Facial Expression Recognition Using Hybrid Discriminant Analysis

We propose a novel set of features extracted using Hybrid Discriminant Analysis (HDA) for use in a facial expression recognition task. HDA is formulated by combining the optimization problems which are solved in Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) methods. By tuning weighting parameters, new sets of discriminant directions are obtained. Using these directions for feature extraction, we show that we can obtain excellent recognition rates of 95.9% on a database of 6 expressions plus neutral which is much better than that obtained using PCA or LDA. The performance of HDA is comparable to that of Gabor wavelet based methods. However, the computation time required to form a HDA feature vector is significantly lower than the time required to form a Gabor feature vector making this method very attractive for real-time applications.