Feature selection based on co-clustering for effective facial expression recognition

Facial expressions are considered to be an effective and non-verbal means of expressing the emotional states of humans in more natural and non-intrusive way. Automatically recognizing the emotions consequently contributes towards the advances in the areas such as human computer interaction, clinical psychology and data-driven animations. Deriving a relevant and reduced set of features is a vital step for effective facial expression recognition. In this paper, we propose a co-clustering based approach to the selection of distinguished and interpretable features to deal with the curse of dimensionality issue. First, the features are extracted from images using a bank of Gabor filters. Then, a co-clustering based algorithm is designed to seek distinguishable features based on their non-inclusive information in co-clusters. Experiments illustrate that the selected features are accurate and effective for the facial expression recognition on JAFFE database and the best recognition rate is obtained by using selected features with SVM for classification. Moreover, we illustrate that the selected features not only reduces the dimensionality but also identify the distinguishable face regions on images amongst all expressions.