Low-resolution Facial Expression Recognition Via Generative Adversarial Network

Facial expression conveys human's emotion and psychological state.How to automatically recognize the expression contained in face images is an important research direction of computer vision.In recent years,with the development of convolution neural networks,the facial expression recognition technology has made gratifying progress.However,in real scenarios,it is often necessary to recognize low-resolution facial expression images due to factors such as shooting distance,device resolution,and computing power,while many current method's recognition accuracy usually decrease much at low resolution.In this paper,we conduct research on the above issues,and the main work includes:1.In this paper,we innovatively introduce the generative adversarial network(GAN)into the problem of low-resolution facial expression recognition.By decoupling this problem into two steps: facial image superresolution and facial action unit recognition,and using GAN to achieve high-resolution reconstruction of low-resolution faces,our method effectively improves the recognition accuracy at low-resolution scenarios.Meanwhile,through multiple experiments,we prove that the GAN-based superresolution method is more suitable for low-resolution expression recognition problem than the CNN-based method.2.Based on previous works,we use an enhanced residual module with a wider number of channels to construct an improved facial image super-resolution model.Compared with existing models,it has achieved a better result in low-resolution expression recognition task.3.In this paper,an improved training method is proposed.We utilize data augmentation and fine-tuning method to help the expression recognition network adapt with the inter-domain difference between the super-resolution image and the natural image,which further improves the recognition accuracy.We conduct a lot of experiments on the BP4D public expression recognition dataset and analyze each improvement point proposed in the paper in detail.Experiment result shows each method is effective.Finally,based on a classroom scene expression recognition dataset constructed by our laboratory,this paper discusses the application of this technology in real scenes.