Research On Facial Expression Recognition Based On Convolutional Neural Network

Human emotional expressions are complicated and diverse.The ways of emotional expressions include movements of facial expressions,changes in tones of language when speaking,physical movements,emotional expressions in text communications etc.Among these expressions,facial expressions are the most comprehensive and complex,and are dominant.Nowadays,internet technology is advancing rapidly and facial expressions are being applied in more and more fields.How to recognize facial expressions more quickly and accurately has become a research hotspot.However,traditional expression recognition based on artificial feature extraction does not meet people's application requirements in terms of its recognition speed and accuracy.With the development of artificial intelligence(AI)technology,AI technology is becoming mature and being applied more and more widely.The deep learning method based on convolutional neural network(CNN)has become a dominant technology in AI area.Its applications in computer vision and big data processing are also becoming more and more mature.Thus,applying CNN to facial expression recognition is of great significance for improving recognition accuracy.This paper mainly contains the following two aspects:(1)Aiming at the cumbersome extraction of artificial feature and low recognition accuracy of traditional expression recognition methods,this paper combining with existing research on deep learning proposes an improved deep learning network structure and designs a neural network serving as a backbone network for training and recognizing facial expressions.The neural network consists of 4 convolution layers and 3 fully connected layers.The convolutional neural network can automatically extract feature information.And the concatenated convolutional layer can extract deeper information.Compared with manual extraction,it is more accurate and effective.At the same time,combined with the Inception proposed by Google,we improve the neural network structure proposed in this paper.We replace the last two convolutional layers with the Inception-V1 structure to enhance the depth and width of network,improve accuracy and reduce overfitting.At last,the experimental results show that the proposed method dose further improve the accuracy of facial expression recognition.(2)A face picture that a single network structure receives serves as an input.However,the picture contains a lot of background noise information such as light intensity and occlusion on face.They exist in the network structure in the form of noises.In order to weaken the bad influence of these factors,this paper proposes a convolutional neural network model with multibranch structure.The feature information extracted by different methods is input into the branch structures and is coalesced after quadratic feature extraction of convolutional neural network.Finally,the information is sorted and output through the fully connected network.The feature information includes facial feature point contour map,LBP feature response graph,and Canny feature response graph,respectively representing facial contour information,texture feature information,and edge feature information.After feature fusion,the information forms a convolutional neural network structure of multi-feature fusion,based on which,a facial expression recognition system is designed.The final experimental results show that the multibranch network model has a high recognition accuracy for facial expressions.