Research On Violent Video Recognition Technology Based On Multi-modal Fusion

With the rapid development of the Internet technology, more and more videos appear on the Internet. However, violence in movies has harmful influence on children. So it is necessary to label this kind of video data. This paper explores the problem from the view of video content, presents the feature extraction method based on text, audio, video and the classification algorithms based on SVM(Support Vector Machines).Therefore, this paper focuses on text category classification and the fusion of the video and audio, as well as its application to violent movie recognition. The main work of this paper is summarized as follows:First, the text, video, audio data of various types of films is collected in the video sharing sites. After the text preprocessed and the shots of films cut, The data obtained form the network video database.Second, a based text violent movie recognition approach is proposed. The main content of this approach is two-fold. First, the violence characteristic thesaurus is collected and built. Second, useful texts around the video are crawled, including the introduction of the videos, the comments of audiences, etc. And the texts are preprocessed, the feature of text is extracted, and each text is represented as a vector space model. Eventually, the classifier is trained, and the test samples are classified into violent or nonviolent texts.Finally, a violent movie recognition method of fusing multi-modal information is studied. The text information is utilized to build a pre-classifier which selects the potential violent movie segments. At a second stage, a classifier is adopted, which combines the visual and audio information, in order to classify the potential violent movie segments as “violent” or “non-violent”. The experiments show that the method based multi-modal fusion greatly reduces the space complexity and improve the recognition accuracy.