An Experimental Comparison of Super-resolution Reconstruction for Image Sequences

Super-resolution reconstruction for image sequences is a promising image processing technology that using complementary information among a set of images to reconstruct a high-resolution image. Several super-resolution reconstruction algorithms have been studied in the literature to reconstruct a high-resolution image. In this paper, first, after presenting a condensed introduction of image registration algorithms including Lucchese algorithm, Vandewalle algorithm and Keren algorithm, we experimentally compare the relative merits of these registration algorithms in terms of registration accuracy and noise reduction. Secondly, we experimentally compare four image reconstruction methods: projection onto convex sets method (POCS), iterative back-projection method (IBP), robust super resolution (Robust SR) and structure-adaptive normalized convolution (Structure-Adaptive NC), mainly in terms of Peak Signal to Noise Ratio (PSNR), in which salt and pepper noise is added in the low resolution image. It is clearly demonstrated that the combination of Keren algorithm and Structure-Adaptive NC can achieve the best performance regarding the Lena image.