AraMS-Restore — Real Damaged Arabic Manuscript Lines

177 line images cropped from real damaged pages of a historical Arabic manuscript (book_09), each with its transcription. This is the evaluation input for AraMS-Restore: the restoration models are trained on synthetic degradation, and these lines are the honest test of whether that transfers to genuine manuscript decay.

There are no clean counterparts and no ground-truth restored images — the damage is what was on the page. Models are therefore scored by CER before vs. after restoration, read by a frozen HATFormer OCR critic, not by PSNR/SSIM.

Contents

Files	177 PNGs (RGB), every one transcribed
Pages	22, book_09_page_009 … book_09_page_039 (2–13 lines per page, median 10)
Size	20.4 MB
Line dimensions	653–977 px wide, 65–112 px tall (median 887 × 78)
Layout	<page>/line_NN.png, plus metadata.csv

metadata.csv

Column	Meaning
file_name	path to the image, e.g. book_09_page_009/line_011.png
page, line_num_file	source page, and the line number in the filename
page_offset	correction from the filename number to the corpus line index (see below)
line_id	the corresponding id in AraMS-28k-HTR
width, height	pixel dimensions
gt_text	the transcription, already normalized

Usage

huggingface-cli download Archatext/AraMS-Restore --repo-type dataset --local-dir real_damage

Run from the root of the code repo, this puts the lines exactly where the scripts expect them:

python scripts/restore_real.py --score     # CER before vs. after, both models
python app/server.py                       # web demo — "Try a random real-damage sample"

Or straight into 🤗 Datasets:

from datasets import load_dataset
ds = load_dataset("Archatext/AraMS-Restore", split="test")

Relationship to AraMS-28k-HTR

These images are not a subset of AraMS-28k-HTR and cannot be reconstructed from it. The same physical lines appear in both, but they are cropped differently:

	this dataset	AraMS-28k-HTR
Crop	raw bounding box	bounding box padded ~9 px, then polygon-masked
Neighbouring lines	fragments of the lines above/below remain	masked away
Coverage	177 lines, 22 pages	the whole corpus, 28,595 lines, 14 manuscripts

The intruding neighbour fragments are a real part of the degradation a restoration model has to cope with, and polygon masking is not reversible. None of these images is byte-identical to its corpus counterpart.

Transcription alignment

line_id is derived from the file path, with a per-page correction carried in the page_offset column (implemented as PAGE_OFFSET in scripts/restore_real.py): pages 009 and 011 are numbered one ahead of the corpus.

book_09_page_010/line_015.png  ->  book_09__book_09_page_010_line015
book_09_page_009/line_011.png  ->  book_09__book_09_page_009_line010   (offset -1)

Every line here resolves to a transcription. An earlier revision also shipped 11 trailing lines from pages 014, 015, 018, 024, 034 and 037 whose indices fell past where the corpus's line segmentation stopped for those pages; they carried no transcription and were dropped, taking pages 034 and 037 with them.

gt_text is copied from the corpus so this set can be evaluated on its own — the corpus remains the source of truth if the two ever disagree.

Leakage

book_09 sits in the test split of AraMS-28k-HTR — an entire manuscript held out from restoration training, and never seen by the OCR critic that scores the output. The before/after CER on these lines is therefore leak-free.

Citation

License