Research On Deep Learning-based Face Image Super-Resolution Methods

Face images extracted by data collection device are important information of identity comparison of people,due to the unevenness of the collection equipment,influence of weather factors and lighting conditions,the collected face images have the problems of blurred image details and low resolution,which can't reach the requirement.At present,it's effective to recover low resolution image to high resolution image by using super-resolution methods based on Interpolation,reconstruction and learning,therefore,face image super-resolution has attracted extensive attention from researchers.This thesis research methods of face image super-resolution based on Deep Learning,including single image super-resolution,face image super-resolution based on Generative Adversarial Networks,face image super-resolution based on piecewise processing and programming face image super-resolution software.The main contents of this thesis are as follow:With respect of single image super-resolution,discuss the model of image degradation and reconstruction,study various methods of image quality assessment,implement methods of image super-resolution based on interpolation and Deep Learning.Several methods are tested on Set5 and Set14 datasets,experimental results show that the methods based on Deep Learning reduce image distortion effective and generate more high-definition HR images.With respect of face image super-resolution based on Generative Adversarial Network,a face image super-resolution method which combining SRGAN and Self-Attention is proposed.Using the advantage of Self-Attention on long-range dependency modeling to reach the purpose of optimizing global outline and local distortion in face super-resolved images by adding Self-Attention module to Generator and Discriminator of SRGAN.The SRGAN and modified SRGAN based on different losses were simulated tested on CelebA datasets,experimental results show that modified SRGAN make super-resolved face image global outline and local features more realistic and better visual effects.With respect of face image super-resolution based on piecewise processing,a piecewise processing Super-Resolution method is studied and propose a face image super-resolution method based on piecewise processing SRGAN.This method divides common 4upsample process into two times 2upsample process,adopt face image datasets including multi-resolution to train neural network and optimize loss function based on relativistic average discriminator.Simulation experiments and property tests on various super-resolution models such as SRCNN,VDSR,SRGAN et.al on CelebA datasets show that the proposed piecewise GAN can reduce face images distortion effectively while maintaining image perceptual quality,fix unnatural patch and local color distortion in image and improve visual effects.With the respect of development of face super-resolution software,adopt PyQt5 as the design of Gui language and Pycharm as the IDE,use Tensorflow as the Deep Learning architecture based on Python,program face image super-resolution software and test corresponding function modules with the usage of face images from both dataset and natural scene images.The software has two main functions: face detection and face image super-resolution.