Research On Character Facial Micro-expression Recognition Based On Transfer Learning

The micro-expression on the face of a character is a kind of expression that is inadvertently revealed under a special stimulus.Because it is uncontrollable,it can be known through the recognition of the micro-expression that the character is hiding.The true emotions of the human body have great application prospects in the fields of clinical treatment,criminal investigation and business communication.However,because micro-expressions have the characteristics of small range of facial muscle movement and short duration,it is difficult to recognize them with the naked eye and the accuracy rate is not high.Therefore,it is of great research significance to use computers to recognize micro-expressions.At present,there are two main problems in the research of micro-expression recognition methods based on computer technology.One is that the micro-expression recognition algorithm is easily affected by noise.Second,the existing deep learning-based methods often use a deep network structure to obtain better model performance by deepening the network,but for micro-expression recognition,they are limited by the amount of data,because the deep network has more layers and parameters,it may overfit the noise and produce overfitting phenomenon.In view of the above problems,this paper first uses the optical flow method to extract micro-expression features,by extracting the start frame and key frame of the micro-expression sequence,68 key points of the face of the person in the picture are detected,and the face of the character is detected by the position of the key point.Alignment and cropping,and finally use the optical flow method to calculate the optical flow characteristics of the micro-expression,which solves the problem that the micro-expression recognition algorithm is easily affected by noise;a lightweight network Dual-channel Shuffle Net(DSNet)is proposed for micro-expression Recognition,by using group convolution and depthwise separable convolution in the network,the size and computational overhead of the model are significantly reduced,and the deep neural network is prone to overfitting in micro-surface recognition.On this basis,this paper proposes two network models based on transfer learning,including parameter transfer-based micro-expression recognition network Tr DSNet V1 and feature transfer-based micro-expression recognition network Tr DSNet V2.Finally,it is verified on two micro-expression datasets using the proposed DSNet,Tr DSNet V1 and Tr DSNet V2.By comparing the classification performance of other models(Bi WOOF,Motion Boundary,and ATNet)on micro-expression recognition,the experimental results show that DSNet can be used for micro-expression recognition tasks.Using the micro-expression recognition network Tr DSNet V1 based on parameter transfer can shorten the convergence time of the model while obtaining similar classification performance as the model without transfer learning.Using the micro-expression recognition network Tr DSNet V2 based on feature migration can enable the model to learn the common features between the source domain and the target domain,so that the transferred model has better classification performance.Experimental results show that with transfer learning,the model shows better performance on the micro-expression recognition task and can achieve higher accuracy with less training time.