A Deep Learning Approach for Subject Independent Emotion Recognition from Facial Expressions

摘要

This paper proposes Deep Learning (DL) models for emotion recognition from facial expressions. We have focused on two "deep" neural models: Convolutional Neural Networks (CNN) and Deep Belief Networks (DBN). For each of these DL neural models, we have chosen several architectures. We have considered both the case of subject independent emotion recognition and also that of subject dependent emotion recognition. One has selected the Support Vector Machine (SVM) as a benchmark algorithm. We have chosen the JAFFE database to evaluate the above proposed models for person independent/dependent facial expression recognition. Using DL approach, we have obtained a subject-independent emotion recognition score of 65.22%, corresponding to an increase of 6% over the best score given by the considered benchmark methods. For person dependent emotion recognition, the DL model leads to the recognition score of 95.71%, representing an increase of 3% over the best of chosen benchmark methods.