Cumulative Differential Gabor Features for Facial Expression Classification

Abstract

Emotions are written all over our faces. Facial expressions of emotions can be possibly read by computer vision and machine learning system. Regarding the evidence in cognitive science, the perception of facial dynamics is necessary for understanding the facial expression of human emotions. Our previous study proposed a temporal feature to model the levels of facial muscle activation. However, the quality of the feature suffers from various types of interference such as translation, scaling, noise, blurriness, and varying illumination. To cope with such problems, we derive a novel feature descriptor by expanding 2D Gabor features for a time series data. This feature is called Cumulative Differential Gabor feature (CDG). Then, we use a discriminative subspace for estimating an emotion class. As a result, our method gains the advantages of using both spatial and frequency components. The experimental results show the performance and the robustness to the underlying conditions.