Abstract: A feature matching method based on Log-Gabor filtering is proposed to address the problem of high-precision matching for multimodal remote sensing images. The method adopts a multi-scale dense matching framework via a coarse-to-fine manner, which avoids the low repeatability problem of feature detectors in multimodal images and is able to extract a large number of accurate correspondences. The method consists of two main steps: first, a feature pyramid robust to non-linear radiometric differences between images is constructed using multi-scale multi-angle Log-Gabor filters; then, the coarse feature map is used for dense template matching to extract a large number of coarse feature correspondences; the feature pyramid is then used to achieve bottom-up refinement of coarse correspondences layer by layer. Furthermore, to address the problem of inefficient sliding window operation for template matching, a fast implementation method of dense template matching is proposed, which effectively reduces the running time of dense template matching. The results show that the proposed method can overcome the influence of non-linear radiation differences between images, and outperforms existing multimodal image feature matching methods in terms of the number of correct matches, matching accuracy and matching precision.

Key words: multi-modal remote sensing image, feature matching, Log-Gabor filter, template matching, nonlinear radiation difference