juktoMala: A Handwritten Bengali Consonant Conjuncts Dataset for Optical Character Recognition Using BiT-based M-ResNet-101x3 Architecture

Abstract

Bengali, the seventh most spoken language in the world by the number of speakers, doesn't have a well-established Optical Character Recognition (OCR) system for handwritten texts. One of the major reasons behind this lacking is contributed to having no complete conjuncts database. No dataset available today covers all the conjunct characters that are used by authors around the globe. In this research, we prepared a complete dataset consisting of 292 consonant conjunct characters, which is the biggest consonant conjunct character dataset to date by the number of classes available in the literature to our knowledge. We applied Big Transfer-based M-ResNet-101x3 Deep Convolutional Neural Network (DCNN) which achieves 91.32% accuracy that outperforms the baseline EfficientNetB7 approach which obtained 81.05% accuracy.