Two U-Net generators that take a degraded Arabic manuscript line and return a
readable one. Same architecture, same data, same schedule — they differ only in
what they optimize:
unimodal (Level 1)
recaware (Level 2)
Loss
L1 + ResNet50 perceptual
the same + a frozen HATFormer OCR critic
Optimizes for
visual fidelity
legibility
Text
not used
used as a loss target, never as an input
The Level-2 critic reads the restored image and is scored against the true
transcription, so gradients flow into the pixels that matter for reading. Code:
github.com/ArchaText/AraMS-Restore.
Results
On real damaged lines from a held-out manuscript — the honest test — the
image-only model does not help, while the recognition-aware one does:
Input to the reader
CER %
WER %
Degraded (untouched)
29.81
59.98
Level 1 restored
30.15
59.98
Level 2 restored
28.15
58.44
The ranking inverts between the two ways of measuring — Level 1 wins every pixel
metric it is optimized for, and loses on reading:
PSNR ↑
SSIM ↑
CER ↓
Degraded
22.76
0.9498
35.11
Level 1
34.66
0.9886
28.12
Level 2
30.66
0.9781
27.13
(Synthetic test set. CER read by the frozen HATFormer reader.)
Each is a state_dict for the generator built by
restoration.models.unet_pix2pix.build_generator: 31.04 M parameters, float32,
RGB in/out, base=64, depth=4, no discriminator (pix2pix off).
A line is resized to height 64 keeping aspect and right-padded to width 1152
(build_strip), RGB in [0, 1], padding pure black and excluded by a mask. 1152 =
3 × 384 tiles cleanly into the HATFormer 384 × 384 RTL-flipped canvas, which is
what makes the Level-2 recognition loss exact. Feeding a differently-shaped or
grayscale image will degrade output quality.
Training
Both models: 20 epochs, AdamW lr=2e-4 (β=0.5, 0.999), batch 8, seed 1337, and 4
distinct degraded variants per clean line per epoch. Damage is synthesized on the
fly from clean train-split lines only — never val/test — by the localized
degradation engine (erasure, tears, worm holes, bleed-through, ink blots, ink
feathering), calibrated against real manuscript decay.
Training data: AraMS-28k-HTR,
train split — 20,103 lines from 9 manuscripts.
Leakage
Splits are manuscript-level disjoint: the evaluation manuscripts (book_03, 05, 09)
are held out of restoration training entirely, and the OCR critic that scores
the output never trained on them either. Both conditions are needed — a critic that
had seen the test books would contaminate the training signal and the metric at
once.
Limitations
Trained on synthetic degradation; real decay is a different distribution, and
the gap shows — the CER gain on real damage (1.7 points) is smaller than on the
synthetic test set.
Single script tradition and a single corpus of 14 manuscripts; no claim of
generalization to other hands, papers or languages.
Level 1 makes real damage marginally harder to read (30.15 vs 29.81 CER) while
scoring far better on PSNR/SSIM — do not select a restoration model on pixel
metrics alone.
Line-level only. Page segmentation is upstream and not part of this release.
