High-quality still images from video frame sequences

Abstract

This article tackles the classic super-resolution (SR) problem Elad99 of obtaining a high-resolution (HR) still image from a sequence of low-resolution (LR) images that have been warped and sub-sampled. The goal here is to recover frequencies higher than the Nyquist frequency by merging the LR information. We focus on the critical step of the SR process before any fusion technique can be applied which consists of the registration of LR images with an arbitrary reference image at sub-pixel accuracy. We propose a registration algorithm for color images, derived from the one described by Djamdji and Bijaoui in Ref. 2. This algorithm achieves automatic feature-based registration at sub-pixel accuracy. It seeks to take advantage of the multi-band (RGB) information in a color image to improve the robustness and accuracy compared to more usual greyscale registration. The fusion of the data from LR images into a higher resolution image is then carried out through thin plate spline interpolation. The results show the algorithm's performance for simulated image sets. The influence of several parameters on the registration algorithm is described.