MenghanXia / InverseHalftoning

An inverse halftoning model, which is applied to recover the 8-bit grayscale image from 1-bit error diffused halftones. [ACCV 2018]

Deep Inverse Halftoning

We run this code under TensorFlow 1.6 on Ubuntu16.04 with python pakage IPL installed.

Network Architecture

TensorFlow Implementation of our paper "Deep Inverse Halftoning via Progressively Residual Learning" accepted to ACCV 2018.

Results

Preparation

• You can apply existing halftone algorithms (e.g., Foyd-Steinberg Error diffusion as did in our experiments) on grayscale images to generate binary halftone version of them, then you obtain training pairs <halftone, grayscale>.
• The patch size is set to 256x256 in the model.py (you may change it to any other size as you like).
• Download the pretrained VGG19 model in here.

Train

• Set hyperparameters in main.py.
• Start training.

python3 main.py --mode train --train_dir 'training_image_dir' --val_dir 'val_image_dir'

Test

• Download the pretrained model and place it in the folder "./checkpoints".
• Start evaluation and the result will be saved in the folder "./output".

python3 main.py --mode test --test_dir 'testing_image_dir'

Copyright and License

You are granted with the license for both academic and commercial usages.

Citation

If any part of our paper and code is helpful to your work, please generously cite with:

@inproceedings{XiaW18,
  author    = {Menghan Xia and Tien-Tsin Wong},
  title     = {Deep Inverse Halftoning via Progressively Residual Learning},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2018}
}