An Effective Approach for Bengali Handwritten Punctuation Recognition by Using a Low-cost Convolutional Neural Network

Bengali handwritten character recognition is considered difficult due to the complex curvatures and diverse variations of Bengali characters. Being the seventh most-spoken language of the world, Bengali has received some worth-mentioning contributions in this domain with the creation of different datasets such as NumtaDB, BanglaLekha, CMATERdb, and so on. However, none of the previous studies solely focused on recognizing punctuation marks in Bengali literature. Punctuations are important in Bengali language as without them the meaning of the whole sentence can be changed drastically. Another challenge in handwritten character recognition is the lack of low-cost convolutional neural networks (CNN). Recent studies suggested remarkable performance utilizing transfer learning approaches. However, due to the high computation of such methods, they are not suitable for recognizing in real-time via mobile devices. To solve this dilemma, in this study, we have proposed a low-cost CNN for the recognition of handwritten punctuations. Experimental results showed that our proposed architecture achieved an overall accuracy of 96.25% and outperformed popular transfer learning approaches such as DenseNet-201, InceptionV3, Xception, and EfficientNet-B6 by a significant margin with the least recognition time and the least number of parameters. Moreover, to solve the lack of dataset problem in this domain, we have collected 50 handwritten samples for each of the 28 Bengali punctuation marks in this study and formed a new dataset called ‘BiramDB’ (In Bengali, ‘Biram’ signs means punctuations). This research will be beneficial for diverse applications such as converting handwritten poems, stories, historical documents, and other writings into machine-encoded texts.