Well-Defined Semantic Templates for Pornographic Images Identification

Skin-color is proved to be powerful to pornographic images identification. However, skin-color is not determinant but heuristic. If we find documents with the same topic and some same keywords, we usually regard them as the same class. Based on this motivation, we proposed a well-defined semantic template, including topic template and bag-of-visual-words, to incorporate global semantics and local semantics for pornographic images identification. Topic template is designed to capture the patterns of whole images, such as areas of skin-color and some objects or scenes related to pornographic content. Visual words are intended to find some patterns of pornographic parts or poses which can be determinant to pornographic images identification. Our semantic templates can be applied for concept understanding and content analysis.