Cai B, Xu X, Jia K, et al. Dehazenet: An end-to-end system for single image haze removal[J]. IEEE Transactions on Image Processing, 2016, 25(11): 5187-5198.
1. Overview
论文提出DehazeNet结构
- Maxout unit to generate almost all haze-relevant feature
- Bilateral rectified linear unit (BReLU)
1.1. 模型结构
- Local Extremum
假设medium transmission是局部常量。
1.2. BReLU
1.3. Maxout
Maxout activation functions can be considered as piece-wise linear approximations to arbitrary convex functions.
1.4. Haze-Relevant Feature
- Dark Channel
Haze-free patches中至少有一个channel中的一些像素值非常低,接近0. 因此,dark channel feature与haze amount高相关,能够用来估计medium transmission.
- Maximum Contrast
Haze transmission会减小对比度。因此,contrast与medium transmission高相关。
Color Attenuation
Prior:Hazy导致saturation下降,brightness上升。Color attenuation feature:
与depth成正比,能够用于transmission estimation.
- Hue Disparity
Original image与semi-inverse image之间的hue disparity能够用于检测haze.
Semi-inverse Image