Remote sensing image registration based on Gaussian-Hermite moments and the Pseudo-RANSAC algorithm

This study deals with an important issue that is often encountered with the registration of remote sensing images which are obtained at different times and/or through inter/intra sensors. Remote sensing images may differ significantly in gray-level characteristics and contrast, among other aspects. Thus, it may be difficult to apply directly area-based approaches which are dependent on image intensity values. In this work, a novel image registration approach based on Gaussian-Hermite moments and the Pseudo-RANSAC algorithm is proposed. The problem of intensity difference commonly incurred in multi-temporal or multimodal remote sensing image registration is tackled using features that are invariant to intensity mapping during the feature point matching process. In particular, the feature points are herein represented by a range of newly introduced Gaussian-Hermite moments, and the corresponding feature points in a certain reference image are sought with the Euclidean distance measure. Moreover, an improved RANdom SAmple Consensus (RANSAC) algorithm is presented, reducing computational time complexity while improving performance in stability and accuracy. The final warping of images according to their refined feature points is conducted with bilinear interpolation. The proposed approach has been successfully applied to register synthetic and real remote sensing images, demonstrating its efficacy with systematic experimental evaluations.