Exploring Semi-Supervised Learning in Facial Expression Recognition

Semi-supervised learning methods are popular in facial expression recognition (FER). It solves the problem of in-sufficient large labeled datasets by using unlabeled data. However, the accuracy of semi-supervised FER is still unsatisfactory. This is because the existence of incorrect labels in pseudo-labels leads to noise. Moreover, the discriminative capability of the features extracted by a standard encoder is not apparent. To address the problem, we explore an SSL framework for FER. Our model improves the quality of pseudo-labels by a reasonable threshold setting. A consistent regularisation method is used instead of the pure pseudo-labeling method to reduce noise interference. Besides, channel attention mechanism is added to the backbone to extract more discriminating features. In that way, we improve prediction accuracy and data utilization. Experimental results on challenging datasets validate that our method is superior to the state-of-the-art methods.