This is the code (in PyTorch) for our paper Single Image Colorization via Modified CycleGAN,accepted in ICIP 2019, which allows using unpaired images for training and reasonably predict corresponding color distribute of grayscale image in RGB color space.
Note: The pkl-weight in the dir /checkpoints
corrupted during the upload. I’m sorry I didn’t check it in time after uploading.
Linux
Python 3
CPU or NVIDIA GPU + CUDA CuDNN
The color domain data in the paper is randomly selected from the PASCAL VOC, and grayscaled color domain data to gray domain data. You can build your own dataset by setting up the following directory structure:
├── datasets
| ├── src_data # gray
| | ├── train
| | ├── test
| ├── tgt_data # color
| | ├── train
| | ├── test
- Training
python colorization.py
- Testing
python test.py
If you find the code useful, please cite our paper:
@INPROCEEDINGS{8803677,
author={Xiao, Yuxuan and Jiang, Aiwen and Liu, Changhong and Wang, Mingwen},
booktitle={2019 IEEE International Conference on Image Processing (ICIP)},
title={Single Image Colorization Via Modified Cyclegan},
year={2019},
volume={},
number={},
pages={3247-3251},
doi={10.1109/ICIP.2019.8803677}}