Facial expression recognition using switch instance-batch normalization

Abstract

Facial expression recognition is widely used in video surveillance, assisted driving, and distance education. With the application of deep learning technology in facial expression recognition, many previous studies have shown good performance, but they are mainly identified in small-scale datasets with limited samples, lacking generalization ability for a wide range of scenes, and can’t meet the practical application needs. Previous studies have shown that Instance Normalization (IN) exhibits strong performance in terms of appearance invariance. In this paper, the normalization layer of facial expression recognition network is studied extensively, and a Switchable Instance-Batch Normalization (SIBN) method is proposed to balance feature appearance variance and content semantic information. The method was verified in three commonly used expression datasets CK+, Oulu-CASIA, and MMI. The experimental results show that SIBN stability improves the recognition accuracy of the model on a single dataset, and greatly enhances the performance of network cross-domain identification. The experimental results demonstrate the superiority of the proposed method.