Super-resolution Image Reconstruction Sina Jahanbin Richard Naething EE381K-14 March 10, 2005
Problem Statement • There is a limit to the spatial resolution that can be • recorded by any digital device. This may be due to: • optical distortions • motion blur • under-sampling • noise
SR image reconstruction is the process of combining several low resolution (LR) images into a single higher resolution (HR) image. Introduction to Super-resolution (SR) Reconstruction Techniques
“Restoration of a Single Superresolution Image from Several Blurred, Noisy, and Undersampled Measured Images”[Elad & Feuer, 1997] • Three main tools in single image restoration • Maximum likelihood (ML) estimator • Maximum a posteriori (MAP) • Projection onto convex sets (POCS) • This paper takes these existing single image restoration techniques and applies them to SR • A hybrid algorithm has been proposed that combines the ML estimator and POCS
“Superresolution Video Reconstruction with Arbitrary Sampling Lattices and Nonzero Aperture Time” [Patti, Sezan, & Murat, 1997] • Uses a model that takes into account details ignored by previous SR models • Arbitrary sampling lattice • Sensor element’s physical dimensions • Aperture time • Focus blurring • Additive noise
“Limits on Super-Resolution and How to Break Them”[Baker & Kanade, 2002] • Assumes image registration has already been accomplished and focuses on fusing step – or combining multiple aligned LR images into HR image • Uses what the authors call a “hallucination” or “recogstruction” algorithm • Claims significantly better results – both subjectively and in RMS pixel error
Future Work • Many papers on SR base their results on subjective viewing of images or use an objective measurement, such as RMS, that in many applications is not meaningful. • We propose to develop an objective measure of SR methods that has a basis in real world application performance.