Pair-wise Mapping Based Single Image Super-resolution Reconstruction

Due to the physical limitations of the imaging device and the influence of theimaging environment, it is difficult to obtain an image or image sequence with highresolution that can show all the details of the original scene. To improve the spatialresolution of the obtained image, researchers conduct two different studies on thisproblem. On one hand, some researchers attempt to upgrade the hardware device ofimaging system; on the other hand, some researchers regard the image as some kind ofsignal to achieve the enhancement on the spatial resolution. As the cost of the upgradedhardware is great, the super-resolution reconstruction (SR), as an economical andefficient technology, has attracted even-increasing attention.This paper focuses on single image SR based on structural information. Forestablishing the mapping relationships between low-resolution (LR) image andhigh-resolution (HR) image, this paper makes a deep research on Gaussian processregression and regularization priori. The major contributions of this paper aresummarized as follow:(1) To construct the nonlinear relationship between LR block and HR pixel, asingle image SR reconstruction algorithm based on Gaussian process regression andimage patches classification is proposed. Considering the requirements of the Gaussianprocess regression on the correlation between training data and test data, we classifyimage blocks with the similar structure into the same category, and then inference themapping relationships on the same category of image blocks by Gaussian processregression. The algorithm improves the existing SR reconstruction based on Gaussianprocess regression and can achieve the magnification and deblurring at the small scalesimultaneously.(2) This paper proposes a single image SR reconstruction algorithm based onconjunct regularization restraint (CRR) with the help of regularization priori. Thismethod mainly constructs three kinds of regularization constraint items. First, sparserepresentation is applied to formulate the regularization items for establishing themapping relationships between LR image and HR image which can generate richhigh-frequency details. To target the problem that inappropriate training samples causeunsatisfactory reconstruction results, we use structural information of the image itself totrain the over-complete dictionary, thereby to recover the lost high-frequency details.Second, taking it into account that the gradient histogram can efficiently deliver the texture information, Gradient histogram preservation (GHP) is used to design theregularization. GHP can remain the consistency of the gradient histogram between theHR image and the denoising image, and then preserve the fne scale texture structureswhile removing noise. At last, the non-local means (NLM) filter is adopted to constructnon-local regularization term with the result of the captured similarity redundancy at thesame scale. The estimated pixel is the mean of the weighted similarity pixels. So NLMcan suppress the apparent artifacts and maintain sharp edges. The effectiveness of theproposed algorithm is verified by the experiment.