A Comparative Analysis for Bengali Handwritten Character & Digit Recognition Using Capsule Network

At the present time handwritten character recognition is the most demandable and interesting topic to work on for many research areas and projects. A lot of work has already been done on Bengali handwritten character recognition. We proposed Bengali handwritten character recognition (HCR) using Capsule Network. It is a deep-learning technique and has never been used in this area except for Bengali handwritten digit recognition. We have tried different Bengali character datasets, among them Bornomala Dataset(Bornomala Fifty) was more reliable as the data was cleaned and processed already. The regularly used machine-learning techniques really struggle with rotated images. Capsule Network works very amazingly in the case of rotated images. We have tested our model with images that are rotated 50 degrees to both the left and right side. We provide a comparative study of recognizing Bangla handwritten character and digit using capsule network. The predicted accuracy is achieved 81.29% for characters, 95.57% for digits and 81% for number and digit combination.