which is a wing of Ministry of ICT, Bangladesh has funded this collection of datasets and made it publicly available for all sort of research. The validation accuracy achieved on full dataset is 92.02%, 97.58% was achieved for 11 vowel class, 94.70% for 39 consonant class, 98.00% for 10 number class and 90.03% for 24 joint letter class. Also, the validation accuracy and loss were later being compared with another contemporary method of Bangla handwritten character recognition system that has used same Banglalekha-Isolated dataset.
