Single Image Haze Removal Using Dark Channel Prior Professor : 王聖智 教授 Student : 戴玉書 CVPR Best Paper AwardBest Paper Award Kaiming HeKaiming He, Dept. of Inf. Eng., Chinese Univ. of Hong Kong, Hong Kong, China Jian Sun Xiaoou TangXiaoou Tang, Dept. of Inf. Eng., Chinese Univ. of Hong Kong, Hong Kong, China
Outline Background What is the Dark Channel Prior? How to estimate ? How to estimate atmospheric light? Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
Background Observed intensity Scene radiance The global atmospheric light The medium transmission,
Outline Background What is the Dark Channel Prior? How to estimate light? How to estimate atmospheric light? Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
Dark Channel Prior Observation on haze-free outdoor images: Observation on haze-free outdoor images: In most of the non-sky patches, at least one color channel has very low intensity at some pixels In most of the non-sky patches, at least one color channel has very low intensity at some pixels
Mainly due to three factors Shadows Colorful of objects or surfaces Dark objects
haze-free imageThe dark channel of haze-free image
Statistics of the dark channels Except for the sky region, the intensity of is low and tends to be zero Except for the sky region, the intensity of is low and tends to be zero
Visually, the intensity of the dark channel is rough approximation of the thickness of the haze haze image The dark channel of haze image
Outline Background What is the Dark Channel Prior? To estimate of light To estimate of atmospheric light Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
To estimate of To estimate of atmospheric light Pick the top 0.1% brightest pixels in the dark channel
Outline Background What is the Dark Channel Prior? How to estimate light? How to estimate atmospheric light? Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
Estimating the transmission
Soft Matting Image matting equation:
L ij : Minimize the following cost function: A. Levin, D. Lischinski, and Y. Weiss. A closed form solution to natural image matting. CVPR, 1:61–68, , 5, 7
Outline Background What is the Dark Channel Prior? How to estimate light? How to estimate atmospheric light? Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
(t 0 =0.1)
Outline Background What is the Dark Channel Prior? How to estimate light? How to estimate atmospheric light? Estimating the transmission t(x) & Soft Matting Recovering the Scene Radiance Result
Result The patch size is set to 15x15 Soft matting: Preconditioned Conjugate Gradient (PCG) algorithm Local min operator using Marcel van Herk ’ s fast algorithm fast algorithm
► Tan's result ► Fattal's result ► Dark channel
► Tan's result ► Fattal's result ► Dark channel
► Kopf et al's result ► Dark channel