Real Time Approaches for Video Genre Classification

This technology for video-genre classification developed at TU Berlin provides applications for personal video recorders and for an automatic classification of audio-visual data for the internet and for search engines. A main challenge in the field of multimedia content analysis is the transformation of human interpretations of audio-visual data to correlating machine processable representations.

This invention analyses such contents with the help of high-level audio-visual descriptors and of classification methods. To give an example: If you wish to distinguish between the genres ‘cartoon’ and ‘non-cartoon’, the performance of several combinations of the newly developed descriptors and classification methods is analyzed. A classification accuracy (CA) of 95.6 % is achieved by using two hidden Markov models (HMM). For the distinction between the features ‘commercial’ and ‘non-commercial’ a CA of 98.4 % can be achieved by using two HMM. A decision tree results in a CA of 95.5 % for music videos and a CA of 91.9 % for news. When distinguishing between ‘sport’ and ‘non-sport’, Bayes’ Theorem achieves a CA of 95.2 %.