Research On Facial Expression Recognition Based On Convolutional Neural Network

Expression is an important mean of information communication between people.Different from other fields of image vision,there are some problems in face expression recognition,such as very small expression features and short expression duration.Based on deep learning,most of the existing researches on facial expression recognition adopt the way of network model exploration and optimization.In this paper,a multi-scale fusion convolutional neural network model is proposed based on light standardization and convolutional neural network feature extraction visualization method.The superiority of the model is demonstrated by design experiments.The main work of the paper is as follows:(1)Explore a light normalization algorithm to deal with light noise in facial expression recognition.Based on the basic principle of light normalization algorithm,the light normalization algorithm with good performance in the fields of gamma correction,Gaussian-Laplacian edge extraction and TT is studied.Using the Alex Net convolution neural network classifier,the FER2013 datasets processed by these algorithms are trained and tested respectively.According to the experimental results,a suitable algorithm is selected as the light normalization algorithm to process the FER2013 datasets.(2)The performance of convolutional neural network layer in different depths for expression feature extraction was studied.Based on the method of visualized convolutional neural network and using Grad-CAM technology,the classification weight heat map is drawn for the features extracted by convolutional neural network layer at different depths in the traditional VGGNet16 network.The deep meaning of the weight heat map of each depth convolutional neural network layer is analyzed,and the information of the network layer which has great effect on face expression recognition accuracy is obtained.(3)Based on the model feature extraction performance,an improved multi-scale fusion VGGNet16 convolution neural network is proposed.The multi-scale fusion VGGNet16 convolution neural network with different scale features in width and depth is constructed by using the results of model feature extraction performance as the theoretical guidance,Inception structure,batch standardization and Dropout technology.The superiority of the proposed model is demonstrated on the FER2013 data set.(4)A face expression recognition system is designed and implemented.Based on the proposed multi-scale fusion VGGNet16 convolutional neural network,a face expression recognition system is designed.The experiments of face expression recognition on static face image and dynamic character video are carried out.It also demonstrates the good application value of the model proposed in the paper in practical scenarios through experiments.