DESCAN-18K

DESCAN-18K is a dataset for the descanning task. This dataset contains 18,360 paired real scanned/original images for recovering original digital images from their real scanned counterparts.

---

---

Overview

DESCAN-18K is a dataset for the descanning task: recovering the original digital image from a real scanned page. This dataset supports that task with large-scale real scanned/original pairs collected from real scanners and original digital pages.

DESCAN-18K's key characteristics:

• Real paired data: each scan/clean pair comes from a real scanned page and its corresponding original digital page
• New task setting: the work explicitly defines descanning as its own restoration problem
• Cross-scanner evaluation: the test set uses scanner models not seen during training
• Mixed degradations: color distortion, bleed-through, halftone, texture distortion, and scanner noise often co-occur
• Scale: 18,360 aligned pairs across four real-world scanners
• Format: RGB TIFF images at 1024 x 1024
• Content diversity: photographs, graphics, layouts, and text-heavy magazine pages

Each sample contains:

Folder	Contents
scan/	Real scanned page image (model input)
clean/	Corresponding original digital page image (target)

The core dataset consists of real scanned/original pairs. Synthetic data appears in the paper only as an additional training augmentation strategy for the DescanDiffusion+ model variant.

---

Descanning Process Overview

The figure below shows the descanning process: a digital image is printed, scanned, and then restored back toward its original digital form through descanning.

---

Dataset Statistics

The table below follows the paper-reported split sizes.

Split	Scanners	Paper-Reported Pairs
Train	scanner3, scanner4	17,640
Valid	scanner3, scanner4	360
Test	scanner1, scanner2	360
Total	-	18,360

---

Scanner IDs

ID	Scanner Model	Split
scanner1	Fuji Xerox ApeosPort C2060	Test only
scanner2	Canon imagePRESS C650	Test only
scanner3	Canon imageRUNNER ADVANCE 6265	Train / Valid
scanner4	Plustek OpticBook 4800	Train / Valid

Training and validation use scanner3 and scanner4. The test set uses scanner1 and scanner2, which are unseen during training and therefore support cross-scanner generalization evaluation.

---

Data Structure

DESCAN-18K-dataset/
|-- Train/
|   |-- scan/    # real scanned page images used as model inputs
|   `-- clean/   # aligned original digital target images
|-- Valid/
|   |-- scan/    # validation scanned images
|   `-- clean/   # validation original targets
|-- Test/
|   |-- scan/    # unseen-scanner scanned images for evaluation
|   `-- clean/   # corresponding original targets
`-- README.md

Each file in scan/ has a same-named counterpart in clean/.

Path	Role
Train/scan	Degraded real scanned inputs for training
Train/clean	Paired original digital targets for training
Valid/scan	Scanned validation inputs
Valid/clean	Paired validation targets
Test/scan	Scanned test inputs from unseen scanners
Test/clean	Paired test targets

---

File Naming

Split	Convention	Example
Train/	Source-document style	BookOfMaking.00001.01.tif
Valid/	Scanner-based	scanner03_000.tif
Test/	Scanner-based	scanner01_000.tif

All images are stored as TIFF files (.tif).

---

Source and Construction

According to the paper, DESCAN-18K was built from 11 magazine types from the Raspberry Pi Foundation, licensed under CC BY-NC-SA 3.0. The pages contain diverse image and text content and had preservation durations ranging from a few days to seven years.

The paper describes the construction pipeline as follows:

1. Magazine pages were manually scanned with four scanner models.
2. Corresponding original PDF pages were collected online.
3. Both scanned and original data were converted to RGB TIFF.
4. Page pairs were aligned with AKAZE registration.
5. Poor matches were manually filtered.
6. The aligned pages were randomly cropped to 1024 x 1024.
7. The cropped pairs were registered again with AKAZE to secure aligned patch pairs.

The paper also states that scanned TIFF images were calibrated under the IT 8.7 / ISO 12641 standard.

---

Degradation Types

The paper identifies six representative degradation types in DESCAN-18K. These often appear in combination within the same sample.

#	Degradation	Description from the paper
1	Color transition	Global color fading, saturation change, or altered color tone
2	External noise	Dots or localized stains caused by foreign substances
3	Internal noise	Scanner-generated crumpled, curved, or laser-like artifacts
4	Bleed-through effect	Back-page contents becoming visible in the scan
5	Texture distortion	Global physical textures or wrinkle-like artifacts
6	Halftone pattern	Printing-induced CMYK dot structures

---

Intended Use

DESCAN-18K is intended for paired descanning research and evaluation:

• Input: scanned image from scan/
• Target: original image from clean/

It can be used for developing, benchmarking, and analyzing methods that recover original digital images from real scanned pages. The scanner-based split is especially useful for evaluating generalization to unseen scanner types.

Why this dataset matters:

• It is built from real scanned pages, not only synthetic corruption pipelines.
• It provides direct paired supervision between scanned inputs and original digital targets.
• It explicitly supports the newly defined descanning task.
• It includes a test split from unseen scanner models.

---

Reported Benchmark Results

The following results are reported in Table 1 of the paper on the original DESCAN-18K testing set.

Method	PSNR (dB)	SSIM	LPIPS	FID
Pix2PixHD	20.58	0.8014	0.057	18.30
CycleGAN	21.52	0.8417	0.050	16.99
HDRUNet	20.90	0.8480	0.055	16.42
Restormer	20.37	0.7915	0.152	25.57
ESDNet	21.22	0.8418	0.088	15.24
NAFNet	22.03	0.8538	0.048	16.00
OPR*	18.09	0.7249	0.158	21.45
DPS*	17.93	0.7354	0.150	41.64
Clear Scan	21.46	0.8183	0.054	18.09
Adobe Scan	15.80	0.6153	0.141	23.55
Microsoft Lens	20.48	0.8013	0.056	18.97
DescanDiffusion	23.40	0.8717	0.042	13.51
DescanDiffusion+	23.43	0.8736	0.044	14.60

* Methods with an asterisk are official pre-trained versions used by the paper authors.

The paper states that the DescanDiffusion variants outperform competing methods and commercial products on the DESCAN-18K test set. Within Table 1, DescanDiffusion+ has the best PSNR and SSIM, while DescanDiffusion has the best LPIPS and FID.

---

Citation

If you use DESCAN-18K in your research, please cite:

@article{cha2024descanning,
  title={Descanning: From Scanned to the Original Images with a Color Correction Diffusion Model},
  author={Cha, Junghun and Haider, Ali and Yang, Seoyun and Jin, Hoeyeong and Yang, Subin
          and Uddin, AFM and Kim, Jaehyoung and Kim, Soo Ye and Bae, Sung-Ho},
  journal={arXiv preprint arXiv:2402.05350},
  year={2024}
}

---

References

Resource	Link
Paper	https://arxiv.org/abs/2402.05350
Code	https://github.com/jhcha08/Descanning
Lab	https://mlvc.khu.ac.kr/

---