Research Of Facial Expression Recognition Based On Convolutional Neural Network

Facial expression recognition,which has been widely used in health care,education,traffic safety and human-computer interaction,is now a hot research topic in the field of pattern recognition,machine vision and artificial intelligence.So there is an important research value and bright business prospects in this technique.The facial expression recognition also has certain research difficulties because of illumination,pose,background,object occlusion and so on.Another difficulty is that the distinction between two expressions is fairly vague,usually a subtle change in one part of the human face is enough to change a facial expression.In recent years,deep convolutional neural network has achieved great success in the field of large-scale object recognition.The 152-layers residual network achieved more than 96% recognition rate on the Image Net database,which demonstrates that the deep convolutional network with training will have amazing recognition ability.With a series of research methods,such as residual learning,being published,making it possible to train a very deep convolutional network.This thesis mainly compare the advantages and disadvantages of different convolution networks in the performances of facial expression recognition through a lot of experiments.Then,on the basis of reconstructing the network,this thesis proposes a multi-level convolution network which can recognize the expression sequence.The main work is as follow:1.This thesis introduces the research significance and application prospect of facial expression recognition,and summarizes the research status of facial expression recognition and convolution neural network.Then making a detail introduction into the principle and structural characteristics of the convolution neural network.And making a deep analysis on the deep networks,such as Alex Net,VGGNet and Res Net.2.This thesis makes improvement on the structure of the shallow net,VGG net and residual net,making it can be applied to facial expression recognition.The shallow net contain 6 layers;VGG net adopts 9-layer,11-layer and 16-layer configuration;20-layer,32-layer and 50-layer residual network are designed to make contrast experiment to observe the effect of depth on the recognition rate.The experiments run on Fer2013 dataset which contains more than 28 thousand training images and 7 thousand testing images.Rotation and fuzzy processing has been taken to increase the original training set number.Stochastic gradient descent with momentum has been taken to train the net.In order to make the network to learn in a right direction,a large number of attempts have been taken.The verification is not only carried out on Fer2013 testing set,but also taken on Jaffe database to observe the effectiveness of trained network.3.The traditional convolution neural network has a disadvantage that it cannot recognize the expression sequence.To solve this problem,this thesis uses a multi-net convergence technology to construct a new network that can recognize the expression sequence.First,a number of copies of one convolution network are constructed,so that each copy can processes one frame image.Second,the results from the first step are merged in the Marge Layer.Finally,through the Softmax Layer output identification results.Experiments are performed on the CK+ facial expression database which contains a lot of expression sequence.One sequence shows the expression changes of a person.Before training,a series of preprocessing procedures are needed,including face detection,scale normalization and gray equalization.3 frames,4 frames and 5 frames sequence data are used to carry out comparative experiments,and the recognition rate is 92.15%?92.24% and 92.88%.