Research On Deep Learning Methods Of Face Super-Resolution In Complex Imaging

Facial images represent a person’s identity attributes and have always been a key research object in the fields of image processing and computer vision.Although deep learning based facial super resolution algorithms have made significant progress,the pre-assumed degradation of existing super resolution models is often inconsistent with the complex degradation in the real world,resulting in most super resolution networks not being suitable for complex imaging environments.As for this issue,this paper,on the one hand,studies the construction of an effective and robust supervised super resolution network,on the other hand,studies the unsupervised optimization model based on discriminative image priors,and discusses how to improve the generalization performance of the face super resolution method under complex unknown degradation conditions.Specifically,this paper conducts the following research work:1.Considering that attention mechanism helps facial super resolution models better focus on the features of restored faces,this paper uses the natural image super resolution model ELAN as a benchmark method to discuss how to optimize its adaptation to facial images.On the one hand,in order to improve the training stability and the accuracy of the learning model,this paper uses Mish activation function instead of Re LU activation function;On the other hand,in order to address the issue of possible deformation of facial features captured in complex scenes,this paper adds a spatial transformation network module into ELAN,and ultimately proposes a facial super resolution model called ELAN+.In order to explore the generalization performance of ELAN+ in blind facial restoration problems,the EDFace-Celeb-1M dataset,which is closer to the degradation mode of real scenes,was further selected for retraining and testing ELAN+.The experimental results show that when the face degenerates into simple down sampling,ELAN+ achieves comparable or even superior results compared to various representative facial super resolution methods.When face degradation is complex,diverse,and unknown,ELAN+ achieves the best pixel domain evaluation results compared to the new nearest human face blind restoration method,verifying the generalization potential of ELAN+ for complex degradation scenes.2.This paper further turns to unsupervised learning,building a blind facial super resolution network with the help of the depth generative model,and effectively improves the estimation accuracy of face image and blur kernel by introducing a new discriminative image prior in the loss function.Specifically considering that the mainstream blind restoration image priors represented by L0+X are non differentiable,this paper combines a class of discriminative redescending potential functions(RDP)in the field of blind deblurring recently.On the one hand,the RDP discriminative priors obtained based on theoretical analysis are applied to unsupervised blind face super resolution to expand the application range of the RDP priors and improve the effectiveness of blind face super resolution;On the other hand,inspired by RDP,this paper then proposes a kind of discriminative prior based on Beta probability density function,which not only enriches the method connotation of RDP prior,but also verifies the rationality and the Beta discriminative prior to unsupervised blind face super resolution.Finally,the experimental results validate the effectiveness and superiority of the unsupervised facial blind super resolution method based on discriminative prior.