Multimodal Remote Sensing Image Matching Considering Structural Saliency

Abstract

Nonlinear radiometric distortion, orientation reversal, and image rotation considerably affect multimodal image matching performance in remote sensing. Furthermore, existing multimodal image matching algorithms are highly complex and inefficient. We addressed these issues by developing a reliable matching method considering structural saliency (RMSS) for multimodal remote sensing images. First, we propose a method to describe structural saliency, which reliably represents both structural intensity and directional information. Second, we achieve rotation invariance in the descriptors by integrating structural information with a directional histogram descriptor framework. Subsequently, we design a coarse-to-fine matching strategy and propose an adaptive fast sample consensus improvement algorithm, which effectively enhances matching stability and accuracy. Finally, we design a two-stage subpixel refinement strategy to further improve localization accuracy and the number of matched points. The comparison of RMSS with seven state-of-the-art algorithms using six types of multimodal remote sensing images as experimental data demonstrates that the RMSS is simple to implement, exhibits high robustness to rotation in multimodal images, and can substantially improve the number and localization accuracy of matching points and the correct matching rate.