Research On Face Super-Resolution Image Generation Algorithm Based On Multi-Task Cooperation

Face images are the core data types for many applications,such as video surveillance,face recognition,old photo restoration and face processing,etc.Images are often acquired at low resolution,and the lack of facial detail information cannot meet the needs of real-world work.Therefore,it is usually necessary to increase the resolution of raw face images to facilitate more successful applications.In recent years,peers have conducted a lot of research aimed at reconstructing high-resolution face images from low-resolution ones,resulting in a hot research direction called face super-resolution.Due to the application of deep learning,face super-resolution has been developed rapidly.This thesis conducts research on face super-resolution image generation algorithms based on multi-task collaboration,aiming to enhance the quality of low-quality face images,so as to obtain clearer face images with more high-frequency detail information.Addressing the limitations of the existing research on face super-resolution generation algorithms,this thesis conducts further research on face super-resolution image generation.The main research topics of this thesis is as follows:1、Addressing the issue that the face prior information used by existing face super-resolution methods can only provide sparse and coarse-grained descriptions of local facial features,we propose a coarse-to-fine face super-resolution with intensive guidance of facial structure.In this model,face structure information extraction and super-resolution image generation are multiple tasks that mutually benefit each other and optimize iteratively.In addition,the model can fully extract face structure information through the combination of the sparse structure extraction module and the dense structure extraction module,and provide intensive guidance for the entire face super-resolution process.Finally,to generate high-fidelity face images,this thesis also introduces a generative adversarial network into the model,whose visual effect is significantly better than that of the mean square error-based network model.This model solves the problem of how to make full use of face structure information to achieve accurate face information reconstruction.The experimental results show that the qualitative and quantitative results of this method are better than the existing face super-resolution methods under the 8 times magnification factor.2、Aiming at the problem of how to achieve super-resolution of different scale factors in one model,we propose a model that explores super-resolution image generation from arbitrary low-resolution images in a unified manner.The feature extraction at multiple anchor resolutions is treated as multiple tasks in this model.It extracts multiple features at multiple anchor resolutions for an input image of any resolution at the same time,and then combines these features to derive the optimal result.This model solves the problem of how to learn a general model that can perform robust information recovery for input images of different resolutions.As far as we know,for the first time in the field of face super-resolution,we clearly pointed out the rigid constraint problem of optimal model and input image resolution,and explored the method to solve this problem.Extensive experiments show that the algorithm achieves state-of-the-art performance on both quantitative and qualitative results on multiple datasets.