Li R, Cheong L F, Tan R T. Single Image Deraining using Scale-Aware Multi-Stage Recurrent Network[J]. arXiv preprint arXiv:1712.06830, 2017.
1. Overview
现实生活中rain的两个显著特点
- Rain streaks of various sizes and directions can overlap each other
- Veiling effect
因此，论文提出Scale-aware Multi-stage CNN
- parallel sub-network to deal with different rain streaks
- veil module to deal with veiling effect
- multi-stage to deal with rain streaks accumulation
(DenseNet能够提高效果)
1.1. Related Work
- DetailsNet
JORDER
Have not been subject to the full force of the tropical heavy rain.
- Have not been tested where the scenes contain a range of depths.
2. Model
2.1. Rain Model
2.2. Framework
DenseNet结构去掉transition layer，不使用down-sampling.
- Veil Module
训练集包含不同的A值。测试阶段将最亮的pixel设为A.
2.3. Loss Function
3. Experiments
3.1. 数据集
- BSD300
- rain size (area). small (0, 60], middle (60, 300], large (300, 600).
- 3300 rain images containing 11 rain streak orientation.
- NYU (depth information)
- Rain12
- 12 synthetic rain image with one type streak.
- Rain12S
- extension of Rain12. various sizes and densities of streak.
- Rain100-COCO
- rander different-sized streak on 100 images from COCO.
- Rain12-Veil
- rander streak and atmosphere veils. 12 images from BSD300.
3.2. 实验结果
3.3. Ablation Study
