Bangla Handwritten Character Recognition using Vision Transformer

Handwritten character recognition (HCR) is a crucial field of research, especially for complex scripts like Bangla, which include a variety of detailed and diverse characters. Hand written character recognition has mostly been approached using Convolutional Neural Networks (CNN) and transfer learning. Though researchers have been actively exploring handwritten character recognition, only a few of them have explored the potential of Vision Transformers (ViT), especially for recognizing Bangla handwritten characters. Vision Transformer is a recent addition to deep neural networks, offering strong performance with lower computational power than existing methods. This paper introduces a Vision transformer with data augmentation techniques to tackle Bangla Handwritten Character Recognition (BHCR) using Vision Transformer (ViT) model with data augmentation techniques. Traditional convolutional neural networks (CNN) are focused on the local pattern while ViT is focused on capturing long-range dependencies by processing images as sequences of patches. The proposed Vision transformer model with 8 transformer layers was trained and evaluated on the CMATERdb 3.1.2 dataset which consists of 15,000 images representing 50 Bangla handwritten characters. Our proposed ViT model with data augmentation techniques, achieved an outstanding accuracy of 95.27% outperforming many existing Bangla handwritten character recognition methods. These remarkable outcomes highlight the importance of data augmentation and the efficiency of Vision Transformers in handling complex scripts like Bangla.