Research On Micro-expression Recognition Based On Dynamic And Static Features

A micro-expression is an involuntary facial expression,usually to conceal one’s emotions.Its exercise intensity is low and the duration is short,which can reveal people’s true emotions and is often used in lie detection and other aspects.Micro-expression recognition is a key task in the micro-expression task,which is to identify emotion categories in a given micro-expression video.Due to the characteristics of micro-expressions,computer vision recognition is more accurate than manual recognition of micro-expressions.Therefore,in recent years,the work of identifying micro-expressions through deep learning has gradually increased,and it has shown great application potential of micro-expressions in clinical medicine and investigative interrogation.Aiming at the problems of redundant micro-expression video frames,weak amplitude changes and short duration in micro-expression recognition,a method of preprocessing micro-expression video information into residual product information was proposed.For the fact that micro-expressions usually appear locally and are relatively subtle,a segmented Gaussian pyramid module is proposed,which segments images of different scales through Gaussian pyramids,and crops images of different scales into multiple overlapping subimages.Features are then extracted through a Crop Net model with incrementally increasing channels.In addition,Crop Net introduces feature fusion module and position embedding module to learn different subgraph features and position information.Aiming at the problem that the pixel residual product is easy to lose some local static expression information without movement,which makes the model unable to fully learn more effective micro-expression features,a micro-expression recognition method based on a two-branch polymorphic fusion network is proposed.The network adopts a dynamic and static dual-branch architecture.First,static information is extracted by depthwise separable convolutions to assist the model in learning more effective micro-expression features.Second,a well-designed two-stage sparse convolution and an adaptive feature fusion mechanism are used to extract deep and shallow dynamic features.Finally,dynamic and static adaptive feature combinations are learned by designing a self-learning dynamic and static residual connection module.The thesis conducts experiments to verify the proposed method.On the MEGC2019 mixed data set,both the Crop Net model and the dual-branch polymorphic fusion network model have achieved competitive performance,with UF1 of 0.911 and 0.871,respectively.Finally,the thesis designs and implements a micro-expression recognition system,and applies the model in the thesis to the cloud through a reasonable and concise humancomputer interaction interface,providing users with a more convenient micro-expression recognition platform.