Research On Face Super-resolution Method Based On Deep Feature Representation Learning

Face super-resolution is a domain-specific image super-resolution that aims to recover the input low-resolution face image into a high-resolution face image.It has important applications in security scenarios such as video surveillance and recognition systems.Different from general image super-resolution,face super-resolution also need to focus on the structure of faces to restore the detailed textures.The main research contents of this paper are as follows.(1)A joint method based on image denoising and face super-resolution is proposed.A joint collaborative learning network is constructed to solve the problems of mask occlusion and low resolution in the face images captured by the video surveillance.In this method,mask occlusion is regarded as image noise,and the denoising module and face super-resolution module are combined by generative adversarial network.The network also introduces carefully designed loss functions and identity information to achieve feasible feature representation and informative feature learning.Jointly performing denoising and face super-resolution tasks can achieve complementary effects.The method can obtain good visual results in both Celeb A and Helen databases.(2)A face super-resolution reconstruction method based on Transformer and face attention integration unit is proposed.The Transformer is introduced into the network as a global supervision to solve the problem of restricted perceptual field size in convolutional neural networks.Transformer has the advantage to model long-term dependencies with shifted window patterns.The method is a mixture architecture dominated by convolutional neural network and supplemented by Transformer,which can reliably recover local details of faces and capture global information while avoiding huge computational consumption.The effectiveness of the proposed method is verified in Celeb A and Helen databases.(3)A face super-resolution reconstruction method based on face edge information and attention fusion mechanism is proposed.The method introduces face edge information as a prior constraint into the network to replace other complex priors,and introduces face attention aggregation module and adaptive weights.The model motivates the convolutional layers to refine features corresponding to the main facial structures flexibly and provides edge information to aggregate local and global features progressively.Benefiting from adaptive weights,the network enables to extract useful information from shallow features in an optimal way to help recover the detailed textures of face images.Qualitative and quantitative evaluations show that this method has an advantage over some advanced face super-resolution methods in restoring textured images,and its superiority is verified in the Celeb A and the Helen databases.