Research On Facial Expression Recognition Based On Generative Adversarial Networks

Facial expression recognition(FER)enables the computer to perceive and recognize human emotions,which is of great significance in human-computer interaction,computer vision,cognitive science and other fields.In recent years,with the rapid development of deep learning and digital image processing,the methods based on deep learning have made a breakthrough in image recognition.However,the lack of facial expression data restricts the expression ability of deep neural networks.In this thesis,a research of FER is carried out based on generation adversarial network(GAN).The aim is to improve the recognition rate and generalization ability in FER from two aspects of data augmentation and model optimization.Data augmentation: A static image data augmentation method is designed for the standard expression database CK+.A multi-domain image-to-image translation model is realize on the basis of StarGAN,improving the reconstruction loss and fusing image information and category information with the depth separable convolution.The generator generates multi-expression facial images from the ones of natural expression to expand the amount of dataset.Compared with the traditional methods,this method can generate paired data,which makes the model pay more attention to the difference between expressions rather than the difference between people and enhances the expressive ability of the model for expression features.Also,this method can be used for reference to solve the problem of data imbalance.Model optimization: The trained discriminator model from GAN is extracted and finetuned on the expanded CK+ database,considering the classification ability of the discriminator mode.The advantage is no need to redesign and retrain a classification model.Inspired by the region of interest(ROI),the facial image is divided into four interest regions: left eye,right eye,nose and mouth,which are fused with the original image to enhance the network's attention to ROI areas.In this thesis,several comparative experiments were carried out on CK+ and JAFFE databases and the results prove the validity of the proposed data augmentation and model optimization methods.