Non-Local Kernel Regression for Image and Video Restoration
by Haichao Zhang et al.
Introduction
The proposed Non-Local Kernel Regression (NL-KR) method exploits both the local structural and non-local similarity information for image and video processing, e.g. super resolution.
The basic idea is illustrated in the following chart:
Non-Local Kernel Regression: (1) Similar patch searching: different colors indicate the similarity (red the highest, green the medium and blue the least); (2) Structural kernel estimation and reweighting: estimate a regression kernel adapted to the structure at each position where the similar patches reside and re-weight them according to similarity; (3) Non-local kernel regression: estimate the value for the query point with both local structural and non-local similar information in raster-scan order.
Results
Some SR results on single images:
More results on single image and videos:
Related Publications:
Image and Video Restorations via Non-Local Kernel Regression
Haichao Zhang, Jianchao Yang, Yanning Zhang and Thomas S. Huang
IEEE Trans.on Systems, Man, and Cybernetics-Part B (IEEE TSMC-B), 2012
Non-Local Kernel Regression for Image and Video Restoration
Haichao Zhang, Jianchao Yang, Yanning Zhang and Thomas S. Huang
The 11th European Conference on Computer Vision (ECCV) 2010
Related Projects: