A novel approach to video transition detection based on a two-phase classification strategy

Video transition detection plays an important role in many tasks of video analysis. Aiming at the target application of commercial detection in news video, this paper tackles the problem in a unified framework and proposes an alternative classification strategy. Our method is made up of two phases. In the first stage, SVM is employed to classify the transitions into three classes: non-transition, cut, and big-transition. In the second stage, we concentrate on the discrimination of the rapid motion situation and gradual transition, which is based on another set of features. Apart from the new classification strategy proposed in this paper, imbalanced data classification issue is also taken into consideration, which has not received sufficient attention by previous work. Experimental results show that our method is effective.