Research On Micro-Expression Recognition Algorithm Based On Convolutional Neural Network

Micro-expression is a kind of facial movement with short duration and low intensity,which can reveal the real emotion that a person tries to hide,so it can be used to guess a person's real psychological state,and has important applications in clinical medicine,criminal investigation and interrogation,public safety,business negotiation,and other fields.However,the development of micro-expression recognition technology faces challenges in terms of tiny differences between samples and insufficient database,and its recognition accuracy is very low in the existing research work.Convolutional neural network,the most widely used deep learning algorithm,has been successful in several fields of artificial intelligence.Therefore,in order to improve the recognition accuracy of micro-expressions,this paper investigates the micro-expression recognition technique based on convolutional neural network,and the main contributions are summarized as follows1.An overview of CASME ? and SMIC micro-expression public databases is described.In order to remove distracting factors in the samples that are not relevant to the study and to ensure the universality of subsequent methods such as feature extraction,a unified pre-processing process for micro-expression databases is proposed,which contains face key point detection,face alignment,face cropping,and Apex frame localization in micro-expression video sequences2.Apex frame micro-expression recognition algorithm based on dual-attention model and migration learning is proposed.Apex frame,as the most expressive frame in the micro-expression video sequence,is the most representative,so it is used as the input of the convolutional neural network.For the problem that there is less useful information in micro-expression images and different regions have different effects on the classification effect,the spatial and channel dual-attention modules are integrated into the ResNetl8 network,and the Focal Loss function is introduced to alleviate the problem of unbalanced micro-expression data samples.Due to the correlation between micro-expressions and macro-expressions,finally transfer learning is adopted to apply the prior knowledge in the field of macro-expression recognition to the Apex frame recognition of micro-expressions to further improve the recognition effect3.A micro-expression recognition algorithm based on optical flow features and an improved MobileNetV2 network is proposed.To address the problem of too much redundant information in micro-expression video sequences,the TV-L1 optical flow algorithm is used to extract the motion features of facial muscles between the Onset frame and Apex frame,and the redundant video frames are ignored.Then the computed optical flow feature maps are input into the MobileNetV2 network for training.Due to the small number of training samples and the small discrimination between classes,the Softmax classification layer in the network is replaced with a support vector machine for micro-expression classification.The proposed algorithms are experimented on both micro-expression databases mentioned above,and the results show that the proposed two algorithms have better recognition performance compared with the dominant algorithms.