A Comparative Study of Facial Emotion Classification

Abstract

Face plays an important role in communication so automatic recognition of facial emotion is an important addition to computer vision research. This paper presents a comparative study of two types of approaches to facial emotion classification on single images. Gabor wavelets is the technique we employ here to extract features from upper face and lower face. Linear discriminant function (LDF) is applied first for classification. In this type, we compare principle component analysis (PCA) and Fisher linear discriminant (FLD). In the second type of neural network, we focus on multi-layer perceptron (MLP), where single big MLP and multiple MLP classifier are compared. The experimental results show PCA outperforms FLD, and multiple MLPs classifier beats single big MLP. In addition, it is also indicated that Gabor coefficients at high frequency and vertical orientation may contain more information about facial emotion.