DE-GAN: A Conditional Generative Adversarial Network for Document deblurment

This is a fork of the original repository of the DE-GAN model. It provides a simple API and CLI for using the model installable via pip. Note that while neural models do most of the old-school hand-crafted image processing they can still be improved by some custom modification. Pre- and post-processing the images is recommended for optimizing the results to your use-case.

Installation

pip install https://github.com/00sapo/degan/releases/download/v0.1.1/degan-0.1.1-py3-none-any.whl

For CLI usage, I recommend using pipx: pipx install https://github.com/00sapo/degan/releases/download/v0.1.1/degan-0.1.1-py3-none-any.whl

For installing pipx, please refer to the official documentation.

Requires Python >=3.8, <3.11. Python 3.10 is recommended.

Usage

API

from degan import DEGAN

# use the official weights distributed by the authors and provided within the pip package:
model = DEGAN()
model.binarize(image)
# input images should be grayscale float-32 numpy arrays with values in range [0, 1]
# You can use load_image and write_image for loading/writing to/from paths
from degan import load_image, write_image

# or use your own weights:
model = DEGAN(bin_weights='path/to/binary/weights.h5', deb_weights='path/to/deblur/weights.h5', wat_weights='path/to/unwatermark/weights.h5')

# the following loads the weights and run the inference:
binarized_image = model.binarize(image)
# and now it won't reload the weights but runs the inference on the same model:
binarized_image_2 = model.binarize(another_image)
# or you can force the loading of weights when you need it:
model.load_weights()
# you can also instantiate only certain models:
model = DEGAN(deb_weights=None, wat_weights=None)
model.load_weights()

# similar for deblur and unwatermark
deblurred_image = model.deblur(image)
watermark_removed_image = model.unwatermark(image)

# you can also compute the PSNR metric:
from degan import psnr
psnr_value = psnr(image, binarized_image)

For training custom weights, please refer to the original repository.

CLI

Note: for using as a CLI, consider installing via pipx

• Default weights: degan binarize image.png
• Custom weights: degan binarize image.png --out_dir ./out_dir/ --bin_weights path/to/binary/weights.h5
• Other subcommands: degan deblur, degan unwatermark
• All options: degan --help, degan - --help

Citation

• If this work was useful for you, please cite it as the original authors' publication:

@ARTICLE{Souibgui2020,
author={Mohamed Ali Souibgui and Yousri Kessentini},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
title={DE-GAN: A Conditional Generative Adversarial Network for Document deblurment},
year={2020},
doi={10.1109/TPAMI.2020.3022406}}