A CNN Framework for Real-Time and Edge Deployment with High Accuracy for Lightweight Deep Learning for Bengali OCR

Abstract

The Bengali script, like other Indic scripts, has a large character set, complex conjunct characters, and varying handwriting styles, making optical character recognition (OCR) a complex task. Although convolutional neural networks (CNNs), ensemble techniques, and CNN-Transformer models have shown high accuracy, their high computational complexity and large input size make them unsuitable for mobile and embedded systems. To make the processing cost-effective and improve the accuracy of handwritten Bengali character recognition, this paper proposes a light-weight CNN with optimized filters and a 40 × 40 input size. Compared to the existing deep learning and hybrid approaches, the proposed approach has shown 98.29\% top-1 accuracy on a 50-class dataset with 12,000 training samples and 2,997 testing samples, reducing the parameter size and computational complexity. A brief literature review is also presented to discuss the evolution of real-time and large-scale Bengali OCR systems and the challenges involved in compound character recognition, efficiency, and system-level validation.