Research On Facial Expression Recognition Based On Convolutional Neural Network

In recent years,the introduction of convolutional neural network(CNN)into the research of facial expression recognition has become the main technical route,and has made good progress and effect,promoting the face expression recognition(FER)gradually from the laboratory to the practical application.Although the existing deep convolution neural network has made important progress in facial expression recognition,it is often at the cost of increasing network complexity.This limits the potential application of the model to a large extent,such as harmonious humancomputer interaction.Based on the analysis and research of facial expression recognition,a lightweight expression network(LENet)based on Attention mechanism and LSTM is proposed in this paper.It realizes the fast and accurate recognition of facial expression.Thesis has completed the following research work:(1)According to the development trend of facial expression recognition and the requirements of the subject,the network model and algorithm of facial expression recognition are understood,and the general design scheme of facial expression recognition in this paper is proposed.(2)Based on the analysis and research of facial expression learning rate,two kinds of deep neural network models,ResNet18 and ResNet34,are built to recognize facial expression.SGD,Adam and Adabound optimization algorithms are applied to these network models,and the 10-fold cross validation method is used to experiment on the CK + data set.Experimental results show that ResNet34 model achieves 99.98% accuracy in CK + dataset.This shows that Adabound has a significant effect on adjusting the learning rate of expression recognition model.(3)Different from the existing FER networks with complex structures,we design a base net containing 6 convolution layers to extract facial expression features,which greatly reduces the model parameters.Then an Attention mechanism based on LSTM is designed to enable the network to focus on the most expressive discrimination areas in the face,and improve the accuracy of facial expression recognition.Finally,the experiments conducted on CK+ and FER2013 expression databases demonstrate that LENet with only 1.3M parameters achieve an average expression recognition rate of 99.98% on CK+ and an average rate of 69.83% on FER2013,and obviously increases the recognition rates of three expressions(anger,fear and sadness),which is comparable or superior compared with the state-of-the-art methods.