Research On Facial Expression Recognition Based On Convolutional Neural Network

With the continuous development of pattern recognition and machine vision,facial expression recognition has gradually become a hot research topic.Through feature extraction and classification of facial expressions,facial expression recognition has been widely used in human-computer interaction,safe driving,intelligent monitoring and case detection,which has important research value.Traditional methods of facial expression recognition have low recognition accuracy and weak generalization ability,but with the development of deep learning,especially convolutional neural network,facial expression recognition technology has also developed rapidly,and recognition accuracy has been greatly improved.The method of facial expression recognition based on convolution neural network are researched in this paper.The main research work is as follows:(1)Facial expression data sets are collected by manual acquisition and crawler,then these data sets are preprocessed as the final input network data set.At the same time,the structure of the classical AlexNet network are introduced,especially the parameters of each layer.Facial expression images are used to train and test AlexNet network.The accuracy of facial expression recognition is 74.91%.(2)The defects of AlexNet network are analyzed.Aiming at the low recognition accuracy of classical AlexNet network for training and testing facial expression images,further improvements are made to the classical AlexNet network,which mainly includes three parts: introducing multi-channel volume set,global average pooling and batch normalization.By using facial expression images to train and test the improved AlexNet network,the accuracy of the final test data set is 88.15%.Compared with the improved network,the accuracy is improved by 13.24%.This shows that the performance of facial expression recognition based on the improved AlexNet network has been greatly improved.(3)Because the collected facial expression image is the expression image of the real scene,there must be the influence of illumination factors.LBP features are insensitive to the change of illumination and have good robustness to the change of illumination,so LBP features are introduced.However,the image transformed by LBP is not suitable to be directly put into the network.In this paper,the original image is directly mapped,and then input into the improved AlexNet network.However,LBP coding exists in the LBP feature mapping,and some useful information is lost in the non-uniform mode of LBP.In order to get more sufficient feature information,the LBP feature mapping is carried out with CNN utilizing feature fusion.Four kinds of LBP feature mappings are used in this paper.They are classical LBP feature mapping,circular LBP feature mapping with radius 1,5 and 10.These four kinds of LBP feature mappings are directly put into convolutional neural network.Among them,circular LBP feature mapping with radius 5 and 10 has better experimental results,with accuracy of 89.87% and 90.08% respectively.Then,feature fusion of LBP feature mapping and CNN with radius 5 and 10 is done.The experimental results show that the face expression recognition effect of LBP feature mapping and CNN feature fusion with radius 10 is better,and the accuracy is 93.15%.At the same time,in these experiments,its convergence speed is the fastest.Finally,150 images are tested by using the scheme of LBP feature mapping with radius 10 and CNN feature fusion,and the correct number of images tested is compared.