Semi-supervised Learning for Compound Facial Expression Recognition with Basic Expression Data

Automatic facial expression recognition serves as a crucial technique in human-machine interaction. Existing works on facial expression recognition mainly focus on recognizing basic expressions while paying far less attention to compound expressions. Even worse, the scale of compound facial expression data remains small. Labeling compound expression data requires the annotators to be equipped with prior psychological knowledge, and it is a time-consuming and labor-intensive task. Fortunately, large-size labeled basic expression databases are available. As basic expression images can potentially be compound expressions, they can be used for compound facial expression recognition. To achieve this goal, in this work, we propose a semi-supervised learning framework to generate pseudo-compound-emotion labels for basic expression data. This approach aims to increase the number of training data and improve model capability for compound facial expression recognition. Our method further explores leveraging the basic labels by introducing a basic-label smoothing mechanism. Experimental results on the RAF-DB and the EmotioNet compound subset demonstrate significant improvement achieved by the proposed method over the baseline methods.