Face Hallucination Based On Generative Adversarial Network

Nowadays,with the rapid development of deep learning,computer vision tasks based on deep learning have been greatly improved and transformed into accessible applications in our daily life.Image super-resolution reconstruction is different from face recognition,object detection and other high-level seman-tic understanding problems.Image super-resolution reconstruction studies the visual presentation effect of images,which is a relatively low-level technol-ogy and one of the basic sub-problems in the field of computer vision.Image super-resolution reconstruction can be applied in many practical fields,such as video monitoring,face hallucinationg,video reconstruction and so on.Com-pared with general image classification and recognition problems,the difficulty of imagesuper-resolution reconstruction is mainly in two aspects.First,the cal-culation of image super-resolution reconstruction network is relatively large,which takes up a lot of memory and relies on hardware equipment.How to achieve high and fast performance of the super-resolution algorithm is a urgent problem to be solved.Second,how to super resolve the blur and tiny image is also a need to break through the difficulty.As for large high quality image,it owns relatively rich details.However,fuzzy and small image has the less information.In order to solve the above difficulty,this paper mainly made the following several contributions:· The image reconstruction based on deep learning method is analyzed,dif-ferent convolutional neural network structures are compared,and the real-time imasge super-resolution reconstruction network structure TLSR is de-signed.·For the problem of super-resolution reconstruction and deblurring,a multi-scale pyramid network structure is designed,and a hierarchical supervi-sion method is used,so that the network can simultaneously reconstruct multiple super-resolution reconstruction images of different scales and re-move the "ringing" artifacts caused by JPEG blur in the images.·For the reconstruction of fuzzy face in surveillance video,an end-to-end architecture based on adversarial generative network is designed to recon-struct small face.Three different residual modules,common,upsample and downsample,are used.When the objective function is designed and optimized,in addition to the loss function generated by the confrontation of general GAN network,the distance loss function between pixel values is also introduced.