Face Super-resolution Method Based On Deep Feature Transfe

Face super-resolution is a technique that extracts high-frequency details from lowresolution face images to reconstruct high-resolution face images.With the continuous development of deep learning technology,many different structures of face super-resolution technology have been proposed,but these technologies can not effectively solve the problem of face super-resolution reconstruction in real scenes.With the improvement of visual effect,the fidelity of face images will be affected.To solve the problems of existing methods,this thesis presents a method to construct a real-world face dataset and two face super-resolution methods.At the same time,this thesis designs and implements a face image super-resolution system software.The specific work in this thesis is as follows:(1)A construction method for real-world face images dataset.This thesis uses a webcam to collect high-resolution images and low-resolution images respectively.After face positioning,the face images are cut out.A series of data processing methods such as deformation compensation and brightness correction are used to process the face images of two resolutions,and the real-world face dataset is obtained through artificial screening.(2)Face super-resolution method based on deep feature transfer.This method uses the semantic information of the low-resolution face image as the guide information,uses the spatial variation linear representation model to model the low-resolution face image,the guide information,and the high-resolution face image that needs to be restored,uses the encoderdecoder structure to learn the spatial representation parameters,and finally realizes the improvement of the super-resolution reconstruction quality of low-resolution face images in real scenes.(3)Face super-resolution method based on identity information.This method uses the channel attention mechanism to improve the ability of the network to use the feature channel of high-frequency information,and introduces the face recognition network into the face super-resolution task to constrain the face image identity information during the network training,the visual effect and fidelity of the high-resolution face image after super-resolution are improved.(4)Face images super-resolution system software.This thesis designs a simple and easyto-use system.It encapsulates two interpolation algorithms and the two algorithms proposed in this thesis,which is convenient for users to reconstruct low-resolution face images.