Research And Application Of Facial Expression Recognition Based On DNN

Facial expression is one of the most important non-verbal expressions in human interaction,and it's a natural way of interaction.Facial expression recognition is widely used in the fields of intelligent video surveillance,service robot and so on.Therefore,the research of facial expression recognition has important theoretical and application value.Firstly,the facial expression recognition system is designed in this thesis.Then,the face detection method is studied and the AdaBoost algorithm is selected for face detection.Then,the detected faces are scaled and normalized.After analyzing the expression recognition algorithm,it choosing the deep neural network method as the focus of subsequent research.In the study of facial expression recognition based on deep neural network,the expression images are small sample set and direct training neural network model is easy to overfit,so an improved second transfer learning training method is proposed.At first,the MobileNets pre-training model is fine-tuned at the first time on the face recognition data set.Then,a new convolutional neural network is designed in combination with the improved transfer learning method,and the second fine-tuning of the convolutional neural network is performed on the face expression set.Finally,the training of convolutional neural networks is implemented.Through experiments,it is proved that the neural network model combined with secondary transfer learning is more effective in extracting facial expression features.And this method can avoid the problem of overfitting.Aiming at the problem of the features extracted by single-channel networks are insufficient,a network based on two-channel convolutional neural network is proposed in this thesis.At first,it combines the second transfer learning to design a new convolutional neural network.Then,the new convolutional neural network is separately fine-tuned by uses the global face sample and the local sample.And two convolutional neural networks with different feature extraction capabilities are obtained.Finally,the two networks are characterized by a neural network fusion algorithm to obtain a two-channel convolutional neural network.Compared with other facial expression recognition algorithms,it shows that the feature representation after fusion is more rich and the facial expression recognition rate is further improved.After completing the theoretical research,a service robot control system based on Facia expression recognition is implemented.The experiments results show that this method is effective in this thesis,and the stability of the service robot is controlled in real time.