Research Of Micro-expression Recognition Algorithm Based On Optical Flow Feature

Micro-expression is a kind of facial expression with weak facial muscle movement and rapid state change.The duration of facial muscle movement of a standard micro-expression is no more than 1/3 of a second.The micro-expression recognition task refers to the emotion recognition of known micro-expression fragments,which is one of the main emotional computing tasks based on the intersection of computer vision and psychology.Micro-expression is the external expression of human spontaneous emotional movement.The emotion expressed by micro-expression can be used as an important basis to judge human subjective feelings.Hence,micro-expression recognition has a wide application prospect in criminal investigation,lie detection,security,and others.Due to the rapid change of micro-expression,how to accurately capture the subtle changes in the local area of the face is one of the difficulties in the research of micro-expression recognition.As an information representation that can capture the subtle movement of objects,optical flow technology has been widely used to recognize micro-expression in the last few years.With the help of optical flow to capture the subtle changes of micro-expression,the recognition accuracy of the model has been significantly improved.Therefore,the micro-expression recognition algorithm based on optical flow features has important research significance and application value.At present,there are a lot of algorithms and research related to micro-expression recognition,especially the micro-expression recognition algorithm based on optical flow has achieved good performance in recent years.Both the deep learning method based on optical flow and the traditional method based on optical flow have achieved good performance in micro-expression recognition.In traditional methods,the design of the optical flow feature descriptor is a key step.However,in feature calculation,to reduce the dimension of features,optical flow often needs to be sparse,which may lead to the loss of some facial motion information.In the deep learning method,we first calculate the optical flow data including facial motion details,and then input the optical flow data into the designed network for feature reconstruction and refinement,to further optimize feature representation.Some of the existing methods directly take the optical flow component as the input of the network and do not take into account the direction differences of various micro-expressions.At the same time,the micro-expression dataset size is limited,and there are many difficulties in the rationality of network architecture design.The improvement of micro-expression recognition accuracy is still a difficult problem to be solved.In summary,considering that the optical flow can capture the slight changes in muscle movement,this paper also uses optical flow to represent the facial dynamic information of micro-expression.However,to highlight the differences of different micro-expressions in the direction of facial muscle movement,this paper does an anisotropic weighting operation on the optical flow component.At the same time,this paper also proposes two different deep learning network models for micro-expression recognition,and designs a multi-scale feature catcher in the network to fully extract the semantic information of the context.This structure ameliorates the performance of the model by combining multi-scale features.Firstly,a micro-expression recognition model based on the channel attention mechanism is designed to automatically learn the weights of different components of optical flow and complete micro-expression recognition.To ameliorate the performance of the above model and further improve the accuracy of the micro-expression recognition model,this paper designs an anisotropic weighted optical flow image and uses the single backbone multi-scale network designed in this paper to extract and classify the micro-expression features.The innovation of this paper is summarized as follows:1.Based on the horizontal and vertical components of optical flow generated by facial muscle movement,we observed that the facial changes of different micro-expressions are different in the horizontal and vertical directions.We believe that different micro-expressions should be given different weights in the two components of optical flow.Therefore,we propose a micro-expression classification network based on a channel attention mechanism.The attention module in the network can automatically learn the weight of each optical flow component to highlight the key facial motion information.At the same time,in order to fully combine a variety of semantic information,a multi-scale feature catcher based on extended convolution is designed in the network to further improve the recognition performance of the model by combining multi-scale features2.To reduce the risk of overfitting and further ameliorate the performance of the network,the channel attention network is simplified and a more concise and efficient micro-expression recognition model is proposed.Specifically,the weight of each optical flow component is calculated according to the modulus of the optical flow component,and then the anisotropic weighted optical flow image is obtained.This method uses the pre-calculation method instead of the automatic learning of the attention module,and effectively decreases the number of model parameters,thus reducing the probability of overfitting of the recognition model.Then the anisotropic weighted optical flow image is input into the single backbone multi-scale network model designed in this paper to complete the micro-expression recognition.In this paper,the performance of the proposed model is verified on four commonly used micro-expression benchmark datasets,and we also compared our method with the advanced methods.The experimental results show that the proposed single backbone classification network model based on anisotropic weighted optical flow images can achieve better recognition performance.