Research On Facial Expression Recognition Based On Deep Learing

Facial expression carries a lot of emotional information,which can better reflect a person’s emotional state.Facial expression recognition has become the key technology of intelligent interaction and emotional computing,and has become an important topic in the direction of artificial intelligence.Facial expression recognition,as an important means of perception in the field of emotional information exchange,is used in patient’s emotional state detection,public danger warning and blind navigation.It has been widely used in many fields.Fine texture features of facial expressions,as well as similar expressions that are difficult to distinguish and confuse,will lead to inadequate feature extraction,inaccurate classification of facial expressions,and affect the recognition effect of facial expressions,resulting in low recognition rate.In view of the above problems,this paper uses deep learning method to conduct in-depth research on facial expression recognition,the main work is as follows:Firstly,aiming at the problem of insufficient feature extraction ability of facial expression recognition methods,a new network Mob＿Inc,which is more suitable for facial expression recognition,is constructed.The network draws lessons from the depthwise convolution(convolution on channel)mode of Mobile Net structure,and on the basis of deep convolution,the Inception idea is introduced to enhance the network through multi-scale convolution and feature fusion of maximum value.The ability of network feature extraction and the problem of over-fitting are solved by combining the Jump-Layer connection,and the minute feature information contained in facial expression is obtained.Secondly,in view of the insufficient extraction of facial expression subtle features,Dense SIFT with good local feature description ability is introduced,global features are extracted by Mob＿Inc,local features are extracted by 128-dimensional descriptor of Dense SIFT,and global features and local features of facial expression are fused to ensure the richness,subtlety and subtlety of features.Overall,it effectively solves the problem of insufficient feature extraction in facial expression recognition task,and improves the accuracy of facial expression recognition.Thirdly,aiming at the problem of low recognition rate of facial expressions which are difficult to distinguish and confuse,this paper analyses the influence of different loss functions on feature classification,and proposes a new loss function for facial expression recognition.The loss function is based on LP loss function and introduces the idea of isolation loss,which combines the advantages of LP loss to reduce intra-class differences and isolation loss to reduce inter-class similarities.It reduces the overlap between classes,improves the discriminability of class features,and improves the recognition rate of facia expressions which are difficult to distinguish and confuse.Finally,in order to verify the validity of the proposed method,experiments are carried out on two more complex facial expression datasets FER2013 and JAFFE,and compared with other facial expression recognition methods.The experimental results show that the recognition method based on deep learning and Dense SIFT feature fusion effectively improves the accuracy of facial expression recognition,and the overall accuracy reaches73.2% and 96.5%.The effects of different loss functions on each kind of expression recognition results are compared.Experiments show that the improved loss function can effectively improve the recognition accuracy of expressions that are difficult to distinguish.