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

Bengali handwritten character recognition is useful for historical documents recognition, bank checks and postcards recognition, Bengali number-plate recognition and so on. However, unlike other popular language of the world, Bengali is more difficult in terms of handwritten character recognition as Bengali handwritten characters have complex curvatures. Moreover, some of the handwritten Bengali characters have multiple writing fashions. With approximately 400 handwritten characters, Bengali also has some extremely similar characters which are difficult to distinguish using machine learning classifiers. Previously, many studies have been conducted for near-accurate recognition of Bengali handwritten character recognition. However, this domain still suffers from two major dilemmas i.e., lack of low-cost convolutional neural network (CNN) and lack of complete and balanced dataset. To resolve the first issue, in this research, we have proposed a low-cost CNN called ‘BornomalaNet’ that utilizes only 1.94 million parameters. In this paper, BornomalaNet has been applied on four popular and publicly accessible datasets and achieved 98.36%, 99.13%, 96.09%, and 98.11% overall accuracy for the CMATERdb, NumtaDB, Ekush and BanglaLekha datasets respectively. Experimental results showed that the proposed model has outperformed the recent studies by a significant margin. To resolve the second issue, a custom dataset have been build called ‘Bornomala’ dataset consisting of 107 classes i.e., 10 numerals, 50 basic characters, 19 modifiers, 27 punctuations and 1 special character. Experimental results showed the proposed approach has achieved 98.08% overall accuracy on the Bornomala dataset which is better than popular transfer learned models as well.