A Comparative study on Bengali handwritten character recognition and prediction using CNN

Abstract

The transcription Bengali text to digital text is neither very efficient nor accurate. This proves to be a problem because most official work in Bangladesh is traditionally done in Bengali, on pen and paper hardcopy documents, which are difficult to transition to digital format. In our thesis, we attempted to solve this problem by improving the process of recognizing and extracting handwritten Bengali text to digital text. To aid us in our research, we have also collected an extensive data set consisting of approximately 25000 samples of around 90 Bengali characters each, including conjunct characters, to help us establish our findings. The main models we have implemented in our paper are- VGG-19, ResNet50, AlexNet, SqueezeNet. The highest training accuracy was 87% and was achieved from AlexNet, and least was 54% from VGG-19. The reliability of our model was validated by F1 score.