Example-based super-resolution via structure analysis of patches

Abstract

To increase the resolution of an image, a lot of studies have been performed for several decades. Super-resolution is a process to produce a higher resolution image from one or more low-resolution (LR) images. Among them, an example-based super-resolution algorithm produces a high-resolution (HR) image from a single LR image based on relationships between LR images and their corresponding HR images. The relationships are examined via a training database which is built by piling pairs of LR and HR patches. This dissertation deals with an example-based super-resolution algorithm. A main factor influencing the performance of example-based super-resolution is in the method to determine appropriate high-frequency (HF) details from a training database. Conventional algorithms determine HF details based on the patch information in an input LR image. Since the information of a LR patch is limited, however, it is difficult to determine an appropriate HF patch. This may introduce unwanted artifacts or blurs in the obtained HR image. In order to alleviate the drawback of a LR patch based approach, this dissertation presents an example-based super-resolution based on a structure analysis of patches. In the algorithm, we utilize the sharpness of HR patch candidates for the reliable determination of HF patches. For each input patch, we first preselect a sufficient number of HF patches from a training database. We then produce HR patches by adding the selected HF patch candidates to the input patch. After examining the HR patches based on a reconstruction constraint in the LR image domain, we remove the outlier HF patches. We finally reselect several HF patches according to the patch characteristic for producing a final HR image. In this selection refinement, a modified local contrast is used for edge patches, while magnitudes and angles of gradients are used for non-edge patches. After the HF patch selection, we apply a pixel-level optimization process based on a robust sta...