A Three-Stage OCR Framework for Ancient Yi Manuscripts via Relation-DETR and ConvNeXt

Digitizing ancient manuscripts by optical character recognition (OCR) presents distinct obstacles, including diminutive character sizes, glyph deterioration, significant class imbalance, and ambiguous homographs. We propose a streamlined three-stage OCR workflow designed specifically for ancient texts to tackle these challenges. Character identification is accomplished by Relation-DETR, utilizing explicit positional encoding and progressive attention refinement for effective localization of small objects. We propose an unsupervised methodology for dataset generation that integrates contrastive feature pre-training (ProtoNCE) with bidirectional retrieval-based clustering, effectively producing a well-balanced and high-quality dataset. Recognition utilizes a streamlined ConvNeXt-based model, refined to address character deterioration and varied writing situations via sophisticated enhancements and regularization techniques. Comprehensive trials validate the efficacy of this pipeline, with a Top-1 recognition accuracy of $\mathbf{9 8 . 1 4 \%}$. The methodical integration and adaptation of advanced methodologies offer a realistic answer for the digitalization of historical documents, greatly enhancing cultural preservation and historical study.