LARGE SCALE IMAGE-BASED ADULT-CONTENT FILTERING
Henry A. Rowley
Google, Inc., Mountain View, CA, USA
Yushi Jing
College of Computing, Georgia Institute of Technology, Atlanta, GA, USA
Shumeet Baluja
Google, Inc., Mountain View, CA, USA
Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA
Keywords:
Image processing and applications, computer vision, adult-content detection, skin color detection.
Abstract:
As more people start using the Internet and more content is placed online, the chances that individuals will
encounter inappropriate or unwanted adult-oriented content increases. This paper presents a practical and scal-
able method to efficiently detect many adult-content images, specifically pornographic images. We currently
use this system in a search engine that covers a large fraction of the images on the WWW. For each image,
face detection is applied and a number of summary features are computed; the results are then fed to a support
vector machine for classification. The results show that a significant fraction of adult-content images can be
detected.
1 INTRODUCTION
As more people start using the Internet and more con-
tent is placed online, the chances that individuals will
encounter inappropriate or adult-oriented content in-
creases. Search engines can exacerbate this problem
by aggregating content from many sites and summa-
rizing it into a single result page. Many existing meth-
ods for detecting adult-content currently attempt to
classify web pages based on their text content. If the
text content of a page is classified as adult-content,
this information can be propagated to linked images
and pages. However, keyword and other text-based
approaches have significant limitations. First, they are
language specific and require a tremendous amount of
manual work to construct (either directly, or by label-
ing training data for all languages). Second, many
adult-content pages do not contain enough text for re-
liable classification. Third, the text on the page may
be intentionally obfuscated (i.e. encoded in an im-
age).
This paper looks at practical ways to detect adult-
content in the images themselves, on a scale which
can be applied to a search engine covering a large
fraction of the images on the WWW. The focus is on
efficient and robust techniques, such as color classi-
fication and face detection, which together can detect
many pornographic images with little computational
cost. The remainder of this paper is organized as
follows. Section 2 describes some previous work in
the field of image-based adult-content detection. Sec-
tion 3 describes our system in more detail, followed
by a description of our training and test data in Sec-
tion 4 and test results in Section 5. A summary is
given in Section 6.
2 BACKGROUND
Previous work on adult-content image detection can
be divided into two broad categories: those that are
based mainly on skin color or texture, and those
which analyze the shapes of skin colored regions to
determine their similarity to human figures. Repre-
sentative work is described below.
Early work on detecting skin color in images made
the observation that once brightness is normalized
out, skin colors across different races form a fairly
tight cluster, which can be modeled with a Gaussian
distribution (Hunke, 1994). Later work (Jones and
Rehg, 1998a) used a much larger set of data to build a
three-dimensional histogram of skin color using hand
labeled pixels in a large number of images collected
from the WWW. One deficiency of this approach is
290
A. Rowley H., Jing Y. and Baluja S. (2006).
LARGE SCALE IMAGE-BASED ADULT-CONTENT FILTERING.
In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 290-296
DOI: 10.5220/0001377002900296
Copyright
c
 SciTePress

that both very bright and very dark pixels must be in-
cluded in the skin color model. Since it is unlikely
that skin color will have both extremes in the same
image, one can build separate color models for differ-
ent average image brightness levels (Zeng et al., 2004)
or illumination conditions, as a weak form of color
constancy. In addition to looking at color, the size
and rough shape of skin colored blobs and their tex-
tures have been used to refine the decision of whether
a region is actually skin (Zeng et al., 2004; Kruppa
et al., 2002; Arentz and Olstad, 2004; Wang et al.,
1998; Zheng et al., 2004).
Other than texture, another way to verify whether
a patch of skin colored pixels corresponds to a per-
son is to analyze the shape of the region. One of the
earliest papers taking this approach was (Fleck et al.,
1996). A number of authors have followed that work
by developing methods to match the variety of shapes
a human figure can take, typically with the assump-
tion that all the parts have been detected by a skin
color classifier, as in (Ioffe and Forsyth, 1999). One
recent piece of work has attempted to segment the
image using color and texture rather than skin color
alone, in order to detect clothed people (Sprague and
Luo, 2002).
3 PROPOSED SYSTEM
Because of the vast number of images that we will
need to process (on the order of 10
9
), a primary con-
straint is the processing speed; the goal in our work
is to develop an efficient image classification sys-
tem. Because many shape classification techniques
can take a few seconds or even minutes per image,
these methods are not currently applicable to such a
large number of images. Below we describe the set
of features we use, along with their motivations. The
contribution of each set of features will be explored in
Section 5.
The features we use can be broadly grouped into
those which make use of color information, and those
which do not. We first describe those which make use
of a skin color model, as these features are most com-
monly used in this domain. We then describe some
additional features which are not dependent on color.
As will be seen, one of these features, the face detec-
tor, can be used to customize the skin color model to
each image.
Many of the features values below are computed
for a region of interest (ROI) in the image, which is
a rectangle centered in the image, inset by 1/6 of the
original image dimensions on all sides. The size of
this ROI was determined empirically.
Figure 1: Example of skin probability map. (Left) Input
image. (Middle) Initial skin map. (Right) After erosion and
dilation (note that thin lines of skin color are removed).
3.1 Skin-Dependent Features
The first set of 9 features are those dependent on color.
Skin Color Detection (2 features) This work used
the skin color model from (Jones and Rehg, 1998a).
The model is built from many images in which pix-
els have been hand labeled as skin or not. We have
not focused on gray-scale images in this work; han-
dling such images would require other segmentation
and image analysis approaches.
To evaluate whether a pixel in the image is skin
color, its skin and non-skin probabilities are com-
puted from the model of (Jones and Rehg, 1998a),
then combined with Bayes’ rule using a prior for
skin color of 20% to generate a skin map containing
P (skin|color), as shown in Figure 1.
Noise pixels and gaps are reduced by performing
gray-scale versions of erosion followed by dilation,
resulting in the skin probability map that will be used
for the remaining features, as shown at the right side
of Figure 1.
Because not all the images we apply the filter to
are color balanced, in some cases the skin pixels will
not be detected. Although not explored in this work,
one possibility would be to use a color constancy al-
gorithm such as (Rosenberg et al., 2003) to correct
for lighting. We will discuss later how a face detector
can be used to improve the skin color model in this
situation.
The first two features computed from the skin map
are the mean and standard deviation of the skin map
in the ROI.
Connected Component Analysis (4 features) To
further clarify which pixels may belong to skin re-
gions, connected components of pixels where the skin
map has a probability over 50% are located. The num-
ber of resulting connected components is recorded as
a feature. In addition, the mean and standard devia-
tion of the skin map within the skin components (and
LARGE SCALE IMAGE-BASED ADULT-CONTENT FILTERING
291

Figure 2: Example of using edge pixels as a texture mea-
sure. (Left) Example input image containing a textured
skin-toned door, which is detected as skin color. (Right)
Edge pixels in this image.
also in the ROI) are recorded as features.
As a simple measure of the compactness of the con-
nected components, the log of the weighted average
perimeter squared over the area of each component
is computed. The hope is that skin colored regions
which may occur in the non-person background of an
image will be less compact, in general, than those that
are actually skin on humans.
In subsequent sections, the skin connected compo-
nents will be referred to as skin blobs for brevity.
Skin Texture Features (2 features) Another use-
ful way to distinguish skin from other similarly col-
ored regions is with texture. In this work, texture is
approximated by the OpenCV implementation (Brad-
ski, 2000) of the Canny edge detector (Canny, 1986)
applied to the gray-scale image, as illustrated in Fig-
ure 2. We record the following ratio:
number of edge pixels in skin blobs in ROI
number of pixels in skin blobs in ROI
(1)
as a measure of skin texture, and the following ratio:
number of edge pixels in skin blobs in ROI
number of edge pixels
(2)
as a measure of how much of the image texture can be
attributed to skin-colored pixels. The specific ratios
used above were arrived at experimentally.
Lines (1 feature) Often, parts of buildings (like
wooden doors or bricks) appear to have skin color.
One way to distinguish these from skin is that they
often have long straight edges which are not typical
of skin.
Figure 3: Example of line extraction. (Left) Example input
image with straight edges on the skin color regions. (Right)
Detected lines.
We use the OpenCV implementation (Bradski,
2000) of the probabilistic Hough transform (Kiryati
et al., 1991) computed on the edges of the skin color
connected components. The number of distinct lines
detected that are not within 5 pixels of the image
edge is recorded. An example of the detected lines
is shown in Figure 3.
3.2 Skin-Independent Features
The remaining 9 features do not use color informa-
tion.
Image Features (3 features) The first set of non-
color-based features describe the image shape and
size.
Images larger than 10 × 10 pixels have a flag fea-
ture set to one. If the image is larger than 10 × 10,we
also compute the remainder of the features described
here. Images that are smaller than 10 × 10 pixels
have the flag and all other features set to zero; most
of the features cannot be computed for such small
images, which are unlikely to contain recognizable
adult-content in any case.
Two additional features contain the log of the num-
ber of pixels in the ROI defined earlier, and the aspect
ratio of the image.
Entropy Features (2 features) One feature that
may be valuable in distinguishing adult-content im-
ages from icons or other artificial images is the en-
tropy of the intensity histogram of the image, as fol-
lows:
−
255

i=0
(P (i) · log
2
(P (i))) (3)
where P (i) is the fraction of pixels in the image
with intensity i. Artificial images tend to have fewer
distinct intensities present, so their entropies will be
lower. We compute the entropy for both the full im-
age and a 10 pixel wide border around the image, to
aid in the detection of framed images.
VISAPP 2006 - IMAGE ANALYSIS
292

Figure 4: Example of using face detection to improve the
skin color model. (Left) Sample input image. (Middle) Skin
map without customization using the face detector. (Right)
More accurate skin map produced by using face detection
(note that the neck is brighter and sweater is darker).
Clutter Features (2 features) Just as an edge de-
tector was used above to measure the amount of tex-
ture in the skin colored pixels, it can also be used as an
overall measure of the amount of clutter in an image.
We record the fraction of pixels in the ROI which are
edges (a measure of central clutter in the image), and
the fraction of edge pixels in the whole image which
are in the ROI (a measure of the peripheral clutter rel-
ative to the central clutter).
Face Detection (2 features) One of the most re-
liable signals that an image contains a person is
whether or not a face is present. For this work, we
used the OpenCV face detector (Bradski, 2000), an
implementation of the detector described in (Lienhart
and Maydt, 2003), which is in turn based on the work
in (Viola and Jones, 2001). From the face detection
results, two features are used: the number of faces in
the image, and the fraction of the image pixels which
are in the largest detected face. For speed, the face de-
tector is run on a half-resolution version of the input
image.
As mentioned earlier, the face detection results can
be used to customize the skin color model to a partic-
ular image, as illustrated in Figure 4. Once a face is
detected, pixels in the face region are used to build a
3-channel Gaussian skin color model. Then, when de-
tecting skin, each pixel’s skin probability is estimated
both from the model of (Jones and Rehg, 1998a) and
the face skin model, and these estimates are averaged.
In the complete system, all of the skin color depen-
dent features are computed twice, once with the face
present in the skin map, and once with the face re-
gion set to zero probability in the skin map, as shown
in Figure 5. The idea here is that large portrait pho-
tos of people’s faces will contain a large amount of
skin color on the face, but this skin should not count
against the image when trying to decide whether it
contains adult-content.
Figure 5: Example of blanking out the face region of the
skin map. (Left) Example input image. (Middle) Skin map
showing the face region. (Right) Skin map after face region
is set to zero.
3.3 Classification
In total, we employ 27 features; 9 which do not use
the skin map, and 9 which do (and are computed for
two versions of the skin map, with and without faces
present). Once the features have been computed for
an image, they are fed into a support vector machine
(SVM) for classification. In this work, the LIBSVM
library was used to train and run the SVM (Chang and
Lin, 2001).
4 TRAINING AND TEST DATA
The training and test data consisted of images found
on the WWW. A set of 17,300 examples was hand la-
beled for training as either adult-content or not. The
images in this set were selected from websites known
to be family-safe and those known to contain adult
content; images from the latter set were then manu-
ally labeled the authors as to whether each was adult-
content or not. Of these 17,300 samples, 812 were
adult-content and 16,488 were not.
The test set consists of 51,960 sample images uni-
formly collected from the Internet and manually la-
beled by the authors. Of these images, 1,331 were
labeled as adult-content, and the remaining 50,629 as
not adult-content. The test data was also manually
labeled by authors. The next section will show the
classification results for this data.
5 EXPERIMENTAL RESULTS
In this section, we describe the training of the SVM,
experiments to evaluate the contributions of the fea-
tures, and present comparisons with related work.
LARGE SCALE IMAGE-BASED ADULT-CONTENT FILTERING
293

5.1 Training
Because the number of positive adult-content train-
ing examples greatly outweighs the negative samples,
LIBSVM’s weight setting for the positive examples
was set to 18 during training. While this does not
exactly match the proportion of ratio of positive to
negative examples for historical reasons, the results
of the training do not seem very sensitive to this pa-
rameter, as long as the positive examples are given a
fairly large weight. RBF kernels with a gamma para-
meter of 0.05 were used, and all other parameters to
LIBSVM were left at their defaults. LIBSVM used a
total of 3,102 support vectors for the final classifier.
5.2 Testing
To see the impact of each set of features on the per-
formance of the system, a number of tests were con-
ducted. We began by training an SVM with only the
basic skin color features (the mean and standard devi-
ation of the skin map in the ROI), and then used the
SVM to produce scores for each of the test images.
From this data, and ROC curve plotting the number
of errors on adult-content vs. the number of errors on
family safe images could be produced. This curve is
the upper curve in the graph of Figure 6a. For this
type of graph, a better result means the curve comes
closer to the origin.
In steps, we then added batches of features, re-
trained the SVM and computed the resulting ROC
curve. The results are shown together in Figure 6a
(full range) and 6b (zoomed). Each curve represents
adding one batch of features (described in a para-
graph in Section 3.1) to the set of features used by
the previous curve, with the exception of the last two
curves. The “Face Detection” curve adds features for
the number of faces and largest face, modifies the skin
color model using the face pixel samples, and also
modifies the skin map by removing the face rectan-
gles. The “All Features” curve uses both versions of
the features, with and without the faces removed from
the skin map, which improves accuracy slightly.
As can be seen from Figure 6, as more features
are added, the accuracy improves. However, there
appears to be diminishing returns from the types of
straightforward features we are using.
We evaluated the speed of the algorithm on a cor-
pus of around 1.5 billion (1.5 · 10
9
) thumbnail images
of less than 150 × 150 pixels. Processing the entire
corpus took less than 8 hours using 2,500 computers,
for an overall throughput of around 20 images per sec-
ond per computer.
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Error Rate on Non-Adult-Content
Error Rate on Adult-Content
Impact of Adding Features
Skin Color Detection
Connected Component Analysis
Skin Texture Features
Lines
Image Features
Entropy Features
Clutter Features
Face Detection
All Features
(a)
0.15
0.2
0.25
0.3
0.15 0.2 0.25 0.3
Error Rate on Non-Adult-Content
Error Rate on Adult-Content
Impact of Adding Features
Skin Color Detection
Connected Component Analysis
Skin Texture Features
Lines
Image Features
Entropy Features
Clutter Features
Face Detection
All Features
(b)
Figure 6: This figure shows the ROC curves as more fea-
tures are added to the overall feature set. (a) shows the
curves at their original scale, while (b) is zoomed in to show
the differences between the feature sets more clearly.
5.3 Comparison with Related Work
Although we have not been able to locate any two pa-
pers which use the same test sets, it may be useful
to examine the reported error rates which have been
published. A numerical summary of related results is
shown in Table 1. Each of these numbers is drawn
from the paper, or estimated from an ROC curve pre-
sented in the paper. The same results are plotted on
a graph in Figure 7, along with the ROC curve from
our system. In addition, the ROC curve for “Zheng
(our data)” evaluates the work of (Zheng et al., 2004)
using its open source implementation in (di Linguis-
tica Computazionale, 2004) applied to our test set.
The results presented by (Arentz and Olstad, 2004)
are interesting because in addition to showing the ac-
curacy for a general set of images, results are also
reported for a test set containing portrait pictures of
people. As might be expected, the false alarm rate on
these pictures is higher, as they contain a significant
VISAPP 2006 - IMAGE ANALYSIS
294

Table 1: A summary of accuracy results (on different test
sets) from related work.
Citation Positive
error
rate
Negative
error
rate
(Ioffe and Forsyth, 1998) 49% 10%
(Fleck et al., 1996)
56% 2%
(Wang et al., 1998)
4% 9%
(Arentz and Olstad, 2004)
(all images)
4.6% 12%
(Arentz and Olstad, 2004)
(portraits)
4.6% 26.5%
(Duan et al., 2002)
19.3% 10%
(Zeng et al., 2004)
23.5% 5%
(Jones and Rehg, 1998b)
14.2% 7.5%
(Zheng et al., 2004)
9% 20%
amount of skin color while not being adult-content.
It is expected that our system’s use of a face detector
may eliminate some of these false alarms.
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Error Rate on Non-Adult-Content
Error Rate on Adult-Content
Comparison with Related Work
Ioffe
Fleck
Wang
Arentz (general)
Arentz (portraits)
Duan
Zeng
Jones
Zheng (their data)
Zheng (our data)
Our System
Figure 7: Graphical summary of accuracy results (on differ-
ent test sets) from related work, as well as the ROC curve
from our system for comparison purposes. The only direct
comparison which can be made is between Our System and
Zheng (our data), and it can be seen that our system has
significantly better accuracy.
As can be seen, there is a wide variety in the re-
ported accuracies of the systems that have been pub-
lished. Given that many of these systems are heav-
ily dependent on skin color models and features that
are very similar across systems, it is quite likely that
the wide range of differences in accuracy can be at-
tributed to differences in the training and test set dis-
tributions. This underscores the importance of cre-
ating a consistent test set with which to measure
progress in this field. Using the WWW as the source
of test images as we did creates a very difficult, real-
world test case. Not only is the content extremely di-
verse, but the quality, resolution, color balance, and
brightness vary greatly per image—increasing the dif-
ficulty of this problem dramatically. It is worth noting
that many of the false negatives in our results are due
to grayscale or cartoon adult-content images, which
cannot be detected by the algorithms described here.
6 SUMMARY AND FUTURE
WORK
The results above indicate that the system is able to
detect roughly 50% of the adult-content images in a
small test set, with roughly 10% of the safe images
being incorrectly marked as adult-content; or at a dif-
ferent threshold detecting 90% of adult-content im-
ages with a false alarm rate of 35%. However, since
safe images significantly outnumber those with adult-
content, this leads to a large number of false alarms,
indicating there is much work remaining to be done.
The system described above has been incorporated
into Google’s adult-content filtering infrastructure,
and is now in active use for image safe-search.
There are a number of potential ways to improve
this work. Face image processing can determine the
age (Lanitis et al., 2004) and gender (Moghaddam and
Yang, 2000; Baluja and Rowley, 2005) of people in
the image, which may turn out to be useful features,
perhaps simply as priors. More elaborate represen-
tations of the shape of the skin color blobs, such as
those used in (Wang et al., 1998) may be helpful with-
out significantly increasing the computational cost.
Better measures of texture beyond simple edge de-
tection may improve the recognition of skin regions.
Finally, image similarity matching against a database
of adult-content image signatures (using ideas similar
to (Tieu and Viola, 2000; Lowe, 2004)) may be use-
ful. Any image-based system can be used to comple-
ment the results of a keyword based filter, since each
approach has its own strengths and weaknesses.
To end on a practical note, because of the ubiquity
of the Internet, search engines, and the widespread
proliferation of electronic images, adult-content de-
tection is an extremely important problem to address.
To improve the rate of progress in this field, it would
be useful to establish a large fixed test set which can
be used by both researchers and commercial ventures.
Another avenue to pursue is the creation of an an-
nual competition for filtering systems. Unfortunately,
the images used in this paper were obtained from
the WWW, and likely contain copyrighted content,
so cannot be redistributed. Nonetheless, the authors
look forward to collaborating with other researchers,
both in terms of the algorithms and approaches taken,
as well as ideas to promote more interest and faster
progress in this field.
LARGE SCALE IMAGE-BASED ADULT-CONTENT FILTERING
295

REFERENCES
Arentz, W. A. and Olstad, B. (2004). Classifying offensive
sites based on image content. In Computer Vision and
Image Understanding.
Baluja, S. and Rowley, H. A. (2005). Boosting sex iden-
tification performance. In Innovative Applications of
Artificial Intelligence, Pittsburgh, PA, USA.
Bradski, G. (2000). Programmer’s Toolchest:
The OpenCV Library. Software available at
http://www.intel.com/research/mrl/
research/opencv/index.htm.
Canny, J. (1986). A computational approach to edge de-
tection. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 8(6).
Chang, C.-C. and Lin, C.-J. (2001). LIBSVM: a library
for support vector machines. Software available
at http://www.csie.ntu.edu.tw/˜cjlin/
libsvm.
di Linguistica Computazionale, I. (2004). POESIA:
Public Open-source Environment for a Safer In-
ternet Access. Software available at http:
//www.poesia-filter.org, alpha 2 release,
poesieasoft_snapshot_25july2003.tgz.
Duan, L., Cui, G., Gao, W., and Zhang, H. (2002). Adult
image detection method base-on skin color model and
support vector machine. In Asian Conference on Com-
puter Vision, pages 797–800, Melbourne, Australia.
Fleck, M. M., Forsyth, D. A., and Bregler, C. (1996). Find-
ing naked people. In European Conference on Com-
puter Vision.
Hunke, H. M. (1994). Locating and tracking of human faces
with neural networks. Master’s thesis, University of
Karlsruhe.
Ioffe, S. and Forsyth, D. (1998). Learning to find pictures
of people. In Neural Information Processing Systems.
Ioffe, S. and Forsyth, D. (1999). Finding people by sam-
pling. In International Conference on Computer Vi-
sion.
Jones, M. J. and Rehg, J. M. (1998a). Stastical color mod-
els with applications to skin detection. International
Journal of Computer Vision, 46(1):81–96.
Jones, M. J. and Rehg, J. M. (1998b). Stastical color models
with applications to skin detection. Technical report,
Compaq Cambridge Research Laboratory.
Kiryati, N., Eldar, Y., and Bruckstein, A. M. (1991). A
probabilistic Hough transform. Pattern Recognition,
2(4):303–316.
Kruppa, H., Bauer, M. A., and Schiele, B. (2002). Skin
patch detection in real-world images. In Annual Pat-
tern Recognition Symposium DAGM.
Lanitis, A., Draganova, C., and Chirstodoulou, C. (2004).
Compariong difference classifiers for automatic age
estimation. IEEE Transactions on Systems, Man, and
Cybernetics, 34(1).
Lienhart, R. and Maydt, J. (2003). Empirical analysis of de-
tection cascades of boosted classifiers for rapid object
detection. In Annual Pattern Recognition Symposium
DAGM.
Lowe, D. G. (2004). Distinctive image features from scale-
invariant keypoints. International Journal of Com-
puter Vision, 60(2):91–110.
Moghaddam, B. and Yang, M.-H. (2000). Sex with support
vector machines. In Neural Information Processing
Systems.
Rosenberg, C., Minka, T., and Ladsariya, A. (2003).
Bayesian color constancy with non-gaussian models.
In
Neural Information Processing Systems.
Sprague, N. and Luo, J. (2002). Clothed people detection in
still images. In International Conference on Pattern
Recognition.
Tieu, K. and Viola, P. (2000). Boosting image retrieval. In
Computer Vision and Pattern Recognition.
Viola, P. and Jones, M. (2001). Rapid object detection using
a boosted cascade of simple features. In Computer
Vision and Pattern Recognition.
Wang, J. Z., Li, J., Wiederhold, G., and Firschein, O.
(1998). System for screening objectionable images.
In Computer Communications Journal.
Zeng, W., Gao, W., Zhang, T., and Liu, Y. (2004). Image
guarder: An intelligent detector for adult images. In
Asian Conference on Computer Vision, pages 1080–
1084, Jeju Island, Korea.
Zheng, H., Daoudi, M., and Jedynak, B. (2004). Blocking
adult images based on statistical skin detection. Elec-
tronic Letters on Computer Vision and Image Analy-
sis, 4(2):1–14.
VISAPP 2006 - IMAGE ANALYSIS
296