Image denoising techniques are traditionally evaluated on images corrupted by synthesized i.i.d. Gaussian noise. We aim to obviate this unrealistic setting with a methodology for benchmarking denoising techniques on real photographs. Based on pairs of images with different ISO values and appropriately adjusted exposure times, we obtain ground truth by carefully post-processing the nearly noise-free low-ISO image. Our Darmstadt Noise Dataset (DND) enables the realistic evaluation of denoising techniques and yields several interesting insights.

This presentation is part of Minisymposium “MS14 - Denoising in Photography and Video (2 parts)”
organized by: Stacey Levine (Duquesne University) , Marcelo Bertalmío (University Pompeu Fabra) .

Authors:

Stefan Roth (Technische Universität Darmstadt )

Keywords:

benchmarking, image denoising, image reconstruction