Deep learning approaches for historical manuscript digitization: A scoping review of emerging trends and research gaps

Abstract

Historical manuscript digitization has been transformed by deep learning, replacing rule-based approaches in binarization, layout analysis, handwritten text recognition, writer identification, and restoration. However, recent evidence remains fragmented across workflows and manuscript traditions. This scoping review aims to characterize deep-learning approaches in manuscript digitization, map the geographic and linguistic distribution of studies, identify methodological trends, and reveal research gaps. Following PRISMA-ScR guidelines and the Population–Concept–Context framework, literature from six databases and three repositories published between 2015 and February 2026 was reviewed. Findings reveal seven major themes: image binarization, layout analysis, handwritten text recognition, writer identification, restoration and virtual unwrapping, OCR-free document understanding, and low-resource transfer learning. Research is dominated by Latin, Arabic, and Chinese manuscripts, while Southeast Asian and other underrepresented scripts receive limited attention. Deep learning has substantially improved digitization workflows, yet challenges remain in annotated corpora, evaluation consistency, reproducibility, and interdisciplinary integration with paleographic scholarship.