Dege-HDIE:A historical document image enhancement dataset based on the true degradation of the Dege Tibetan text 《Kangyur》

Dege-HDIE Dataset 1. Overview Dege-HDIE contains 200 pixel-aligned pairs of high-resolution color imagesof the Tibetan woodblock-printed Dege Kangyur, accompanied by page-matchedTibetan text files. The observed pages contain natural paper-colorvariations and fiber textures alongside historical stains and dark spots.The enhancement references aim to reduce contamination while retaining theaged appearance, natural paper tones, existing text structures, andpage layout. 2. Files and correspondence `degraded.zip` — 200 original degraded input images.- `clean1.zip` — the first 60 corresponding enhancement reference images.- `clean2.zip` — the remaining 60 corresponding enhancement reference images. - `clean3.zip` — the last 80 enhanced reference images. - `txt.zip` — 200 corresponding Tibetan text files encoded in UTF-8. All images are RGB PNG files with a resolution of 10417 × 2084 pixels(width × height). The release contains 400 images and 200 text files.Each archive contains its corresponding `degraded/`, `clean/`, or `txt/`folder. `clean1.zip` and `clean2.zip` each contain 100 images from the`clean/` reference set. After extraction, `clean1.zip` 、 `clean2.zip` and `clean3.zip’together provide the complete set of 200 clean reference images; they maybe extracted or merged into the same `clean/` directory. Match files by their shared page identifier, for example: `degraded/k91-1.png` <-> `clean/k91-1.png` <-> `txt/k91-1.txt` The file `clean/k91-1.png` is provided in either `clean1.zip` 、`clean2.zip` and `clean3.zip’, depending on its position in the clean reference set. TXT files contain page-associated Tibetan text, without charactercoordinates or pixel-level stain annotations. These archives do not includetrain/validation/test split manifests. When defining a split, keep allpatches from the same page in the same subset. 3. Reference construction The construction pipeline comprises limited manual repair using matchingglyph patches from the same page and removal of manually identified text-region specks; physical paper-boundary detection; foreground-awarebackground-field estimation and initial enhancement; local backgroundcompensation, multiscale cloud-like stain attenuation, and shallow-staincorrection; and localized refinement of text-region specks and darkbackground objects using fixed references and neighboring donor regions. The pipeline retains the original image canvas without page-wide geometricrectification or text rearrangement. Input and reference images thereforeshare the same spatial coordinates. The clean images are constructedenhancement references, not observations of the pages before historicaldegradation. Local repairs may alter text pixels or background textures. 4. Use and citation The dataset supports historical document image enhancement, backgroundstain attenuation, and related evaluation. Apply cropping and geometrictransformations consistently to each image pair. Follow the licensespecified in the Zenodo record and cite its dataset DOI.