Li X, Wu J, Lin Z, et al. Recurrent squeeze-and-excitation context aggregation net for single image deraining[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 254-269.
1. Overview
In this paper, it proposed RESCAN
- dilated convolutional
- SE block to achive different alpha-values to various rain streaks layers
- multi-stage with RNN. useful information for rain removal in previous stages can guide the learning in later stages
2. Methods
2.1. Rain Model
- A. global atmospheric light
- α_0. scene transmission
- α_i. brightness of a rain streak layer or a haze layer
2.2. Architecture
- depth = 6
- Dilation of L1 to L3 (1, 2, 4)
- no BN. rain streaks in different layers have different distributions, and remove 40% memory
2.3. Recurrent
2.3.1. Recurrent Version
- ConvRNN
ConvGRU
ConvLSTM
2.4. Prediction
2.4.1. Additive Prediction
2.4.2. Full Prediction
3. Experiments
3.1. Details
- patch 64x64
3.2. Ablation Study
3.3. Comparison
