Document denoising and binarization are fundamental problems in the document processing space, but current datasets are often too small and lack sufficient complexity to effectively train and benchmark modern data-driven machine learning models. To fill this gap, we introduce ShabbyPages, a new document image dataset designed for training and benchmarking document denoisers and binarizers.
ShabbyPages contains over 6,000 clean "born digital" images with synthetically-noised counterparts ("shabby pages") that were augmented using the Augraphy document augmentation tool to appear as if they have been printed and faxed, photocopied, or otherwise altered through physical processes.
In our paper, we discuss the creation process of ShabbyPages and demonstrate the utility of ShabbyPages by training convolutional denoisers which remove real noise features with a high degree of human-perceptible fidelity, establishing baseline performance for a new ShabbyPages benchmark.
To see the ShabbyPages in action, check out this notebook 
ShabbyPages is a corpus of born-digital document images with both ground truth and distorted versions appropriate for supervised learning use in training models to reverse distortions and recover the original clean documents. This state-of-the-art dataset with synthetically-generated real-world representations can be used to improve document layout detection, text extraction and OCR processes that depend on denoising and binarization preprocessing models.
Often, training data is not accompanied by clean ground truth sources, which leads to inaccurate training and severely-limited volumes of available training data. This dataset was created using the latest version of Augraphy (8.1.0) to produce a synthetic yet realistic dataset based on ground truth documents.
 |
|---|
This repository contains the following scripts for producing the dataset:
letterfit.py, which defines a class that can fit images to a 8.5"x11" Letter page, similar to a document scanner.shabbypipeline.py, which contains a parametrized default Shabby Pages pipeline.generate_kaggle_set.py, which produces the full dataset for the Kaggle competition.remove_blank_pages.py, which removes images with >99% white pixels from the competition set.make_submission.py, which produces the submission file for the Kaggle competition.daily_build.py, which produces a small test set every day.tweet.py, which tweets an example image from the daily build.azure_file_service.py, which manages connections to Azure Files.example_shabby_pipeline_generation.ipynb, which is an example to generate shabby image from Augraphy and shabby pipeline using single input image.
An Augraphy pipeline was applied to ground truth documents to generate printed, scanned, copied and faxed versions of documents encountered in the real world. In order to preserve a pixel-level mapping between ground truth and distorted versions of documents, geometric transformations that skew or warp document images were avoided.
Follow @AugraphyProject to check out the each day's randomly generated shabby page. The ShabbyPages pipeline is used with the latest version of Augraphy each day to generate a ShabbyPage-of-the-Day image posted on Twitter like the following:
Below are related datasets that offer either real-world scanned documents or a combination of ground-truth and distorted versions.
-
RVL-CDIP dataset consists of 400,000 B/W low-resolution (~100 DPI) images in 16 classes, with 25,000 images per class https://www.cs.cmu.edu/~aharley/rvl-cdip/
-
Tobacco3482 dataset from Kaggle offers 10 different classes of forms, letters, reports, etc. https://www.kaggle.com/patrickaudriaz/tobacco3482jpg
-
FUNSD (Form Understanding Noisy Scanned Documents) dataset on Kaggle comprises 199 real, fully annotated, scanned forms that are noisy and vary widely in appearance. https://www.kaggle.com/sharmaharsh/form-understanding-noisy-scanned-documentsfunsd
-
NoisyOffice dataset from University of California, Irvine contains noisy grayscale printed text images and their corresponding ground truth for both real and simulated documents with 4 types of noise: folded sheets, wrinkled sheets, coffee stains, and footprints. For each type of font, one type of Noise: 17 files * 4 types of noise = 72 images. https://archive.ics.uci.edu/ml/datasets/NoisyOffice
-
DDI-100 (Distorted Document Images) is a synthetic dataset by Ilia Zharikov et al based on 7000 real unique document pages and consists of more than 100000 augmented images. Ground truth comprises text and stamp masks, text and characters bounding boxes with relevant annotations. https://arxiv.org/abs/1912.11658
-
NIST-SFRS (Structured Forms Reference Set) consists of 5,590 pages of binary, black-and-white images of synthesized documents from 12 different tax forms from the IRS 1040 Package X for the year 1988. These include Forms 1040, 2106, 2441, 4562, and 6251 together with Schedules A, B, C, D, E, F, and SE. https://www.nist.gov/srd/nist-special-database-2
The synthetic distortions in this dataset were generated by The Augraphy Project using a custom Augraphy pipeline to create realistic old and noisy documents from "born digital" sources. This simulation of realistic paper-oriented process distortions creates large amounts of training data for AI/ML processes to learn how to remove those distortions.
Augraphy is a Python library that creates multiple copies of original documents though an augmentation pipeline that randomly distorts each copy -- degrading the clean version into dirty and realistic copies rendered through synthetic paper printing, faxing, scanning and copy machine processes.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
If you used ShabbyPages in your research, please cite the project's dataset.
BibTeX:
@data{ShabbyPages2023,
author = {The Augraphy Project},
title = {ShabbyPages: A Reproducible Document Denoising and Binarization Dataset},
year = {2023},
url = {https://github.com/sparkfish/shabby-pages},
version = {2023}
}
Copyright 2023 Sparkfish LLC
ShabbyPages is a free and open-source dataset and software recipe distributed under the terms of the MIT license.