Character Segmentation in Brahmi Script: Object Detection and KNN Fusion Approach

Abstract

Automated word recognition has seen significant advancements, and character segmentation remains a crucial step in this process. Although several studies have focused on the segmentation of popular modern scripts, there has been relatively little attention on ancient scripts due to their infrequent use and complex structural features. Brahmi, an ancient script with a rich history, presents unique challenges for segmentation, including disconnected dot components and complex compound characters, which existing methods fail to address effectively. This study proposed a novel two-stage segmentation framework that combines contour-based object detection with a K-Nearest Neighbors (KNN)-based reattachment mechanism. The novelty of this approach lies in its ability to accurately reattach disconnected components (e.g., isolated dots) to their corresponding base characters, a problem not adequately solved in previous studies. The process comprises two main parts: line segmentation and character segmentation. The performance of the proposed approach was evaluated using printed Brahmi texts, achieving 99.81% accuracy for line segmentation and 97.32% for character segmentation, resulting in an overall average accuracy of 98.19%. These results demonstrate that the proposed hybrid framework not only surpasses prior Brahmi segmentation efforts but also provides a generalizable solution for ancient Indic scripts with similar structural challenges.