Automatic Detection Of Micro-expression Based On Double Region Feature Selection From Long Videos

Facial expression is an extremely important way to convey emotional information.Micro-expression is a leaked facial expression when people try to hide or suppress their real emotions,with the characteristics of short duration and low intensity of facial muscle movement.Micro-expressions can truly express people's inner emotional states.Therefore,it is more reliable to identify people's emotions through micro-expressions.At present,the research on micro-expression work focuses on the recognitio n,and less micro-expression spotting research has been conducted.Micro-expressi on spotting means locating the moment of micro-expression occurrence from a pi ece of video,which is the pre-step of micro-expression recognition.There are tw o problems in the existing work related to micro-expression spotting: depend ver y much on labelded database;affected by macro-expression and other forms of movement(traditional detection algorithms tend to regard the frame withthe large st feature difference from the reference frame as the apex frame of micro-expres sion).Aiming at solving the problems in the existing research,this thesis proposes a double region feature selection method to detect the spontaneous micro-expressions from long videos.This method is independent of frame labels of the database,and can locate the moment of micro-expression occurrence from long videos.It firstly extract the image features: selecting the image texture features(LBP)and image motion information(optical flow characteristics);Then,a double region feature selection method was proposed according to the characteristics of the facial muscle changes of microexpression in the duration of time: the difference of facial features being used for feature selection.Experimental results on databases CASME?,SMIC-E-HS,SMIC-E-VIS,and SMIC-E-NIR show that the proposed method can achieve ASR(apex frame rate)of 18.86% ? 26.75% ? 49.29% and 43.69%,higher than traditional method(Optical Strain).Meanwhile,the method shows good performance on different features(LBP and Optical Strain)and different databases.