TV advertisement detection for news channels using Local Success Weighted SVM Ensemble

Research in advertisement detection (in news broadcast videos) has mainly focused on development of efficient audio-visual features. These features are used with standard machine learning algorithms to automatically segregate the advertisements. However, the discriminative abilities of features are local and may not have uniform performance through out the input space. Thus, using a fixed set of features for entire input space limits the performance of classifiers. In this paper, we use a Local Success Weighted Ensemble of SVMs (LSWE-SVM) for advertisement detection. The LSWE-SVM ensures the diversity in errors of component SVMs by training them on individual features with different similarity measures (kernels). The weight of each SVM is determined by an instance dependent “success prediction function”. The success prediction functions predict high values for a particular exemplar, if the corresponding base SVMs have high likelihood of predicting the correct label of the exemplar. During training, the success prediction functions are estimated using support vector regression (SVR) trained on exemplars from cross validation sets of respective SVMs. The target for SVRs is set to 1.0 for the successfully classified exemplars and 0.0 otherwise. Given a test pattern, SVMs having high likelihood of predicting the correct label for the pattern are only allowed to contribute in the ensemble decision, thus suppressing the false decisions. Experimentations on over 150 hours of TV Broadcast news dataset have shown the superiority of LSWE-SVM over other baseline methods in terms of balanced F-score and generalization capability.