The Bengali script, like other Indic scripts, presents significant challenges for optical character recognition (OCR) due to its extensive character set, complex conjunct characters, and diverse handwriting styles. While convolutional neural networks (CNNs), ensemble models, and CNN-Transformer approaches have achieved promising accuracy, their high computational complexity and large input sizes limit practical deployment on mobile and embedded devices. To address these limitations, this study proposes a lightweight CNN architecture with optimized filters and a compact 40 × 40 input size for cost-effective handwritten Bengali character recognition. The proposed model achieves 98.29% top-1 accuracy on a 50-class dataset containing 12,000 training samples and 2,997 testing samples, while significantly reducing computational complexity and parameter size. Additionally, a concise literature review highlights the evolution of Bengali OCR systems and key challenges in compound character recognition, efficiency, and real-time system validation.

https://doi.org/10.63503/j.ijssic.2026.239

Article URL: https://submissions.adroidjournals.com/index.php/ijssic/article/view/239