Optical character recognition for ancient scripts: a case study on Syloti Nagri using deep learning models

This study proposes a novel OCR system for the Syloti Nagri script, addressing the complexities of segmenting and classifying its unique characters. Given the limited annotated data and intricate structures, a tri-level segmentation approach was applied, involving line, word, and character segmentation. Deep learning models-VGG16, VGG-19, ResNet-50, MobileNet-v3, and Xception, were evaluated, with VGG-19 achieving the highest scores in accuracy, precision, recall, and F1 metrics. The segmentation process attained 98% accuracy for word and 94% for character segmentation, while the classification model reached a test accuracy of 99.78%, demonstrating robust recognition of complex patterns. This work advances OCR technology and supports the preservation and accessibility of the Nagri script.