CharNetwork Acter-Level Bengali OCR Using a Custom Convolutional Neural
Year: 2026
Type: preprint
Abstract: Optical Character Recognition (OCR) for the Bengali language remains a challenging research problem because of the script's complex morphology, visually similar characters, compound strokes, and diacritical marks. Although recent deep learning approaches have significantly improved recognition performance, many existing models rely on computationally intensive transfer learning architectures that are unsuitable for real-time or resource-constrained environments. This study proposes a lightweight Convolutional Neural Network (CNN) specifically designed for isolated Bengali character recognition using the Bengali Character Recognition Dataset (BCRD). The proposed framework incorporates image preprocessing techniques, including grayscale conversion, Otsu thresholding, normalization, resizing, and data augmentation, followed by a compact CNN architecture employing Batch Normalization and Dropout for stable training and improved generalization. Experimental evaluation demonstrates that the proposed model achieves 97.1% test accuracy, with corresponding Precision, Recall, and F1-score values of 97.0%, 96.8%, and 96.9%, respectively, while maintaining a substantially lower computational complexity than many large pre-trained networks. The small performance gap between the training and testing datasets further indicates excellent generalization capability. These findings demonstrate that lightweight CNN architectures can provide an effective balance between recognition accuracy and computational efficiency, making the proposed approach well suited for practical Bengali OCR applications, mobile devices, embedded systems, and intelligent document digitization platforms.
Source: Preprints.org
Authors: Khadiza Akter, Kazi Abdul Mannan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204484599