Studies On Micro-expression Recognition Algorithm Using Convolutional Neural Network

Micro expression is a special expression with very short duration and low expression intensity.It usually appears involuntarily when people try to hide their real thoughts.Because micro expression usually stays on the face for only 1/25 to 1/3 of a second,it is difficult for most people to realize its emergence.This special expression is considered to be related to people’s self-defense mechanism and reflects people’s real thoughts.Accurate recognition of micro expression can help people make appropriate judgments and decisions,so researchers began to pay attention to the research related to micro expression recognition.Considering that convolutional neural network is widely used in the field of image classification,this thesis uses the same method to study the task of micro expression recognition.1.A new image feature extraction method is proposed to solve the problems of insufficient attention to salient features of existing image features and the large computational complexity of optical flow feature extraction.By performing difference and accumulation operations on the pixel matrix of the original image sequence in the dataset,the effect of highlighting the facial action area and amplifying the expression change is achieved.Afterwards,a shallow CNN network is used to classify the processed images,and a dual attention mechanism is introduced to make the model distinguish the importance of the extracted features.The verification accuracy of this method on the public micro expression data set reaches 82.11%,which has obvious advantages compared with the existing models.The results of model ablation analysis and complexity analysis also demonstrate the feasibility of the model.2.In order to solve the problem of insufficient attention to face detail information in existing models,a dual stream convolution neural network based on multi-dimensional feature fusion is designed.To compensate for the lost facial details of the model,a feature fusion module is constructed,which combines visual features and abstract features for model classification.For the extracted high-dimensional fusion features,the channel attention module gives different weights to the channel,so that the model pays more attention to the channel with high contribution.The unweighted F1 score and unweighted recall rate of the proposed model on the public micro expression data set casme Ⅱ reach0.9406 and 0.9502 respectively.The classification effect comparison with the existing models shows that the proposed model has excellent recognition performance.The ablation analysis and complexity analysis of the model also show the rationality of the model.