Research On Facial Expression Recognition Technology Based On Data Augmentation

Facial Expression Recognition(FER),as the primary processing method for nonverbal intentions,is an important and promising field of Computer Vision and Pattern Recognition.Computer-based FER research aims to make machines understand human emotions,and is of value in both theoretical research and actual practice in the field of human-computer interaction.Obtaining large-scale and high-quality datasets is always one of the difficulties in FER researches.Optimizing existing raw datasets is an effective solution to this problem.We can divide it into two sub-tasks: one is to expand the original dataset;the other is to improve the quality of the image samples.Based on the deep learning technology,this paper designs two corresponding FER models from the perspective of quantity and quality of data,respectively.Aiming at the problem of data quantity,a FER method based on meaningful-cropping is proposed.The proposed data augmentation method based on meaningful-cropping can expand the original dataset with 4 times while preserving facial expression features,and the cropped patches have high similarity with the side view of human faces.Additionally,the view-aware model is proposed.We interact positively with the difference of view as additional information to drive the Dense Net networks.Finally,the experimental evaluation results show that the proposed MCVNN model has a significantly higher recognition accuracy than other existing FER approaches.Aiming at the quality of data,a FER method based on ESRGAN image augmentation is proposed.This paper addresses the problem of lack of high-quality datasets in the field of FER from the perspective of image augmentation.Using ESRGAN to superresolute the original expression images with low resolution,obtain generated images with significantly improved resolution.After that,utilize the generated images for training deep networks(e.g.,VGG,Res Net and Dense Net)and the original images for testing The final experimental evaluation results show that the image augmentation based FER method can improve the resolution of facial expression images and is effective for the recognition accuracy.