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-igniteTo 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>
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>
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 |