Abstract
Registration errors between two images produce artefacts when their per-pixel difference is used to detect changes. These artefacts constitute a source of noise hampering image interpretation. In this article, an adaptive filtering approach for misregistration artefact reduction is presented. Both univariate and multichannel images are considered. The proposed filters rely on robust statistics, switching mechanisms and a misregistration-induced change estimation model. An evaluation performed on a synthetic image confirms (1) the high efficiency of the approach in both reducing misregistration artefacts and preserving real changes and (2) the advantages of the method over the Knoll and Delp filter when applied to a difference image. Experiments conducted on real images show, from careful visual analysis, that the adaptive filters are able to remove misregistration noise with good capability whilst preserving real changes. The methods are designed for handling misregistration errors in the order of one to two pixels.
Acknowledgement
I thank Robert Fraser for supplying the data used in §4.2.3.