Basic to Compound: A Novel Transfer Learning Approach for Bengali Handwritten Character Recognition

Transfer learning is widely used in various character recognition tasks. In this paper, we propose a transfer learning approach with convolutional neural network (CNN) for Bengali handwritten character recognition. When children learn the Bengali scripts, they first learn basic characters (vowels and consonants) and then go for compound characters (consonant conjuncts). Without prior knowledge of basic characters, it would be quite difficult for them to learn compound characters. In our approach, the machine mimics this human child learning process. Our study shows that CNN trained on basic characters is well capable of recognizing compound characters with minimal retraining. It performs better and also trains much faster than CNN fully trained on compound characters. Similarly, CNN trained on digits easily recognizes basic characters with a short period of training. Furthermore, pretrained CNN consistently outperforms the randomly initialized CNN while training only last few layers.