Handwritten Bengali Digit Classification Using Deep Learning

Bengali Character Recognition has recently become a field of increased academic interest due to its wide and varying applications. In this paper, a recognition system based on transfer learning has been proposed for robust recognition of Bengali numeric digits. Several preprocessing techniques have been used on the images to aid in the training process. ResNet-18 has been used for the transfer learning. All deep learning applications have been achieved by the extensive use of the PyTorch library. The dataset used has been chosen for its unbiasedness, which allows the proposed model to perform better on digits it has never encountered before. The proposed model has shown state of the art performance, with an accuracy of 92% in just 25 epochs. Thus, the proposed model doesn’t only provide a decent accuracy, but it does so with lesser number of parameters than other leading approaches along with taking fewer epochs to reach the final result.