Fast super-resolution based on weighted collaborative representation
Hailiang Li, Kin‐Man Lam · 2014
Recently, collaborative representation (CR) has been proposed as an l2-norm least-square solution for image super-resolution with significantly less computation than the l1-norm version of the Sparse-Coding-based Super-Resolution (ScSR) without any sacrifice in terms of image quality. In this paper we propose a novel weighted collaborative representation (WCR) instead of the original CR model for single image super-resolution. Our proposed method can achieve more than a 0.2~0.3 dB gain without requiring any additional cost compared to the original CR model. Moreover, we devise a hierarchical-clustering KD-tree searching scheme which can reduce the computational complexity on searching part in our WCR model from O(n) to O(n1/m), where n is the atom count and m is the number of layers, without any compromise of image quality.