Stroke-Based Data Augmentation for Enhancing Optical Character Recognition of Ancient Handwritten Scripts

This paper presents a novel stroke-based data augmentation technique for enhancing the recognition accuracy of ancient handwritten scripts, focusing on Vattezhuthu characters. With the lack of large standardized datasets for such scripts, the proposed method generates realistic variations in stroke thickness, directionality, and structure to mimic natural handwriting differences. Unlike traditional geometric augmentation techniques, this approach offers fine-grained control over character modifications, resulting in a 7-10% accuracy improvement in recognition tasks. A comparative analysis demonstrates the superiority of the stroke-based method over other state-of-the-art augmentation techniques, such as GANs and Neural Style Transfer (NST), which may introduce artifacts or require extensive computational resources. The study concludes that stroke-based augmentation preserves the integrity of handwritten characters while providing sufficient diversity to enhance model performance. Future work will explore extending this method to other scripts, combining it with GANs, and incorporating adaptive augmentation strategies to further optimize recognition models.