phamquiluan/DISE-2021

DISE 2021: Document Image Skew Estimation Dataset

DISE 2021 is a benchmark dataset for document image skew estimation, introduced in the ICIP 2022 paper Adaptive Radial Projection on Fourier Magnitude Spectrum for Document Image Skew Estimation.

• Paper: https://arxiv.org/abs/2603.05942 (IEEE: https://ieeexplore.ieee.org/document/9897910)
• Code: https://github.com/phamquiluan/jdeskew
• Zenodo mirror (DOI): https://zenodo.org/records/12570649

Contents

File	Description
dise2021_15.zip	Skew angles within ±15° (10,764 PNG images)
dise2021_45.zip	Skew angles within ±45°

The ground-truth skew angle of each image is embedded in its filename in square brackets, e.g. 00771163[6.24].png has a skew angle of 6.24°.

import re
from pathlib import Path

def get_angle_from_name(path):
    return float(re.search(r"\[(.+)\]", Path(path).stem).group(1))

Usage with jdeskew

pip install jdeskew

from jdeskew.estimator import get_angle
from jdeskew.utility import rotate

angle = get_angle(image)
output_image = rotate(image, angle)

License

This dataset is built upon three other datasets: DISEC 2013, RVL-CDIP, and RDCL 2017. Please respect the licenses and terms of use of these upstream datasets when using DISE 2021.

Citation

@inproceedings{pham2021dise,
  title={Adaptive Radial Projection on Fourier Magnitude Spectrum for Document Image Skew Estimation},
  author={Luan Pham, Hao Hoang, Toan Mai, and Tuan Anh Tran},
  booktitle={2022 29th International Conference on Image Processing (ICIP)},
  year={2022},
  organization={IEEE}
}