Shen L, Yue Z, Chen Q, et al. Deep joint rain and haze removal from a single image[C]//2018 24th International Conference on Pattern Recognition (ICPR). IEEE, 2018: 2821-2826.
1. Overview
1.1. Motivation
- Rain streak correspond to high-frequency
- Haze correspond to dark channel
Paper proposed Deep Joint Rain and Haze Removal Network (DJRHR-net)
- Haar wavelet transform
- Extract dark channel as input
1.2. Wavelet Transformation
- LL. background
- HL. vertical
- LH. horizontal
- HH. diagonal
1.3. Simple Rain Removal Network (SRR-net)
Spectrum of an image loses a lot of great properties such as local receptive field, which makes it difficult to use convolutional neural network
Process
Loss function
1.4. DJRHR-net
It is more effective to add the artificial feature directly than the features learned by the deep network.
Process
Loss Function
1.5. Dataset
1.5.1. TrainSetA
- Directly used Deep Detail Network’s
- Without haze veil
- 12 type rain streaks
1.5.2. TrainSetB
- 1449 RGBD from NYU Depth V2
- 12 type rain streaks
2. Experiments
2.1. Training Setup
- Dense Block
- Remove BN and pooling to get better result
2.2. Metric
- PSNR
- SSIM
- NIQE. the lower the better
2.3. Synthetic
2.4. Real
2.5. DenseBlock parameters
