Ancient-book restoration gets 100,000 paired pages: AncRestore routes visual priors by damage type to 26.70 dB PSNR
Synopsis
Ancient-book pages often carry paper ageing, water-induced distortion, insect holes and mold at once, while existing public benchmarks are limited in scale and stack isolated degradations, which weakens supervised restoration. AncBench supplies 100,379 pixel-aligned clean–degraded pairs in which chemical, physical and biological damage are composed in sequence; trained on it, AncRestore first estimates page-wide and local degradation, then dynamically selects and distributes visual priors according to those cues. On the AncBench test set AncRestore reaches 26.70 dB PSNR, 0.9191 SSIM and 0.1346 LPIPS, with 89.58% character-recognition accuracy.
Interpretation
Ancient-book image restoration now has a paired benchmark at the hundred-thousand scale, where degradations are no longer stacked as separate types but each step acts on the page state left by the previous one. Earlier cultural-heritage benchmarks mostly target inscriptions or character-level reconstruction, and their synthetic data are organized around single degradation types or built from mask-based defects and local perturbations, with limited scale and diversity. AncBench releases 100,379 aligned clean–degraded pairs, 99,879 for training and 500 for testing, and every degraded image contains damage from the chemical, physical and biological families.
The restoration network first reads where a page is damaged, then decides which visual prior to call on and where to weight it more heavily. Generic restoration methods typically model degradation as a global condition or select from fixed degradation priors, which limits them on composite degradation coexisting within one page. A damage representation block extracts a global degradation vector, local degradation tokens, a soft damage map and a multi-scale damage-prior pyramid; dynamic expert routing weights four DINOv3 visual priors by both image-level and location-wise scores and injects them into a half-quadratic-splitting restoration backbone.
Each module alone yields limited gains; only together do they widen the reconstruction margin. Against the baseline, adding dynamic expert routing alone raises PSNR from 25.56 dB to 25.80 dB, and adding the damage representation block alone reaches only 25.61 dB. With both enabled, AncRestore reaches 26.70 dB PSNR, 0.1346 LPIPS and 45.2042 MUSIQ on AncBench, improvements of 1.14 dB, 0.0201 and 0.2956 over the baseline.
Perspective
The demonstrated scope is full-page restoration on the synthetic paired AncBench data, plus external evaluation on the public HDR28K and HanInpaint historical-document datasets; the intended reader is someone working on document image restoration and cultural-heritage digitization.
AncBench degradations are produced by a synthesis pipeline that the authors describe as approximating how real ancient-book degradation forms rather than collecting real pairs directly; on AncBench, AncRestore ranks behind IR-SDE on MUSIQ, which the authors attribute to differing preferences of perceptual and reconstruction-oriented metrics.