Shape-based facial expression classification using Angular Radial Transform

Facial expression classification research has been heavily dominated by appearance-based methods. However, changing facial expressions influence shape of the face and facial features. Therefore, exploiting shape information is of paramount importance for developing robust facial expression classification systems. In this study, to benefit from shape information efficiently, we investigated the use of an advanced shape descriptor, namely the Angular Radial Transform (ART), in classification of six prototypical emotional expressions. In the work of Hung-Hsu Tsai [1] they already used ART combined with gabor filtering in the feature extraction step only for an edge-based image representation. We however applied ART on two different and more advanced image representations: regions around 83 manually annotated points on the face, and regions around eyes and mouth. The experiments conducted on the BU3DFE database have shown that employing ART to extract shape-based features provides very high performance. This indicates that, while utilizing shape information for facial expression classification, it is essential to use an advanced shape descriptor-in contrast to use simply the normalized vertex locations as in the previous studies.