HUMAN SKIN COLOR DETECTION AND APPLICATION TO
ADULT IMAGE DETECTION
Ryszard S. Chora´s
Institute of Telecommunications, University of Technology & Life Sciences
Kaliskiego Street 7, 85-796 Bydgoszcz, Poland
Keywords: Skin color, Y C
b
C
r
color space, skin likelihood image, adult images.
Abstract: In this paper, we aimed at the detection of adult image. The methods of detection mainly focus on the de-
tection/identification of skin region. Skin detection is of the paramount importance in the detection of adult
images. Our algorithm is designed to detect human skin color in Y C
b
C
r
color space. The proposed system
finds skin regions and then generates the skin likelihood image. Since the skin likelihood image contains shape
information as well as skin color information, we used the skin likelihood image to classify the adult images.
1 INTRODUCTION
The rapid development in the field of digital media
has exposed us to huge amounts of non-textual in-
formation such as audio, video, and images. The re-
trieval and classification techniques have been and are
continuously being developed and improved to facili-
tate the exchange of relevant information.
But images can be able to contributing nude content.
Effective filtering and blocking these images is of
paramount importance in an wide computer networks
(e.g. Internet). An image analysis technique is needed
in order to classify adult (nude, objectionable) images
and further block accessing to objectionable sites.
The detection of images containing nudity and
pornography is based on the identification of human
skin and usually based on color with texture as a ad-
dition feature (Fleck, M.M., Forsyth, D.A., Bregler,
C., 1996),(Jones, M.J., Rehg, J.M., 1998),(Chan, Y.,
Harvey, R., Bagham, J., 2000).
There is an extensive literature (J.Z. Wang, J. Li,
G. Wiederhold, O. Firschein, 1998), (A. Bosson, G.
C. Cawley, Y. Chan and R. Harvey, 2002) on the de-
tection of naked images by using features such as
color histograms, texture measures and shape mea-
sures. Bosson et al. finds skin blobs, and then com-
putes the area, centroid, length of the axes of an el-
lipse, eccentricity, solidity, and extent of skin blobs.
The skin color model used by Fleck et al. (Fleck,
M.M., Forsyth, D.A., Bregler, C., 1996) consists of
a manually specified region in a log-opponent color
space. Detected regions of skin pixels form the input
to a geometric filter based on skeletal structure.
Skin-color detection has been employed in many
applications such as face detection, gesture recogni-
tion, human tracking, pornographic image filtering,
etc.
Skin color has proven to be a useful and robust
cue for face detection, localization and tracking. Nu-
merous techniques for skin color modeling and recog-
nition have been proposed during several past years.
Most existing skin segmentation techniques involve
the classification of individual image pixels into skin
and non- skin categories on the basis of pixel color.
The main goal of skin color detection is to build
a decision rule that will discriminate between skin
and non-skin pixels. Identifying skin colored pixels
involves finding the range of values for which most
skin pixels would fall in a given color space. Various
color spaces are used for processing digital images.
For some purposes, one color space may be more ap-
propriate than others.
Jiao et al. (F. Jiao, W. Gao, L. Duan, G. Cui, 2001)
presented an adult image detection method. They first
use the skin color model to detect naked skin areas
roughly. Then the Gabor filter are applied to remove
those non-skin pixels. Some simple features are ex-
tracted to detect adult images. Several representative
features induced from the naked images are used to
verify these skin areas.
388
S. Chora
´
s R. (2008).
HUMAN SKIN COLOR DETECTION AND APPLICATION TO ADULT IMAGE DETECTION.
In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 388-392
DOI: 10.5220/0001939003880392
Copyright
c
 SciTePress

2 COLOR
Each image is represented using three primaries of the
color space chosen. Most digital images are stored in
RGB color space. RGB color space is represented
with red (R), green (G), and blue (B) primaries and
is an additive system. RGB color space is not per-
ceptually uniform, which implies that two colors with
larger distance can be perceptually more similar than
another two colors with smaller distance, or simply
put, the color distance in RGB space does not repre-
sent perceptual color distance.
RGB is one of the most widely used colorspaces,
however, high correlation between channels, signif-
icant perceptual non-uniformity, mixing of chromi-
nance and luminance data make RGB not a very
favorable choice for color analysis and colorbased
recognition algorithms.
Normalized RGB is a representation, that is eas-
ily obtained from the RGB values by a simple
normalization procedure: r =
R
R+G+B
; g =
G
R+G+B
; b =
B
R+G+B
. The three normalized com-
ponents r, g and b are called pure colors; they con-
tain no information about the luminance. Because
(r+g+b = 1), the third componentdoes not hold any
significant information and can be omitted, reducing
the space dimensionality. It is enough to use only two
components r and g to completely describe the skin
color space.
Y C
r
C
b
is an encoded nonlinear RGB signal for
image compression work. Color is represented by lu-
minance, computed from nonlinear RGB (Poynton,
1995), constructed as a weighted sum of the RGB
values, and two color difference values C
r
and C
b
that
are formed by subtracting luminance from RGB red
and blue components.
Y C
b
C
r
color space has been defined in response
to increasing demands for digital algorithms in han-
dling video information, and has since become a
widely used model in a digital video. It belongs
to the family of television transmission color spaces.
These color spaces separate RGB into luminance and
chrominance information.
Y = 0, 299R + 0, 587G + 0, 114B
C
r
= 0, 713(R − Y ) (1)
C
b
= 0, 564(B − Y )
HSI, HSV, HSL - Hue, Saturation, Intensity
(Value, Lightness), color spaces describe color with
intuitive values. Hue defines the dominant color (such
as red, green, purple and yellow) of an area, satura-
tion measures the colorfulness of an area in propor-
tion to its brightness (Poynton, 1995). The ”inten-
sity”, ”lightness” or ”value” is related to the color lu-
minance. The intuitiveness of the color space com-
ponents and explicit discrimination between lumi-
nance and chrominance properties made these color
spaces popular in the works on skin color segmenta-
tion (Zarit, B.D., Super, B.J., Quek, F.K.H., 2002),
(Sigal, L., Sclaroff, S., Athitsos, V., 2000).
H = arcsin(
(C
r
− 128)
128 · S
); V =
Y
256
S =
p
(C
r
− 128)
2
+ (C
b
− 128)
2
128
(2)
3 SKIN-COLOR MODEL
3.1 Explicitly Defined Skin Region
The statistical skin-color model is generated by means
of a supervised training, using a set of skin-color re-
gions, obtained from a color human body database.
Such images were obtained from people of different
races, ages and gender,with varying illumination con-
ditions.
One method to build a skin classifier is to define
explicitly (through a number of rules) the boundaries
skin cluster in some color space. For example (Peer,
P., Kovac, J., Solina, F., 2003) (R, G, B) is classified
as skin if:
R > 95 and G > 40 and B > 20 and
max{R, G, B} − min{R, G, B} > 15 and
|R − G| > 15 and R > G and R > B.
We have found that skin-color region can be iden-
tified by the presence of a certain set of chrominance
(ie C
r
and C
b
) values that is narrowly and consis-
tently distributed in the Y C
r
C
b
color space. We de-
note RC
r
and RC
b
as the respective ranges of C
r
and
C
b
values that correspond to skin color, which sub-
sequently define our skin-color reference map. The
ranges that we found to be the most suitable for all
the input images that we have tested are RC
r
=
[133, 173] and RC
b
= [77, 127]. This map has been
proven, in our experiments, to be very robust against
different types of skin color (Figure 1).
With this skin-color reference map, the color seg-
mentation can now begin. Since we are utilizing only
the color information, the segmentation requires only
the chrominance component of the input image. The
output of the color segmentation, is a skin bitmap SM
described as
SM (x, y) =

1 if C
r
(x, y) ∈ RC
r
∩ C
b
(x, y) ∈ RC
b
0 otherwise
(3)
HUMAN SKIN COLOR DETECTION AND APPLICATION TO ADULT IMAGE DETECTION
389

Figure 1: The histogram of Y, C
b
, C
r
components of white
(a) and black (b) skin colors.
The output pixel at point (x, y) is classified as
skin-color and set to 1 if both the C
r
and C
b
values
at that point fall inside their respective ranges, RC
r
and RC
b
. Otherwise, the pixel is classified as non-
skin-color and set to 0 (Figure 2).
Figure 2: Original image (a) and skin map (b).
Figure 3: Skin color distribution.
The skin color model based on C
b
and C
r
values
can provide good coverage of all human races. De-
spite their different appearances, these color types be-
long to the same cluster in C
b
C
r
plane (Figure 3). We
classify the popular skin colors of naked images, in-
cluding various lighting conditions, into several cat-
egories. The corresponding compactly chroma his-
togram of each category is compiled in advance.
3.2 Feature Extraction
The performance of the naked image detection ex-
tremely depends on the accurate skin segmentation.
Since the skin property is very smooth, we utilize the
roughness feature to further reject confusion from non
skin stuff with skin like chroma.
3.2.1 Texture Feature from Gabor Wavelet
Transform
For further improved skin detection blobs we propose
skin color filter based on texture features. The Gabor
filter will be eliminated non-skin pixels.
Gabor wavelet based texture is robust to orienta-
tion and illumination change, It is a powerful tool to
extract texture features. Gabor functions are Gaus-
sians modulated by complex sinusoids. In two dimen-
sions they take the form:
g(x, y) = (
1
2πσ
x
σ
y
) exp[−
1
2
(
x
2
σ
2
x
+
y
2
σ
2
y
)+ 2πjW x] (4)
G(u, v) = exp{−
1
2
[
(u − W )
2
σ
2
u
+
v
2
σ
2
v
]} (5)
where σ
u
=
1
2π σ
x
, σ
v
=
1
2π σ
y
.
Gabor wavelets can be obtained by appropriate di-
lations and rotations of g(x, y) through the generating
function:
g
mn
(x, y) = a
−m
g(x
′
, y
′
); a > 1 m = 0, 1, . . . , S − 1
(6)
x
′
= a
−m
(x cos θ+y sin θ), y
′
= a
−m
(−x sin θ+y cos θ)
(7)
θ =
nπ
K
k = integer.
Where k is the number of orientations, given an
image I(x, y), its Gabor wavelets transform is calcu-
lated as:
W
mn
(x, y) =
XX
I(x
1
, y
1
)g
∗
mn
(x − x
1
, y − y
1
) (8)
Where g
∗
is the complex conjugate of g, the tex-
ture feature is given as:
T (x, y) =
q
(
XX
W
2
mn
(x, y)) (9)
In our experiment, we use three scales (m = 0, 1, 2)
and four orientations (n = 0, 1, 2, 3) (Figure 4).
Figure 4: Real (a) and imaginery (b) parts of Gabor
wavelets and Gabor kernels with different orientations (c).
Then texture image is computed from equation 9,
the texture mask image is obtained as:
T M (xy) =

1 if T (x, y) ≤ th
0 if T (x, y) > th
(10)
where th is the texture threshold. Since human skin
is smooth, its texture feature value is relatively low.
Using texture mask, pixels reassemble skin in color
but has high texture feature value will be filtered.
SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications
390

For each obtained skin color region (blob) we de-
fine a ”rectangular box” with the aspect ratio ar de-
fined as
ar =
|x
right
− x
lef t
|
|y
top
− y
down
|
(11)
where x
right
, x
left
, y
top
, y
down
indicate the
smallest and largest x- and y-coordinates, respec-
tively.
The rectangular box contain a skin color region as
well as several non-skin color region. The skin color
ratio scr is defined as
scr =
T he number of sk in color pixels
Area rectangular box
(12)
Next, we apply the ellipse model for skin body
blobs (Figure 5). The central position (x
c
, y
c
) of the
blob are estimated as x
c
=
M
10
M
00
; y
c
=
M
01
M
00
where
M
00
, M
10
, M
01
are the first order moment calculated
from the blob pixels.
Figure 5: Blob ellipse model.
The ellipse model for a candidate skin body region
is defined by area, major and minor axis of the ellipse
with following equations
axis
major
=
q
6(p + r +
p
q
2
+ (p − r)
2
) (13)
axis
minor
=
q
6(p + r −
p
q
2
+ (p − r)
2
) (14)
where p =
M
20
M
00
− x
2
c
, q = 2(
M
20
M
00
− x
c
y
c
),
r =
M
02
M
00
− y
2
c
.
The parameters of the ellipse are used to decide
the body skin area.
The nudity detection algorithm works in the fol-
lowing manner:
1. Calculate the corresponding Y C
b
C
r
values from
the RGB values. Label each pixel as skin or non-
skin and identify connected skin pixels to form
skin regions.
2. Use the Gabor filter to improve skin regions and
eliminate non-skin pixels.
3. Define a ”rectangular box” for the largest skin
blobs and calculate ar, scr. Fit the skin blobs
using a simple geometric shape such as ellipse.
Calculate the parameters of the ellipse.
4. If the percentage of skin pixels relative to the im-
age size is less than 15 percent, the image is not
nude. If the ranges of ar and sar are ar =
[0.35, 0.85] and scr = [0.75, 0.85] the image is
nude.
4 EXPERIMENTAL RESULTS
AND CONCLUSIONS
In this paper, we use the color information of image
to detected the regions skin in nude image. Because
the shape and texture information of image is useful
information, in classifying nude image is necessary.
The detection rate and false alarm rate are ex-
pressed as a percentage, which represent in equation
15
DR =
T P
T P + F N
; F AR =
F P
T P + F P
(15)
where T P = True Positive, F P = False Positive
and F N = False Negative.
The detection rates and false alarm rates for the test-
ing set are respectively DR = 93% and F AR = 7%.
These results show that the proposed segmentation al-
gorithm could provide the face segmentation more ef-
ficient and it also has less affect from they fast or slow
moving object in video sequences.
REFERENCES
A. Bosson, G. C. Cawley, Y. Chan and R. Harvey (2002).
Non- retrieval:blocking pornographic images. In Pro-
ceedings of the International Conference on the Chal-
lenge of Image and Video Retrieval, volume 2383
of Lecture Notes in Computer Science, pages 50–60.
Springer, London.
Chan, Y., Harvey, R., Bagham, J. (2000). Using colour fea-
tures to block dubious images. In Proc. Eusipco.
F. Jiao, W. Gao, L. Duan, G. Cui(2001). Detecting adult im-
age using multiple features. In Proceedings of IEEE
International Conference on Info-tech and Info-net,
pages 378–383.
Fleck, M.M., Forsyth, D.A., Bregler, C. (1996). European
Conference on Computer Vision, volume II, chap-
ter Finding naked people, pages 593–602. Springer-
Verlag.
HUMAN SKIN COLOR DETECTION AND APPLICATION TO ADULT IMAGE DETECTION
391

Jones, M.J., Rehg, J.M. (1998). Statistical color models
with application to skin detection. Technical Report
CRL 98/11 11, Compaq Cambridge Research Labora-
tory.
J.Z. Wang, J. Li, G. Wiederhold, O. Firschein (1998). Sys-
tem for screening objectionable images. Computer
Communications, 21(15):1355–1360.
Peer, P., Kovac, J., Solina, F. (2003). Human skin colour
clustering for face detection. In EUROCON 2003 -
International Conference on Computer as a Tool.
Poynton, C. A. (1995). Frequently
asked questions about colour. In
ftp://www.inforamp.net/pub/users/poynton/doc/colour/.
Sigal, L., Sclaroff, S., Athitsos, V. (2000). Estimation and
prediction of evolving color distributions for skin seg-
mentation under varying illumination. In Proc. IEEE
Conf. on Computer Vision and Pattern Recognition,
volume 2, pages 152–159.
Zarit, B.D., Super, B.J., Quek, F.K.H. (2002). Comparison
of five color models in skin pixel classification. In
ICCV99 Intl Workshop on recognition, analysis and
tracking of faces and gestures in Real-Time systems,
pages 58–63.
SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications
392