Fast Direct Super-Resolution by Simple Functions

Abstract

The goal of single-image super-resolution is to generate a high-quality high-resolution image based on a given low-resolution input. It is an ill-posed problem which requires exemplars or priors to better reconstruct the missing high-resolution image details. In this paper, we propose to split the feature space into numerous subspaces and collect exemplars to learn priors for each subspace, thereby creating effective mapping functions. The use of split input space facilitates both feasibility of using simple functions for super-resolution, and efficiency of generating highresolution results. High-quality high-resolution images are reconstructed based on the effective learned priors. Experimental results demonstrate that the proposed algorithm performs efficiently and effectively over state-of-the-art methods.

Single-Image Super-Resolution: A Benchmark

“…The formation is compatible with most SR methods [16,11,17,30,46,43,36] where the reconstructed images are well aligned with the ground truth images.…”

“…For fair comparisons, we evaluate the methods using the original binaries or source code [27,46,17,5,43,36]. In addition, we implement four state-of-the-art algorithms when the source or binary code is not available [16,30,11,9].…”

“…We note these methods are implemented in different programming languages. For algorithms where the blur kernel width is an adjustable parameter [16,27,11,46,5,43], we set the same values as used in the LR image formation. We only evaluate the SSXS, KK, and FF methods [30,17,9] under scaling factors 2, 3, and 4 because the released code or priors only support these scaling factors.…”

“…From the ground truth HR images, we generate LR test images using various settings of scaling factor and blur kernel width. We generate the SR images by the originally released code [27,17,46,5,43,36] or our implementation [16,30,11,9] if the code is not available. The generated SR image are used to evaluate the performance of SISR algorithms and quality assessment metrics.…”

“…While domain-specific SISR algorithms focus on specific classes of images such as faces [35,42], scenes [33], and graphics artwork [18], generic SISR algorithms [10,38,8,3,27,30,11,46,9,34,32,12,5,39,44,43] are developed for all kinds of images where the priors are typically based on primitive image properties such as edges and segments. In order to evaluate the performance of a SISR algorithm, human subject studies or ground truth images are used [33,43]. In this work, we focus on performance evaluation of state-of-the-art SISR algorithms under different settings based on a set of ground truth images.…”