AI-Powered Sinhala Character Recognition and Digital Transformation

Abstract

This research focuses on developing an effective system for recognizing and converting handwritten and printed Sinhala text into digital format. As the primary language of Sri Lanka, Sinhala presents unique challenges for handwriting recognition due to its intricate strokes and complex character structures. Existing methods often fall short in accurately interpreting Sinhala characters, highlighting the need for a tailored solution. The proposed system employs Convolutional Neural Networks to classify and recognize Sinhala characters with high precision. A key innovation is error-guided preprocessing, applied iteratively to images misclassified during the initial training phase. Failed images are processed using methods such as blurriness detection, dynamic contrast adjustment, noise removal with bilateral filtering, and morphological operations for stroke enhancement. This approach ensures improved image quality and meaningful feature extraction for subsequent retraining. Additional techniques like contour analysis and gradient- based feature extraction further enhance the system's recognition capabilities. To optimize performance, strategies such as data augmentation, hyperparameter tuning, and model ensembles are explored, improving the system's adaptability and robustness. The system is evaluated on a diverse dataset of handwritten and printed Sinhala text, demonstrating significant improvements in recognition, accuracy and efficiency. Its applications include optical character recognition, document digitization, and automated form processing. This thesis contributes a comprehensive, CNN-based methodology tailored to the complexities of Sinhala script, offering a promising solution for advancing Sinhala language technologies.