New OCR Framework Improves Chinese Historical Document Transcription

Preserving and analyzing Chinese historical documents is challenging due to their complex layouts, rare characters, and unique reading orders, which hinder traditional OCR methods. A new framework, TongGuOCR, has been developed to overcome these obstacles, enabling more accurate digital transcription of these valuable cultural artifacts. TongGuOCR operates in two main stages. First, a Layout-Aware Preprocessing module intelligently structures recognition blocks to maintain local context while minimizing interference from surrounding regions. Second, a Token-Augmented Recognition module enhances transcription by expanding the character vocabulary to directly represent rare glyphs and by incorporating discrete spatial displacement tokens to guide the decoder through intricate reading paths. Evaluations on two benchmarks demonstrate TongGuOCR's superior performance compared to both traditional OCR models and advanced multimodal large language models. Notably, it achieved significant improvements in accuracy and error reduction on the challenging M5HisDoc benchmark, making vast collections of historical texts more accessible for computational analysis.