A Comprehensive Survey on Advanced Ancient Document OCR

This paper reviews recent advancements in Optical Character Recognition (OCR) for ancient documents to establish a robust digital archiving framework. Historical records present challenges including physical degradation, non-standard layouts with complex interlinear notes, and a scarcity of labeled data. We delineate the paradigm shift from local feature extraction in CNNs to global context modeling using Transformers and Vision-Language models. For preprocessing, structural restoration via Hierarchical Networks and quality enhancement through Diffusion-based binarization are examined. Layout analysis techniques, such as DBNet for non-linear text flow and Graph Convolutional Networks for logical structure recovery, are evaluated to ensure hierarchical integrity. To overcome data scarcity, we explore diverse learning paradigms: weakly-, semi-, and self-supervised learning (e.g., Masked Autoencoders), along with transfer learning for domain adaptation. Furthermore, we review specialized datasets like HUST-OBC and AI Hub that facilitate effective model benchmarking. This review serves as a technical roadmap for improving the robustness of East Asian document digitization, bridging the gap between historical preservation and advanced machine intelligence.