A Single Image Super-Resolution Algorithm Using Non-Local-Mean Self-Similarity and Noise-Robust Saliency Map

Hui Jung LEE
Dong-Yoon CHOI
Kyoung Won LIM
Byung Cheol SONG

IEICE TRANSACTIONS on Information and Systems   Vol.E100-D    No.7    pp.1463-1474
Publication Date: 2017/07/01
Publicized: 2017/04/05
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2016EDP7291
Type of Manuscript: PAPER
Category: Image Processing and Video Processing
self-similarity,  super-resolution,  self-example,  NLM,  soft-switching,  

Full Text: PDF(3.6MB)>>
Buy this Article

This paper presents a single image super-resolution (SR) algorithm based on self-similarity using non-local-mean (NLM) metric. In order to accurately find the best self-example even under noisy environment, NLM weight is employed as a self-similarity metric. Also, a pixel-wise soft-switching is presented to overcome an inherent drawback of conventional self-example-based SR that it seldom works for texture areas. For the pixel-wise soft-switching, an edge-oriented saliency map is generated for each input image. Here, we derived the saliency map which can be robust against noises by using a specific training. The proposed algorithm works as follows: First, auxiliary images for an input low-resolution (LR) image are generated. Second, self-examples for each LR patch are found from the auxiliary images on a block basis, and the best match in terms of self-similarity is found as the best self-example. Third, a preliminary high-resolution (HR) image is synthesized using all the self-examples. Next, an edge map and a saliency map are generated from the input LR image, and pixel-wise weights for soft-switching of the next step are computed from those maps. Finally, a super-resolved HR image is produced by soft-switching between the preliminary HR image for edges and a linearly interpolated image for non-edges. Experimental results show that the proposed algorithm outperforms state-of-the-art SR algorithms qualitatively and quantitatively.

open access publishing via