Super-resolution (SR) is a technique to obtain a higher resolution image (frame) by fusing multiple low-resolution (LR) images (frames) of the same scene. In a typical super-resolution algorithm, image registration is one of the most affective steps. The diffculty of this step results in the fact that most of the existing SR algorithms can not cope with local motions because image registration in general assumes global motion. Moreover, modeling SR noise including image registration error has great influence on the performance of the SR algorithms. In this paper, we report that Laplacian distribution assumption is good selection for global and slow motions that can be easily registered, while for fast motion sequences that contain multi-moving objects, Gaussian distribution is better for error modeling. Based on these results, we propose a cost function with weighted L₂-norm considering the SR noise model where the weights are generated from the error of registration and penalize parts that are inaccurately registered. These weights serve to reject the outlier image regions. Both the objective and subjective results demonstrate that the proposed algorithm gives better results for slow and fast motion sequences.

Date Published: 28 January 2008
PDF: 9 pages
Proc. SPIE 6822, Visual Communications and Image Processing 2008, 682222 (28 January 2008); doi: 10.1117/12.768747

Published in SPIE Proceedings Vol. 6822:
Visual Communications and Image Processing 2008
William A. Pearlman; John W. Woods; Ligang Lu, Editor(s)