Performance analysis on image super-resolution using kernel resolution synthesis based on SVR
Jacob Thomas, S. Mohan, Jesna Anver · 2011
Image super-resolution is the process by which additional information is incorporated to enhance a low resolution image thereby producing a high resolution image. In the simplest case super-resolution of a single image is a process of obtaining high-resolution image with more number of pixels with more resolving power. Therefore the super-resolved image should demonstrate an improvement in the perceived detail content compared to that of the low-resolution images. This will typically involve restoration of the high-frequency content. Along with the original information inherent within the low-resolution image, the additional information may come in several forms: a group of several shifted versions of the low resolution image, a collection of optimally estimated filters selected for specific image content i.e., a relationship representing a training set that contains low and high resolution image pairs. The common image interpolation functions, like the cubic spline for example, often gives blur edges and image details, and that such analytical methods cannot reproduce details in textured regions well. The new training-based super-resolution approach using the support vector regression, which maintains all the main features that characterize the maximal margin algorithm used for classification to regression analysis, such as duality, sparseness, kernel and convexity, produces better results.