Research On Facial Expression Recognition Based On VGGNet Deep Convolution Features

Expression is unique and the most direct and comprehensive expression of human mind.For this reason,people always expect to be able to recognize others facial expressions accurately by machine.With the rapid development of artificial intelligence technology,it is possible to identify facial expressions more accurately.It can be widely used in the fields of human-computer interaction,robot,public security system,medical system and other related fields,meanwhile it has been paid close attention by scholars at home and abroad.However,there are many problems in the process of facial expression recognition in traditional machine learning.For example,there are some shortcomings in the artificial feature extraction,such as poor characterization ability,less feature data,and information loss;in the algorithm,the training model is required stricted training data and it is easy to appear over fitting / less fitting and so on.The initialization of the model is zero,which leads to the convergence of the weight of the model training process and fall into the local optimization,so that the model is difficult to train,even the training fails...To solve these problems,three aspects of feature extraction,parameter optimization and recognition classification are studied in facial expression recognition system.First,feature extraction: the selection of meaningful feature input is very important for the algorithm and model training of deep learning.The feature extraction using VGGNet network structure is a convolution neural network based on supervised learning.It uses more small size convolution kernel structure,even this network structure can also be used under different micro expression images.We can get enough and abstract features for image.Second,parameter optimization: parameter optimization not only can determine the success or failure of the training model,but also can determine the recognition rate of the training model.It initializes the parameters by using random values and constructing the optimal learning rate function through the linear composition in the iteration provides theoptimal step length to adjusts the parameters properly.Thus it is effectively solving the problem of local optimal value.Third,classification recognition: combining the advantages of VGGNet network to feature extraction with the characteristics of convolution neural network classification recognition,this thesis proposes a method for facial expression recognition based on VGGNet deep convolution features.After optimization,facial expression recognition system runs on the deep learning framework platform of Tensorflow,to carry out model training and recognition test in Sugon Server.On the FER-2013 dataset,the recognition accuracy is up to 85.4%.Compared with other methods,the recognition rate has been improved.It shows that the VGGNet network has a good effect on improving the recognition rate and robustness in the abstraction feature level of the image.