Facial Expression Recognition Based On Feature Fusion Of Parallel Convolutional Neural Network

Facial Expression Recognition(FER)is a part of human-computer emotional interaction research,which has potential application value in many fields such as medical,criminal investigation,education and entertainment.In recent years,with the deep research of deep learning,Convolutional Neural Networks(CNN)with powerful automatic feature extraction and classification capabilities have been widely studied in the field of FER.The structure optimization of CNN is studied to improve the feature extraction ability of CNN model,so that the model has higher recognition accuracy under a certain degree of complex conditions and practical application value.This paper first analyzes the frequently-used methods for feature extraction and feature classification of facial expressions,and the improved CNN structure is utilized for FER.Aiming at the problem that the conventional CNN is prone to over-fitting due to the large number of parameters,and the gradient disappearance or explosion problems are easy to occur during the error backpropagation.This paper proposes a special layer batch normalization convolutional neural network model.Continuous convolution enhancement feature extraction is applied to the front of the network model.Then the batch normalization is used to adjust the distribution of feature data at the partial layer of the network,effectively solving the gradient problem during error backpropagation.To reduce the possibility of over-fitting,the sparse connection is implemented at the fully connected layer by optimizing the Dropout connection probability values.The experimental results show that the improved convolutional neural network model has faster training speed and better recognition effect.Aiming at the facial expression feature extraction with illumination,gesture and whether it is spontaneous expressions complex conditions,the special layer batch normalization convolutional neural network model can not achieve the ideal recognition effect,this paper proposes a model based on parallel convolution channel feature extraction and feature fusion.The difference value channel fusion is used for the different channel convolution feature maps,and then the feature maps are internally fused to combine the feature information and reduce the number of feature maps to effectively reduce the fully connected layer parameters.In the output layer,the method of integrating the global average pooling and the neurons fully connected output decision is adopted,which effectively enhances the model classification decision-making ability.The parallel convolutional neural network model is applied to the JAFFE,CK+ and USTC-NVIE datasets to achieve recognition accuracy of 98.93%,98.78% and 96.19%,respectively.Experiments show that the optimized model has a better recognition effect on facial expressions under complex conditions.