Zhang H, Sindagi V, Patel V M. Joint transmission map estimation and dehazing using deep networks[J]. arXiv preprint arXiv:1708.00581, 2017.
1. Overview
大多数现有的方法假设constant atmosphere light,包含两个步骤
- 基于prior-based方法估计transmission map
- 近似解计算haze-free image
论文提出multi-task结构
- Relax constant atmosphere light assumption, joint estimate transmission map and de-hazing
- Introduce GAN
- Introduce perceptual loss
1.1. Related Work
- DehazeNet
- Multi-scale Net
1.2. Model
- PReLU
- Generator使用U-Net结构
1.3. Loss Function
- Transmission Map Loss
- Dehazing Loss
Perceptual loss (VGG-16 relu3_1)
1.4. 速度
512x512. 18 fps
1.5. 数据集
α ∈ [0.5, 1.2]. β ∈ [0.4, 1.6]
1.5.1. NYU Depth Dataset
- Training Set. 600 images x 4
- Testing Set. 300 images x 4
- Real Set. 30 images
2. Experiments
2.1. AblationStudy
Adversarial Loss
Perceptual Loss
Euclidean Loss
Transmission Map