Super-resolution algorithms generate high-resolution (HR) imagery from single or multiple low-resolution (LR) degraded images. In this paper, an efficient single image super-resolution (SR) algorithm using higher-order regression is proposed. Image patches extracted from HR image will have self-similar example patches near its corresponding location in the LR image. A higherorder regression function is learned using these self-similar example patches via. sparse representation model. The regression function is based on local approximations and henceforth estimated from the localized image patches. Taylor series is used as local approximation of the regression function and hence the zeroth order regression co-efficient will yield the local estimate of the regression function and the higher-order regression co-efficient will provide the local estimate of the higher-order derivative of the regression function. The learned higher-order regression mapping function is applied to LR image patches to approximate its corresponding HR version. The proposed super-resolution approach is evaluated with standard test images and is compared against state-of-the-art SR algorithms. It is observed that the proposed technique preserves sharp high-frequency (HF) details and reconstructs visually appealing HR images without introducing andy artifacts.
Digital Object Identifier (DOI)
Jino Hans, W.; Venkateswaran, N.; Narayanan, Srinath; and Ramachandran, Sandeep
"Estimation of Higher-order Regression via. Sparse Representation Model for Single Image Super-resolution Algorithm,"
Applied Mathematics & Information Sciences: Vol. 10
, Article 39.
Available at: https://dc.naturalspublishing.com/amis/vol10/iss5/39