Research On Face Super-Resolution Based On Prior Knowledge And Low-Resolution Face Recognition Based On Sample Expansion

In recent years,with the continuous development of Internet technology and artificial intelligence technology,digital images have already become one of the important carriers for people to obtain information.In different scenarios,image data has different functions.For example,doctors can use MRI images to diagnose a patient's condition,and police can use surveillance video images to determine the identity of a suspect.Image data has become an indispensable part of people's lives,and face images are a very important part of image data.When the resolution of a face image is extremely low,the performance of many methods(including face alignment,face recognition,etc.)deteriorates sharply.For the super-resolution method of the face image,the input of the model is only a part of the lowfrequency information contained in the real high-resolution face image,and the generated high-resolution face image lacks its high-frequency face information,defined as the difference between the real high-resolution face image and the interpolated low-resolution face image),which will cause the performance of the super-resolution method to decrease,and the reconstructed face image will not perform well.To this end,we propose a face super-resolution method based on prior knowledge.For the face recognition method in the real scene,the resolution of the face image captured by the camera is extremely small.What's more,most cameras are in an open environment,and the face images they shoot have noises such as lighting and angles,which causes the video or images captured by their surveillance cameras to be of low quality and low resolution.The low-resolution face dataset captured under surveillance video not only needs to be detected from the surveillance video,but also needs to be intercepted and annotated.It is particularly difficult to form a better and complete data set.To this end,we propose a low-resolution face recognition method based on data augmentation.(1)Aiming at the problem that when the resolution of the input image is extremely low,the reconstructed face image lacks its high-frequency information and the reconstruction effect is not good,we propose a face super-resolution method based on prior knowledge,which uses face attributes prior knowledge,such as face points and key points,can increase the high-frequency information of the reconstructed image and improve the super-resolution effect of the face.This method estimates prior information such as face attributes and key points during the training of the model,and adds the prior information to the training of the model.The experimental results on the Celeba dataset show that the method proposed in this paper has obvious effects in the reconstruction process of low-resolution face images.(2)Aiming at the problems of insufficient low-resolution data sets and low performance of face recognition in the real environment,we propose a low-resolution face recognition method based on data augmentation.Firstly,this method uses data augmentation method based GAN to expand the low-resolution face image dataset.Secondly,a highresolution face recognition model is trained using the MS-Celeb-1M dataset.Finally,transfer learning is used to migrate the face recognition model trained in the second stage to the low-resolution face images captured in the actual scene.This can enhance the applicability of the model and improve the performance of the low-resolution face recognition method.