Filtering objectionable contents for blocking nude images based on Bayesian color classification.
Skin detection refers to separation of image or video pixels that corresponds to human skin. It is very useful for the detection of human body parts such as faces, hands, etc. in numerous applications, such as face detection, face tracking, filtering of objectionable web images, and human-computer interaction to mention a few. While multimedia data are becoming popular in wireless environments, some methods try to group such multimedia data based on their semantic contents. Skin color can be an appropriate option for this type of semantic grouping or clustering. Internet has become readily accessible in most organizations such as schools, colleges, homes, office. The freedom of access makes the World-Wide Web (WWW) as the most popular place where people can search, deliver and exchange information. At the same time, the problem with pornography through the internet access in the workplace, at home and education has considerably increased. In recent years, problems on lack of control and regulation over what information can be made available, especially over the internet. Unknowingly images that would be illegal can be easily transferred to home through the internet, causing young person to see those offensive images intentionally or unintentionally. Therefore, the way of effectively blocking or filtering out pornography has awakened a serious concern in related research areas. Dealing with pornography in the workplace is a serious issues for many large organizations but hiring a block-all-images email policy no longer provides a feasible solution. Email is more media-based and it is common for business mail to contain images such as logos, publicity shots etc. In a commercial environment, an image analysis is required to automatically classify embedded or attached images as acceptable or inappropriate. The problem therefore is the non-retrieval of certain types of image. The rest of the paper is organized as follows. In section 2, the work done by other researchers for detecting objectionable objects are given. The proposed algorithm is specified in section 3. The results analysis of the proposed method is shown in section 4. Section 5 concludes the paper.
Shu Lee et al (2007) proposed a novel method that detects nude image based on adaptive and extensible skin color model, which determines the image skin chroma distribution online so that it can tolerate the chromatic deviation coming from special lighting without increasing false alarm.The performance of the naked image detection extremely depends on the accurate skin segmentation. There may be a lot of objects, possessing skin like chroma in an image, e.g. wood, foods etc. It is almost impossible to distinguish them from the human skin by using the chroma property. However, smoothness is a very important feature for skin and the roughness of the featuresare utilized to reject the non-skin objects with skin like chroma. There are many approaches for measuring the texture roughness in image processing (Fourier power spectrum, etc). Intensity of the image will be greater in the rougher object's surface. Three common properties are induced for naked images through investigating a mass of naked images. First, the naked body usually occupies a significant portion of an image. Second, the aspect ratio of the naked body is usually in a reasonable range. Third, the position of the naked body is close to the image center to harmonize with the frame. The features based on these properties are extracted to judge if the smooth skin region contains naked bodies. The location feature is used to express if the examined area is close to the image center. If the face-to-skin ratiois larger than a threshold value, the test image is viewed as a mug shot. Otherwise, the test image is considered a naked image. In terms of computational complexity, the face detection stage indeed increases the system's computation load. However, it can effectively filter out those false positives resulting from mug shots in the classification stage. The detection rates of the naked images and the non-naked images are 71.7% and 84.2%, respectively. A.A.Zaidan et al (2014) proposed a new method for detecting skin pornography based on the multi-agent learning neural and bayesian methods. The Bayesian method is used with a grouping histogram technique again to extract the features from the skin detection based on YCbCr and RGB color spaces and the back propagation neural network method using shape features extracted again from skin detection. The anti-pornography system works on four machine learning methods in two different stages namely skin detector stage and pornography classifier stage. A multi-agent learning is used twice. In the first stage, a multi-agent learning method is proposed that combines the Bayesian method with a grouping histogram (GH) technique and the back-propagation neural network with a segment adjacent-nested (SAN) technique based on the YCbCr and RGB color spaces to extract skin regions from the image accurately by taking into consideration the problems of the light-changing conditions, skin-like color and reflection from glass and water. In the second stage, the features from the skin are extracted to classify the images into either pornographic or non- pornographic. Accurate classification occurs when different image sizes are used in the existing anti- pornography systems. The classification of the pornographic images becomes more robust to the variation in images sizes. The result shows that the proposed multi-agent learning system for skin detection has produced a significant rate of true positives (TP) (i.e., 98.44%). In addition, it has achieved a significant low average rate for the false positives (FP) (i.e., only 0.14%) while the proposed multi-agent learning for pornography classifier has produced significant rates of TP (i.e., 96%). Moreover, it has achieved a significant low average rate of FP (i.e., only 2.67%).
Ahmad Ali Abin et al (2009) presented a new skin detector algorithm using anew dynamic cellular learning automata. Because of the various image processing tasks, there is no optimum method that can perform properly for all applications. A novel skin detection algorithm combines color and texture information of skin with cellular learning automata to detect skin-like regions in color images. The output of this part is fed into a texture analyzer which extracts texture features from the skin-like region. Previous works achieve a high TP rate but with a high FP rate, but this proposed skin detector obtains candidate skin regions with a TP of about 86.3% versus a FP rate of 9.2% regions on Compaq skin database which shows its efficiency. Rehanullah Khan et al (2014) demonstrated a systematic skin segmentation algorithm by merging spatial and non-spatial data. The detected faces are used as foreground seeds for calculating the foreground weights of the graph. If local skin information is not available, they opt for the universal seed. With this setup, robust skin segmentation is achieved, outperforming off-line trained classifiers. Experiments on two datasets with annotated pixel-level ground truth show that the systematic skin segmentation approach outperforms other approaches and provides robust skin detection.
Alison Bosson et al (2002) extended their earlier work on detecting pornographic images. The image classifier described in this paper is integrated into a mail-based security product MAILsweeper which is a content security solution that sits at an SMTP gateway, assessing email traffic entering and leaving a company and protecting the organization from mail-borne threats such as viruses, breaches of confidentiality, offensive email content, legal liability and copyright infringement etc. MAILsweeper disassembles emails into their components, for example, zipped email attachments will be unzipped. These are then analyzed according to user-defined policies which may be company-wide, department-wide or unique to an individual employee. Jinfeng Yang et al (2007) proposed a method for classifying human skin based on geometric feature. Content-based image filtering is particularly essential for providing a healthy Internet environment for our society, since the images always conveniently convey some un-healthy information from websites. Different algorithms have therefore been developed for content-based skin image classification. First, a nonparametric skin color classifier is used to skin detection. Second, a region splitting scheme has been designed to generate adaptive grids on images. Based on the grids the outline of skin regions are constructed using a curve evolution method based on adaptive grids. After that the geometric features are extracted from the contours, and the cosine similarity measure is adopted for image classification. The algorithm is tested on a large scale database consisting of 6000 images. Experimental results illustrate that the proposed method perform well in classifying skin images.
Luca Giangiuseppe Esposito and Carlo Sansone (2013) proposed a method for detecting a naked human bodies in images using a multiple classifier approach. Content-based image retrieval methods are capable of analyzing image visual features like colors and shapes. Millions of people upload multimedia contents on the web in their day to day life. Images are published without control and only if users report a problem then offensive content are removed. So, there is a need for the system that automatically processes such kind of offensive data. As compared to other state-of-the-art proposals, it demonstrated the effectiveness of the proposed approach. Abadpour and S. Kasaei (2005) proposed a technique for filtering pornography using pixel-based skin detection method. Twenty one color spaces were used for pixel-based skin detection in pornographic images. Consequently, this paper holds a large investigation in the field of skin detection, and a specific run on the pornographic images. Each color space is considered in all its seven possible representations. The examination includes the measurement of the best performance for classifying the skin pixels in the training dataset plus the real performance in highlighting skin areas in the samples of a large pornographic datasets. Two approaches, ordinary and Baysian LUT-based skin detection are evaluated. The results shows that the Nbr is the best choice for ordinary LUT-based skin detection. Also, it was observed that the best solutions for Baysian LUT-based skin detection are 1-D color spaces utilizing the Baysian approach is proved to result in higher correct rate classification, while it also increases the possibility of false positive classification.
Huicheng Zheng et al (2004) demonstrated a way of blocking adult imagesusing statistical skin detection aiming in the detection of adult images that appear in Internet. Skin detection is important in the detection of adult images so a maximum entropy model called the First Order Model is built. Parameter estimation as well as optimization cannot be tackled without approximations. With Bethe tree approximation, parameter estimation is removed and the Belief Propagation algorithm permits to obtain exact and fast solution for skin probabilities at pixel locations. Plenty of experimental results are presented including pictures and a ROC curve calculated over a test set of 5,084 pictures, which show stimulating performance for such that was based on the Bayesian decision theory and developed using a large training set of skin colors and non-skin colors. Two homogeneity measures for skin regions that took into account both global and local image features were also proposed. The proposed skin segmentation technique performed better than fixed threshold pixel-level skin color segmentation. In 2003, Chengjun Liu (2001) presented a novel Bayesian Discriminating Features (BDF) method for multiple frontal face detection.
The Bayesian decision rule for minimum cost was a well-established technique in statistical pattern classification (Phung, 2001; Vladimir Vezhnevets Sazonov, 2003). The class-conditional probability density function was estimated using histogram or parametric density estimation techniques. The value of Pskin(c) computed in (1) was actually a simple features. To improve the results one can use a face detector. Manynon-parametric methods had been proposed for skin color modeling. Besides histograms often used non-parametric methods are skin probability maps and neural networks. In 2000, D. Chai and A. Bouzerdoum (2000) described an image classification technique that used the Bayes decision rule for minimum cost to classify pixels into skin color and non-skin color. The Bayesian model to skin color classification was discussed along with an overview of YCbCr color space. Using the Bayes decision rule for minimum cost, the amount of false positives and false negatives could be controlled by adjusting the threshold value. The results showed that this approach effectively identifies skin color pixels and provide good coverage of all human races. In 2000, D. Chai and A. Bouzerdoum (2005) presented a Bayesian approach to skin color classification. In 2002, M. J. Jones and J. M. Rehg (2002) described the construction of color models for skin and non-skin classes from a dataset of nearly 1 billion labeled pixels. Color distributions for skin and non-skin pixel classes learned from web images could be used to fashion a surprisingly accurate pixel-wise skin detector. Visualization studies showed a surprising degree of separability in the skin and non-skin color distributions. They also revealed that the general distribution of color in web images was strongly biased by the presence of skin pixels. A pixel-wise skin detector could be used to detect images containing naked people, which tended to produce large connected regions of skin. As compared to other skin detection techniques, the proposed skin classifier was much faster. In 2003, S. L. Phung, et al., (Phung, 2001) proposed anew skin segmentation technique for color images. The proposed technique used a human skin color model conditional probability
P(c|skin) - a probability of observing color c, knowing that a skin pixel was seen:
[p.sub.skin](c) = skin(c)/Norm (1)
where skin(c) gave the value of the histogram bin, corresponding to the color vector c and Norm was the normalization coefficient (sum of all histogram values, or maximum bin value present. The normalization value of the lookup table bins constituted the likelihood that the corresponding color corresponded to skin.
A more appropriate measure for skin detection was p(skin|c) - a probability of observing skin, given a concrete c color value. To compute this probability, the Bayes rule was used:
p(c/skin) = [(2[pi]).sup.-d/2] [[absolute value of ([C.sub.S])].sup.1/2] exp [-1/2[(c - [m.sub.s]).sup.T] [C.sup.-1.sub.s] (c - [m.sub.s])] (2)
p(c|skin) and PC|nonskin) were directly computed from skin and non-skin color histograms.
The prior probabilities p(skin) and p([??]skin) were estimated from the overall number of skin and non-skin samples in the training set. An inequality p(skin|c) >= [tau], where t is a threshold value, used as a skin detection rule. This means that p(skin) value affected only the choice of the threshold [tau].
The proposed algorithm is tried with eleven color spaces CIE, HSV, LMS RGB, YES, XYZ, YCgCr, YCC, YCbCr, YDbDr and YPbPr using the Bayesian model producing the better results. Of all the color spaces used YPbPr produces the best results which is shown in the table I. Skin detection was an important process in computer vision and used in different applications such as human face detection face tracking and adult content filtering. The results of their implementation showed that the likelihood Bayesian method was a good approximation of the best pixel based skin detector. Currently they were trying to analyze other pixel based skin detectors in this model. The comparative analysis of skin detection results using Bayesian color model are shown in the table I.
[FIGURE 1 OMITTED]
The proposed algorithm is tried with eleven color spaces CIE, HSV, LMS, RGB, YES, XYZ, YCgCr, YCC, YCbCr, YDbDr and YPbPr using the Bayesian model producing the better results. Of all the color spaces used YPbPr produces the best results which is shown in the table I.
[FIGURE 2 OMITTED]
[FIGURE 3 OMITTED]
To evaluate the performance of this proposed novel skin segmentation techniques, simulation is carried out on XM2VTS Extended Multimodal Face Data base containing more than 100 color images together with images downloaded from for skin segmentation. In addition, manually prepared ground-truth images are used for skin segmentation and face detection. Many non-parametric methods have been proposed for skin color modeling. The improper selection of color space and threshold values lead to false positives and true negatives. Hence, a statistical model like Bayesian model can be explored to overcome these problems. The Bayesian color model produces good results. This algorithm need to be fine tuned with regard to the threshold values for the specific database being used to produce better results rather than using the threshold values specified in the developed algorithms. The future researchers can address this open research problem on the ideal threshold values suitable for the database at hand.
Received 23 July 2015
Accepted 28 August 2015
Available online 25 September 2015
Abadpour, A., S. Kasaei, 2005. Pixel-Based Skin Detection for Pornography Filtering, (July 2004).
Ali, A., A. Mehran, F. Shohreh, 2009. A new dynamic cellular learning automata-based skin detector, pp: 309-323. doi:10.1007/s00530-009-0165-1
Bosson, A., G.C. Cawley, Y. Chan, R. Harvey, 2002. Non-retrieval : Blocking Pornographic Images, pp: 50-60.
Chai, D. and A. Bouzerdoum, 2000. "A Bayesian Approach to Skin Color Classification in YCbCr Color Space.", Intl. Proceedings IEEE Region Ten Conference (TENCON'2000), 2: 421-424.
Esposito, L.G., C. Sansone, 2013. A Multiple Classifier Approach for Detecting Naked Human Bodies in Images, pp: 389-398.
Jones, J. and J.M. Rehg, 2002. "Statistical Color Models with Application to Skin Detection," Int'l J. Computer Vision, 46(1): 81-96.
Khan, R., A. Hanbury, R. Sablatnig, J. Stottinger, F. Ali, K. Fakhri, A. Khan, 2014. Systematic skin segmentation : merging spatial and non-spatial data, pp: 717-741. doi:10.1007/s11042-012-1124-y
Lee, J., Y. Kuo, P. Chung, E. Chen, 2007. Naked image detection based on adaptive and extensible skin color model, 40: 2261-2270. doi:10.1016/j.patcog. 2006.11.016
Phung, S.L., A. Bouzerdoum, D. Chai, 2005. "Skin Segmentation Using Color Pixel Classification: Analysis and Comparison", IEEE Transactions On Pattern Analysis and Machine Intelligence, 27(1).
Phung, S.L., D. Chai and A. Bouzerdoum, 2001. "Skin Color Based Face Detection", Seventh Australian and New Zealand Intelligent Information System Conference, 18-20 November 2001, Perth Western Australia, pp: 171-176.
Vladimir Vezhnevets Sazonov, V. and A. Andreeva, 2003. "A Survey on Pixel-Based Skin Color Detection Techniques", Proc. of Intl. Conf. in GraphiCon'2003, Moscow, Russia.
Yang, J., Y. Shi, M. Xiao, 2007. Geometric Feature-Based Skin Image, pp: 1158-1169.
Zaidan, A.A., N.N. Ahmad, H. Abdul Karim, M. Larbani, B.B. Zaidan, a. Sali, 2014. On the multi-agent learning neural and Bayesian methods in skin detector and pornography classifier: An automated anti-pornography system. Neurocomputing, 131: 397-418. doi:10.1016/j.neucom.2013.10.003
Zheng, H., M. Daoudi, B. Jedynak, 2004. Blocking Adult Images Based on Statistical Skin Detection, 4(2): 1-14.
Dr. Ravi Subban
Dept. of Computer Science, School of Engineering and Technology, Pondicherry University, Pondicherry
Corresponding Author: Dr. Ravi Subban, Dept. of Computer Science, School of Engineering and Technology, Pondicherry University, Pondicherry
Table I: Comparative Analysis of Skin Detection Results using Bayesian Color Mode Methods/Color Space YCbCr CIE HSV RGB YIQ YUV YES Jones 90 Brand 93 S.L. Phung 90 Phung 84 Phung 91 91 91 L. Sigal 99 Brand 93 Leonid Sigal Jayaram 91 91 91 91 Shahreza 91 Phung 76 Our Method 99 94 98 94 98 99 97 Methods/Color Space NRGB CIE-XYZ YCgCr KL CIE-Lab YESRGB Jones Brand S.L. Phung Phung Phung 91 91 L. Sigal Brand Leonid Sigal 99 91 Jayaram 91 Shahreza Phung Our Method 98 99 99 95 98 99 Methods/Color Space YESRBYB YPbPr YDbDr Jones Brand S.L. Phung Phung Phung L. Sigal Brand Leonid Sigal Jayaram Shahreza Phung Our Method 98 99 93
|Printer friendly Cite/link Email Feedback|
|Publication:||Advances in Natural and Applied Sciences|
|Date:||Aug 1, 2015|
|Previous Article:||Detecting wrinkles and sagginess of skin from human face photographs.|
|Next Article:||Analytical comparison of waterfall model and object-oriented methodology in software engineering.|