Blind Single Image Super-resolution Based On Cross-scale Low Rank Regularization

Although spatial resolution is an important factor to determine the image quality,the inherent sampling frequency of the sensor limits the spatial resolution of the image.Highresolution image is profile to its further analysis and processing.Image super-resolution technology aims to recover or reconstruct reasonable high-frequency components for low-resolution images by using one or many low-resolution images.It breaks through the limitation of sensor inherent sampling frequency,so as to achieve the purpose of improving image spatial resolution.In practice,it is difficult to obtain a series of lowresolution images of the same scene at the same time,so that the single image superresolution is more universal.Although most of the existing image super-resolution methods assume the blur kernel is known,the blur kernel in real scene is much more complicated,which leads to severe drop of the reconstructed image quality when the assumed blur kernel is inconsistent with the real one.Single image blind super-resolution aims to reconstruct high-resolution image from low-resolution image with unknown blur kernel,which is a severely ill-posed inverse problem.The key to solve the problem is to introduce the image prior knowledge to provide extra information for estimating the blur kernel and high-resolution image.In this thesis,we introduce the image cross-scale self-similarity prior,low rank prior and image pyramid structure to research the blind super-resolution regularization methods.The main work and innovation points of this thesis include the following two aspects:1.In this thesis,we propose a single image blind super-resolution method based on cross-scale low rank prior,which is combining cross-scale self-similarity with low rank prior.According to the cross-scale self-similarity among high-resolution image,lowresolution image and its down-sampled image,we construct the cross-scale similar image patch group.By using the low rank matrix approximation to explore the low rank structure of this group,the reconstructed high-resolution image gradually approximates the real high-resolution image,which is enforced the estimated blur kernel to approximate the ground-truth kernel during the iterations.In addition,the low rank matrix approximation elegantly indicates the global structure of data,which can improve the robustness to noise.Both qualitative and quantitative experiments demonstrate that,compared with other methods,the proposed method can estimate both the more accurate blur kernel and more visually favorable high-resolution image.2.In order to solve the problem that the down-scaled factor is too large to rarely cross-scale similar image patches between the down-sampled image and the highresolution image.In this thesis,we proposed a single image blind super-resolution method based on pyramid cross-scale low rank prior,which uses a large number of cross-scale similar image patches in image pyramid to provide more additional information for reconstructing high-resolution image.The proposed method establishes multiple corresponding relationships among high-resolution image,low resolution image and its down-sampled image in the image pyramid.We construct cross-scale similar image patch group in each corresponding relationship,and the low rank matrix approximation is applied to explore the latent low rank structure of this group.In the iterative process,the blur kernel and high-resolution images are estimated effectively.Both qualitative and quantitative experiments demonstrate that,compared with other methods,the proposed method can estimate both the more accurate blur kernel and more visually favorable highresolution image when the size of the down-sampled image is much smaller than that of the high-resolution image.