Facial Micro-expression Recognition with Adaptive Video Motion Magnification

Abstract

Facial micro-expression commonly has an extremely short duration and subtle motion. At the same time, the micro-expression databases are rarely available for research. These are not conducive to directly introduce the deep neural network which is currently outstanding in the field of image recognition into facial micro-expression recognition. Recently, some researchers use the video motion magnification technology to magnify micro-expressions and achieve a good performance of enhancing the recognition results. This paper follows this idea and introduces an innovative way to adjust the amplification rate adaptively instead of the previous manual way. In addition, we use the extended continuous frames, instead of a single apex frame, to extract a concatenate feature map for the final classification. We have demonstrated through a series of experiments on the CASME II database that our method can effectively improve the accuracy of facial micro-expression recognition.