Development of Real Time CNN-Based System for Telugu Word Recognition on Raspberry Pi

The recognition of handwritten regional scripts remains a significant challenge in the domain of Optical Character Recognition (OCR), particularly for complex languages such as Telugu. This paper presents a real-time Telugu word recognition system based on Convolutional Neural Networks (CNNs), implemented on the Raspberry Pi 3B+, a cost-effective and portable embedded computing platform. The system utilizes the IIIT-H Indic Handwriting Dataset, which contains a wide variety of wordlevel handwritten samples collected from native speakers. To improve recognition performance, comprehensive preprocessing techniques are applied, including resizing, grayscale conversion, normalization, noise reduction, and data augmentation. The CNN model is trained using supervised learning and optimized for embedded deployment through quantization techniques, enabling efficient inference on the Raspberry Pi 3B+. Real-time image acquisition and processing are handled using OpenCV, facilitating the recognition of both printed and handwritten Telugu words. Experimental evaluation demonstrates that the system achieves a recognition accuracy of 94 %, indicating its effectiveness and reliability. The proposed implementation offers a lightweight and scalable OCR solution suitable for applications such as document digitization, assistive technologies, and educational tools. This work contributes to the advancement of regional language processing by combining deep learning with accessible embedded systems.