Research On Facial Expression Recognition Based On CNN

In recent years,with the rapid development of convolutional neural network(CNN)and facial expression recognition(FER),automatic facial expression recognition system has been widely used.At the same time,this thesis designs two kinds of facial expression recognition methods to improve the accuracy of expression recognition.The experiment is divided into two parts The first part,multi-task experiments of facial expressions recognition(FER)and Facial Action Units(AUs)recognition are conduct to verify the auxiliary effect of facial AUs on expression recognition.The second part,for the confused expressions and weak-intensity expressions,an attention mechanism is added to facial expression recognition,and a convolutional neural network based on a pyramid structure is designed to improve the accuracyThe research covers both controllable facial expression data sets and non-controllable facial expression data sets;In multi-task experiments,the AUs data are input into the network as the second type of input and expression data of the network.The network simultaneously performs two kinds of task learning and then separately classifies.In the experiment of the attention mechanism,we use key points generate a Gaussian distribution map as a manual mask for spatial attention;A CBAM attention convolution module is added to the network to obtain an automatic mask for spatial attention.Manual mask and automatic mask jointly give corresponding weights to different regions of the feature map.Key points respectively location features,facial key area features,and complete facial image features are used to classify expressions,which is a pyramid structure model formed for facial expression recognition;In this thesis,in the AUs-assisted experiment,the network conducts multi-task learning,with the help of facial expression-related tasks to assist facial expression classification.The amount of training data has been expanded,and the addition of the AUs task has increased the accuracy of facial expression recognition by 1.1%;Using a convolutional neural network with a pyramid structure,using a manual and automatic spatial attention model to strengthen the attention of the effective area,classifying from three aspects:point,block,and surface features,the recognition accuracy of weak expressions and confusing expressions go higher,and the experiment improves the accuracy rate by 2.14%.