FourBi

This repository contains the official implementation for our paper Binarizing Documents by Leveraging both Space and Frequency. If you find it useful, please cite it as:

@inproceedings{pippi2023handwritten,
  title={{Binarizing Documents by Leveraging both Space and Frequency}},
  author={Quattrini, Fabio and Pippi, Vittorio and Cascianelli, Silvia and Cucchiara, Rita},
  booktitle={International Conference on Document Analysis and Recognition},
  year={2024},
  organization={Springer}
}

Setup

To run this project, we used python 3.11.7 and pytorch 2.2.0

conda create -n fourbi python=3.11.7
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip3 install opencv-python wandb pytorch-ignite

Inference

To run the model on a folder with images, run with the following command

python binarize.py <path to checkpoint> --src <path to the test images folder> 
--dst <path to the output folder>

Training

The model is trained on patches, then evaluated and tested on complete documents. We provide the code to create the patches and train the model. For example, to train on H-DIBCO12, first download the dataset from http://utopia.duth.gr/~ipratika/HDIBCO2012/benchmark/. Create a folder, then place the images in a sub-folder named "imgs" and the ground truth in a sub-folder named "gt_imgs". Then run the following command:

python create_patches.py --path_src <path to the dataset folder> 
--path_dst <path to the folder where the patches will be saved> 
--patch_size <size of the patches> --overlap_size <size of the overlap>

To launch the training, run the following command:

python train.py --datasets_paths <all datasets paths> 
--eval_dataset_name <name of the validation dataset> 
--test_dataset_name <name of the validation dataset>

Models

We release the pre-trained weights for the FourBi variants trained on DIBCO benchmarks.

	Testing data	URL	PSNR
0	H-DIBCO 2010	model	23.37
1	DIBCO 2011	model	22.26
2	H-DIBCO 2012	model	24.29
3	DIBCO 2013	model	24.17
4	H-DIBCO 2014	model	25.18
5	H-DIBCO 2016	model	19.74
6	DIBCO 2017	model	19.66
7	H-DIBCO 2018	model	20.92