This repository contains the datasets and code for the paper ColDBin: Cold Diffusion for Document Image Binarization by Saifullah Saifullah, Stefan Agne, Andreas Dengel, and Sheraz Ahmed.

Requires Python 3+. For evaluation, please download the data from the links below.

Approach:

Qualitative Results:

Quantitative Results

Dataset	FM	p-FM	PSNR	DRD
DIBCO 2009	94.19	96.52	20.65	2.58
DIBCO 2010	95.29	96.67	22.06	1.36
DIBCO 2011	95.23	96.93	21.53	1.44
DIBCO 2012	96.37	97.41	23.40	1.28
DIBCO 2013	96.62	97.15	23.98	1.20
DIBCO 2014	97.89	98.10	24.38	0.66
DIBCO 2016	89.50	93.73	18.71	3.84
DIBCO 2017	93.04	95.12	19.32	2.29
DIBCO 2018	89.71	93.00	19.53	3.82

Prepare dibco datasets

Download the datasets from the link: Use the example dataset preparation script provided for DIBCO 2013 dataset:

./scripts/prepare_dataset.sh

Train

Train a diffusion model in cold manner using the example training script for DIBCO 2013 dataset:

./scripts/train.sh

Test:

Test the trained model using the example testing script for DIBCO 2013 dataset:

./scripts/test.sh

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.