Anti-pornography algorithm based on multi-agent learning in skin detector and pornography classifier

An automated computerized algorithm for identifying and blocking pornographic content was designed. Primitive information on pornography is studied and used to determine if a given image falls under the pornographic category. In this thesis the pornography image is defined as image that contains human body exposed between neck and knee area. Skin regions are extracted from images as the first stage. The skin color has been used to detect skin as it is quite a simple and straightforward task. In addition, color has processing time advantage, since color processing is faster compared to other features. However it is not robust enough to deal with complex image environments, such as the light-changing conditions, skin-like colors and, reflection from glass and water. These factors could create major difficulties for pixel-based skin detector especially when the color feature is used. Thus, in this part of the research a novel multi-agent learning is proposed using Bayesian method with grouping histogram technique and back propagation neural network with segment adjacent-nested (SAN) technique based on YCbCr and RGB color spaces respectively, to improve the skin detection performance. Then, the features from the skin are extracted to classify the images as either pornographic or non-pornographic. Inaccurate classification occurs when different image sizes are used in the existing anti-pornography algorithms. Thus, in this part of the research a novel multi-agent learning is proposed between the Bayesian method using color features extracted from the skin detection based on YCbCr color space and the back propagation neural network method using shape features extracted again from skin detection. The classification of the pornographic images becomes more robust to overcome the problems in relation to variation in images sizes and this attainment was previously not accomplished by others.