Abstract
Ancient Tamil scripts preserved in stone inscriptions represent valuable cultural, historical, and social heritage. However, extracting meaningful information from these inscriptions remains a challenging task due to severe degradation, erosion, noise, and complex character structures. Word spotting (WS) has emerged as an effective alternative to traditional optical character recognition (OCR) systems for information retrieval from such degraded documents. This paper proposes a hybrid deep learning framework for word spotting in ancient Tamil inscription images. The proposed system consists of pre-processing, feature extraction, script identification, and word spotting stages. To capture comprehensive information from inscription images, both local and global features are extracted. Local visual features, such as stroke patterns and character shapes, are extracted using a self-attention convolutional neural network (SA-CNN), while global contextual features are learned using a stacked long short-term memory and bidirectional gated recurrent unit (S-LSTM–Bi-GRU) model to capture sequential dependencies. The extracted local and global features are fused through a concatenation layer to form a unified feature representation, which is then classified using a softmax layer for accurate word spotting. The effectiveness of the proposed approach is evaluated using a realtime dataset of ancient Tamil inscriptions. Experimental results demonstrate that the proposed framework achieves an accuracy of 96.6% and a precision of 98.1%, outperforming several existing word spotting methods. The results indicate that integrating local and global feature representations significantly enhances the robustness and reliability of word spotting in degraded inscription images, making the proposed approach suitable for real-world epigraphically analysis.
Keywords: Ancient Tamil Scripts, Global Features, Local Features, Self-Attention Convolutional Neural Network, Word Spotting.