Related Experiment Video
Updated: Jan 9, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.3K
Enhancing Bangla handwritten character recognition using Vision Transformers, VGG-16, and ResNet-50: a performance
A H M Shahariar Parvez1, Md Samiul Islam2, Fahmid Al Farid3
1Department of Software Engineering, Daffodil International University, Dhaka, Bangladesh.
Frontiers in Big Data
|December 1, 2025
Summary
Vision Transformer (ViT) achieves 98.26% accuracy for Bangla Handwritten Character Recognition (BHCR), outperforming CNNs like VGG-16 and ResNet-50. This demonstrates ViT
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Bangla Handwritten Character Recognition (BHCR) faces challenges due to complex scripts and handwriting variability.
- Existing deep learning models often struggle with the nuances of Bangla characters.
Purpose of the Study:
- To comparatively evaluate Vision Transformer (ViT), VGG-16, and ResNet-50 for BHCR.
- To assess the effectiveness of transformer-based architectures in improving BHCR accuracy and generalization.
Main Methods:
- Utilized the CMATERdb 3.1.2 dataset with 24,000 images of 50 basic Bangla characters.
- Implemented and compared three deep learning architectures: Vision Transformer (ViT), VGG-16, and ResNet-50.
- Analyzed model behavior, including overfitting in Convolutional Neural Networks (CNNs) and character-level misclassifications.
Main Results:
- Vision Transformer (ViT) achieved a state-of-the-art accuracy of 98.26%.
- ViT significantly outperformed VGG-16 (94.54%) and ResNet-50 (93.12%).
- ViT demonstrated superior ability in capturing global context and long-range dependencies, leading to better generalization.
Conclusions:
- Transformer-based architectures, particularly ViT, show significant potential for robust and accurate BHCR.
- ViT's performance establishes a new benchmark for future research in Bangla character recognition.
- The study provides valuable insights into model behavior and misclassification patterns in BHCR.