Facial Expression Recognition Based On Bayesian Inference

With our enhanced science and technology,"intelligence" has penetrated into our daily life.Facial expression recognition is a technology of emotional intelligence,which is an important part in the research of artificial intelligence.The application demand of facial expression recognition in the fields of public security,medical services and human-computer interaction is constantly improving,so image-based facial expression recognition is one of the research focuses.Based on Bayesian inference theory,this paper studies the extraction and classification of facial expression features,which is of great theoretical and practical value.The main work is as follows:Firstly,two methods to extract facial expression features are designed.The first one gets the traditional features based on semantic patches.Semantic patches are generated with the multi-scale sampling strategy,then the fusion features of Dense SIFT and LBP_HF are extracted.The second method gets the deep features based on the convolutional neural network,which adopts the improved VGG19 network to achieve effective feature learning.Secondly,based on the traditional features of semantic patches,the facial expression recognition method based on Bayesian topic model is proposed,in which the traditional features are reprocessed hierarchically.First,a visual dictionary is constructed based on the local features,and the expression word features are extracted.Then through the learning of the facial topic model,the potential expression topic features are mined,which are used as the final recognition vectors.This method gains an accuracy rate of 94.9% in JAFFE dataset,and also achieves good results in CK+ dataset both on 6-class FER and 7-class FER.Finally,based on the deep features of convolutional neural network,the facial expression recognition method based on Bayesian neural network is proposed.In view of the lack of uncertainty in traditional networks,the portable and effective VGG19_BNN network is constructed under the framework of Bayesian neural network,which is improved on the existing network,and both epistemic uncertainty and arbitrary uncertainty are added.At the same time,the objective function composed of cost loss and risk loss is designed.Compared with the original network,VGG19_BNN achieves higher accuracy and better robustness,and reduces the uncertainty of prediction.In the “PublicTestSet” and “PrivateTestSet” of FER2013 dataset,the recognition rate of VGG19_BNN reached 71.5% and 73.1%,which shows great advantages in the comparative experiments.