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

net/publication/266012529

Static Filtered Skin Detection

Article  in  International Journal of Computer Science Issues · March 2012

CITATIONS READS
7 133

4 authors, including:

Rehanullah Khan Muhammad Aamir

Qassim University Federal Urdu University of Arts, Science & Technology
80 PUBLICATIONS   830 CITATIONS    41 PUBLICATIONS   246 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Distributed Deep Learning View project

street lights control and office automation using FPGA View project

All content following this page was uploaded by Rehanullah Khan on 23 August 2015.

The user has requested enhancement of the downloaded file.

IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 3, March 2012
ISSN (Online): 1694-0814
www.IJCSI.org 257

Static Filtered Skin Detection

Rehanullah Khan 1 , Zeeshan Khan 2 , Muhammad Aamir 1 , Syed Qasim Sattar 1

1
Sarhad University of Science and Information Technology, Peshawar
2
University of Engineering and Technology, Peshawar

ABSTRACT ters in the IHLS and CIELAB color spaces. The two new
A static skin filter defines explicitly (using a number of rules) static filters and four state-of-the-art static filters in YCbCr,
the boundaries the skin cluster has in a color space. Single or HSI, RGB and normalized RGB color spaces are evaluated
multiple ranges of threshold values for each color space com- on the two datasets DS1 and DS2, on the basis of F-measure.
ponent are created and the image pixel values falling within Experimental results reveal the feasibility of the developed
these range(s) for all the chosen color components are defined static skin filters.
as skin pixels. In this paper, we investigate and evaluate static Singh et al. [19] presents a detailed discussion on various
skin filters for skin segmentation. As a contribution, two new color spaces and skin detection techniques. In their work, they
static skin filters for the IHLS and CIELAB color spaces are mainly consider three color spaces; RGB, YCbCr and HSI.
developed. The two new static filters and four state-of-the- They have compared various algorithms based on these color
art static filters in YCbCr, HSI, RGB and normalized RGB spaces and have combined them for face detection. However,
color spaces are evaluated on the two datasets DS1 and DS2, the algorithm fails when sufficient non face skin is visible in
on the basis of F-measure. Experimental results reveal the the images. In [21], the effect of using various color spaces
feasibility of the developed static skin filters. We also found in skin detection is discussed. In addition, the authors state
that since the static filters use static boundaries, any shift that excluding color luminance from the classification process
of skin color ranges from the static boundaries will result cannot help achieving better discrimination between the skin
in varying performance. Therefore, the F-measure rankings and non skin areas/pixels in the image.
of the color spaces are different for the datasets DS1 and DS2. In [25] and [26], skin information based image filters are
described. The first step in their approach is skin detection.
Keywords: Static skin filters, color space thresholding The distribution of skinness in the input image is modeled
using the Maximum entropy modeling. Then, a first order
1. INTRODUCTION model is built that introduces constraints on color gradients
of neighboring pixels. The output of the skin detection step is
Locating and tracking patches of skin-colored pixels through a gray scale skin map with the gray levels being proportional
an image is a tool used in many face recognition and gesture to the skinness probabilities. False alarms are raised in the
tracking systems [19][9]. Skin information contributes much skin detection process when the background color matches
to object recognition [24]. One of the usage of skin color the human skin color. According to [2], using two or more
based tracking, locating and categorization could be blocking color spaces shows better performance in comparison to a sin-
unwanted video contents on World Wide Web [10]. On ded- gle color space which may limit the performance of the skin
icated websites, people can upload videos and share it with color filter.
the rest of the world. There are uploaded adult videos, which Jae et al. [12] discuss elliptical boundary model for
may not be allowed by the service providers. Therefore, how skin color detection. They investigate the characteristics of
to effectively categorize and block such videos in real-time skin and non-skin distributions in terms of chrominance his-
has been arousing a serious concern for the service providers. tograms to devise an appropriate model for skin detection
The advantage of static filters is the simplicity of skin de- and don’t employ the method of combining different color
tection rules. This results in the construction of a classifier spaces. In [22], a method to detect body parts in images
which is computationally favorable [21]. For the static filters, is presented. The algorithm presented in this work is com-
we need to find both a good color space and adequate deci- posed of both content-based and image-based classification
sion rules empirically. Generally, the true positive rate can be approaches. In the content-based approach, color filtering
increased by tuning but at the same time the false positive rate and texture analysis is used to detect the skin region in an
is also affected [5, 7, 11]. image and its classification depends on the presence of large
In this paper, we investigate and evaluate static skin fil- skin bulks. In the image-based approach, color histogram
ters. As a contribution, we introduce two new static skin fil- and coherence vectors are extracted to represent the color and

Copyright (c) 2012 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 3, March 2012
ISSN (Online): 1694-0814
www.IJCSI.org 258

spatial information of the image. chosen color components are defined as skin pixels. The
According to [13] and [15], the quality of skin color mod- advantage of this method is the simplicity of skin detection
eling is influenced by the selection of color space. The pix- rules and the computational efficiency because it is pixel
els belonging to skin region in all human races exhibit sim- based. The main difficulty achieving high recognition rates
ilar Cb and Cr chromatic characteristics, therefore the skin with this method is the need to find both a good color space
color model based on Cb and Cr values cover almost all hu- and adequate decision rules empirically [21]. Generally, the
man races. Accordingly, despite their different appearances, True Positive (TP) rate is high but at the same time due to the
these color types belong to the same small cluster in Cb-Cr large boundary of the static filter, the False Positive (FP) rate
plane. A typical viewer perceives a difference in skin colors is also high.
mainly based on the darkness or fairness of the skin. These Chai and Ngan [1] exploit the spatial distribution of hu-
features are reflected as a difference in the brightness of the man skin color in images. A static skin filter is derived and
color, which is governed by Y component rather than Cb and uses the chrominance components of the image for skin pixel
Cr components. It provides an effective separation into lu- detection. It is assumed that the different skin colors that
minance and chrominance channel and generates a compact are perceived in the image cannot be differentiated by the
skin chroma distribution. Yang et al. [24] have introduced a chrominance information of the corresponding image region
new Gamma Correction method to reduce the effects of illu- and therefore, skin color can be represented by the static val-
mination on images. They have also introduced a new RGB ues of C b and C r component of the YCbCr color space. The
non-linear transformation to describe the skin and non-skin ranges for the static filters are found by testing on a large num-
distributions. Khan et al. [11][5][7] use face detection for ber of images and then tuning the corresponding values in
adapting to the varying illumination conditions for detecting case of violations. The final values reported are,
skin in videos. The authors in [6] introduce the usage of De-
cision Trees for pixel based skin detection and classification. C bmax = 127, C bmin = 77, C rmax = 173, C rmin =
A global seeds based skin detection method is introduced in 133
[8]. (1)
Color is a low level feature, which makes its usage com- A pixel is skin, if it lies between these values.
putationally inexpensive and therefore suitable for real-time Peer et al. [16] advocate the usage of the RGB color space
object characterization, detection and localization [13]. The for face detection. They specifically deal with the problem
major aim of skin color detection or classification for skin of varying illumination and compensate for lighting correc-
contents filtering is to establish a decision criterion that will tion using the Gray World algorithm and Color by Correla-
discriminate between skin and non-skin pixels in an image. tion technique. Classification of skin color is performed by
Identifying skin colored pixels in a given color space involves heuristic rules taking into account two different illumination
finding the range of values which most skin pixels would pos- conditions: Uniform daylight and lateral illumination. A filter
sess. This may be as simple as classifying a pixel as skin pixel for uniform daylight illumination is:
based on its association to an explicit set or range of values
R > 95, G > 40, B > 20 (2)
of Red, Green and Blue color channels’ distribution. On the
other hand, a range of complex prediction techniques can be
employed which use Neural Networks and Bayesian methods (M ax{R, G, B} − min{R, G, B}) > 15
[13].
The rest of the paper is organized as follows: Section
2 explains in detail the four state-of-the-art static filters and |R − G| > 15, R > G, R > B
the two new static filters developed. Section 3 explains the A filter for daylight lateral illumination (flashlight) is:
datasets and an evaluation and Section 4 concludes.
R > 220, G > 210, B > 170 (3)
2. COLOR SPACE THRESHOLDING
|R − G| ≤ 15, B < R, B < G
Human skin color can be approximated in a well defined
cluster given a color space, if the recording conditions A static filter for the normalized RGB color space is re-
for the images remain consistent (illumination controlled ported in [3]. The paper describes a new constructive in-
environment)[23]. Based on this idea, one method is to build duction algorithm for creating adequate attributes to consti-
a static skin classifier. A static skin classifier defines explic- tute the skin map. Using a simple set of operators and the
itly (using a number of rules) the boundaries the skin cluster three normalized RGB components, a model for skin detec-
has in a color space. Single or multiple ranges of threshold tion is presented with a combination of different rules. The
values for each color space component are created and the Restricted Covering Algorithm (RCA) is used for selective
image pixel values falling within these range(s) for all the learning during the training phase. RCA is based on selec-
tion of single well defined separable rules. RCA performs its

Copyright (c) 2012 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 3, March 2012
ISSN (Online): 1694-0814
www.IJCSI.org 259

Fig. 1. Example frames from the annotated video dataset (DS1).

search in parallel for finding a single set of rules. There are

different combinations of rules reported and the highest pre- following rules are adopted:
cision and accuracy is reported for the following rule: iHmax = 50, iHmin = 0, iSmax = 0.9, iSmin = 0.1
(7)
nr
> 1.185 (4) where iHmax and iHmin are the upper and lower boundary
ng
values for the hue component, iSmax and iSmin are the upper
nr.nb
> 0.107 and lower boundary values for the saturation component of
(nr + ng + nb)2
the IHLS color space. For CIELAB, we also built a static
nr.ng
> 0.112 filter using the same procedure as that of the IHLS color space
(nr + ng + nb)2 and finally using the following rules:
where nr, ng and nb correspond to normalized coordinates. amax = 14, amin = 2, bmax = 18, bmin = 0.7 (8)
Skin detection is used as a cue for face detection in [17] us-
ing the HSI color space. A binary skin map is generated for
oriental face detection for locating multiple faces in natural 3. EVALUATION
scenes. A clustering-based splitting algorithm is used to sep-
In an experimental setup, the two new static filters and four
arate facial and non-facial regions in the skin color map. The
state-of-the-art static filters in YCbCr, HSI, RGB and nor-
HSI color space is favored because of its stable behavior in
malized RGB color spaces are evaluated on the two datasets
non-uniform lighting conditions. Skin is segmented in the
DS1 and DS2, on the basis of F-measure. The dataset (DS1)
HSI color space based on the following rules:
is created mainly by Christian Liensberger [14], spanning 25
I > 40 (5) YouTube videos (see Figure 1). The second dataset (DS2) is
provided by Sigal et. al. [18] (see Figure 2).
13 < S < 110, 0 ◦ < H < 28 ◦ Or 332 ◦ < H < 360 ◦
Or
I > 40
13 < S < 75, 309 ◦ < H < 331 ◦

A static skin filter in the HSV color space is reported in

[20]. The authors argue that skin color can be accurately char-
acterized by hue and saturation. The thresholds used in the
HSV color space for skin segmentation are:

V ≥ 40 (6)
0.2 < S < 0.6;
0 ◦ < H < 25 ◦ or 335 ◦ < H < 360
◦ Fig. 2. Example frames from dataset (DS2) consisting of 21
video sequences. (Source: [18]).
The V component filters out dark colors. The range of satu-
ration S excludes pure red or dark red colors. The hue H and In an evaluation setup, for YCbCr, the static filter reported
saturation S account for slightly varying lighting conditions. in [1] is used. For HSI, the static filter of [17] is used. For
We developed two new static filters for the IHLS and the RGB color space, the static filter reported in [16] is used.
CIELAB color spaces. For the IHLS color space, we built For normalized RGB, static filter of [3] is used. For IHLS
a static filter from the skin distribution in Weka [4] and re- and CIELAB color spaces, we use the newly developed static
fined the corresponding values on test images. Finally, the filters.

Copyright (c) 2012 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 3, March 2012
ISSN (Online): 1694-0814
www.IJCSI.org 260

shift of skin color ranges from the static boundaries will re-
sult in varying performance. Therefore, the F-measure rank-
ings of the color spaces are different for the datasets DS1 and
DS2.

5. REFERENCES

[1] D. Chai and K.N. Ngan. Locating facial region of a

head-and-shoulders color image. In Int. Conf. Automatic
Face and Gesture Recognition, pages 124–129, 1998.

[2] Moheb R. Girgis, Tarek M. Mahmoud, and Tarek Abd-

El-Hafeez. An approach to image extraction and accu-
rate skin detection from web pages. In Proceedings of
Fig. 4. F-measure for DS1 and DS2 based on static skin filters
World Academy of Science, Engineering and Technol-
in six color spaces.
ogy, pages 367–375, 2007.

[3] Giovani Gomez and Eduardo F. Morales. Automatic fea-

Figure 3 shows the output of these static skin filters on ture construction and a simple rule induction algorithm
two images. Figure 3 (second row) shows an actor from a for skin detection. In ICML, pages 31–38, 2002.
movie scene with items having skin like colors. The static fil-
[4] G. Holmes, A. Donkin, and I.H. Witten. Weka: a ma-
ter in the IHLS color space (Figure 3(k)) reports fewer false
chine learning workbench. In Second Australian and
positives compared to the other five static filters. Figure 4
New Zealand Conference on Information Systems, pages
shows the F-measure for static filters of the six color spaces
357–361, nov-2 dec 1994.
for datasets DS1 and DS2. For DS1, The highest F-measure
of 0.50 is reported by the CIELAB static filter and the low- [5] Rehanullah Khan, A. Hanbury, and J. Stöttinger.
est F-measure of 0.43 by the normalized RGB. The second Weighted skin color segmentation and detection using
highest F-measure of 0.49 for DS1 is reported by the HSI and graph cuts. In Proceedings of the 15th Computer Vision
the RGB static filters. YCbCr achieves an F-measure of 0.45. Winter Workshop, pages 60–68, February 2010.
For the dataset DS2, the highest F-measure of 0.50 is reported
by the normalized RGB static filter and the lowest F-measure [6] Rehanullah Khan, Allan Hanbury, and Julian Stoet-
of 0.33 by YCbCr. The second highest F-measure of 0.49 is tinger. Skin detection: A random forest approach. In
reported by the HSI static filter. IHLS, CIELAB and RGB ICIP, pages 4613 – 4616, 2010.
achieve F-measures of 0.38, 0.37 and 0.45 respectively. It
can be seen that in Figure 4, varying (non-uniform) perfor- [7] Rehanullah Khan, Allan Hanbury, and Julian Stöttinger.
mance is reported for different static filters. The two datasets Augmentation of skin segmentation. In International
DS1 and DS2 represent skin colors in different lighting con- Conference on Image Processing, Computer Vision, and
ditions, resulting in different skin locus (skin color ranges in a Pattern Recognition, pages 473–479, 2010.
color space) for the corresponding dataset (and color space).
Since, the static filters use static boundaries, any shift of skin [8] Rehanullah Khan, Allan Hanbury, and Julian Stöttinger.
color ranges from the static boundaries will result in varying Universal seed skin segmentation. In International
performance. Therefore, the F-measure rankings of the color Sym- posium on Visual Computing, pages 75–84, 2010.
spaces are different for the datasets DS1 and DS2.
[9] Rehanullah Khan, Allan Hanbury, Julian Stttinger, and
Abdul Bais. Color based skin classification. Pattern
4. CONCLUSION Recognition Letters, 33(2):157 – 163, 2012.
In this paper, we investigated and evaluated six static skin fil- [10] Rehanullah Khan, Asad Maqsood, Zeeshan Khan,
ters for skin segmentation. We introduced two new static skin Muhammad Ishaq, and Arsalan Arif. Skin-color based
filters for the IHLS and CIELAB color spaces. The two new videos categorization. IJCSI, 9(1):473–477, 2012.
static filters and four state-of-the-art static filters in YCbCr,
HSI, RGB and normalized RGB color spaces are evaluated [11] Rehanullah Khan, Julian Stöttinger, and Martin Kampel.
on the two datasets DS1 and DS2. Experimental results re- An adaptive multiple model approach for fast content-
veal the feasibility of the developed static skin filters. We also based skin detection in on-line videos. In ACM MM,
found that since the static filters use static boundaries, any AREA workshop, pages 89–96, 2008.

Copyright (c) 2012 International Journal of Computer Science Issues. All Rights Reserved.
IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 3, March 2012
ISSN (Online): 1694-0814
www.IJCSI.org 261

(a) Original (b) YCbCr (c) HSI (d) IHLS (e) CIELAB (f) RGB (g) nRGB

(h) Original (i) YCbCr (j) HSI (k) IHLS (l) CIELAB (m) RGB (n) nRGB

Fig. 3. Example results of skin detection using static skin filters in different color spaces. Black shows non-skin.

[12] Jae Young Lee and Suk Yoo. An elliptical boundary [21] Vladimir Vezhnevets, Vassili Sazonov, and Alla An-
model for skin color detection. In ISST, pages 579–584, dreev. A survey on pixel-based skin color detection tech-
2002. niques. In GraphiCon, pages 85–92, 2003.

[13] Jiann-Shu Lee, Yung-Ming Kuo, Pau-Choo Chung, and [22] Shilin Wang, Hong Hui1, Sheng hong Li, Hao Zhang,
E-Liang Chen. Naked image detection based on adap- Yong yu Shi, and Wen tao Qu. Exploring content-based
tive and extensible skin color model. PR, 40(8):2261– and image-based features for nude image detection. In
2270, 2007. Fuzzy Systems and Knowledge Discovery, pages 324–
328, 2005.
[14] Christian Liensberger, Julian Stöttinger, and Martin
Kampel. Color-based and context-aware skin detection [23] Jie Yang, Weier Lu, and Alex Waibel. Skin-color mod-
for online video annotation. In MMSP, pages 1–6, 2009. eling and adaptation. In ACCV, pages 687–694, 1997.

[15] Yung ming kuo, jiann-shu lee, and pau-choo chung. The [24] Jinfeng Yang, Zhouyu Fu, Tieniu Tan, and Weiming Hu.
naked image detection based on automatic white bal- Skin color detection using multiple cues. In Proceedings
ance method. In 2006 ICS International Computer Con- of the 17th International Conference on Pattern Recog-
ference, pages 990–994, 2007. nition, pages 632–635, 2004.

[16] Peter Peer, Jure Kovac, and Franc Solina. Human skin [25] Huicheng Zheng, Mohamed Daoudi, and Bruno Jedy-
colour clustering for face detection. In EUROCON, nak. Adult image detection using statistical model and
pages 144–148, vol.2, 2003. neural network. Electronic Letters on Computer Vision
and Image Analysis, 4(2):1–14, 2003.
[17] Hsieh I. Sheen, Fan K. Chin, and Chiunhsiun Lin. A
statistic approach to the detection of human faces in [26] Huicheng Zheng, Mohamed Daoudi, and Bruno Jedy-
color nature scene. PR, 35(7):1583 – 1596, 2002. nak. Blocking adult images based on statistical skin de-
tection. ELCVIA, 4(2):1–14, 2004.
[18] L. Sigal, S. Sclaroff, and V. Athitsos. Skin color-
based video segmentation under time-varying illumina-
tion. PAMI, 26(7):862–877, July 2004.

[19] Sanjay Kr. Singh, D. S. Chauhan, Mayank Vatsa, and

Richa Singh. A robust skin color based face detection
algorithm. Tamkang Journal of Science and Engineer-
ing, 6(4):227–234, 2003.

[20] Sofia Tsekeridou and Ioannis Pitas. Facial feature ex-

traction in frontal views using biometric analogies. In
EUSIPCO, pages 315–318, 1998.

View publication stats

Copyright (c) 2012 International Journal of Computer Science Issues. All Rights Reserved.