Research On Algorithms For Micro-expression Feature Extraction

Micro-expression is a brief and momentary facial expression which reveals genuine emotions involuntarily even when people want to conceal at some high-stakes situations,leading to its strong applied value in many areas of life.The research on micro-expression has been a very challenging work for its three distinguished characteristics,which are short duration,low intensity and typically local movements.So this paper has studied and proposed some related methods for micro-expression feature extraction and applied them into micro-expression detection and recognition.The main work is summarized as follows:(1)Explored and analyzed four classic feature extraction algorithms:LBP-TOP,STLBP-IP,DiSTLBP-RIP and MDMO.LBP-TOP is an effective dynamic texture feature extraction algorithm which contains both spatial and temporal information.STLBP-IP extracts appearance and subtle motion features using LBP operator and integral projection based on difference-image.DiSTLBP-RIP extracts RPCA-based discriminate spatiotemporal LBP features for micro-expression recognition.MDMO extracts a main directional optical flow feature for micro-expression recognition.(2)A micro-expression apex frame spotting method based on RHOOF feature is proposed.In order to extract more effective features,this proposed method first detects facial landmarks and partitions the facial region into some specific ROIs according to the coordinates of facial landmarks,then extracts HOOF feature to get the motion statistics in each ROI.The apex frame of a micro-expression sequence is detected by the change of motion direction.Finally,the experimental results have verified the accuracy of this proposed method in micro-expression apex frame spotting.(3)A micro-expression recognition method based on MHOOF feature is proposed.Considering that the onset and apex stages can describe micro-expression more effectively and less impact of head movements will be generated with shorter duration,this proposed method extracts features just from onset to apex of a micro-expression sequence by the apex frame detection algorithm.Due to the intensity of micro-expression is very low and the change of optical flow between frame and frame will be very subtle especially recorded in high-speed cameras,it extracts the salient optical flow features between each frame and the first frame,then normalizes the extracted features by mean pooling for SVM recognition.Experimental results have verified the validity of this proposed method.Finally,a demonstration system for micro-expression apex frame detection and emotion recognition is set up to visualize the motion and features in facial regions.