A locally adaptive semi-dense matching method for multimodal remote sensing images

Non-rigid deformation between multimodal remote sensing images makes it difficult to adopt a fixed model to filter and augment matching points. Moreover, nonlinear radiometric variations between images lead to low matching precision, a limited number of matching points, and poor spatial distribution quality. To address these issues, we propose a locally adaptive semi-dense matching method (LADM) for multimodal remote sensing images. First, we employ multimodal image matching algorithms, such as the weighted phase histogram, to perform initial matching and obtain a set of initial matching points. Second, we pre-detect the spatial structure of Delaunay neighbouring points using angular similarity and directional consistency, and we apply a local affine transformation model for consistency statistics to rapidly remove obviously incorrect matching points, yielding sets of local matching points. Subsequently, we adopt a reverse pruning strategy to detect medium-confidence matching points from high-confidence matching points, thereby eliminating false matches that satisfy local consistency. In addition, we utilize a locally adaptive model to detect feature points, which increases the number of matching points and improves their spatial distribution quality according to local image deformation characteristics. Finally, from the perspective of global consistency, we employ the least squares method to remove matching points with large errors and obtain the final accurate matching points. We conducted experiments on 24 multimodal image pairs involving translation, rotation, and scaling, and compared LADM with five state-of-the-art algorithms. The results demonstrate that the proposed algorithm achieves superior performance across multiple evaluation metrics, significantly improving matching precision, yielding a large number of matching points, and enhancing their spatial distribution quality.