Affective classification in video based on semi-supervised learning

Abstract

In the previous work of affective video analysis, supervised learning methods are frequently used as classifiers. However, labeling abundant examples is time consuming and even impossible for it needs annotation from human beings. While unlabeled video clips are easy to be obtained and they are adequate. In this paper, we present a semi- supervised approach to recognize emotions from videos. Firstly, visual and audio features are extracted. Then bivariate correlation is used to select sensitive features. After that, low density separation, a semi-supervised learning algorithm, is adopted as the classifier. The comparative experiments on our own constructed database showed that the semi-supervised algorithm performs better than supervised one, illuminating the effectiveness and feasibility of our approach. © 2011 Springer-Verlag.