TV news reviews are of strong interest in media and communication sciences, since they indicate national and international social trends. To identify such trends, scientists from these disciplines usually work with manually annotated video data. In this paper, we investigate if the time-consuming process of manual annotation can be automated by using the current pattern recognition techniques. To this end, a comparative study on different combinations of local and global features sets with two examples of the pyramid match kernel is conducted. The performance of the classification of TV new scenes is measured. The classes are taken from a coding scheme that is the result of an international discourse in media and communication sciences. For the classification of studio vs. non-studio, football vs. ice hockey, computer graphics vs. natural scenes and crowd vs. no crowd, recognition rates between 80 and 90 percent could be achieved.
展开▼
