Swami K, Das S K. Candy: Conditional adversarial networks based fully end-to-end system for single image haze removal[J]. arXiv preprint arXiv:1801.02892, 2018.
1. Overview
现有的去雾方法关注intermediate parameter (transmission map)，并没有将haze-free image quality考虑到optimization framework中。而intermediate parameter的估计误差会进一步影响haze-free image的质量。
因此，论文提出CANDY (Conditional Adversarial Networks based Dehazing of hazY images)结构
- First work of end-to-end de-hazing to generate haze-free image
- First work of introducing GAN for de-hazing
1.1. 速度
On GPU.
- 256x256. 35ms
- 1024x1024. 53ms
- Model size 3MB
1.2. Related Work
估计intermediate parameter，没有将image quality考虑到optimization framework中。
- DehazeNet
- Multi-scale Net
1.3. Model
1.3.1. Generator
- 6 Conv + 6 Deconv. 3x3 kernel size, 64 channels, 1 stride, 1 padding
- Down-Sampling会导致图片特征丢失
- PReLU
1.3.2. Discriminator
- 7 Conv. 3x3, 2s, 1p, double channel
- Leaky ReLU. λ=0.2
1.4. Loss Function
- Content Loss
- Feature Reconstruction Loss
使用VGGNet提取
- 9(relu2_2)
- 16(relu3_3)
- 23(relu4_3)
1.5. 数据集
- Make3D Depth
- BSDS500
- MeddleBury
- NYU Depth只包含indoor images
- 论文使用CVPR 2015的single image depth estimation方法估计图片的depth
- α = [k, k, k], k ∈ [0.7, 1], β ∈ [0.5, 1.5]
1.5.1. Training Set
- 700 x 3 images. 500 from Make3D, 200 from BSDS500
1.5.2. Validation Set
- 40 images
1.5.3. Testing Set
- Test-Synthetic-A. 90 from Make3D and BSDS500
- Test-Synthetic-B. 23 from Middlebury
- Test-Real-500
2. Experiments
2.1. Baseline
- GEN. L2 loss + 9th L2 loss
- CANDY_L1_9P. 500 iteration GEN initialized
- …
2.2. 模型选择
- Smooth L1比L2更有效，并且能够稳定GAN训练
- Lower Layers feature reconstruction的结果更好。可能是因为higher layer preserve spatial structure，而忽略了texture and color
最终选择CANDY_L1_9P。
2.3. 实验结果
2.4. Night Hazy Image
Although train on daytime hazy image, it can work on night hazy image.
