Answers are grounded in this paper and cite their sources.
Demonstrates enhanced script classification in Sanskrit manuscripts, indicating robust integration of visual and textual information.
Ancient Sanskrit manuscripts constitute an important source of historical and scientific knowledge, but their automated processing remains challenging due to script diversity, handwriting variability, image degradation, and limited annotated data. From an engineering perspective, these challenges demand robust and scalable solutions capable of operating under real‐world constraints. This paper presents a multimodal engineering framework for Sanskrit script classification that integrates complementary visual and textual information through an attention‐based fusion strategy. The proposed system adopts a dual‐branch architecture in which a transformer‐based visual module extracts hierarchical structural features from manuscript images, while a textual module captures linguistic representations derived from recognized script content. An attention‐driven fusion mechanism combines these heterogeneous feature streams into a unified representation, improving classification reliability across varying manuscript conditions. The framework is evaluated on a dataset of 2846 manuscript images encompassing three major Sanskrit scripts: Bengali, Devanagari, and Grantha. Experimental results demonstrate that the proposed approach consistently outperforms visual‐only, textual‐only, and conventional convolutional baselines in terms of accuracy, precision, recall, and F1‐score. The findings highlight the effectiveness of multimodal feature integration for engineering robust document analysis systems and support the development of automated tools for large‐scale manuscript digitization and archival applications.
No takes yet. Share an insight, caveat, or question.
Yohoshiva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: