Multimedia genre characterisation with fuzzy embedding classifiers

Multimedia classification is a key issue in modern data management, where the number of available items is dramatically growing and there is an increasing demand for access to distributed multimedia data. Selection by genre is a simple and effective mechanism for most of the users interested in these applications. In this paper, we present a feature extraction architecture and a novel learning algorithm for multimedia genre characterisation. We show how genre classification can be regarded as a sub-case of this general task, for which we give a complete solution. Our extracted features were designed to offer a reduced semantic gap, trying to take into account structural and cognitive content descriptors, rather than low-level features. Our learning algorithm is based on fuzzy set theory, and makes use of fuzzy C-means (FCM) algorithm as the kernel to learn concepts configurations from data. We tested our learning framework on a test database of over 100 hours of TV broadcast programmes belonging to 7 different common genres. Experimental evaluations showed the effectiveness of our approach. Additionally, we compared our technology with neural networks applied on the same task, in terms of training accuracy. We also compared the generalisation performances of our technique with neural networks and support vector machines.