Identifying Emotion in Static Face Images

We use combinations of feedforward networks trained to recognize emotions in face images to achieve excellent generalization. Networks trained with an input encoding of face features (eyes and mouth) achieved about an 84% generalization rate on novel faces. A similar encoding technique applied to the entire face with the same number of parameters achieved only a 60% generalization rate. This suggests that the actual representational scheme used by the brain to identify emotions may consist of face features rather than the entire face.