| Micro-expressions are low-intensity and brief motion on the human face.They have the characteristic of being unforgeable and can truly reflect the psychological state of the human.In recent years,micro-expression recognition has received extensive attention from researchers due to its important application value in criminal investigation,psychological intervention,education and medical treatment.The information of micro-expressions in space and time is very sparse,so the combination of apparent feature and time dynamic feature can effectively improve the performance of micro-expression recognition algorithms.The main work of this article includes:(1)We investigated a micro-expression recognition algorithm that combines dimensionality reduction methods and LBP extension algorithm with optical flow.This method employ the optical flow method to process the original video to enhance the movement feature of the micro-expression,and use LBP-TOP,LBP-SIP,STLBP-IP to encode the optical flow field.The dimensionality reduction method is used to improve the effectiveness of features.Experimental results prove that this combination method can effectively improve the accuracy.(2)A micro-expression recognition method based on PCANet+is proposed.This method uses the PCA filter bank to extract the information in the optical flow field.The SVM is employed to classify the micro-expression features extracted by PCANet+.Experimental results show that the method is superior to traditional manual methods and comparable with some excellent feature learning methods.(3)We researched a micro-expression recognition method based on LEARNet with multi-stream input.We feed the dynamic maps of the video and optical flow into the LEARNet to enrich the spatio-temporal dynamic feature.Experimental results show that the recognition performance of LEARNet with multi-stream input exceeds of most general convolutional networks,and prove that multi-stream input improves the performance of LEARNet. |
PDF Full Text Request