Facial Expression Recognition Based On Deep Learning

Facial expression recognition technology is an interdisciplinary study that crosses many fields such as artificial intelligence,psychology,biology,etc.It also has a wide range of applications in sentiment analysis,human-computer interaction,business forecasting.Expression recognition technology can give robots the ability to understand human emotions,so that intelligent devices like robots can provide intimate services to humans in a timely and accurate manner.After studying related papers deeply,this dissertation summarizes the shortcomings of current facial expression recognition algorithms,and proposes corresponding improvement methods from the perspective of model improvement and feature selection.In the meanwhile,this dissertation verifies the effectiveness of the algorithm through experiments.The main work of this paper is as follows.First,it overviews the development of facial expression research from the dynamic and static inputs,and then starting from the theoretical significance and practical application,it details the current research methods,research progress and actual application situation of facial expression recognition technology,and analyzes the problems and limitations in related work.Finally,it briefly describes the difficulties and challenges faced by facial expression recognition technology.Secondly,considering the current large number of dynamic facial expression recognition algorithms using ensemble methods,the entire network is huge,and it is difficult to train the models.An expression recognition framework based on convolution GRU is proposed.The algorithm deploys CNN to extract the abstract features of the expression pattern.It overcomes the difficulty of inconsistency between the 3D feature map output by CNN and the traditional GRU network input by changing the form of the traditional GRU input data and the internal operation mechanism,and uses the convolution operation to retain the correlation between pixels in the three-dimensional feature map.Therefore,the proposed algorithm is able to encode the three-dimensional expression feature map in time series directly and improves the accuracy of facial expression classification under a single model.Finally,after reviewing the mathematical principle and application of the attention mechanism,summarizing several variants of the attention mechanism,the attention mechanism applied in the channel domain and spatial domain is proposed for expression recognition tasks.The attention mechanism utilizes convolution and pooling operations to distinguish the importance of spatial domain features and channel domain features in the feature map,giving features related to expression higher importance.Therefore,the generated decision vector can better represent the expression pattern and experiments on the RAF,FER2013 and SFEW datasets prove that the proposed model can capture key areas of expression patterns.