Micro-expression Spotting Based On Main Directional Maximal Difference Analysis

Micro-expressions are facial expressions with the feature of short duration(generally con-sidered within 0.5 seconds,low intensity of muscle movement and involuntary appearence.They are regarded as significant cues to reveal one's hidden emotions.Although methods for spotting and recognition of general facial expressions have been discussed by numerous re-searchers,research on automatic spotting and recognition of micro-expressions is still relatively small,especially,little progress has been made in spotting micro-expressions.In this paper,a robust and effective Main Directional Maximal Difference Analysis(MD-MD)feature is proposed to obtain more accurate features of the movement of facial expres-sions for automatically spotting micro-expressions and macro-expressions from videos.Facial calibration,cropping and division mainly are completed by non-reflective similarity transfor-mation and shape model based on local landmarks.This MDMD method gains both the tem-poral and spatial locations of the facial movements based on block structured facial regions which considers action units.The computational complexity is reduced because the dimen-sion of the MDMD feature is small.The length of a MDMD feature vector for recognition of micro-expressions is only 72 and the length of a MDMD feature vector for spotting of micro-expressions is only 12,when the block structure is 6 x 6.The workload of micro-expression database coding is large,researchers need to extract a small number of valid micro-expression fragments from thousands of videos.This paper realizes the assistant coding system of micro-expressions,detecting the possible frame number of micro-expressions in the video,which reduces the waste of manpower and time.Evaluations on the CAS(ME)2 database containing micro-expressions and macro-expressions and the CASME database containing only micro-expressions show that MDMD is more robust,accurate than LBP algorithm.At last,a SVM classifier is trained to recognize micro-expression with the MDMD feature.Evaluation on two spontaneous databases called CASME 2 and CASME,demonstrating that the MDMD can achieve better performance than LB P-TOP feature.