See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3920157

Optimum color spaces for skin detection

Conference Paper · February 2001

DOI: 10.1109/ICIP.2001.958968 · Source: IEEE Xplore

CITATIONS READS

181 246

3 authors, including:

Luis Torres
UPN
43 PUBLICATIONS   954 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Luis Torres on 16 April 2019.

The user has requested enhancement of the downloaded file.

 OPTIMUM COLOR SPACES FOR SKIN DETECTION

Albert0 Albiolt, Luis Torres$, Edward J. Delpfj

Politechnic University of Valencia, Spaint

Politechnic University of Catalonia, Spain$
Purdue University, USA§

ABSTRACT have the same performance since the separability of the skin
The objective of this paper is to show that for every color and no skin classes is independent of the color space cho-
space there exists an optimum skin detector scheme such sen. That is, the color space does not have any influence,
that the performance of all these skin detectors schemes is provided that the optimum skin detector for that color space
the same. To that end, a theoretical proof is provided and is used. A mathematical proof is given in section 3 and ex-
experiments are presented which show that the separability amples are provided to verify this statement in section 4.
of the skin and no skin classes is independent of the color
space chosen. 2. THEORETICAL BACKGROUND

In this section some definitions are provided which are use-

1. INTRODUCTION
ful to proof our statement.
Skin detection can be defined as the process of selecting Definition: Let Z be an image, p a pixel of the image
which pixels of a given image correspond to human skin.
Z and x p a vector containing the values of the pixel p in
the color space C. A function D ( x p )is called a skin color
Skin detection is useful in, for example, face detection and
face tracking for security and video indexing applications, detector in C if
model-based video coding, etc. [l,2,3]. There are some dif- 1 if x p is considered skin
ficulties when detecting skin pixels. Skin color is affected
by ambient light which is unknown in many situations; dif-
D(XP) = {0 if x p is considered no skin (1)

Definition: Given a set of images Ifl and a skin color

ferent cameras produce different colors, even from the same
detector D (xP) ,the performance of D (x p )can be evaluated
person, under the same illumination conditions; and finally,
with the following two parameters:
skin colors change from person to person. Several color
spaces have been proposed in the literature for skin detec-
tion applications. YCbCr has been widely used since the
skin pixels form a compact cluster in the Cb-Cr plane. As
YCbCr is also used in video coding and then no transcoding where PD is the detection rate, PFAis the false alarm rate
is needed, this color space has been used in skin detection and S and N are the subsets of skin and non-skin pixels in
applications where the video sequence is compressed [ 1,2]. Ifl with populations Ps and PN respectively.
The HSV color space has been proposed in [4,5] because it Definition: A skin color detector D ( x p )is said to be
is more related to human color perception. In [6] two com- optimum when it has the highest detection rate PD for a
ponents of the normalized RGB color space (rg) have been given fixed false alarm PFA.
proposed to minimize luminance dependence. And finally
CIE Lu*v* has been used in [7]. However, it is still not 3. STATEMENT
clear which is the color space where the skin detection per-
formance is the best. In addition, it is said many times in Let D ( x p )be an optimum skin detector defined in C and
the literature that the use of a specific color space improves T ( x p )be an invertible function that transforms the pixels
the performance of the skin detectors, [6,81 which is not values x p from the color space C into a vector xb in a dif-
exactly true. The objective of this paper is to show that for ferent a color space C', then there exists another optimum
all color spaces their corresponding optimum skin detectors skin detector D'(xk) in C', with the same detection rate and
This work was partially supported by the grants TIC 98-0442, TIC 98-
false alarm rate, given by:
0335 of the Spanish Govement and Program Hispano-Norteamericano
de Cooperaci6n Cientifica y Tecnol6gica D'(xk)= D ( T - l ( x L ) ) (3)

0-7803-6725-1/01/$10.00 02001 IEEE 122

 than 200 CIF images. These images were randomly selected
180 x 50 x 33 from the ViBE video database [IO]. One important issue in
85 x 85 x 85 order to estimate p ( x , / H o ) and p ( x P / H 1 )is the number
YCbCr 85 x 128 x 128 of bindchannel of the histograms. This number is a trade
CbCr 128 x 128 off between poor accuracy for few bins and over-fitting for
too many bins. Table 1 shows the number of binslchannel
Table 1. Number of bins/chanel used to estimate P(Xp/Ho) which have given the best results in our experiments for
and p ( x P / H 1 for
) each color space. each color space. Figure 1 shows the ROC (Receiver Op-
erating Characteristic) curves obtained for each skin detec-
tor. 1; can be seen that the ROC curves corresponding to
This means that we can always find a skin color detector
the three dimensional color spaces are practically the same.
with the same performace no matter the specific color space
In [ 111 6822 images were used in order to estimate the con-
chosen, provided that there exists an invertible transforma-
ditional probabilities, however the results that they obtained
tion between the compared color spaces.
are very close to ours which proves that our dataset is rep-
Proof: Since x p = T - l ( x ; ) , the detection and false
resentative enough. Therefore it can be seen that the perfor-
alarm rates for D ( x p )and D’(xa)are exactly the same. If
mance of the optimum skin detectors is independent of the
there exists another skin detector D“(xb)in C’ which has a
color space. The small differences are consequence of the
higher detection rate a new skin detector D ’ ” ( x p )in C can
quantization of the color space. One important restriction
be constructed:
is that the transformation between the color spaces T ( x p )
D”’( x , ) = D ” (T ( x p ) ) has to be invertible. Figure 1 also shows the ROC curve
(4)
for the CbCr color space. It can be noticed that the per-
which would have a detection rate Po > PA in C, which is formance is lower since the transformation from any three
wrong by hypothesis. dimensional color space to the bidimensional CbCr color
space is not invertible. In [12] more examples are presented
using other non-invertible color space transformations. As a
4. RESULTS AND EXPERIMENTS conclusion, in this paper we have proven that if an optimum
skin detector is designed for every color space, then their
In order to prove the theoretical statement of the previous
performace will be the same.
section three skin detectors have been designed for the RGB,
YCbCr and HSV color spaces using the Neyman-Pearson
test [9]. To that end, the skin detection is formulated as a 5. REFERENCES
classical detection problem where two hypothesis are tested.
The two possible hypothesis for the pixel values of x p are: [ l ] A. Albiol, L. Torres, C.A. Bouman, and E. J. Delp,
“A simple and efficient face detection algorithm for
HO = x p is skin H1 = x p is not skin (5) video database applications,” in Proceedings of the
In any decision problem to make a choice implies some IEEE International Conference on Image Processing,
costs. The Neyman-Pearson test is useful in situations where Vacouver,Canada, September 2000, vol. 2, pp. 239-
these costs are difficult to quantify, since it works directly 242.
with the false alarm and detection probabilities. One im-
portant property of the Neyman-Pearson test is that it is [2] H. Wang and S-F. Chang, “A highly efficient system
optimum in the following sense: given a fixed false alarm for automatic face region detection in mpeg video,”
IEEE Transactions on circuits and system f o r video
it provides the detector with the maximum detection rate.
technology, vol. 7, no. 4, pp. 615-628, August 1997.
Let p ( x P / H o )and p ( x p / H 1 )be the conditional probability
density functions given the classes Ho and H I respectively.
The decision rule using the Neyman-Pearson test is formu- [3] D. Chai and K. N. Ngan, “Foreground background
lated as follows: video coding using H.261,” in SPIE Visual Communi-
cations and Image Processing, San Jose, CA, January
1998, pp. 434-445.

[4] K. Sobottka and I. Pitas, “Face localization and facial

where X(xp)is the likelihood ratio and X is a suitable thresh- feature extracion based on shape and color informa-
old chosen to fulfill the constraint of a fixed false alarm. In tion,” in Proceedings of the IEEE International Con-
our experiments p ( x P / H o )and p ( x p / H 1 )have been esti- ference on Image Processing, Lausanne, Switzerland,
mated from the normalized histograms obtained from more September 1996, vol. 3, pp. 236241.

123
 [5] D. Saxe and R. Foulds, “Towards robust skin identifi-
cation in video images,” in Proceedings of the Second
International Conference on Automatic Face and Ges-
ture Recognition, Killinton, Vermont, October 1996,
pp. 379-384.

[6] J. G. Wang and E. Sung, “Frontal-view face detection

and facial feature extraction using color adn morpho-
logical operators,” Pattern recognition letters, vol. 20,
no. 10, pp. 1053-1068, October 1999.

[7] Ming-Hsuan Yang and Narendra Ahuja, “Detect-

ing human faces in color images,” in Proceedings
of the International Conference on Image Processing,
Chicago, IL,October 4-7 1998, pp. 127-130.
[8] C. H. Lee, J. S. Kim, and K. H. Park, “Automatic
human face location in a complex background using
motion and color information,” Pattern recognition,
vol. 29, no. 11, pp. 1877-1889,1996. ROC curves for different color spaces

[9] M. D. Srinath, P. K. Rajasekaran, and R. Viswanathan,

Introduction to statistical signal processing with ap-
plications, Prentice-Hall, 1996.

[lo] J. Chen, C. M. Taskran, A. Albiol, C. A. Bouman,

and E. J. Delp. “Vibe: A video indexing and browsing
environment,” in IEEE- SPIE Conference on Multime- I
dia Storage and Archiving Systems IV, Boston (USA),
September 1999.
07

[113 M. J Jones and J. M. Rehg, “Statistical color models

with application to skin detection,” Technical report,
Cambridge Res. Lab., Compaq Comput. Corp., 1998. 06 I ,
01 02 03 04 05
PFa
[12] J. Brand and J. S. Mason, “A comparative as-
sessment of three approaches to pixel-level human
Fig. 1. ROC curves for different color spaces
skin-detection,” in Proceedings of the International
Conference on Patter Recognition, Barcelona, Spain,
September 2000, pp. 1056-1059.

124

View publication stats