• Embed Doc
  • Readcast
  • Collections
  • CommentGo Back
Download
 
Research on the Discrimination of Pornographic and Bikini Images
Yushi Wang
1
, Weiqiang Wang
2
, Wen Gao
1, 2
1
 Dept. of Computer Science and Engineering, Harbin Institute of Technology, Harbin,China, 150001
2
 Institute of Computing Technology, Chinese Academy of Science, Beijing, China, 100080
 
 E-mail: {yswang, wqwang, wgao}@jdl.ac.cnTel: +86 10 58858300 Fax: +86 10 58858300-399
 
Abstract
Our research extends the general technologiesdetecting pornographic images to prevent the benignimages whose content is approximate with the pornographic ones from being screened. This paper  presents a multiple-step method to distinguish the decent bikini photos from pornographic photos. The proposed approach utilizes the information about the body shapeand face to determine the feature region where the specific patterns of the skin distribution of bikini model’sbody are represented, and uses a neural network to learnthese patterns and recognize the bikini photos.Theexperimental results show that more than 80% normalbikini photos will be recognized from the pornographicimages by the method, and at the same time few pornographic images are mistaken as bikini photos. Theresults indicate it is possible to distinguish the bikini photos from pornographic ones.
1. Introduction
Nowadays the Internet is infested with a large volumeof objectionable contents, such as pornographic images,which are very harmful to people’s bodies and mind,especially to teenagers. It is a meaningful and urgent task to develop an effective detection and filtering system toprevent the access to unexpected pornographic imagesthrough analyzing image content. Much research effortsabout the topic can be found in the recent literatures. Theearliest work was done by Forsyth and Fleck [1]. Theyfirst located those skin regions with the shape of cylindersafter skin detection, and then checked whether thecylinders as body parts can be reasonably assembledtogether to form a nude human body. Wang et al. [2]developed a multi-filter system. Their system firsteliminated icons, graphs or the photos with few skin colorregions, then extracted the edge distribution informationbased on the high frequency coefficients of theDaubechies-3 wavelet transform on an image, and finallya 28-dimensional moment feature vector derived from theedge information was exploited to compare with those of training images. Zeng et al. [3] also developed apornographic image filter, Image Guarder, and acombination of the skin feature, texture feature and shapefeature was used by the system. Bosson [4] presented afiltering system called PORNsweeper. Their system firstidentified the skin pixels with high skin probabilities, thena region-growing algorithm used them as the seeds toobtain skin blobs. Five informative features wereevaluated and four different classifiers’ performanceswere compared and discussed. Jeong et al. [5] exploitedthe appearance information of porn images. They used theskin likelihood ratio of every pixel in a resized image asthe feature and a SVM classifier was applied.It is obvious that most detection systems of pornographic images are based on low-level visualfeatures, but generally these low-level features will fail todistinguish benign images with large skin regions frompornographic images. The experiments in our researchshow that the portrait images with large skin regions are asignificant part of the images that are wrongly classifiedas pornographic images. Until now, no literature is foundresearching on distinguishing those benign images withlarge skin areas and pornographic photos specially, andsome researchers dealt with that problem incidentallywhen they designed their filters. For example, one featureused by Bosson’s system was the fractional area of skinthat is accounted for by a face [4], which was helpful toeliminate the portrait pictures, especially the headportraits.Among portrait images, the bikini photos are
 
very common on the Internet and also the mostchallenging kind as they are visually very close to thosepornographic images. Since the conventional approachesusually classify bikini photos as pornographic images, asystem that is able to distinguish bikini images from theporn ones will further the capabilities of the technology toaccurately detect pornographic images.This paper presents an approach to distinguish thebikini photos with nearly frontal postures frompornographic images. This work can be integrated into apornographic image filter as a module. In thepornographic image filtering, the module will help toprotect the commercial activities having relations to bikinishows, such as the advertisements of bikinis or thepropaganda of beauty competitions, andlower the falsepositive rate in detection. As a beginning classifier, thiswork can’t recognize the complex cases far from frontalpostures well. However, the future work may process thecomplicated cases based on it, which will be discussed inthe last section. It needs to be pointed out that the bikiniimages in which the bikini models’ postures are socomplicated that some bikini pieces are covered by thebody are out of our consideration, because it’s hard forthe machine to tell whether the bikinis are worn or not. Inaddition, in most conditions the target of the filteringsystem is to find the potential pornographic sites on theInternet, and whether a certain potential adult image isclassified correctly is not crucial. When a considerableproportion of the photos from a site are bikini photos withnormal postures, it can be looked on as a benign web sitehaving relations with bikini shows.In the proposed method,after normalizing theorientation of the human body, a feature region is firstlylocated through analyzing the information about the bodyshape and face, and then related features are extractedfrom the feature region to characterize skin spatialdistribution. Aneural network is learned to recognize thebikini photos. The remainder of this paper is organized as follows.Section 2 presents the details of the proposed algorithm.The experimental results are given in section 3. In section4, we conclude this work.
2. The Proposed Algorithm
The proposed algorithm is directed to recognize thosebikini images from the images that are classified aspornographic by one pornographic image filter similar tothose systems mentioned above, to refine its systemperformance as a post-processing module. Some benignimages which are not bikini model pictures may also beclassified as pornographic by the filter. That some of these images are classified as bikini ones coincidentallyby our module will also help the filter to detectpornographic images more accurately, so we needn’t careabout those images and pay most attention todistinguishing bikini photos from the true pornographicimages.We outline the working procedure of the proposedalgorithm in Fig. 1. The module first detects the skinpixels to form the skin regions in the image so that thebinary skin mask image can be obtained. When themodule is integrated in a filtering system, this step can beskipped. If we find the model’s body is slantwise throughthe miniature of the skin mask image that reflects theskeleton of the body at a very low resolution, we have torotate the body into the upright direction. Subsequently, afeature region will be located on theorientation-normalized skin image through a projectionhistogram reflecting skin vertical distribution, the faceinformation and the holes detected in the body region if necessary. Then the classification features can beextracted from the informative region containing the mainpart of the body. A 256-dimensional feature vector is usedto characterize spatial distribution of skin pixels within it.Finally we train a neural network to recognize bikiniimages from those pornographic images in which afeature region exists. The proposed approach can onlyclassify the bikini photos in which models are in nearlyfrontal postures, but the whole bodies of models aren’tnecessarily fully seen in bikini photos.Skin region segmentationPosture normalizationFeature region Faceselection detectorFeature extractionClassification by a neural net
Figure 1.
 
The flow chart of recognizing the bikiniphotos
 
2.1 Skin Segmentation
 
To make use of the information about the bikinimodel’s body, the proposed approach first segments skinregions in the image. Any robust skin detection algorithmcan be used in the module, but we have chosen theadaptive skin detection algorithm presented by Zhang etal. [6], since the experimental results show the algorithmhas a lower computation cost compared with otheradaptive skin detection algorithms. It not only has an
 
87.1% true positive rate and a 10.3% false positive rate,but also has the post-processing that tends to form skinregions with large areas, which is preferred by our module.After skin region segmentation, we obtain a binary skinmask image
skin
 M 
, on which most work in the followingsteps is based.
2.2 Normalize Body Orientation
 
Sometimes the orientation of model’s body in thebikini image is not upright, for example, it leans againstsome place. To make the extracted features be invariant tothe body orientation, we normalize the orientation of skinregions to the upright posture. First, we downsample thebinary skin mask image
skin
 M 
into a miniatureimage so that only the global shape information is left.Next, the morphological closing operator is used to fill upthe holes caused by the covering of bikinis that are nearthe centroid of skin regions in the miniature image. Thenthe Hough transform is performed on the miniature image.Fig. 2(d) gives an example of the result of Houghtransform, the Hough transform matrix
 H 
, and we candeduce the main direction of body by analyzing thematrix.
×
1616In the matrix
 H 
, each column corresponds to an angle,and each element
 H(d,a)
is the number of the skin pixelswhich distribute on the line whose distance to the imageorigin is
, and the angle between the line’s normal andthe X axis is
a
. We use the following formulae todetermine the main direction of the body.
=
()(((,)))__
i
caiHiaskinpixelnum
()(((),)(
w
clusteraHcaiaweighti
= +
(1)))
iw
=−
(2)_argmax(())
a
maindirectionclustera
=
(3)The
skin_pixel_num
is the number of skin pixels in theminiature image after morphological closing operation.The
c(a)
corresponds to the position of body’s centroidwhen you look at the body in the direction
°
+
90
a
, and itis actually a distance to the origin. A window whosewidth is
2w+1
is used to eliminate the effect of the skinpixels far from
c(a)
. The
weight(i)
whose index is from
 –w
to
w
will decrease rapidly as
i
increases like theGaussian function where
w
is concerned with the skinarea. We can deduce the main direction of body directlyfrom the function , as is shown in Fig. 2(e).When the main direction is obtained, we rotate theoriginal skin image to allow the main direction of thebody to be upright. Now, we have the model’s body withan upright posture.()
clustera
 To obtain the
 BW 
, the estimation of the width of thehuman body, we calculate the mean of skin pixel numbersof all the columns in the image. The column number of the first column from left to right in which the sum of skinpixels is bigger than the mean is called
left_skin
, and the
Figure 2. An example of normalizing the bodyorientation. (a) The original skin mask image. (b)The miniature image. (c) The miniature imageafter morphological closing operation. (d) TheHough transform matrix
, where the brightnessrepresents the element values. (e) The selectionof the main orientation of the body.
2.3 Selection of the Feature Region
 
Intuitively the difference between bikini images andpornographic images lies in whether sexual organs areexposed. So we only consider a specific square region thatis expected to cover the chests and lower abdomen. Wecall the square region a feature region, since classificationfeatures are only extracted from it. For the rotated skinmask image, the algorithm first estimates the positions of sexual organs based on the histogram of verticaldistribution of skin pixels, which can be obtained throughcounting skin pixels on each row of the skin mask image,namely each row corresponding to a bin of the histogram.For bikini images two distinct valleys are expected to beseen in the histogram due to the covered sexual organs, asindicated in Fig. 4(d). A function
V p
is defined tomeasure the distinctness level of a valley at the position
 p
 in the histogram.() 
α 
+< =
()()()()max((),())()0
 LVpRVpifLVpRVpLVpRVpVpotherwise
 (4)
= − − ∫ 
()((()())(()())())
 p p
 LVpUfxfpfxfpGxpdxB
 (5)
()((()())(()())())
 p p
 RVpUfxfpfxfpGxpdxB
+
= − − ∫ 
 (6)
=
__
 BWrightskinleftskin
(7)
=>
0()10
U
0(8)
of 00

Leave a Comment

You must be to leave a comment.
Submit
Characters: ...
You must be to leave a comment.
Submit
Characters: ...