A novel GAN-transformer framework for early Brahmi script generation and recognition

Abstract

Recognizing ancient scripts is crucial for understanding the historical, cultural, and linguistic context of past civilizations. However, the recognition of Early Brahmi letters from Sri Lankan inscriptions faces significant challenges due to scarce digitized data, degradation of inscriptions, and visual similarity among characters. This study introduces BrahmiGAN, a novel Generative Adversarial Network (GAN) designed to generate realistic synthetic Early Brahmi letters, addressing data limitations that hinder recognition model training. Using a benchmark dataset of 73 inscriptions, 21,195 high-quality synthetic images were generated from 888 real samples. These synthetic images demonstrated high fidelity, validated through feature-based, raster-based, and vector-based evaluations, and achieved a 92.15% approval rate from human experts. Furthermore, a vision Transformer-based ensemble model integrating Pyramid Vision Transformer and Swin Transformer is proposed for Early Brahmi letter recognition. A classification accuracy of 96.06% was attained by the ensemble model when trained on combined synthetic and real images, outperforming existing methods and surpassing the same model trained exclusively on real images. The generated dataset is publicly available to support future research (https://zenodo.org/records/14961074).