Single Image Haze Removal

Using Dark Channel Prior

Kaiming He
Jian Sun
Xiaoou Tang
The Chinese University of Hong Kong
Microsoft Research Asia
The Chinese University of Hong Kong
Hazy Images
Low visibility
Faint colors
Goals of Haze Removal
Scene restoration
Depth estimation
depth
Haze Imaging Model
) 1 ( t t + = A J I
Transmission Scene radiance Hazy image
Atmospheric light
Transmission
Haze Imaging Model
) 1 ( t t + = A J I
Depth
t d ln | =
Ambiguity in Haze Removal
input
scene
radiance
depth
.
Previous Works
Using additional information
Polarization filter [Shwartz et al., CVPR06]
Multiple images [Narasimhan & Nayar, CVPR00]
Known 3D model [Kopf et al., Siggraph Asia08]
User-assistance [Narasimhan & Nayar, CPMCV03]
Previous Works
Single image
Maximize local contrast [Tan, CVPR 08]
Previous Works
Single image
Maximize local contrast [Tan, CVPR 08]
Previous Works
Single image
Maximize local contrast [Tan, CVPR 08]
Independent Component Analysis [Fattal, Siggraph 08]
Previous Works
Single image
Maximize local contrast [Tan, CVPR 08]
Independent Component Analysis [Fattal, Siggraph 08]
Priors in Computer Vision
Smoothness prior
Sparseness prior
Exemplar-based prior
Ill-posed
problem
well-posed
problem
prior
Dark Channel Prior
Dark Channel
min (rgb, local patch)
Dark Channel
min (rgb, local patch)
min (r, g, b)
min (r, g, b)
Dark Channel
min (rgb, local patch)
min (r, g, b)
min (local patch) = min filter
15 x15
darkest
dark channel
Dark Channel
min (rgb, local patch)
min (r, g, b)
min (local patch) = min filter
dark channel
)) ( J min ( min ) ( J
c
} b g, r, { c ) ( y
y x
x e O e
=
dark
J
c
: color channel of J
J
dark
: dark channel of J
Dark Channel
min (rgb, local patch)
min (r, g, b)
min (local patch) = min filter
dark channel
) J min ( min J
c
c O
=
dark
J
c
: color channel of J
J
dark
: dark channel of J
A Surprising Observation
Haze-free
A Surprising Observation
Haze-free
A Surprising Observation
Haze-free
A Surprising Observation
Haze-free
A Surprising Observation
Haze-free
A Surprising Observation
Haze-free
A Surprising Observation
0
0.2
0.4
0.6
0.8
1
0 64 128 192 256
Prob.
Pixel intensity of dark channels
86% pixels
in [0, 16]
5,000 haze-free
images
Dark Channel Prior
For outdoor haze-free images
0 ) J min ( min
c
c

O
What makes it dark?
Black object
Colorful object
Shadow
Dark Channel of Hazy Image
The dark channel is no longer dark.
hazy image dark channel
Transmission Estimation
) 1 ( t t + = A J I
Haze imaging model
t t + = 1
A
J
A
I
c
c
c
c
Normalize
t t +
)
`

=
O O
1 )
A
J
min ( min )
A
I
min ( min
c
c
c
c
c
c
Compute dark channel
t t +
)
`

=
O O
1 )
A
J
min ( min )
A
I
min ( min
c
c
c
c
c
c
Compute dark channel
Transmission Estimation
0 ) J min ( min
c
c

O
Dark Channel Prior
0
Compute dark channel
t t +
)
`

=
O O
1 )
A
J
min ( min )
A
I
min ( min
c
c
c
c
c
c
Transmission Estimation
Estimate transmission
)
A
I
min ( min 1
c
c
c O
= t
Transmission Estimation
input
I t
estimated
)
A
I
min ( min 1
c
c
c O
= t
Estimate transmission
Transmission Optimization
) 1 ( t t + = A J I Haze imaging model
) 1 ( o o + = B F I Matting model
+
Refined
transmission
+
tri-map
o
Transmission Optimization
L - matting Laplacian [Levin et al., CVPR 06]
Constraint - soft, dense (matting - hard, sparse)
Lt t t t t
T
2
~
) ( + = E
Data term Smoothness term
Transmission Optimization
before optimization
Transmission Optimization
after optimization
hazy image dark channel
brightest pixels
Atmospheric Light Estimation
brightest pixel
A: most hazy
Scene Radiance Restoration
) 1 ( t t + = A J I
Scene radiance Transmission Hazy image
Atmospheric
light
Results
input
Results
recovered image
Results
depth
Results
input
Results
recovered image
Results
depth
Results
input
Results
recovered image
Results
depth
Comparisons
input [Fattal Siggraph 08]
Comparisons
input our result
Comparisons
input [Tan, CVPR 08]
Comparisons
input our result
input our result
[Kopf et al, Siggraph Asia 08]
Comparisons
Results: De-focus
recovered scene radiance
input
depth
input
depth
Results: De-focus
de-focus
Results: Video
output
input
Results: Video
output
input
input
our result
transmission
Inherently white or grayish objects
Limitations
Haze imaging model is invalid
e.g. non-constant A
input our result
Limitations
Summary
Dark channel prior
A natural phenomenon
Very simple but effective
Put a bad image to good use
Thank you