Research On The Single Image Super-Resolution Reconstruction Method Based On Low-Rank Constraints

As the basis of human vision,images are an important source of human understanding of the world.With the rapid development of computer multimedia technology,people expect higher and higher image quality.Due to factors such as the imaging equipment and the external environment,the resolution of images usually does not meet the needs of practical applications.How to improve the resolution of images is an important challenge in the field of image processing.Super-resolution reconstruction technology can break through the limitation of imaging equipment and realize the reconstruction of high-resolution images from the software aspect,which is a research hotspot of common concern for scholars at home and abroad in recent years,and is widely used in remote sensing,medicine,military public security,robotics and other fields.Single image super-resolution algorithms are more challenging because there is only one input image and limited image information available to be mined.After compressed sensing,the low-rank model is another important theory in data acquisition and representation in recent years.It can effectively eliminate noise and divide all sample data into their respective subspaces,thus mining the intrinsic structure of data and contributing to data analysis,and its application to the superresolution reconstruction problem is an effective way to improve the reconstruction quality.Therefore,aiming at the difficulties in single image super-resolution reconstruction,this thesis conducts an in-depth study on optimizing the performance of the reconstruction algorithm based on low-rank constraints.The main contents and innovations are:(1)A single image super-resolution reconstruction method with edge-preserving based on low-rank representation is proposed.Data redundancy is an important challenge in some existing sample learning based image super-resolution reconstruction methods.The super-resolution reconstruction methods based on lowrank representation can fully explore images’ global information and effectively divide the sample data into each subspace.However,they ignore the local prior information in the image,which easily lead to edge blurring and key details missing in the reconstructed images.To solve this problem,this dissertation extracts the gradient-domain guided image filtering from high-resolution dictionary atoms and introduces it into low-rank representation as high frequency prior information to enhance image edge details,and a structure-constrained low-rank representation model with edge preserving ability is constructed.The model not only considers the sparsity and correlation of the image,but also improves the local details of the image while fully mining its essential structure,achieving a full and effective combination of local prior information and global information.(2)An image super-resolution reconstruction method combining edge-preserving regularization and low-rank constraint is proposed.The non-local self-similarity reconstruction methods based on sparse representation show good performances because they consider the non-local self-similarity of images based on sparse representation.However,the nonlocal self-similarity constraint may produce edge artifacts and local structure blurring when encountering high noisy situations,which cannot effectively protect the key details of the image.To solve this problem,from the perspective of the similarity of the edge structure between the reconstructed image and the original image,this dissertation applies the side window gradient-domain guided filtering to the low-resolution input image and the degraded reconstructed image,combines the difference between the two as a constraint on edge consistency into the non-local self-similarity model of sparse representation,and constructs a sparse coding model that can enhance image patch details.In addition,the low-rank matrix recovery technique is used to remove potential sparse noise from the reconstructed initial super resolution image,which enhances the image detail and improves the robustness of the image to noise.(3)A locally regularized and adaptive low-rank constraint for image superresolution reconstruction method is proposed.The method based on collaborative representation combines the idea of sparse representation and neighbor embedding to improve the algorithm effectiveness.However,the methods based on collaborative representation may not be accurate in describing the mapping relationship between low-resolution space and high-resolution space.Meanwhile,this kind of algorithms are not very suitable for some complex images and are prone to noise interference.In view of this,this dissertation utilizes the correlation between dictionary atoms and neighborhood samples to construct a local structural prior,constraining the representation coefficients of collaborative representations to optimize the projection matrix.Moreover,the shape adaptive low-rank constraint is used to explore the nonlocal self-similarity between image patches.A locally regularized and adaptive lowrank constraint reconstruction model is constructed.This model can more accurately represent the nonlinear mapping relationship between low-resolution space and highresolution space,reduce the influence of noise on the quality of reconstructed image,and improve the visual perception effect of reconstructed image.