Robust Image Enhancement with DRL

Description:

Example of image enhancement with Proximal Policy Optimization (PPO) algorithm, using MIT-Adobe FiveK dataset. The environment is wrapped into OpenAI Gym format.

Dependencies:

• tensorflow >= 2.0.0
• scikit-image >= 0.16.2
• Pillow >= 6.2.1
• tqdm >= 4.36.0

Usages:

1. Download MIT-Adobe FiveK Dataset.
2. Process raw images and retouced images to JPEG format with quality 100 and color space sRGB by Adobe Lightroom.
3. Resize images so that the maximal side consists of 512 pixels.
4. Split dataset into training and validation, and generate train_pairs and valid_pairs files where each line consists of {raw_image_path}\t{retouched_image_path}.
5. Run python train.py --mode=train with specific args.

Citing

@book{deepRL-2020,
 title={Deep Reinforcement Learning: Fundamentals, Research, and Applications},
 editor={Hao Dong, Zihan Ding, Shanghang Zhang},
 author={Hao Dong, Zihan Ding, Shanghang Zhang, Hang Yuan, Hongming Zhang, Jingqing Zhang, Yanhua Huang, Tianyang Yu, Huaqing Zhang, Ruitong Huang},
 publisher={Springer Nature},
 note={\url{http://www.deepreinforcementlearningbook.org}},
 year={2020}
}