Research On Super-Resolution Algorithm For Face Images Collected In Unconstrained Scenarios

With the development of economy and technology in the modern society,equip-ment for collecting images of unconstrained scenarios such as mobile phones,drones and surveillance cameras are becoming more and more popular.Face recognition ap-plications based on such equipment and computer vision algorithms gradually entered people's lives.Compared with the clear,high-resolution frontal face images collected in the restricted scenarios,the low resolution face images with noise,blurring and other degrading factors collected from unconstrained scenarios have greatly reduced the dis-crimination and the amount of information of the face images,which requires the face recognition algorithm to have high robustness.So the face recognition system for un-constrained scenarios needs to be specially optimized for image degradation factors such as low resolution,noise,and blur.The task of super-resolution of low-resolution face images in unconstrained scenarios is of great significance for large-scale applications of face recognition algorithm.In recent years,thorough research has been carried out on the problem of single image super-resolution and face image super-resolution.With the introduction of deep convolutional neural networks and generative adversarial networks to the problem of face image super-resolution,the performance of super-resolution algorithms is becom-ing more and more powerful.But most of the algorithms are mainly aimed at theoretical scenarios to achieve better results,and it is difficult to apply to real-world unconstrained scenarios.However,a small number of existing algorithms for unconstrained scenarios are mostly used under harsh conditions or high time complexity,and training methods are difficult to apply.This thesis mainly analyzes and studies the problem of iden-tity information recovery,arbitrary scaling factors and real-world unknown noise and degraded for face images super-resolution in unconstrained scenarios,and proposes al-gorithms to solve the corresponding problems.The main research results of this thesis on the problem of super-resolution of face images in unconstrained scenarios include:1)A face image super-resolution algorithm based on the identity prior information and supervised pixel-wise generative adversarial networks is proposed.This algorithm introduces identity prior information into the discriminator network which uses the dis-criminator to enhance the texture details related to face recognition in super-resolution images.And different from the traditional unsupervised generative adversarial net-works,we propose a novel supervised adversarial loss function to force each pixel of the generated face image to be close to the corresponding pixel of the real high-resolution face image in the perceptual field.In order to enhance the supervision strength of the adversarial loss function,we also propose a pixel-wise discriminator network struc-ture.Due to the pixel-wise discriminator structure and the supervised mechanism of the adversarial loss function,this algorithm can generate more accurate and realistic high-resolution face images,and effectively avoid the artifacts introduced by the unsu-pervised adversarial loss function.2)A bilateral upsampling network is proposed for single image super-resolution with arbitrary scaling factors.the existing single image super-resolution algorithms need to train different models for each scaling factor,and mainly focus on integer scal-ing factors.In order to solve the problem of arbitrary scaling factor with a single model,we propose a bilateral upsampling network.It uses the scaling factor and the con-tent information of the input image to predict the weight of the bilateral upsampling filter.Then,the bilateral upsampling filter is provided to the depth-wise feature up-sampling convolution layer as the weights of the convolution kernel to upsample the low-resolution feature map to the high-resolution feature space.Since the weights of the bilateral upsampling filter can be learned adaptively according to different image content and scaling factors,this algorithm can obtain both texture enhancement and structurally accurate super-resolution results.3)A degradation variational autoencoder is proposed for face image super-resolution in unconstrained scenarios.The existing learning-based super-resolution methods mostly rely on supervised learning manner.They need paired low-resolution and real high-resolution image datasets to learn the mapping function.In order to simulate real-world noise and degradation,we propose an image degradation model based on variational autoencoder,and train the model based on an unsupervised cycling training strategy,so that the model can learn the real-world noise distribution and degradation mode from real-world low-resolution images in the process.By degrading high-quality and clear images,we can use a supervised method to train a super-resolution model that can be robust against real-world noise and degradation.