Micro-expression Recognition Based On Three-stream Convolution Neural Network With Feature Enhancement Network

Facial expression contains rich emotional information,which is an important part of human daily life.Facial expression can be divided into macro-expression and micro-expression.The macro-expression lasts for 3/4 to 2 seconds and can be easily detected on the whole face.However,macro-expression can be easily controlled and giving false emotional information to people.In contrast,micro-expression only lasts for 1/25 to 1/5 of a second.It appears unconsciously on face,cannot be concealed,and represents the real emotion.Therefore,the recognition of micro-expression is of great significance.Because of the short duration and the low intensity of micro-expression,it is very difficult to recognize micro-expression with naked eyes.To automatically recognize micro-expression is of significance in theory and practice.At present,most of micro expressions recognition research extracts information from the whole face.However,in most cases,micro-expressions only have a facial muscle movement on a small part of the face.Therefore,in the process of micro-expression recognition,a lot of information that is not related to micro-expressions is extracted,which affects the recognition result and increase the computation load.In order to solve these problems,this paper proposes a method based on three-stream convolution neural network with feature enhancement network(SETFnet)to recognize Spontaneous micro-expression.Three streams in the network input images of three different local facial regions(left eye + left eyebrow,right eye + right eyebrow and mouth),which can reduce the amount of computation and reduce the information irrelevant to micro-expression.In addition,we add squeeze-and-excitation(SE)networkblock in the network,which can enhance the effective features and suppress the useless features,so as to improve the recognition results.Finally,we test our proposed method on SMIC and CASME? databases,and the results show that the proposed method can effectively improve the recognition result.