• Embed Doc
  • Readcast
  • Collections
  • CommentGo Back
Download
 
 DETECTING PORNOGRAPHIC IMAGES
 Liang K.M., Scott S.D., Waqas M.
Asia Pacific Institute Of Information TechnologyLot 6, Technology Park Malaysia,Bukit Jalil, 57000 Kuala Lumpur, Malaysia.Email: liangkm, simon, waqas@apiit.edu.my 
ABSTRACT
This paper describes a system for determining whether images are “pornographic” or not. The system locates all the skin pixels within an image and groups them together as skin blobs by using colour, texture and edgeinformation. Multiple features are then extracted from the skin blobs and used as an input to a MLP neural network classifier. The accuracy of the classifier and theimportance of the individual features in the classification process, as investigated through a series of experiments,are discussed.
 
1. INTRODUCTION
Advances over the last decade in digital image captureand data repository technologies, along with widespreadInternet access, have significantly increased theinformation flow amongst the general populace. Althoughthere are numerous benefits to be derived from suchinformation sharing, some less ennobling material, suchas pornography, is also free available. To ensure thatcomputing resources are being used constructivelysystems that accurately identify and block such materialwithout manual intervention are required.There are many commercial systems designed to prevent access to pornographic material. Most of thesesystems, such as NetNanny, CyberSitter, CyberPatrol andChildWebGuardian block websites based on comparisonof IP addresses / URLs and text contained within the siteswith a long list of objectionable addresses and texts intheir system databases. Whilst this approach is effectivefor blocking well-known porn sites and pages of  pornographic links, it fares less well in blocking pagescontaining galleries of pornographic images since thesefrequently do not contain links to other pages or objectionable text.It is has been shown that there is a strong positivecorrelation between the percentage of skin within animage and its pornographic content [1]. Although the useof the percentage of skin within an image as the solemeans of determining whether an image contains pornographic content is not very accurate, such anapproach has been implemented in commercial systemssuch as ScreenShield, Snitch, System Recon andEnologic NetFilter Home. These systems are effective asmonitoring software in which suspect images andassociated information, such as the user and the time of the event, are stored for inspection by a humanadministrator after the event. In other words thesesystems are designed to deter the viewing of  pornographic material, rather then actively blocking thematerial. By including features other than the percentageof skin in the image it is envisaged that the classificationaccuracy of images as pornographic or non-pornographiccan be increased, and so decrease the amount of humanintervention required in such systems.This paper is organised as follows. Section 2addresses related work by other researchers in detecting pornographic images. Section 3 details the skin blobdetection and feature extraction algorithms used in thesystem. Section 4 describes the construction of theoptimum classifier and evaluates the effectiveness of thisclassifier as well as the individual features based on theFalse Acceptation Rate (FAR) and False Rejection Rate(FRR) results. Section 5 concludes the paper and outlinesthe current direction of our research.
2. RELATED WORK 
Given the strong correlation between the percentage of skin and pornographic content within an image [1], thefirst step in all pornographic image classification systemshas been the identification of skin. However, there are anumber of potential barriers to accurate skinidentification [2, 3]:(a) Poor image quality; for example extremely lowcontrast.(b) Existence of non-human objects in image whosecolour closely resembles human skin.
 
(c) Desaturation of skin area caused by the skinreflectance and illumination problems.
 
To overcome the above barriers, a number of competing algorithms based on different colour spaceapproaches such as YCbCr, HSV, RGB, CIE Lu*v* andLog opponent have been proposed for the identificationof skin coloured pixels [1, 2, 3, 4, 5, 6]. However, it has been proved that for all colour space approaches,optimum skin detectors have the same performance sincethe separability of the skin and non-skin pixels isindependent of the colour space chosen [7]. Thus,choosing the best colour space so as to optimise theaccuracy of skin detector is no longer an issue. Whilst the proposed algorithms do represent an improvement over simply defining a static set of pixel colour values as beingskin colour, they are unable in themselves to overcomethe above barriers.A number of groups have proposed additionalfeatures to support the identification of pornographicimages. Forsyth and Fleck [2] used a body geometricfilter to detect the presence of human structures such aslimbs from the regions of skin pixels. Images containingsufficiently large skin-coloured groups of possible humanstructures are labelled as pornographic. The WIPE systemdeveloped by Wang et al. [3] utilised moment descriptorsto capture shape information within images. The momentfeatures were computed using edges derived fromDaubechies’ wavelet transform.
3. OVERVIEW OF THE PROPOSED SYSTEM
Skin Detection ModuleImageDecision ResultFeature Extraction ModuleDecision Classifier ModuleSize and Palette Analysis Module
 Figure 1: The four modules within the system.In the proposed system thirteen features are extractedfrom the detected regions of skin in the image and areused to train a MLP neural network in order to derive anoptimum and generalised classifier for the purpose of determining whether pornographic content is present or not. Although some of the extracted features are similar to those in related studies [1, 2, 3], this is the first timethat all the features have been combined in the detectionof pornographic images. The system itself is divided intofour modules: Size and Palette Analysis, Skin Detection,Feature Extraction and Decision Classifier Modules, asillustrated in Figure 1.The Size and Palette Analysis Module acts as a fastand basic filter. If the image dimension is below a pre-determined threshold, for example 32x32 pixels, then theimage is automatically labelled as non-pornographic sinceit is probably an icon or bullet image. Similarly, an imagethat contains only a limited number of colours (less thanone hundred) is labeled non-pornographic since it isunlikely to be photographic image. Only images that passthrough the first module are tested in the subsequentmodules.The Skin Detection Module uses colour, texture andedge information to identify potential skin pixels.Connected skin pixels are grouped together, forming blobs (regions) of skin and non-skin. The details of thismodule are described in Section 3.1.In the Feature Extraction Module thirteen featuresare extracted from the detected skin blobs. The details of these features and the approaches to their extraction aredescribed in Section 3.2. Prior to the Decision Classifier Module, an optimum classifier has to be determined. Inorder to obtain this classifier the features derived from amanually labelled image data set are used to train theMLP neural network classifier.The optimum classifier is used in the DecisionClassifier Module for determining whether the image is pornographic or not.
3.1 Skin Blob Detection
The six steps used to detect skin blobs are shown inFigure 2. First, image contrast enhancement throughhistogram normalisation is applied [8]. This step is usefulas it significantly improves the accuracy of subsequentsteps, particularly where the images are washed out or aretoo dark.
Image EnhancementSkin Tone Colour DetectionSkin Texture DeterminationSkin Region ExpansionSkin Region SegmentationSkin Blob Detection
 Figure 2: The six steps for detecting skin blobs.Second, pixels are labelled as skin pixels if their RGB values lie in a skin-tone look-up table. The tableitself was derived by the manual selection of skincoloured pixels in a test set of images.
 
Third, the texture of the pixels is determined in order to eliminate those pixels with skin colour but high textureamplitudes (skin regions are generally smooth and sohave small texture amplitudes). The co-occurrence matrixtechnique was used as it represented the best trade off  between accuracy and computation time [8]. Figure 3illustrates the effectiveness of texture determination inobtaining accurate skin pixels for pornographic and non- pornographic images. From the second column of thisfigure it may be seen that the skin tone colour detector wrongly labels some clothing and background regions asskin pixels in the images. Incorporation of the texturedeterminer re-labels non-skin pixels appropriately, as can be seen in the third column of this figure.(a)(b)Figure 3: Application of the skin tone colour detector and texture determiner on (a) a pornographic image,and (b) a non-pornographic image. Probable skin pixels are highlighted in black. The first columnshows the original images; the second column depictsthe probable skin pixels based on the skin tone colour detector; and the third column depicts probable skin pixels after application of the texture determiner.Fourth, the “probable skin” pixels based on thetexture determination are expanded to include neighbour  pixels with similar colour and texture properties. Thisstep allows the inclusion of the following into the skinregions:(a) Skin coloured pixels that are mislabelled in thetexture determination due to their proximity toedges within the image.(b) Non-skin coloured pixels from valid human skinregions that have been mislabelled due todesaturation.
 
Fifth, the skin regions are segmented by applyingadaptive Canny-edge detection and labelling pixels thatlie on edges as non-skin. This is done to ensure thatadjoining skin regions, for example an arm lying on topof a torso, remain distinct when constructing skin blobs.Finally, sixth, skin blobs are formed by groupingtogether the connected skin pixels. Small skin blobs, andsimilarly non-skin blobs, are removed by re-labelling the pixels as non-skin and skin respectively.
3.2 Feature Extraction
If skin blobs are detected in the image then the thirteenfeatures, as shown in Table 1, are extracted. If no skin blobs can be found then the image is classified as non- pornographic.Feature 1 is simply the percentage of pixels that have been labelled as skin within the image at the end of theskin blob detection module. Feature 2 is the percentage of colours in the image that are classified ‘skin’ colours as inthe look-up table. It is believed that images containing alarge amount of human skin would contain a greater number of ‘skin’ colours compared to images with littleor no human skin. For efficiency, this measure iscomputed using a 64x64x64 array of buckets, each bucketrepresenting a 4x4x4 region in the RGB colour space.These first two features were used as approximations of the features proposed in related work by HP labs [1]. Feature 3 is simply the number of skin blobs in theimage. It was felt that images with large number of skin blobs would probably correspond to group scenes and so probably not be pornographic.Feature 4 represents an approximation of Forsyth andFleck’s body geometric filter, which indicates the presence of human structures [2]. In order to determinethe number of limb like objects in the image a fastthinning algorithm was used on the skin blobs to revealtheir underlying skeletons. Potential limbs, defined as thelong straight parts of the skeleton whose length anddistance value ratios are similar to human limbs, werethen counted. The examples of limb objects located usingthis technique are shown in Figure 4.Features 5-13 relate to the size and shape of thelargest skin blob. If a naked person is present then it isassumed that the largest skin blob will correspond to that person. Features 5 and 6 relate to the ratio of the widthIndex Feature Remarks1 PercentageSkinPercentage of skin pixelsin the image2 PercentageSkin Colour Percentage of coloursassociated with skin tones3 Blob Number  Number of blobs in theimage4 Limb Number  Number of limb objectsfound in image5 Ratio BlobHeightRatio of height of the largest blob to the image height6 Ratio BlobWidthRatio of width of the largest blob to the image width7 - 13 Hu Moments Hu Moment (1 –7)Table 1: The thirteen features extracted from theimages.
of 00

Leave a Comment

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