Micro-expression Recognition Based On 3D Residual Convolutional Network

A micro-expression is an expression that occurs very quickly.The duration is generally only 1/25s?1/5s.The movement amplitude is small when it occurs,and it is difficult for people to perceive it directly with the naked eye.Micro-expression is also an uncontrollable and unforgeable spontaneous expression.It is more suitable as a basis for human's true psychological thoughts and has stronger applicability in the fields of psychological research,public safety,and business negotiation.At present,microexpression recognition methods still rely on manual feature extraction,which is timeconsuming,labor-intensive and has limited features that can be extracted,resulting in unsatisfactory final recognition results.With the development of technology,some researchers have introduced deep learning algorithms into the field of micro-expression recognition,but they generally only focus on the characteristics of the spatial domain,while ignoring the time-domain information possessed by the continuous action itself,resulting in room for improvement in the recognition effect..At the same time,the existing micro-expression data sets are scarce,and the existing data sets also have problems such as inconsistent frame numbers and irregular images.This article conducts research on these two issues,the main work content is as follows:(1)A micro-expression data expansion method based on geometric transformation of the region of interest is proposed.Through prior experience combined with microexpression expression characteristics,based on the key points of the face,the region of interest sensitive to micro-expressions is divided.By comparing the existing geometric transformation methods,combined with the characteristics of weak changes in microexpression features,the scaling,mirroring,and combined transformation of the two are selected as expansion methods.The general process is to perform face detection on the image samples,mark the key points of the face,crop and normalize the image,and then perform geometric transformation of the region of interest based on the vertex frames corresponding to the image sequence to generate new image samples.Experimental results prove that this method can generate a large amount of new image data with little noise impact;at the same time,this paper also uses a variety of hash algorithms to compare the similarity between the generated picture and the original picture,and all have achieved good results,proving the effectiveness of the new sample.(2)Most of the deep learning currently applied to micro-expression recognition only focuses on the spatial features of the samples,and does not effectively extract the time-series features.In view of this situation,this paper proposes a new method based on 3D-Res Net network structure to complete micro-expression recognition.Based on the three-dimensional convolutional neural network,combined with the residual network idea,a new three-dimensional residual module is obtained,and a new network structure is correspondingly constructed.According to experiments,combined with different micro-expression data expansion methods,results of 68.78% and 70.38% were achieved,achieving better results than before.