Segmentation of Ancient Brahmi Inscription Image Using Deep Learning Approach

Abstract

Segmentation of inscriptions from Ancient Brahmi scripts poses a significant challenge due to image quality, demand for greater data accessibility, and inadequate segmentation of inscriptions. Ancient Brahmi scripts serve as a progenitor for many modern-day scripts and are inscribed on stones, pillars and tablets in India and neighboring countries. Traditional image segmentation methods utilizing thresholding, edge detection, and clustering have considerably poor performance on inscriptions with artifacts such as cracks, uneven illumination, and varying surface geometries. In order to address the above limitations, we developed a segmentation method using the U-Net model with transfer learning based on pre-trained VGG-16 with heavy data augmentation. Images of inscriptions were divided into patches of size 128×128 to accommodate memory constraints. The proposed method showed significantly improved segmentation of Brahmi characters, with a total of 42 characters correctly segmented, compared to 34 characters with U-Net, and 29 characters with traditional image segmentation methods. Patch-wise and section-wise segmentation were carried out to evaluate the robustness of the segmentation model on real world illumination conditions and varying surface geometries. The proposed method achieved a Dice score of 0.714. In addition, high precision (0.893), recall (0.800), and F1-score (0.844) demonstrated that the segmentation method is robust on inscriptions with cracks and noise. The segmentation method combines transfer learning and data augmentation for segmentation of ancient scripts, and offers a potentially reliable and effective solution for the digitization of archaeological texts and advanced studies in computational epigraphy.