A new radial basis function network classifier for holistic recognition of universal facial expressions

摘要

According to psychologists there are six types of universal facial expressions namely, "fear", "surprise", "anger", "sad", "disgust" and "happy". Holistic recognition of these facial expressions from static images requires nonlinear classifiers capable of operating on noisy high-dimensional feature spaces. Often radial basis function networks (RBFN) are used for classification in these applications. Conventional RBF networks however, in spite of their capabilities in working with high-dimensional feature spaces, often fail to deliver satisfactory performance in these scenarios due to small training sample sets, noisy features and/or features not following the required class structure. This paper presents an improved RBFN architecture that overcomes these problems through asymmetrical scaling of feature axes according to specific requirements of the class structure of the classification problem. The scaling factors are computed automatically from the available training samples, without any explicit analysis of their multivariate statistical properties. The proposed network yielded an overall recognition rate of over 92% for the 6 expression classes, and a smaller network size compared to other types of RBFN classifiers.