Recognizing Microexpression as Macroexpression by the Teacher-student Framework Network

Yaqi Song, Wei Zhao, Tong Chen, Shigang Li, Jianfeng Li

Abstract: Microexpressions have great research value because of their high reliability in revealing human emotions. However, microexpression recognition is challenging due to its low amplitude. In this paper, we propose a teacher-student framework for microexpression magnification. First, we input the macro-expression samples into an action transfer model first order motion model (FOMM) for image animation to extract the variation features from neutral expression to macroexpressions as the teacher model. Similarly, we extracted features from neutral expressions to microexpressions. Then, we designed a dual-channel encoder-decoder network, which was guided by the middle-layer feature map of the teacher model, to learn the jump from microexpression features to macroexpression features. To ensure the magnification effect and image quality, we introduced a self-attention mechanism and loss functions. Finally, an end-to-end microexpression magnification model (MEMM) was constructed. We fed our magnified images directly into a simple ResNet50 network for recognition, achieving a competitive score under the MEGC2019 standard, compared with recent complex recognition networks. Our model can also be applied in the wild for microexpression magnification.