Research On Super-resolution Image Reconstruction Algorithms Based On Generative Adversarial Networks

Image Super-Resolution(SR),which is an important problem in the field of image processing,refers to the process of restoring Low-Resolution(LR)image with less details to High-Resolution(HR)image with more texture details.Most of the existing SR algorithms based on deep learning do not make full use of the global features of the image in the process of super-resolution,which causes some unreasonable local details and the lack of image-level in the SR image.At the same time,the low-frequency information and highfrequency information are not considered separately,which leads to the lack of texture details and high-frequency information when SR algorithms are applied to the real scene.In order to solve these problems,this thesis proposes two new SR algorithms for real scenes,which are based on the Generative Adversarial Networks(GANs)in deep learning.The main tasks of this thesis is as following:(1)A super-resolution algorithm based on the self-attention GANs is proposed.In this algorithm,A GAN model is used to learn the degradation process from the HR image to LR image,which help us to obtain a pair of HR and LR images.In the SR process,The proposed self-attention layer which can excavate the global features is added to the G network and D network,to enhance the rationality of the SR image's details and the image-level.In training,we introduce the contextual loss based on feature contrast,which make the SR image partial texture clear and closer to the HR image.The experimental results show that the texture details of ours SR images are rich and the visual effect is better.(2)A super-resolution algorithm in image stationary wavelet domain based on GANs is proposed.In this algorithm,We also use a GAN model to learn the degradation process from the HR image to LR image,so as to obtain a pair of HR and LR images.In the SR process,the low-frequency and high-frequency components of the image obtained by 2D Stationary Wavelet Transform(SWT)are processed respectively.For low-frequency sub-band,a Lowfrequency Enhance Network(LEN)composed of residual blocks is designed to enhance it,in order to obtain the accurate low-frequency sub-band.For high-frequency sub-band,a High-frequency GAN(HGAN)is designed for super-resolution,the HGAN's G network is consists of dense blocks,which can provide more sufficient multi-scale detail information for high-frequency sub-band.In addition,The algorithm also uses the combined loss function in the image domain and the wavelet domain to guide the HGAN training.The numerical results on various datasets show that our algorithm can effectively improve the high-frequency information in SR image,and our SR image's texture details are rich and the evaluation index is better.