Suppression of artifacts in super-resolved images.

Abstract

Super-resolution of imagery is an image restoration problem in which the goal is to recover attenuated spatial frequencies from diffraction-limited images. In this dissertation we consider the limits on super-resolution performance in terms of usable bandwidth of the restored frequency spectrum. Based on a characterization of the spectral extrapolation errors (viz. null objects), we derive an expression for an approximate bound on accurate bandwidth extension for the general class of super-resolution algorithms that incorporate a-priori assumptions of a non-negative, space-limited object. For super-resolution of sampled data, we show that it may be possible to achieve significant bandwidth extrapolation depending on the relationship between sampling rate, optical cutoff frequency, and object extent. Simulation results are presented which substantiate the derived bandwidth extrapolation bounds. Given the existence of oscillatory restoration artifacts, we present several artifact suppression techniques as adjuncts to a Poisson maximum a-posteriori (MAP) super-resolution algorithm developed by Hunt. It is shown that "resolution kernels" can be used to accomplish artifact-free restoration with limited bandwidth extension. To achieve increased bandwidth extension, a smoothness constrained MAP estimator is derived which demonstrates substantial artifact reduction in the restoration of natural scenes. It is shown that the constrained Poisson MAP estimator combines a Poisson image observation model with a specific form of Markov random field object prior. Several simulation results demonstrate the artifact suppression capabilities of the regularized algorithms.