Semi-supervised learning of deep difference features for facial expression recognition

Abstract

Facial expression recognition (FER) is an important means of detecting human emotions and is widely applied in many fields, such as affective computing and human-computer interaction. Currently, several methods for FER heavily rely on large amounts of manually labeled data, which are costly and not available in real-world applications. To address this problem, this paper proposes a semi-supervised method based on the deep difference features. First, a cascaded structure is introduced to the original safe semi-supervised SVM (S4VM) to solve the multi-classification task. Then, multiple deep different features are fed to the cascaded S4VM to train the six basic facial expressions using the information of the unlabeled data safely. Extensive experiments show that the proposed method achieved encouraging results on public databases even when using a small labeled sample set.