color transfer between high-dynamic-range images
TRANSCRIPT
Color transfer
between high-
dynamic-range
images
H. Hristova, R. Cozot,
O. Le Meur, K. Bouatouch
University of Rennes 1
Rennes, France
Outline
● Introduction
- Main objective
- Contributions
● Extension to the HDR domain of a color
transfer method
● Results and evaluation
● Generalization for state-of-the-art color
transfer methods
● Conclusion
2
Main goal
● Carrying out a color transfer between two HDR images
directly in the HDR domain
Input Reference
3
● Solution: apply color transfer methods to stylize an HDR
image with regards to a reference image
Why do LDR color transfer methods need to
be extended to the HDR domain?
● LDR color spaces
- well predict the color gamut for luminance levels
between zero and the display white point
- uncertain applicability to HDR images
● Color trend above the perfect diffuse white
4
Why do LDR color transfer methods need to
be extended to the HDR domain?
● Assumption: a unique multivariate Gaussian distribution
● HDR domain: to fit the high range of lightness of HDR
images we need to assume mixture of Gaussian
distributions
5
Why do LDR color transfer methods need to
be extended to the HDR domain?
● Lightness - approximated by luminance in the LDR
domain
● HDR domain - distinguish between the absolute
luminance and the lightness (the L channel of CIE Lab)
6
Contributions
● Adaptation of [Hristova et al., 2015] color
transfer method to HDR images
- HDR color spaces
- Modifications of the clustering step and of the
image classification
● Cluster-based local chromatic adaptation
transform
● Generalization for state-of-the-art color transfer
methods
7
8
Extension to HDR images
• Linear search for significant peaks in the image hue histogram
- Colors-based style images: more than one significant color cluster
- Light-based style images: one significant color cluster
Input and
reference
images
Color
space
conversion
Image
classification
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final result
[Hristo
va
et a
l., 2
01
5]
• The number of significant peaks determines the number of clusters
- Colors-based style images: hue histogram
- Light-based style images: luminance histogram
Extension to HDR images
Input and
reference
images
Color
space
conversion
Image
classification
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final result
9
[Hristo
va
et a
l., 2
01
5] LDR
imagesCIE Lab
L channel of
CIE Lab
HDR
imageshdr-CIELab
Log-
luminance
● Dashed line: cubic function of L channel
(CIE Lab)
● Solid line: Michaelis-Menten function by
which we replace the cubic function of L
channel (CIE Lab)
● hdr-CIELab color space [Fairchild et al.,
2004]
[Fairchild et al., 2004]
Mo
dific
ation
sM
od
ifie
d
Extension to HDR images
10
LDR
imagesCIE Lab
L channel of
CIE Lab
L-based
clustering
HDR
imageshdr-CIELab
Log-
luminance
Log-
luminance
clustering
[Hristo
va
et a
l., 2
01
5]
Lo
ga
rith
mic
tra
nsfo
rm
Input and
reference
images
Color
space
conversion
Image
classification
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final result
Mo
difie
dM
od
ific
ation
s
Extension to HDR images
11
Local CAT
Cluster-
based local
CAT
[Hristo
va
et a
l., 2
01
5] LDR
imagesCIE Lab
L channel of
CIE Lab
L-based
clustering
HDR
imageshdr-CIELab
Log-
luminance
Log-
luminance
clustering
Input and
reference
images
Color
space
conversion
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final resultImage
classification
Mo
difie
dM
od
ific
ation
s
Extension to HDR images
12
Ga
ussia
n lo
w-p
ass filt
er
(h)
(m)
(sh)
(sh) (m) (h)
Local CAT
Cluster-
based local
CAT
[Hristo
va
et a
l., 2
01
5] LDR
imagesCIE Lab
L channel of
CIE Lab
L-based
clustering
HDR
imageshdr-CIELab
Log-
luminance
Log-
luminance
clustering
Input and
reference
images
Color
space
conversion
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final resultImage
classification
Mo
difie
dM
od
ific
ation
s
Extension to HDR images
13
Input ReferenceCluster-based local CAT
Local CAT
Cluster-
based local
CAT
[Hristo
va
et a
l., 2
01
5] LDR
imagesCIE Lab
L channel of
CIE Lab
L-based
clustering
HDR
imageshdr-CIELab
Log-
luminance
Log-
luminance
clustering
Input and
reference
images
Color
space
conversion
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final resultImage
classification
Mo
difie
dM
od
ific
ation
s
Extension to HDR images
14
Input ReferenceCluster-based local CAT
Local CAT
Cluster-
based local
CAT
[Hristo
va
et a
l., 2
01
5] LDR
imagesCIE Lab
L channel of
CIE Lab
L-based
clustering
HDR
imageshdr-CIELab
Log-
luminance
Log-
luminance
clustering
Input and
reference
images
Color
space
conversion
Clustering
and
mapping
Color
transfer
Chromatic
adaptation
transform
Final resultImage
classification
Mo
difie
dM
od
ific
ation
s
Objective evaluation of the results
15
● 10 image pairs
● Two tone-mapping operators: [Durand et al., 2002] and
[Reinhard et al., 2002]
● SSIM and Bhattacharya coefficient
Results
16
Input
Reference
Color transfer with CAT
Color transfer without CAT
Color transfer with CAT
Color transfer without CAT
[Hristova et al., 2015] HDR extension
Generalization and results
17[Reinhard et al., 2001] - global method [Tai et al., 2005] - clustering (local
transformations)
Inp
ut
Re
fere
nce
Generalization and results
18[Pitié et al., 2007] - CIE Lab [Pitié et al., 2007] - hdr-CIELab
Inp
ut
Re
fere
nce
Generalization and results
19[Bonneel et al., 2013] - luminance clustering [Bonneel et al., 2013] - log-luminance
clustering
Inp
ut
Re
fere
nce
Conclusion
● Extension of a novel local color transfer
method [Hristova et al., 2015]
- Modifications to CIE Lab -> hdr-CIELab
- Luminance/Lightness -> Log-luminance
● Generalization to state-of-the-art methods
● Future work
- Need for a more precise color
mapping/color transformation between two
HDR images
- Need for better HDR color spaces
20
Thank you for your attention!
21