Micro-expression Detection Based On Convolutional Neural Network And Sliding Window Processing

Facial expression conveys rich emotional information.Micro-expressions are special facial expressions that are leaked when people fail to hide or suppress their real emotions.Micro-expressions have the characteristics of short duration,low intensity and only involving part of the face.Since micro-expressions are not under people's autonomous control,they are more reliable for judging human real emotions.Therefore,micro-expressions have potential applications in lie detection,psychological treatment,public security and other fields.Nowadays,most of the works related to micro-expression focus on micro-expression recognition,while the research on micro-expression spotting lags behind.Micro-expression spotting means locating the moment of micro-expression occurrence from a piece of video,which is a previous step of micro-expression recognition.In addition,automatic micro-expression spotting can also overcome the problems of time-consuming when manually labeling micro-expressions.At present,the related work of micro-expression spotting uses manual design features,which relies heavily on the experience and knowledge of researchers.What is more,traditional detection algorithms tend to regard the frame with the largest feature difference from the reference frame as the apex frame of micro-expression,which tends to affected by other forms of facial movement(such as blinking).To solve the aforementioned problems,this paper proposes a novel method based on convolutional neural network and sliding window processing to detect the spontaneous micro-expressions from long video.This is the first time that deep learning is used to explore micro-expressions detection.Firstly,convolutional network(SMEConvNet)is trained to automatically extract feature vectors from long video.Then,combined with the temporal characteristics of micro-expression presentation,the feature matrix processing method based on sliding window was used to detect the apex frame to avoid the influence of blinking and other movement for micro-expression detection.Experimental results on CASME Ⅱ show that compared with the traditional detection algorithm(LBP,Optical Strain),the proposed method can achieve MAE(Mean Absolute Error)12% and 18.5% higher than LBP and Optical Strain.