A Pornographic Video Recognition Algorithm Based On Semi-supervised Learning On Graphs

With the rapid development of network transmission technology and multimedia technology, people can easily and conveniently access to a variety of information resources, which changed people’s daily lives. However, the presence of a large number of porn video will have a great physical and mental stimulation on network users, especially for health and values of young people formed a serious negative impact. Therefore, how to effectively detect and block the bad information of the network has attracted more and more public attention, while the research of the related algorithms are also attracted more and more researchers’ attention.In the field of pornographic video detection, it is relatively difficult to acquire the complete foreground region in video shots. It is necessary for further research in order to study the difficulty comprehensively. Therefore, this paper presents an identification method of pornographic video based on semi-supervised learning on graphs. With the help of the continuity of the time between frames, the algorithm can acquire foreground region in the video accurately, and then detected the skin color and private parts of the foreground region in order to make the identification of video. The algorithm can effectively avoid the interference of background color which close to skin color, then improve the accuracy of detecting the inappropriate content of the video.This algorithm includes the following steps:first, make video shot segmentation,then obtains key frames from each video shot, and computes the inter-frame difference between key frames and adjacent front and rear multi-frame image, then merges these multi-frame image, extracts extract the moving foreground area section. Secondly, considering the moving foreground area section as priori information, it uses semi-supervised learning on graph method (linear neighborhood propagation) to extract a complete foreground region. Finally, establish a skin color model, then segment the skin color area from the full foreground region. By detecting the pornographic content of the skin color area to make a identification of key frame category and with the test results of the key frame to judge the video category. Found through experiments that the algorithm exhibits better robustness in the detection of poor video and can correctly detect pornographic video with a high accuracy (up to90%). Therefore, it can effectively detect and block undesirable video content in the network transmission.