Research On Static Facial Expression Recognition Based On Convolutional Neural Network

Facial expression recognition(FER)enables the computer to perceive and understand human emotion.It has great research and application value in computer vision,human-machine interaction and affective computing.The rapid development of deep learning has brought new opportunities for breakthroughs in various fields.Different from traditional methods based on manual feature extraction,deep learning can learn the high-level semantic features from the big data automatically.In this thesis,we carry out the research of FER based on the convolutional neural network(CNN).The aim of this study is to improve the generalization ability of CNN from two aspects of data optimization and model improvement.The main works of this thesis are as follows:Convolutional neural network is prone to be the problem of overfitting and has poor generalization ability due to the serious lack of data.This thesis constructs a data augmentation mechanism on open standard dataset,named artificial face.The disadvantage of the traditional data augmentation method is that the transformation form is single and the ability to expand the database is limited.The artificial face data augmentation method can effectively expand the CK+ dataset by integrating the feature regions of different faces,so as to give the model a better ability to learn semantic features.Inspired by the active areas and converged network,this thesis proposed a deep learning training improvement scheme based on region of interesting(ROI).Based on the facial expression prior knowledge,the face image was divided into several interest regions as the input to the neural network.This method can actively guided the CNN focus on the areas associated with the expression,and improve the distributed representation ability of model,which helps to intensify the reliability of the predicted targets.Meanwhile,a variety of ROI decision fusion methods are introduced in the testing stage.A test image was identified by combing the output of CNN on each ROI region,which further improveed the generalization of the model.In this thesis,we employed pre-trained AlexNet and GoogLeNet to evaluate the performance of our method.Several experiments were conducted on the CK+?JAFFE and FER2013 databases.The experimental results verify the generalization ability of the proposed artificial face augmentation method and ROI-based convolutional neural network method.