This is the repository of paper DeepWSD: Projecting Degradations in Perceptual Space to Wasserstein Distance in Deep Feature Space. Related Quality Assessment results and Optimization results are in `results' folder.
- DeepWSD is a perceptual FR-IQA that performs excellently not only in Quality Assessment Tasks but also in optimization tasks.
- DeepWSD is a complete metric that satisfies the metrics definition and may be used as a perceptual loss.
- DeepWSD is completely free of training and can achieve nice performance on several datasets.
2024/7/24 Our new work has been accepted by IEEE Transactions on Image Processing, which contains the updated DeepWSD plus its deep distribution measure (DeepSKLD and DeepJSD) siblings. Please refer to the new website.
2022/10/3: Find problems on batch form DeepWSD. Now, substitute the original single-image comparison form. Please compare images one by one. The batch form for optimization will be released as soon as possible.
2022/8/29: Updating batch DeepWSD.
- imageio==2.9.0
- matplotlib==3.5.0
- numpy==1.20.1
- Pillow==9.2.0
- POT==0.8.1.0
- torch==1.8.0
- torchvision==0.9.0
- For DeepWSD.py:
python DeepWSD.py --ref images/Lena.jpg --dist images/white.jpg
- For recover.py:
python recover.py --ref_path images/Lena.jpg --pred_path images/white.jpg
- Note 'utils.py' contains supporting function for DeepWSD, make sure they are in same folder.
We thanks a lot for work of 'dingkeyan93' and work of DISTS. DeepWSD is mostly inspired by the insightful idea from him.