Algorithm Research Of Facial Expression Recognition Based On Deep Learning

In our daily life,facial expression with us no matter where we go or what we do.Even we stay alone,someone can judges ours emotions by our facial expression.Compared with language and character,it contains more truly information about us.As one of the objective indicators of understanding others psychological situation.Facial expression not only can be used for interactive applications to improve human's living environment,but also for medical diagnosis to understand their mood condition.With the construction of wisdom city,the development of artificial intelligence is to speed up the pace,and facial expression recognition as a application in artificial intelligence.It is beyond reproach that research of facial expression recognition will bring important significance to the development of society.This master's thesis has completed the following work in the application of facial expression recognition based on machine learning methods:(1)It makes a research on stacked sparse auto-encoder for feature extraction and classification.First,it analysis and studies of hidden layer neurons number influence on feature extraction through network visualization experiments,and then using the classification experiment for further validation.In addition,it is studied respectively that the impact of identity caused by network layers and numbers of iteration.A series of experiments based on ORL face database show that the less number of hidden neuron units will result in interference on feature extraction and recognition rate decreases;as the network layers and number of iterations increase,the training time increase,but the recognition rate with increase of the former and decrease of the latter.(2)Stacked sparse auto-encoder was applied in facial expression recognition.Since this network can automatically extracts features from samples,which reduce labor for feature extraction.And according to the above research,7 stacked sparse auto-encoder responding to 7 facial expression were constructed.Finally two kinds of classifiers based on two assumptions--the networks extract global features or local features--were proposed.Experiments on JAFFE and CK+ expression databases make it clear thatstacked sparse auto-encoder can extract features of expression automatically.These extracted features not only a little better than features made by artificial,but also prefer to local characteristics which further proofed by visualization experiments of networks.In addition,multiple network is better than a single network in terms of facial expression recognition rate.(3)Extreme learning method was applied to facial expression recognition.A new pruning method was put forward to determine the number of neurons in hidden layer.When the network structure is done,parameters in network determined by training network repeatedly.And final recognition results were made by voting mechanism with more above networks.Experiments of person-dependent and person-independent in JAFFE show that: the new pruning method make the number of hidden layer neurons stable in the process of pruning;and training time is greatly reduced comparing with neural network of deep learning method,but recognition rate stability is poorer.