A Comparison of Single and Multi-Class Classifiers for Facial Expression Classification

This paper compares various Kernel, neural network and statistical approaches for the task of novel facial classification. In addition to multi-class classifiers, it explores the use of the single-class methods known as autoassociators, including the recently proposed Kernel Autoassociator which reconstructs feature vectors from Kernel feature space to input space. Comparisons are made using feature vectors composed of image pixel values, as well as image points convolved with Gabor filters. Results show an advantage to using multi-class methods, with linear disriminant analysis and Kernel based approaches providing the best results.