The Research Of Facial Expression Recognition Algorithm Based On Deep Learning

Facial expression recognition is a very popular research topic in many fields of our society,especially in the field of pattern recognition and artificial intelligence.If we want the computer to express emotions like humans and to have the intelligence,the computer must be able to understand human emotions first.Facial expression contains a lot of personal emotion information,so it is an essential part of computer intelligence.This field has attracted a large number of researchers.Because of the high complexity of facial expression recognition research,many scholars put forward many algorithms to improve the recognition rate,but there are still many problems in the field of expression recognition.The traditional method of facial expression recognition needs the manually definition of the features and to extract the appropriate features.In this way,a large number of human factors are increased,which brings uncertainty to the expression recognition.Deep learning is a kind of artificial neural network.With the development of computer,the problem of training multi-layer neural network has been solved,and various network structures are proposed and proved in the field of image classification and recognition,speech recognition and so on.Deep learning has set off a new upsurge of research.Deep learning has been proved to be able to find the complex structure of high-dimensional data,so it will be widely used in the fields of science,business,medical and other fields.The main work and innovation of this paper are as follows:1.The classic LeNet-5 convolutional neural network is applied to facial expression recognition,but there is no general network architecture can solve all problems well.Although the LeNet-5 convolutional neural network performs well on the digital dataset,it performs poorly on the expression dataset.It can only get the correct rate of 61.97%and 32.32%on JAFFE expression dataset and CK+ expression dataset.2.In this paper,we improve the LeNet-5 network,and propose an algorithm of facial expression recognition based on Cross-connect LeNet-5 Network,and this algorithm combines the low-level features with high-level features that are extracted from the network structure to construct the classifier.It can get the correct rate of 94.37%and 83.47%on JAFFE expression dataset and CK+ expression dataset.3?In order to understand the process of the recognition and classification of convolutional neural network,and what features are extracted from the deep propagation process of convolutional neural network,deconvolution method is used to reconstruct and understand the convolutional neural network in this paper.The effects of two different deconvolution methods are compared and finally the deconvolution method proposed by Zeiler et al.has better effect on the reconstruction and understanding of the network.