Research On External Patch Prior And Global Low-rank Based Single Image Super-Resolution

With the advanced development of modern intelligent technology,the high resolution images are significant as a transmission in many fields,such as remote sensing,medical imaging,and scientific research.Due to the limitations existing in current hardware equipment of imaging system and the unpredictability of sampling circumstance while obtaining visual images,the visual images or image sequences could degrade into their low-resolution version.Therefore,researchers have proposed the image super-resolution reconstruction technology which improves image resolution by digital signal processing method.With the era of big data coming,the image super-resolution reconstruction technology becomes the research focus for its economical and meaningful features in reality engineering.Based on the sparse representation for single image super-resolution reconstruction method,this paper proposes a new dictionary learning strategy and reconstruction model to overcome the problem that the current reconstruction methods always cause the blurring along the edge structures.The new method combines the external patch prior and global low-rank regularization for image super-resolution,and improves the initial reconstruction image quality both in visual quality and objective index assessment.The main focus of this dissertation includes four parts:1 Research the dictionary learning strategy.The traditional sparse representation model adopts the K-means algorithm in patch clustering for dictionary learning,however this clustering algorithm could not be a good choice for patch structure measure for its depend on the Euclidean distance.To handle this problem,this paper proposes the dictionary learning strategy guided by external patch prior with the Gaussian mixture model.We first choose the training image data which includes various categories of images with rich texture and crop the feature images into patches.Then,we learn the Gaussian mixture model from the training patches to form the Gaussian subspace in which the training patches can be clustered by spatial matching degree guided by the Gaussian mixture model prior.Finally,with these cluster of patches,the sub-dictionary set which can describe image structures thoroughly is trained.2 Develop the image super resolution reconstruction model.Nonlocal self-similarity,which searches the similar patches in a neighborhood,cannot use theself-similarity effectively for inaccuracy of the similar patches searched.Besides,the weighted average operations in nonlocal self-similarity may smooth the high frequency information and blur the edge region of the image.In order to solve these problems,this paper proposes a model with the global low-rank constraint.First,we choose the sub-dictionary by the likelihoods of the patch belonging to each Gaussian space.Then,we search the global similar patches in a certain Gaussian space by Mahalanobis distance,and group the similar patches into a global similar patch.Finally,we make low-rank matrix approximation operation on the global similar patch.3 Solve the new model.We convert the new model into the weighted nuclear norm minimization problem,and solve it by the linearized alternating direction method with adaptive penalty.Furthermore,we utilize the regional redundancy to explicitly quantify the degree of patch redundancy and improve the weighted nuclear norm minimization to discriminate the different impact of different similar patch.4 Conduct the experiments.A large number of experiments are performed on different datasets with different upscaling factors,different upscaling methods to verify the performance of the proposed method and we compare the experiment results with several typical methods.The results show that the external patch prior and the global low-rank based single image super-resolution reconstruction model in this paper can effectively reconstruct the high-resolution image,both in the visual quality and the objective indices.