HAZE REMOVALPhil Morley
The Problem Fog, Haze, or Smog Want a clear image Weather could be common in areas
The MethodOutlined in paper: Single Haze Removal Using Dark Channel Prior by Kaimin He, Jian Sun, and Xiaoou Tang
What is haze?
I(x) = J(x)t(x) + A(1 − t(x))
I(x): ImageJ(x): Scene RadianceA: Atmospheric Lightt(x): Transmittance
Dark Channel Prior Objects of interest have low values in at
least one color channelGreen leafCar ShadowDark building
Haze has a high pixel intensity
Compute Atmospheric Light
High values in Dark Channel Take top 0.1% Pull Values from original image Average
I(x) = J(x)t(x) + A(1 − t(x))
Estimating TransmissionShuffling the Haze Equation and taking min’s gives you:
)))((min(min1)(~)( A
yIxtc
c
xyc
Which is simply:
][ )(1)(~AyIxtdark
Refine Transmission with Soft Matting Estimated Transmission is blocky
Want to take into account fine detail Haze Equation is alpha matting Therefore can use Soft Matting as
shown by Levin et al.
Soft Matting
)~()~()( ttttttt TT LE
Minimize Cost Function:
Has Closed Form Solution:
tt ~)( UL
wjik
kjk
kk
T
kij
k
Uwiw
jiL),|(
3
1
)))()()(1(1),(
II
U3 = 3x3 Identityλ = 0.0001
Things to improve Performance
Processing TimeMemory Allocation
Settings
Things To Expand Depth Map
From Transmittance3D Model
Image EnhancementHistogram Equalization
Current Results