Theoretical performance assessment and empirical analysis of super-resolution under unknown affine sensor motion.

Abstract

This paper deals with super-resolution (SR) processing and associated theoretical performance assessment for under-sampled video data collected from a moving imaging platform with unknown motion and assuming a relatively flat scene. This general scenario requires joint estimation of the high-resolution image and the parameters that determine a projective transform that relates the collected frames to one another. A quantitative assessment of the variance in the random error as achieved through a joint-estimation approach (e.g., SR image reconstruction and motion estimation) is carried out via the general framework of M-estimators and asymptotic statistics. This approach provides a performance measure on estimating the fine-resolution scene when there is a lack of perspective information and represents a significant advancement over previous work that considered only the more specific scenario of mis-registration. A succinct overview of the theoretical framework is presented along with some specific results on the approximate random error for the case of unknown translation and affine motions. A comparison is given between the approximated random error and that actually achieved by an M-estimator approach to the joint-estimation problem. These results provide insight on the reduction in SR reconstruction accuracy when jointly estimating unknown inter-frame affine motion.