Research On Micro-expression Recognition Based On Attention Mechanism

Micro-expression is brief and small facial expression changes,usually lasting less than1/4 second,which is difficult to be detected by the human naked eye.Since micro-expression is one of the important expressions of human emotions and mental states,it is important to study micro-expression both in theory and in practice.However,micro-expression is characterized by short duration,small expression changes,and complex temporal relationships,which make micro-expression recognition face many difficulties and challenges.However,the development of efficient and accurate micro-expression recognition algorithms is of great significance to promote the development of related fields and improve the level of artificial intelligence technology,which can be applied to emotion recognition,identity authentication,virtual reality,and other fields.With this in mind,the objective of this thesis is to investigate the automatic recognition of Micro-expression in human facial expressions using deep learning.The following paragraphs outline the research scope and innovative results of this study:(1)Micro-expression recognition algorithm based on improved Dense Net(Double Attention Dense Net,DA＿Dense Net).The limited number of micro-expression samples can lead to the problem of gradient vanishing and gradient exploding as the network model’s depth or width increases.Hence,Dense Net is employed as the backbone network to reinforce feature reuse.The small amplitude and short duration of micro-expression motion determine the difficulty of feature extraction,so the en-sc SE(enhanced sc SE)module proposed in this thesis is added at the appropriate position of Dense,which can learn adaptively for important channels and important spaces.The model also uses Triplet Attention to focus on the interaction information of channel and spatial dimensions.The feature refinement module is added before the model classification,allowing it to generate features specific to each type of micro-expression,which improves the recognition performance of Micro-expression.(2)Multi-Scale Feature Extraction Network(MFE-Net)algorithm based on attention mechanism.To solve the drawback that it is difficult to capture motion information from micro-expression static images,this algorithm extracts optical flow information from the onset frame and the apex frame as the input of the network model.To achieve improved feature extraction,the SE channel attention mechanism is employed to enhance features.This mechanism assigns greater weights to relevant channel features and eliminates unnecessary channel information.Multi Head Self Attention(MHSA)is used for the calculation of selfattention to focus on local and global relations in the feature ground.To extract features at multiple scales,a three-branch structure is utilized.Each branch extracts features at different scales,which are then combined to enhance the accuracy of micro-expression recognition.