Enhanced Brahmi Script Recognition with GAN-based Inpainting and Attention-Guided Feature Learning

Abstract

This paper introduces a more advanced recognition system that integrates stroke restoration, attention-based learning of features, and hybrid classification. A Generative Adversarial Network (GAN) inpainting technique is used for the reconstruction of missing or deteriorated character strokes so that visual integrity in the input data is enhanced. After restoration, a ResNet50 model directed by attention captures subtle spatial and channel dependencies after enhancement by Convolutional Block Attention Modules (CBAM) and Squeeze-and-Excitation (SE) blocks. Ultimately, a new hybrid classification approach combines deep feature extraction using ResNet50 with the power of an ensemble decision made using a Random Forest classifier to increase accuracy and interpretability. The suggested pipeline outperforms state-of-the-art baseline models on a manually curated Brahmi script dataset significantly, showing strong resistance to erosion, noise, and structural loss. This effort pushes the field of automated recognition for historical scripts forward and is a contribution toward digital preservation work in epigraphy.