Existing halftoning algorithms usually drop colors and fine details when dithering color images with binary dot patterns, which makes it extremely difficult to recover the original information. To dispense the recovery trouble in future, we propose a novel halftoning technique that dithers a color image into binary halftone with decent restorability to the original input. The key idea is to implicitly embed those previously dropped information into the binary dot patterns. So, the halftone pattern not only serves to reproduce the image tone, maintain the blue-noise randomness, but also represents the color information and fine details. See the examples illustrated below.

1. Requirements:

   • Basic variant infomation: Python 3.7 and Pytorch 1.0.1.
   • Create a virutal environment with satisfied requirements:

     conda env create -f requirement.yaml

2. Training:

   • Place your training set/validation set under dataset/ per the exampled file organization. Or download our [preprocessed full dataset](coming soon).
   • Warm-up stage (optional):

     python train_warm.py --config scripts/invhalf_warm.json

     If this stage skipped, please download the pretrained warm-up weight and place it in checkpoints/, which is required at joint-train stage.
   • Joint-train stage:

     python train.py --config scripts/invhalf_full.json

3. Testing:

   • Download the pretrained weight below and put it under checkpoints/.
   • Place your images in any accesible directory, e.g. test_imgs/.
   • Dither the input images and restore from the generated halftones

     python inference_fast.py --model checkpoints/model_best.pth.tar --data_dir ./test_imgs --save_dir ./result

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

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

@inproceedings{xia-2021-inverthalf,
	author   = {Menghan Xia and Wenbo Hu and Xueting Liu and Tien-Tsin Wong},
	title    = {Deep Halftoning with Reversible Binary Pattern},
	booktitle = {{IEEE/CVF} International Conference on Computer Vision (ICCV)},
	year = {2021}
}