Research On Pornographic Image Detection Fusing Image Semantics And Text Information

While people enjoy the convenience of modern life brought by such a wealthof information, there are some potential negative influence along with it. Forexample, the obscene information has a serious influence on the mental health ofteenagers. For the reason that pornography is the most important carrier of obsceneinformation, the detection technology of pornographic information has become animportant aspect of internet information security.Although the direct content analysis method is widely used in networkpornographic image detection, it is still not an effective way, because theidentifying information of pornography could be existed in the content of images,or corresponding text such as file name and page information. These informationcan be used to express the semantic of pornography better and improve theaccuracy of detection. Up to date, the research in this field is limited and lack ofessential fusing of image content information and image-related text information.For the reason that most common solutions to pornographic image detectionproblem are inadequate, a new algorithm based on both text information andsemantic features of pornographic image is proposed. Based on Bag-of-Visual-Words model, this algorithm fuses the text features and visual features of images toenrich word bags, and then uses Support Vector Machine classifier to classifypornographic images from normal images.This paper carries out the research in two aspects: feature extraction fromimage content and image-related text information. Image content features aregenerated by analyzing visual elements, such as color, texture, local shape, and theskin detection. Image-related text features are generated by matching keywordsfrom premade keyword library. An algorithm of fusing these two kinds of featurewith variable weight is proposed. Finally, the complete pornographic imagedetection algorithm is proposed. In addition, a pornographic image detectionprototype system based on the proposed algorithm is designed and implemented.Through the comparing experiment using image set collected online andmanually labeled, the proposed algorithm achieves a higher accuracy ofpornographic image detection than the algorithms presented in existing literature.