GCNet

Single Image Reflection Removal based on GAN with Gradient Constraint

Input real image	Image generated by our method

The sample image is provided by SIR² benchmark dataset.

Requirements

Python

• Pytorch (torch & torchvision)
• numpy
• skimage
• tqdm

Usage

Put input images into images/<your_dataset_name>/input/. Processed images are saved in images/<your_dataset_name>/output/.

If you have ground truth images, put them into images/<your_dataset_name>/gt/. PSNR and SSIM will be calculated. The file name of ground truth images should match with those of input images.

Run python3 demo.py --dataset_name=<your_dataset_name>.

Citation

Please cite this paper if you use this code.

@ARTICLE{abiko2019reflection,
author={R. {Abiko} and M. {Ikehara}},
journal={IEEE Access},
title={Single Image Reflection Removal Based on GAN With Gradient Constraint},
year={2019},
volume={7},
number={},
pages={148790-148799},
keywords={Generative adversarial networks;Training;Generators;Feature extraction;Correlation;Glass;Task analysis;Image restoration;deep learning;reflection removal;image separation;generative adversarial network},
doi={10.1109/ACCESS.2019.2947266},
ISSN={},
month={},}

For further information, please contact: {abiko, ikehara}@tkhm.elec.keio.ac.jp