Research On Deep Models Of Face Superresolution Based On Unsupervised Learning

Face image is one of the important carriers to reflect people’s appearance and identity information,and it has applications in all aspects of social life.Affected by various external factors in the face imaging process,the image degenerates into a low-definition degraded image.In this paper,the degraded face images are taken as the research object.In the two problems of large-scale face super-resolution and small-scale face blind super-resolution,the deep generative prior and image discriminative prior are the starting points.The super-resolution algorithm of face images is discussed under the theoretical framework of supervised learning.The main work is as follows:On the problem of large-scale face image super-resolution,it is difficult for m GAN to reconstruct realistic face image by applying GAN Inversion theory in face super-resolution task.This paper designs a more robust face super-resolution method,termed called m GAN+.On the one hand,through the experimental analysis of the knowledge representation of different layers in PGGAN,it is found that the semantic level biases of different layers are different.Therefore,the multiple latent codes are grouped in different layers for iterative optimization.Combined with the coupling feature penalty term proposed in this paper,it enriches the multiple latent codes semantic information,which in turn enhances the realism of the reconstructed face.On the other hand,in order to ensure that the identity of the reconstructed face image does not change,Arc Face face recognition loss is introduced to constrain the semantic information of multiple latent codes,which effectively maintains the identity feature of the face image.The superiority and robustness of m GAN+ compared to m GAN are verified experimentally under three different sizes of super-resolution factors.On the problem of blind super-resolution of small-scale face images,the blind deblurring baseline model based on deep generative prior is not enough to solve the problem of blind superresolution of faces under complex blur.In this paper,the discriminative prior knowledge is integrated into the image restoration process in the form of a loss function,and the image prior combined with the deep prior prior jointly drives the blind super-score model to restore face images with high quality.Specifically,starting from the demand of blind deblurring,this paper analyzes what characteristics a good discriminative prior should have from the perspective of partial differential equations,and proposes a new discriminative prior RDP-Lec.On the two tasks of face blind deblurring and blind super-resolution,the importance of the discriminator prior to the unsupervised face blind superresolution model is verified by experiments.