Zhang H, Patel V M. Density-aware single image de-raining using a multi-stream dense network[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 695-704.
1. Overview
1.1. Motivation
- Non-uniform rain densities
- Existing method do not consider the density of rain drops (lead to over de-rain or under de-rain)
[6]. Deep Detail Network. CVPR 2017
[33]. Deep Joint Rain Detection and Removal. CVPR 2017
In this paper, it proposed DID-MDN
- joint (two stage) rain density estimation and de-raining (guided by the estimated rain-density label)
- multi-stream (different scale feature)
- use residual to represent rain-density feature
- create a dataset containing rain-density label (heavy, medium, light)
1.2. Related Work
- video-based
- prior-based (over-smooth details)
- Multi-scale feature (U-Net, FCN, skip-connection)
1.3. Dataset
- 12,000 for training
- 1, 200 for testing
- use PS to get different level rain density (noise level [5%, 35%], [35%, 65%], [65%, 95%])
- link
2. Architecture
2.1. Residual-Aware Rain-Density Classifier
single network may not be sufficient enough to learn different rain-densities occurring in practice.
- used for guiding the de-raining process
- residual can better represent the rain feature
- observe that fine-tune pre-trained model is not an efficient solution.
high-level feature focus on localizing the discriminative object (not small rain-streak), so it’s not good for density classify.
2.1.1. step
- estimate residual (residual feature extraction network)
- train classifier with residual (classifier)
2.1.2. Loss
- firstly train residual network
- then train classifier
- finally joint optimized
2.1.3. Classifier (low, medium, high)
2.2. Multi-Stream Dense Network
- Concat multi-stream feature and density label
- Refinement
2.3. Multi-Stream
- smaller rain-streak can be capture by small-scale feature
- longer rain-streak can be capture by larger-scale feature
For each stream
- six dense blocks + six transition layer
- short path better for convergence
- Dense1. 3-down + 3-up (7x7)
- Dense2. 2-down + 2-no + 2-up (5x5)
- Dense3. 1-down + 4-no + 1-up (3x3)
2.4. Total Loss
3. Experiments
4. Dataset
- Train1. 12,000
- Test1. 1,200
- Test2. 1,000 (from Deep Detail Network)
4.1. Detail
- random crop, horizontal flip
- batch size 1
- λF = 1
4.2. Ablation Study
VGG vs Residual classifier
Modules
- Single. single stream without label fusion
- Yang-Multi. multi-stream (dilated)
- Multi-no-label. multi stream without label fusion
- DID-MDN. multi-stream with label fusion
- Over de-rain with blur. Single and Yang-Multi
- Leave some rain-streak. Muti-no-label
4.3. Comparison
- Synthetic
[33, 41] leave some rain-streak.
[6] remove some details.
Real
- long-thin rain-streak
- heavy rain
- small round rain
- medium reain