Research On Face Super-resolution Reconstruction Based On Improved Generative Adversarial Network

The resolution of images is a key factor affecting the application of face images.It directly affects the application prospects of face images in face detection,face recognition and face analysis.However,the disadvantages of low SNR and high frequency detail loss still exist in face image acquisition.Therefore,it is widely used to improve the resolution of face images acquired by imaging systems.Improving the image resolution by improving hardware facilities such as sensors is the easiest and straightforward way,but there are limitations such as high price,long investment cycle and limited space for improvement.Therefore,this thesis focuses on the superresolution reconstruction technology to improve the resolution of face images,and draws on the super-resolution reconstruction model of the adversarial network.The single-frame/multi-frame face image super-resolution reconstruction algorithm based on improved generative adversarial network is proposed.A prototype system based on end-to-end face image super-resolution reconstruction is constructed,which develops and implements related functions of the system.The specific research content is as follows:(1)Propose a single-frame face image super-resolution reconstruction model based on improved generative adversarial networkIn this paper,the super-resolution reconstruction model of the generative adversarial network is applied to the face image,and it is improved.A single frame face image super-resolution reconstruction model based on the improved generative adversarial network is proposed.The model improves the generation network,discrimination network and loss function of the original generation countermeasure network super-resolution reconstruction model.And compared with other popular single-frame super-resolution reconstruction algorithms on the relevant face datasets,based on the improved super-resolution reconstruction model of the generative adversarial network,more external sample knowledge can be obtained.Experimental results show that compared with other super-resolution reconstruction algorithms,it has a better reconstruction effect.(2)Propose a multi-frame face image super-resolution reconstruction model based on improved generative adversarial networkSuper-resolution reconstruction of multi-frame face images introduces time-space correlation between sequence images,which can obtain more high-frequency information.Therefore,based on the super-resolution reconstruction model of singleframe face image,this paper introduces motion compensation and weight representation to extend the network based on the improved generation-oriented single-frame face image super-resolution reconstruction network to multiple frames.Improve the multiframe face image super-resolution reconstruction model that generates the adversarial network.Compared with other commonly used multi-frame image super-resolution reconstruction algorithms and improved generation-oriented network single-frame face image reconstruction algorithms,high-resolution face images with better reconstruction effects are obtained.(3)Construct an end-to-end face image super-resolution reconstruction prototype systemBased on the previous research,an end-to-end face image super is proposed by cascading the improved single-frame/multi-frame face super-resolution reconstruction algorithm based on the improved generative adversarial network.Resolution reconstruction network model.On this basis,a C/S mode end-to-end face superresolution reconstruction prototype system is built.The related functions of the system are developed and implemented in the Windows environment.