P. Satheesh et al. / IJAIR
ISSN: 2278-7844
Pornographic Image Filtering Using Skin Recognition Methods
P.Satheesh 1, B.Srinivas 2 And R.V.L.S.N. Sastry3
1 Associate Professor, M.V.G.R College of Engineering ,Dept of CSE,VZM, [email protected]
2 Assistant Professor, M.V.G.R College of Engineering ,Dept of CSE,VZM, [email protected],
3 M.Tech
from JNTU Kakinada, M.V.G.R College Of Engineering, surya77@g mail.co m
Abstract
In this paper, we describe various skin
detection methods, image filtering methods and
comprehensive comparative study among these
methods proposed to detect adult classified images. It
is based on concept of Co mputer Vision algorith ms
and pattern recognition techniques. First the images
are changed to identify areas with low color intensity
by the color model. In the next part of the proposed
system specify that the image is filtered using skin
detection. The main aim has to segment a person or
people within the image. By counting all p ixels with
identical skin tone, then we can t reat the image with
porn content
The development of Internet make changes
in dramatically falling costs of data storage and
improvement in coding technology are generating
wide variety of wall papers, images, animation,
graphics, sound and video [1]. Now a day’s it is
obvious that the computers with internet connection.
There are facilit ies in large amount of adult images
for free to access. This kind of med ia is availab le for
children and making problem for many parents.
The searching principal of Internet browser
programs to avoid adult classified content should be
filtering of images. There are ways available to stop
adult classified content images on computers. This
process is carried out by blocking unwanted sites or
identifying images that show explicit content. There
are some programs available in the foreign market
that allow blocking sites on Internet with offensive or
explicit content like: Cyber Patrol, Content Protect,
Net Nanny, Family.net and K9 Web Protection
[2]these programs are provided to safeguard their
children fro m using the Internet through having
parental control. There are programs to detect
pornographic images within the computer such as:
Surf Recon that offers a program for this purpose,
and despite being a tool of computer forensic, helps
to detect images with exp licit content.
© 2012 IJAIR. ALL RIGHTS RESERVED
Keywords
RGB, HSV, YCbCr, Human Skin filter, Image Retrieval,
Color model.
I. Introduction
Images are part of World Wide Web in this
visual world. The statistics of more than 10 million
HTM L webpage reveal that 80%of web pages
contain images and that on average there are about
30.2 images per HTML web page [23]. These images
are used to make effect ive and eye catching Web
contents.
However, images are also contributing to
harmful (e.g. pornographic) or even illegal Internet
content. So effective filtering of images plays vital
role in a web image filtering solution. To block adult
content, some representative companies as Net
Nanny and Surf Watch, operate by maintaining lists
of URL s and newsgroups and require constant
manual updating. Now a day’s tons of literature is
available in the web. Detection based on image
content analysis has the advantage to process equally
all the images without the need for updating, so will
produce more effective filtering.
The fact is that there is a relation between
images with maximu m p ixels of skin and images of
adult content. So we need to develop a skin detector.
Skin color offers an effective and efficient way to
detect the adult image content. There are so many
people already worked in this area. The WIPE [4]
system developed by Wang, Li, Wiederhold and
Firschein uses a manually -specified color histogram
model as a prefilte r in an analysis pipeline. The
images with less percentage of skin pixels are treated
as non-offensive. Images that contain considerable
294
P. Satheesh et al. / IJAIR
skin pass on to a final stage of analysis where they
are classified using wavelet features. Forsyth s [5]
research group has designed and imp lemented an
algorith m to screen images of naked people. Their
algorith ms involve a skin filtering method and human
features matrix. The skin color model used by Fleck,
Forsyth and Breg ler consists of a manually specified
region in a log-opponent color space. Detected
regions of skin pixels fro m the input image to a
geometric filter based on skeletal structure.
Skin detection describes recognition of
human skin pixels fro m an image. It p lays an
important role in various functionalities such as face
detection, searching the images and filtering the
content on the web. Skin color tone discrimination
approaches can be defined in two basic types. They
are physical-based approaches and statistical
approaches. Further, Statistical approaches can be
divided into two models, paramet ric approaches [28]
[29] [3] and non parametric approaches [30] [31].
Parametric model approaches uses the skin color
distribution in parametric form, such as Gaussian or
Gaussian mixture [1][3]. In nonparametric model
approaches histograms, used to represent density in
color space. Jones and Rehg [6] proposed model for
skin color detection by estimating the intensity of
skin and non-skin color pixels in the color space
using labeled training data. Both parametric and nonparametric statistical models usually perform colorsegmentation in color spaces that reduce the varying
illu minant. The most common co lor spaces have been
used are, normalized RGB and HSV.
II. SKIN D ETECTION
Skin detection helps in detecting various
parts of human. Such as detect a human limb, torso,
or face within a picture. In past days many methods
of skin identification within a dig ital image had been
developed. Skin color has proved to be a most
successful method for face detection, localization and
tracking. There have been a number of scholars who
have looked at using color informat ion to detect skin.
Jones and Rehg [11] has implemented a color model
using histogram-learning techniques at RGBcolor
space. Yang and Auhuja [12] has computed
probability density function of human skin color
using a finite Gaussian mixture model. Where
parameters are estimated through the EM algorith m.
There are other researchers who have developed
several papers on different models of skin detection
as Vezhnevets et al. [13], Kaku manu et al.[14], Kelly
et al. [15].
© 2012 IJAIR. ALL RIGHTS RESERVED
ISSN: 2278-7844
III. COLOR MODELS
A. RGB Color Model
The RGB color model is a pro minent color model in
which the primary colors are red, green and b lue light
are added together in various combinations to
reproduce a broad range of colors. The name comes
fro m the three basic colors Red, Green, and Blue.
The RGB color stack is illustrated in the following
Figure 1.
Figure 1. RGB Color Model
The main objective of the RGB color stack
is used for various functionalities like recognizing,
visualizing, sensing and display of images in
electronic gadgets like televisions, PCs and high end
mobiles.
The RGB color model is an exemp lar in the
sense that three light beams are combined together to
get a final co lor. The format ion of colors in RGB is
carried out by three colored light beams (one red, one
green, and one blue) should be superimposed. A
beam of each basic color is called as a component of
that color. Where each can have arbitrary intensity
either fro m fu lly off or fu lly on is in that mixture.
Zero intensity for each component results as darkest
color (no light in turn treated as black) and full
intensity of each gives a white.
A color in the RGB color model is described
by indicating mixture of red, green, and blue. It is
included in each component which can vary from
zero to a defined maximu m value. This depends on
the functionality of respective color. The component
values are often stored as numerical values in the
range 0 to 255.
295
P. Satheesh et al. / IJAIR
B. HSV Color Model
HSV color model is a no lineal
transformation of the RGB space color. The colors
are obtained by a combination of the three values: the
Hue (H), Saturation or color quantity (S), and itself
value (V). These values are represented in a circular
diagram, as illustrated in the follo wing Figure 2.
ISSN: 2278-7844
person could be represented in a different color fro m
those seen from different functionalit ies in the
respective image. An example of this is shown in
Figure 3, which is a sample transformation to the
HSV color model o f an image in RGB co lor model.
(a)
(b)
Fig 3. (a) RGB picture (b) HSV picture
Figure 2. HSV Color Model
The three magnitudes combination can be of
following values
 Hue: The kind of color (e.g. red, green, or
yellow). These are represented as a degree
of Angle. The possible values are ranges
fro m 0 to 360° (although for some
applications are normalized fro m 0 to
100%).


Saturation: Is represented as the distance
fro m the axis of the black-wh ite glow. The
range values may vary fro m 0 to 100%.
Value: Represents the height in the blackwhite axis. The range values may vary fro m
0 to 100%. 0 is always considered as black.
100 could be specified white or a more or
less saturated color based on saturation.
As specified in above discussion is of
having an advantage of using this particular color
model is that we can rule out many objects using a
simp le filter. In such situations we use skin detection
to get only the areas of skin wh ich are useful for our
objective.
In this paper a new solution using the HSV
color model is been proposed. Which has similarities
of the RGB color model, The process starts with
change of color model has been made, the next step is
to proceed with pixel detection in hu man skin. Th is is
achieved by looking at several images. The images
are threshold where most people with different skin
color within the image can be segmented.
To determine the threshold, it is mandatory
to make an analysis of the histograms in the HSV
color model. As illustrated in Figure 4 an image of
the face of a g irl identified with major clarity the
threshold that we need.
By using this color model as an input image is
converted using the mathematical exp ressions (1) to
(3) that are shown below.
Once the transformat ion of the input image
has been made, it was observed that the skin tone of a
Figure 4. Histograms of HSV Color Model
© 2012 IJAIR. ALL RIGHTS RESERVED
296
P. Satheesh et al. / IJAIR
ISSN: 2278-7844
The histograms specified in Figure 4 helps
in having idea about the values, which could be
considered to select a threshold able to take the skin
values. This would be considered if only detecting
people with the same skin color of the girl. Th is is
used as metric desired but in Internet there exists a
large amount of images that not only contain people
with a specific skin color, but also people with
different skin color. So after exhaustive analysis the
threshold considered as the following:
H >0 and H<0.25
S>0.15 and S<0.9
V>0.2 and V<0.95
Where H, S, V is in the range fro m 0 to 1.
The major purpose of finding naked people
in this process is to implement parental control over
children. There are types of features like the
percentage of pixels detecting similar skin color. The
process based on these features with a procedure of
segmentation is carried out in co lor images. So me
examples are illustrated in Figure 5 which shows
people with different skin color and can be seen that
the threshold used works appropriately.
(a)
(b)
Figure 5. Skin Color Segmentation
(a) Original Image, (b)Segmented Images
The specified threshold in this section helps to detect
skin color zones effectively.
C. Gaussian models
There are two approaches of Gaussian
models that are single Gaussian model (SGM ) and
Gaussian mixture models (GMM). The Single
Gaussian model is used to estimate skin color
probability or to estimate the Gaussian density
© 2012 IJAIR. ALL RIGHTS RESERVED
function in the two dimensional color planes. The
Single Gaussian model can only simulate a hu man
race with a skin color which is not suitable for
images where there are d ifferent light. The Gaussian
mixtu re model estimates more than a single Gaussian
skin color probability density function. Because it
can estimate the density distribution of arbitrary
shape. The Gaussian mixture model can be used to
achieve better results than the single Gaussian model
when it is used to estimate the skin color space
distribution. Gaussian mixture model is used
frequently. For examp le Jeong et al used a Gaussian
mixtu re model based on Bayesian inference to detect
skin region [3]. Hassan pour et al proposed an
adaptive skin color model based on the Gaussian
mixtu re model to handle the changing lighting or
imaging conditions in which EM algorith m is used to
initially estimate the number and weights of skin
color clusters[4]. Difficulty in Gaussian mixture
modeling method is the need for achieving number of
single Gaussian distribution and how to determine the
optimal nu mber.
D. Content Based Image Retrieval
technique (CBIR)
Content Based Image Retrieval (CBIR)
technique is described to recognize human images.
The performance is better than the previous works
which are mainly based on image understanding
techniques. A database consisting of human-related
images is established. The feature of the target image
is compared with the features in the database. For the
10 most similar images from the database, we
calculate the mean of their distances with the target
image. The mean distance is used to measure the
similarity between the target image and the humanrelated images, or in other words, how much the
target image is likely to have human in it.
Two kinds of image features are used in our
work, i.e., color features and shape invariant features
[7], [8].these two features based on the HSV color
space and denoted by histograms. Since shape
features are more important for classifying the
images, the parameters of shape and color features
are set to 80% and 20% respectively. To speed up the
retrieval p rocess, we also adopt a fast k nearest
neighbor search (KNNS) method [1].
ZHOU et al present image content-based
filtering gateway technology in which the filter for
pornographic images was achieved by analyzing
local characteristics and whole characteristics in
image [24].
297
P. Satheesh et al. / IJAIR
ISSN: 2278-7844
Chen et al used matching algorith m based
on the scale invariant feature transform(SIFT)
features to detect spam images by taking into account
the image shift, rotation, affine transformation and so
on[25]. Deselaers et al used bag-of visual- words
approach and task specific visual vocabulary to detect
skin region [26]. Th is method is simple, but its
performance was poor. Literature [27] achieved the
classification by extracting image features of interest
region.
Arentzetal firstly used genetic algorith m to
determine the importance of feature vector wh ich
contains color histogram of components Cb and Cr as
well as the shape descriptors, then threshold of Cb
and Cr components was determined to extract the
skin region [28]. The key for filtering technology
based on the key features is the accuracy of feature
extraction, not relating to the human body
indescribable shape or contour, but inaccurate
decision may occur when the light changes or the
extracted characteristics is inco mplete. In addition to
static images, there are many video on the web.
IV. Experimental Results:
In the first stage of our proposed system, we
collect the skin pixels(sp) fro m the input image and
also identify the total pixels(tp) in the image.
We check the sp and tp with various color models.
Here we checked with four color models and the
percentage of skin pixels identified by each models is
given in the below Tab le 1 and Table 2.
Then in the second stage, we calcu late
percentage of skin pixels (sp) in the input image.
Then we apply various image filtering algorith ms to
check whether it is a pornographic image or not.
Table 1:
Colour Model
% of Skin
Detection
RGB
HSV
CBIR
Gaussian Mixture Model
96.1%
98.7%
93.6%
94.5%
Results of various color models comparison
© 2012 IJAIR. ALL RIGHTS RESERVED
Table 2:
Classi fier
true positive
rate
false positive
rate
RGB
HSV
Gaussian
0.683
0.408
0.348
0.064
0.048
0.052
Results RGB, HS V, Gaussian color models computed
from test dataset.
V. CONCLUSIONS
This paper describes a comparative study of
all color models to detect images with an adult
classified content in color images. The experiment
using the RGB, HSV, YCBIR and Gaussian color
models for recognition of skin in the given image.
These models work effectively although in some
images. There could be some fault tolerance due to
the image lighting conditions and type of image when
taken for experiment. There is another factor that can
be by a bad interpretation of the system.
HSV colo r model is able to decrease all the
lighting problems that the image could be had.
Moreover, using this color model is mo re visib le the
skin tone than other color models. By th is reason,
HSV co lor model to be able to do skin detection. The
HSV colo r model gives an output skin pixel of 98.7%
that only shows color skin pixels within the image.
The importance of the comparison among
the color models was done to know efficient color
model to recognize skin pixe ls in given content. This
way know whether the input image is a porn content
image or not, at final could prove that the system
carry out effectively. Hence, we could conclude that
HSV color model has greater edge than other color
model.
VI. REFERENCES
[1] A. de Carvalho, A. Brayner et al., ―Grand Challenges in
Computer Science Research in Brazil – 2006-2016,‖
Brazilian Computer Society, T ech. Rep., May 2006.
[2] S. Agarwal and A. Awan, ―Learning to detect objects in
images via a sparse, part-based representation,‖ IEEE
Transactions on Pattern Analysis and Machine
Intelligence, vol. 26, no. 11, pp. 1475–1490, 2004,
member-Dan Roth.
[3] L.M. Bergasa and M.Mazo.and A. Gardel and M.A. Sotelo
and L. Boquete, .Unsupervised and adapative
Gaussian skin-color model.,Image and Vision Computing,
2000,18,987-1003.
298
P. Satheesh et al. / IJAIR
[4] J.Z. Wang, J. Li, G. Wiederhold, O. Firschein, .System for
Screening Objectionable Images., Computer
Communications (21)15:1355.1360, 1998.
[5] M.M. Fleck, D.A. Forsyth, C. Bregler: .Finding na ked
people.,Proc. European Conf. on Computer Vision,
B. Buxton, R. Cipolla, Springer-Verlag, Berlin, Germany,
2:593.602, 1996.
[6] M.J. Jones, J.M. Rehg, .Statistical color models with
application to skin detection.,Computer Vision and
Pattern Recognition, 274.280, 1999
[7] A. Bosson, G.C. Cawley, Y. Chian, R. Harvey, .Non-retrieval:
blocking pornographic images.,Proc. Intl.
Conf. on the Challenge of Image and Video Retrieval,
Lecture Notes in Computer Science, Springer-Verlag,
London, 2383:50.60, 2002.
[8] E. Jaynes, Probability Theory: The Logic of Science,
http://omega.albany.edu:8008/JaynesBook
[9] A. Berger, S.D. Pietra, V.D. Pietra, .A maximum entropy
approach to natural language processing., Computational
Linguistics, 22:39.71, 1996.
[10 ] S.C. Zhu, Y. Wu, D. Mumford, .Filters, Random Fields and
Maximum Entropy (FRAME): towards a
uni_ed theory for texture modeling., International Journal
of Computer Vision, 27:107.126, 1998.
[11] C. Wu, P.C. Doerschuk, .Tree Approximations to Markov
Random Fields., IEEE Transactions on PAMI,
17:391.402, April, 1995.
[12] J.S. Yedida, W.T. Freeman, Y. Weiss, .Understanding Belief
Propagation and its Generalisations.,Technical
Report TR-2001-22, Mitsubishi Research Laboratories,
January, 2002.14 H. Zheng et al. / Electronic Letters on
Computer Vision and Image Analysis 4(2):1-14, 2004
[13] D. Geman, B. Jedynak, .An Active Testing Model for
Tracking Roads in Satellite Images., IEEE Trans.on PAMI,
18(1):1.14, January, 1996.
[14] J. Besag, .On the Statistical Analysis of Dirty Pictures.,
Journal of the Royal Statistical Society, Series B,
48(3):259.302, 1986.
[15] F. Divino, A. Frigessi, .Penalized pseudolikelihood inference
in spatial interaction models with covariates
.,Scandinavian Journal of Statistics, 27(3):445-458, 2000.
[16] J. Zang, .The Mean Field Theory in EM Procedure for
Markov Random Fields., IEEE Transactions on Signal
Processing, 40(10):2570.2583, October, 1992.
[17] G. Celeux, F. Forbes, N. Peyrard, .EM Procedures Using
Mean Field-Like Approximations for Markov Model-Based
Image Segmentation.,Pattern Recognition, 36(1):131.144,
2003.
[18] L. Younes, .Estimation and annealing for Gibbsian
elds.,Annales de l'Institut Henry Poincar ´e, Section B,
Calcul des Probabilit´esetStatistique, 24:269.294, 1998.
[19] B. Jedynak, H. Zheng, M. Daoudi, .Statistical Models for Skin
Detection., IEEE Workshop on Statistical Analysis in
Computer Vision, in conjunction with CVPR 2003 Madison,
Wisconsin, June 16.22, 2003.
[20] B. Jedynak and H. Zheng and M. Daoudi and D. Barret,
.Maximum Entropy Models for Skin Detection.,
publication IRMA, Universit´e des Sciences et
Technologies de Lille, France, 2002, Volume 57, number
XIII.
[21] R.M. Haralick, L.G. Shapiro, Computer and Robot Vision,
1:639.658, 1992.
[22] Y.-H. Pao, Adaptive Pattern Recognition and Neural
Networks, Reading, Addison-Wesley, Massachusetts,
121.129, 1989.
© 2012 IJAIR. ALL RIGHTS RESERVED
ISSN: 2278-7844
[23] B. Starynkevitch, M. Daoudi et al., POESIA Software
Architecture De_nition Document, http://www.poesia_lter.org/pdf/Deliverable 3 1.pdf, Deliverable 3.1:7.9,
December, 2002.
[24] D.E. Rumelhart, G.E. Hinton and R.J. Williams, .Learning
internal representations by error propagation ., In D.E.
Rumelhart and J.L. McClelland (Eds.), Parallel Distributed
Processing: Explorations in the Microstructures of
Cognition, MIT Press, Cambridge, MA., 1:318.362, 1986
[25] J. Pearl, Probabilistic Reasoning in intelligent systems:
networks of plausible inference, Morgan Kaufmann, 1988.
[26] D. Brown, I. Craw, J. Lewthwaite, .A som based approach to
skin detection with application in real time systems. In
Proc. of the British Machine Vision Conference, 2001.
[27] L. Sigal, S. Sclaroff and V. Athitsos, .Skin Color-Based Video
Segmentation under Time-Varying Illumination . IEEE
Trans. on PAMI, 26(7):862-877, July, 2004
[28] T. Darrell, G.G. Gordon, M. Harville and J. Wood_ll,
.Integrated Person Tracking Using Stereo, Color, and
Pattern Detection. In Proc. IEEE Conf. Computer Vision
and Pattern Recognition, pp. 601-607, 2001
[29] W. Hafner and O. Munkelt, .Using Color for Detecting
Persons in Image Sequences.Pattern Recognition and
Image Analysis, Vol. 7, No. 1, pp. 47-52, 1997
[30] S.T. Birch_eld, .Elliptical Head Tracking Using Intensity
Gradients and Color Histograms. In Proc. IEEE Conf.
Computer Vision and Pattern Recognition, pp. 232-237,
1998
[31] K. Schwerdt and J.L. Crowly, .Robust Face Tracking System
for Sign Language Recognition.In Proc. Int'l Conf.
Automatic Face and Gesture Recognition, pp. 90-95, 2000
VII. Authors Profile
P.Satheesh received M.Tech in computer
Science and Technology in 2006 from Andhra
University; he has ten years of teaching
experience. He is currently employed as a an
Associate professor in CSE department,
MVGR College of Engineering. He has more
than ten papers in journals.
Srinivas Baggam received M.Tech in
(Computer Science & Engineering) from
R.V.R & J.C college of Engineering,
Guntur, Affiliated to Acharya Nagarjuna
University. Currently working as an
Assistant Professor in M.V.G.R. College of
Engineering. He got two and half years of
Industrial and Three and half years in
teaching Experience.
RVLSN Sastry Graduate from Andhra
University in Bachelor of Science and
received Master of Science and Master of
Computer Applications in 2005 and 2007
from Annamalai University. Currently
pursuing M.Tech in CSE from JNTU
kakinada. He got 6 years of industrial and 5
years of teaching experience.
299