Content area
Full Text
ABSTRACT
Internet pornography (Internet Porn) is addictive to teenagers and kids around the world. The normal practice is to block those websites or filter out pornography from kids. Many research papers have been published how to detect a human pornographic image on web pages. A new technique was proposed here to classify pornography images using range of YCbCr (colour feature) and three new measurements: %Face_Area, %AHB and Rmax. A computer algorithm such as C4.5 was applied to construct a decision tree. The main contribution of this paper is simplicity that retains high accuracy within an acceptable processing time. The accuracy of experimental results was 85.2% with an average processing time of 0.21314 seconds per image.
Keywords: Pornographic images, Data Classification, Skin Detection, Internet Porn
1. INTRODUCTION
Pornographic images on the Internet has become a serious problem for parents around the world. To address this issue, the normal practice is to block websites or filter out nude images from web pages. This paper presents a strategy to detect pornographic images. Firstly, if nudity means the state of being without or having little coverage with clothing, then range of YCbCr is a good way to detect human skin in images.
Detecting a face using color information is the second step. According to Chai and Ngan1, face location can be identified by performing region segmentation with the use of a skin color map. The presence of Cr values within, 136 and 156, and Cb values within 110 and 123 indicates a human face region. After performing face detection, if the largest face area is greater than the specified threshold, the image is classified as a non-nude image2. In this paper, we classify this type of picture as a Selfie/close-up face type. The first measurement called %Face_Area is defined in step 2 (shown in section 3.2). If it is not a close-up image, then we can calculate a %AHB (Area of Human Body) value by expanding the scope of YCbCr to a full range as shown in section 3.3.
This program was developed by utilizing MATLAB software, Version 7.1 on 15,000 sample images with an average size of 405,318 pixels or 86 Kb. The experimental results showed that accuracy was 85.2%. Average processing time was 213.14 msec. per...