A Lightweight Real-time Facial Expression Recognizer Using Residual Attention Mechanism

The rapid development of AI worldwide and its pervasiveness in everyday life have increased demand for interaction between computers and people.If computers could understand and express emotions,they will fundamentally change the relationship between people and computers,so that computers can better serve humans.Facial expressions are an important form of emotional state and mental state.About 55% of the information in interpersonal communication is conveyed through expressions.In order for a computer to understand and express emotions,it is first necessary to enable computers to understand human expressions.Recently,deep learning and convolutional neural networks have been widely used in the field of image classification,leading to more and more research using CNN to learn the discriminative feature representation of human face for expression recognition.However,due to the increasing depth of the convolutional neural network and the increasing number of parameters,the trained model requires very high performance for the computer.Therefore,research for lightweight facial emotion recognition model becomes the research objective of this paper.In the field of facial expression recognition,several neural network architectures have been proposed,most of which are based on VGG or Resnet.Compared with the traditional methods,the accuracy of these structures is improved,but there are too many parameters in the model and the forward propagation derivation time is too long,which is difficult to be used in many mobile devices.This paper attempts to solve the problems of facial expression recognition and proposes a constructive solution:(1)A fast downsampling feature extraction model is proposed to improve the ability of facial expression feature extraction.(2)Design a lightweight and relatively accurate facial expression recognition algorithm,and by adding the residual attention mechanism to obtain the key features of the expression features that can be classified.(3)This paper propose a new fusion loss algorithm and improve the accuracy of facial emotion recognition by combining the advantages of L2_SVM's generalization ability with the advantage of center loss gathering distance within the class.The general flow of this paper is as follows: Frist,the fast downsampling feature extraction structure is used to extract the expression features at the front end of the network.Then the VGG-like network is used and the backbone network is changed to the Xception structure,and the residual attention mechanism is added as the mask branch.The back end of the network uses a new loss fusion algorithm.Finally,the experimental results show that the design of the backbone does reduce the parameter size of the deep convolution network,which greatly reduces the size of the model and effectively improves the accuracy of expression recognition by using the fusion loss function.