Content-based Porn Image Detection Technique

In recent years, the unlimited Internet pornography has become a prominentproblem that will make much harm to the healthy growth of the teenagers and socialstability. Therefore, conduct indepth research on pornography is necessary. Con-tent-based methods can effectively overcome the shortcomings of traditional methodsbased on text filtering. Thus, content-based methods have become worldwide focusand mainstream research direction.At present, in the field of porn image detection, methods based on Bag-of-Wordsmodel (BoW model) have achieved relatively ideal results, however, there are alsosome problems in this kind of methods:1) On one hand, many methods take little ad-vantage of skin color information in the pornography image, on the other hand, due tothe inherent shortcomings of BoW model, the image representation vector cannot re-flect the spatial information of local features, in other words, the BoW image repre-sentation is unordered.2) In traditional methods based on machine learning, pornog-raphy image detection is treated as a binary classification problem, however, the im-balance of training data limits the further improvement of such methods. In allusion tothe problems mentioned above，this paper put up two methods within the frameworkof BoW model. The main innovations and characteristics of this paper are as follows:1. Propose a feature extraction and image representation method which mixestogether with skin color information and local features’ spatial information. Firstly, tobring in skin information, divide the image into dense regular grids and perform skincolor detection on these grids before feature extraction. Secondly, in the step of imagerepresentation，to bring in spatial information of local features, use spatial pyramidmatching to instead of basic BoW model.2. Propose a porn image detection method based on Support Vector Machine(method I). Firstly, extract features on training sets that consists of porn and normalimages. Secondly, train a two-class classifier on the training sets using SVM. Exper-imental results show that proposed method has excellent performance. What’s more,both skin and local features’ spatial information are helpful to the improvement of theaccuracy in porn image detection.3. Propose a pornography image detection method based on one-class classifica-tion methods (method II). In this methods, after analysis the peculiarity of pornogra- phy image detection, it can be found that one-class methods is suitable to solve thetask of pornography image detection. In other words, method II try to train a one-classclassifier on training set that only contain pornography image. Firstly, in the step ofimage feature extraction, use the same strategy as in method I. Secondly, in the step ofimage representation, the visual dictionary is trained by pornography images, which isdifferent from method I. Thirdly, after get the original BoW representation vector ofimages, use random forests to compute the importance of variables in the vector andkeep the index of important variables. Fourthly, perform vector optimization on orig-inal BoW vector and get the optimized vector. At last, we train a one-class classifieron optimized vectors. Experimental results show that the one-class classifier trainedby a small-scale porn image training set can achieve good performance.