Abstract
It has been widely acknowledged that learning-based super-resolution (SR) methods are effective to recover a high resolution (HR) image from a single low resolution (LR) input image. However, there exist two main challenges in learning-based SR methods currently: the quality of training samples and the demand for computation. We proposed a novel framework for single image SR tasks aiming at these issues, which consists of blind blurring kernel estimation (BKE) and SR recovery with anchored space mapping (ASM). BKE is realized via minimizing the cross-scale dissimilarity of the image iteratively, and SR recovery with ASM is performed based on iterative least square dictionary learning algorithm (ILS-DLA). BKE is capable of improving the compatibility of training samples and testing samples effectively and ASM can reduce consumed time during SR recovery radically. Moreover, a selective patch processing (SPP) strategy measured by average gradient amplitude |grad| of a patch is adopted to accelerate the BKE process. The experimental results show that our method outruns several typical blind and non-blind algorithms on equal conditions.
Highlights
Single image super-resolution has been becoming the hotspot of super-resolution area for digital images because it generally is not easy to obtain an adequate number of low resolution (LR) observations for SR recovery in many practical applications
The parameter settings in blurring kernel estimation (BKE) stage are partially the same with Refs. [16] and [25], i.e., when scale factor s = 2, the size of small query patches pi and candidate patches qisj of nearest neighbors (NNs) are typically set to 5 × 5, while the sizes of “parent” patches qij are set to 9 × 9 and 11 × 11; when performing ×3 SR, query patches and candidate patches do not change size but “parent” patches are set to be 13 × 13 patches
We proposed a novel single image SR processing framework aiming at improving the SR effect and reducing SR time consumption in this paper
Summary
Single image super-resolution has been becoming the hotspot of super-resolution area for digital images because it generally is not easy to obtain an adequate number of LR observations for SR recovery in many practical applications. Chang et al [8] proposed a nearest neighbor embedding (NNE) method motivated by the philosophy of locally linear embedding (LLE) [9] They assumed LR patches and HR patches have similar space structure, and LR patch coefficients can be solved through least square problem for the fixed number of nearest neighbors (NNs). Yang et al [14] proposed another joint dictionary training approach to learn the duality relation between LR/HR patch spaces. L2 norm regularization was used to substitute L0/L1 norm constraint so that latent HR patch can be mapped on LR patch directly through a mapping matrix computed by LR/HR dictionaries This strategy is similar with ANR [17], but we employed a different dictionary learning approach, i.e., ILS-DLA, to train LR/HR dictionaries.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
