Shape registration for remote-sensing images with background variation

Abstract

Image registration is a basic and key processing art for multi-sensor and/or multi-temporal remote-sensing images. Generally speaking, multi-temporal images covering the same region may contain different textures and/or changed backgrounds where the classical autonomous registration methods usually fail, especially those before and after disasters. The main reason is that a single pixel with surrounding region becomes unstable and cannot be applied for descriptor generation. However, the topology (or shape) between points should be viewed as steady description. In this article, we introduce a level line and consequent shape recognition algorithm (level line descriptor (LLD)) from close-range applications and propose an improved LLD for remote sensing. We present our analysis based on multi-temporal images with moderately changed backgrounds. The numerical experiments illustrate the operator's potential to become a relatively robust alternative for shape-based remote-sensing registration.

Acknowledgements

We are grateful to David Lowe, Rob Hess, and Pablo Muse for sharing their source code. All experiments were carried using the publicly available software MegaWave (authors: Jacques Froment and Lionel Moisan). Special acknowledgements must be given to Libo Yin and Jordan, who proofread and polished some of the text. This research was supported by New Century Excellent Talents in University, grant number NCET-08-36.