Professional Documents
Culture Documents
Algorithms
Jason Su
Estimation Techniques
Gray world
“Simplest color balance” – histogram
normalization with saturation [3]
Robust auto white-balance [2]
Sensor correlation illuminant estimation [6][7]
Correction Techniques
von Kries’s method, the chromatic adaptation
transform (CAT) [4]
D65 desired canonical illuminant/white point
d / s 0 0
1
M MA 0 d / s 0 M A
0 0 d / s
s X WS d X WD
M Y M Y
s A WS d A WD
s Z WS d Z WD
Correction Techniques
Choice of MA is the subject of research as well [1]
Can be optimized to various criteria, such as
mean ΔE or statistical distribution testing [5]
von Kries
Bradford
Sharp – based on sharpened sensors, min. XYZ errors
CMCCAT2000 – fitted from all available color data sets
CAT02 – optimized for minimizing CIELAB differences
Gray World
Estimate the color cast using the mean RGB value
across the whole image
Convert to xy-chromaticity assuming image
already lives in the sRGB color space
Apply a chromatic adaptation transform to reach
D65
Go back to RGB for display
Simplest Color Balance
Used in Adobe Photoshop’s auto levels command
Determine the histogram of values for each RGB
channel
Cut off the outlying values by saturating a certain
percentage of the pixels (adjustable, 1% used)
Scale the saturated histogram to span the full 0-
255 range
Robust Auto White-Balance
Convert to YUV color space
Estimating illuminant by finding pixels that are
similar to gray within a threshold:
(|U|+|V|)/Y < T = 0.3
Gain the R or B channels accordingly
Loop until convergence or gray has been
achieved
Alternatively apply a CAT instead of R/B gain
Sensor Correlation
Pre-process to eliminate outliers
Compute the color gamut of the image as the
convex hull of its (R,B) channels
Correlate image gamut with a priori sensor
response gamuts under different temperature
illuminants to estimate color cast
Simulated sensor response gamuts with ISET
13 equally spaced from 188 to 400 mireds
Correct with a CAT
Original
Gray World – Bradford
Simplest Color Balance
Histograms
Robust Auto White-Balance
Plots
Sensor Correlation – Bradford
Sensor Correlation – von Kries
Sensor Correlation – Sharp
Sensor Correlation – CMCCAT2000
Sensor Correlation – CAT02
Original
Gray World – Bradford
Simplest Color Balance
Robust Auto White-Balance
Sensor Correlation – Bradford
Original
Gray World – Bradford
Simplest Color Balance
Robust Auto White-Balance
Sensor Correlation – Bradford
Original
Original – ISET Simulation of 3268K
Gray World – Bradford
Simplest Color Balance
Robust Auto White-Balance
Sensor Correlation – 3034.9K
Original
Gray World – Bradford
Simplest Color Balance
Robust Auto White-Balance
Sensor Correlation – Bradford
Thoughts & Issues
Choice of CAT is not clear, perhaps general linear
transform?
Simplest color balance tends to exaggerate
highlights and shadows
Robust AWB is slower than other algorithms
when implemented in software
Sensor correlation requires knowledge of the
particular hardware
A winner?
References
[1] S. Bianco and R. Schettini, "Two New von Kries Based Chromatic
Adaptation Transforms Found by Numerical Optimization."
[2] J. Huo, Y. Chang, J. Wang, and X. Wei, "Robust Automatic White Balance
Algorithm using Gray Color Points in Images," 2006, pp. 541-546.
[3] N. Limare, “IPOL Algorithm: Simplest Color Balance,”
http://www.ipol.im/pub/algo/lmps_simplest_color_balance/
[4] B. Lindbloom, “Chromatic Adaptation,”
http://www.brucelindbloom.com/index.html?Eqn_ChromAdapt.html
[5] S. Susstrunk and G.D. Finlayson, "Evaluating Chromatic Adaptation
Transform Performance."
[6] S. Tominaga, A. Ishida, and B.A. Wandell, "Illuminant Estimation of Natural
Scene Using the Sensor Correlation Method," Image (Rochester, N.Y.), vol.
4421, 2002, pp. 918-921.
[7] S. Tominaga, S. Member, and B.A. Wandell, "Natural Scene-Illuminant
Estimation Using the Sensor Correlation," Psychology, vol. 90, 2002.
Questions?