Research On Facial Generation For Expression Recognition

Facial expression is the most intuitive and effective way to convey human emotions,and plays an important role in improving human-computer interaction.In the past decades,facial expression recognition(FER)has always been a research hotspot in the field of computer vision,which is widely used in safe driving,distance education,auxiliary medical care,and so on.However,the existing facial expression datasets generally have the problems of insufficient data and unbalanced categories,which leads to the phenomenon of over-fitting,limiting the development of FER.Therefore,how to effectively expand the existing facial expression datasets has attracted a lot of attention from researchers.Based on the above background,our paper proposes a generation method for facial expression recognition.The main work is as follows:(1)Aiming at the problem of insufficient training samples due to the small number of facial expression datasets,we propose a facial expression expansion method based on disentangled representation learning.The proposed method employs the disentangled representation learning to decouple facial pose and identity information,effectively reducing the impact of pose variations and identity biases on FER.Finally,we use the generated images to classify facial expressions.Experimental results on Multi-PIE(Multi Pose,Illumination,Expressions)and RAFD(Radboud Faces Database)show that the proposed method can obtain high quality generated images and effectively improve the recognition rate of facial expressions.(2)Based on the above method,a facial image and coarse-grained pose label generation with cross-class feature transfer are proposed to solve the problem of uneven distribution of face labels and the lack of pose labels in a large number of samples.The proposed method can find labeled images in other classes that are similar to unlabeled images in the target class by matching facial angle features,and then synthesize new face images with unlabeled images in the target class using the generative adversarial network.Finally,the target image pose label is completed by utilizing the pose feature of the generated image with the minimum semantic distance from the target image.Experimental results on Vox Celeb dataset show that the proposed method can effectively achieve facial pose migration between different individuals and generate high-quality face images,obtaining more accurate coarse-grained facial pose labels.