Keyword [AOD-Net] [MS-SSIM] [SSIM]
Liu Y, Zhao G. PAD-Net: A Perception-Aided Single Image Dehazing Network[J]. arXiv preprint arXiv:1805.03146, 2018.
1. Overview
1.1. Motivation
- L2 norm implicitly assumes a white Gaussian noise, which is an oversimplified case that is not valid in general dehazing cases
- L2 treats the impact of noise independently to the local characteristics
In this paper, it replaced the L2 loss with perceptually derived loss function
- SSIM
- MS-SSIM
- prposed Perception-Aided SIngle Image Dehazing Network (PAD-Net). AOD-Net + Loss Function
2. Loss Function
2.1. L2 Loss
2.2. L1 Loss
The derivative of L1 loss is not defined at 0. If the error is 0, do not need to update the weight. So use the convention that sign(0) = 0
2.3. SSIM
- μ_x. view as estimates of the luminance of x
- σ_x. contrast of x
- μ_{xy}. structural similarity
- l(p). measure the comparisons of the luminance
- cs(p). combination of contrast and structure similarity
- p. the pixel
- involved the μ and σ of the Gaussian filter on the pixel p
- p’. center pixel of patch P
2.3.1. Derivation
2.4. MS-SSIM
- smaller σ_G loses the ability to preserve the local structure and reintroduce splotchy in flat region
- larger σ_G tends to keep the noises in the proximity of edges
- Multi-scale σ_G
- α = β_j = 1
- j = 1, …, M
3. Experiments
3.1. System
MS-SSIM + L2
MS-SSIM + L1
3.2. Dataset
- ITS. indoor training set
- OTS. outdoor training set
- random select 10,000 images from ITS + OTS (2,790 IST + 7,210 OST)
- random select 1,000 non-overlapping set as validation set
- test on SOTS (500 indoor + 500 outdoor)
3.3. Directly Train
3.4. Fine-tuing on AOD-Net
