Li J, Skinner K A, Eustice R M, et al. WaterGAN: unsupervised generative network to enable real-time color correction of monocular underwater images[J]. IEEE Robotics and Automation Letters, 2018, 3(1): 387-394.
1. Overview
论文
- 基于SimGAN的思想,提出WaterGAN模型。通过un-supervised方法利用WaterGAN将in-air image渲染成underwater image
- 基于渲染得到的underwater image,通过supervised方法利用Restoration Network将underwater image恢复成in-air image
1.1. Related Work
- CNN
- SimGAN
- RenderGAN
1.2. 模型
1.2.1. WaterGAN
输入大小:48x64
- Attenuation
η:wavelength-dependent attenuation coefficient estimated by network
r_c:range from camera to scene
λ:color channel
约束η≥0:确保颜色衰减而不增强
- Scattering
- 输入:48x64 depth map + 100 noise (project, reshape, concat)
- 计算:3个并行CNN
- 输出:48x64x3 mask
- Camera Model
a, b, c, k estimated by network.
约束条件
1.2.2. Restoration Network
使用segNet中的non-parametric upsampling layer (uses the index information from corresponding max-pooling layers).
Depth Estimation Network
- 输入:56x56x3 downsampled
- 输出:56x56x1
- Loss:L2
Color Restoration Network
- 输入:480x480x1 upsampled and then padded to 512x512x1
- 下采样:128x128 AvgPooling
- Core Component
- 上采样:512x512 DeConv (initialized by bilinear interpolation)
- Loss:L2
1.3. 数据集
- Synthetic
4 Kinect dataset (B3DO, UW RGB-D Object, NYU Depth, Microsoft 7-scenes). Total 15000 RGB-D images (12000 training, 3000 validation). - MHL (University of Michigan’s Marine Hydrodynamics Laboratory)
7000 underwater images.
- Port Royal
6500 underwater images, maximum depth 1.5m. - Lizard Island
6083 underimages, maximum depth 2.0m.
1.4. Metric
- Color Accuracy
- Color Consistenct
1.5. 实验结果
MHL
Validation
- Skip Connection