RegNetY-16GF Classifies Historical Pages with 99.16% Accuracy

A team led by Kateryna Lutsai at Charles University developed an automated page image classifier for historical document archives, addressing the bottleneck of manual sorting in large-scale digitization projects. They created a dataset of over 48,000 annotated page images from century-old Czech archaeological archives, refined through four stages of expert review. Using an 11-category label scheme (e.g., text, tables, graphics), they compared hand-crafted feature baselines with fine-tuned deep learning architectures: EfficientNetV2, RegNetY-16GF, Vision Transformers (ViT-large), Document Image Transformers (DiT), and multimodal CLIP models.

The results showed that image-only models dramatically outperformed the feature-based baseline (75% accuracy). RegNetY-16GF achieved 99.16% Top-1 accuracy on the held-out test set, while ViT-large reached 99.12% and CLIP ViT-B/16 hit 99.14% with optimized text prompts. When applied to 649,508 unlabeled archival pages, image-only models showed over 90% inter-model agreement. However, CLIP exhibited under 65% agreement with image-only models on unlabeled data, making it less reliable for deployment. The authors release the final models, annotated dataset (48k images), and software under open-source licenses for use in humanities and digital libraries.