Studies in Skin Colors
Download
Report
Transcript Studies in Skin Colors
Simon Fraser University
Computational Vision Lab
Lilong Shi, Brian Funt and Tim Lee
Studies of factors affecting skin colour
Simple and linear model of skin
Modelling Skin appearance under lights
Applications:
Estimate melanin and hemoglobin concentrations
Correct imaged skin tones for lighting conditions
Skin tone
correction
Tone
correction
Preserve
melanin
Melanin/Hemoglobin separation
Appearance of human skin determined by
Biological factors
▪ pigmentation, blood microcirculation, roughness, etc..
Viewing conditions
▪ Inducing lights
Acquisition devices
▪ Cones in retina, RGB sensors of CCD digital cameras
Two-layered Skin Model [2]
Epidermis Layer: melanin absorbance
Dermis Layer: hemoglobin absorbance
A layer has properties of an optical filter
Various skin colour <= melanin + hemoglobin
Genetic: Race
Temporary:
▪ Exposure to UV
▪ Hot bath
Mixture varying by 2 independent factors
Analyse melanin and hemoglobin factors
Estimate melanin and hemoglobin concentration
Independent Component Analysis (ICA)
– Statistical technique for revealing “hidden” factors
– To “unmix” or “separate” signals composed of
multiple sources
– Independent and linear mixing
– Related to Eigen-vector analysis
Original Source Signals
Mixing
Observed Signals
v1
70%
s1
0%
30%
20%
s2
v2
80%
100%
v3
s
×
A
=
v
Melanin
Melanin
Hemoglobin
Hemoglobin
Skin
samples
Typical skin spectrum
Visible wavelength 400nm – 700nm
Extract skin bases from observed spectrum by ICA
ICA
(left) 33 skin spectrum
after normalization;
(right) two independent
basis spectrum – the
melanin and
hemoglobin, and the
spectrum of
chromophores other
than melanin and
hemoglobin pigments.
Arbitrary skin spectrum can be approximated
mσm hσh c
m , h
are variables
constru
Human vision
▪ 3 types of Photoreceptors
L, M and S Cones
Digital Cameras
▪ 3 sensors
Red, Green, and Blue
Reflectance spectrum recorded by 3 sensors
=> three values (R, G, B) for a skin colour
pixel mσ m hσ h c
Given a pixel from
skin, compute m , h
by projecting
log(R,G,B) onto σ m , σ h
Possible skin colours
lie within plane
mσm hσh c
m
Melanin
Image
h
Hemoglobin
Image
Input Image [3]
- Inverse melanin concentration
- Inverse hemoglobin concentration
Skin appearance greatly affected by lights
Reveal true skin colour by removing illum.
Common lights blackbody radiation
e.g. tungsten/halogen lamps, sunrise/sunset, etc
Varying colour temperature T
▪ Redish -> white -> bluish
Colour: illumination times reflectance
In log space, multiplication => addition:
Π( m , h , ) mσ m hσ h τω c
Illum. basis
In practice
Π( m , ) mσm τω
Drop hemoglobin basis
▪ Small angle between Illum and hemoglobin axes
Ignore brightness
Skin colour varying by T and
m
384 real skin reflectances times
67 real light sources
=> 25728 samples
Π( m , ) mσm τω
Skin tone correction example (UOPB DB [4])
16 different
illumination +
camera settings
Tone
correction
Preserve
melanin
20
• Skin tone correction example (UOPB DB [4])
Skin colour modelling:
Melanin and Hemoglobin concentration
Linear model in logarithm space
Estimation by Independent Component Analysis
Skin appearance + Light modelling:
Estimates light source
Preserves skin colour by melanin value
Applied to digital images from CCD cameras
[1] Shi, L., and Funt, B., "Skin Colour Imaging That Is Insensitive to
Lighting," Proc. AIC (Association Internationale de la Couleur) Conference
on Colour Effects & Affects, Stockholm, June 2008
[2] Angelopoulou, E., Molana, R., and Daniilidis, K. “Multispectral skin
color modeling,” In IEEE Conf. on Computer Vision and Pattern
Recognition, volume 2, pages 635-642, Kauai, Hawaii, Dec. 2001.
[3] Shimizu, H., Uetsuki, K., Tsumura, N., and Miyake, Y. Analyzing the
effect of cosmetic essence by independent component analysis for skin
color images. In 3rd Int. Conf. on Multispectral Color Science, pages 65-68,
Joensuu, Finland, June 2001.
[4] Martinkauppi, B. “Face color under varying illumination-analysis and
applications,” Ph.D. Dissertation, University of Oulu, 2002.