Performance Evaluation of Super-Resolution Reconstruction Algorithms Based On Linear Magnifications

Super Resolution based Reconstruction of images produces a High Resolution (HR) image from multiple Low Resolution (LR) images by estimating the motion parameters and shifts in the LR images. The problem can be divided into two parts: an image registration part in which the motion parameters and shift between different frames of the same scene are estimated and the reconstruction part in which an HR image is reconstructed from the registered images. In this paper, we consider the second part of the problem: The reconstruction step. We have compared six different reconstruction algorithms which are Bi-Cubic Interpolation method, Iterated Back Projection (IBP) algorithm, Points onto Convex Sets (POCS), Robust Super-Resolution (RSR), StructuredAdaptive Normalized Convolution (SANC) and PopulisGerchberg (PG) approach. The results are compared using Histogram Comparison Index (HCI) based on BHATTACHARYYA [23] distance which is a popular metric for color images comparisons. From an experimental evaluation, we find that SANC, POCS and Bi-Cubic Interpolation methods produce convincing results both under high and low magnification compared to other methods. On the other hand, PG algorithm and RSR degrade image quality on higher magnification.