2013, pp 471-479
Date: 03 Aug 2012

Mutual Information-Based Emotion Recognition

* Final gross prices may vary according to local VAT.

Get Access


Emotions that arise in viewers in response to videos play an essential role in content-based indexing and retrieval. However, the emotional gap between low-level features and high-level semantic meanings is not well understood. This paper proposes a general scheme for video emotion identification using mutual information-based feature selection followed by regression. Continuous arousal and valence values are used to measure video affective content in dimensional arousal-valence space. Firstly, rich audio-visual features are extracted from video clips. The minimum redundancy and maximum relevance feature selection is then used to select most representative feature subsets for arousal and valence modelling. Finally support vector regression is employed to model arousal and valence estimation functions. As evaluated via tenfold cross-validation, the estimation results achieved by our scheme for arousal and valence are: mean absolute error, 0.1358 and 0.1479, variance of absolute error, 0.1074 and 0.1175, respectively. Encouraging results demonstrate the effectiveness of our proposed method.