Abstract:Image matching serves as a fundamental prerequisite for numerous tasks in computer vision, playing a central role in 3D reconstruction, image registration, multi-source image fusion, and geographic information extraction. However, geometric deformation, illumination variation, and noise interference caused by complex environmental factors pose challenges to the stability and robustness of feature matching methods. To this end, this study proposes a random sample consensus-based outlier rejection algorithm—DPCSAC—driven by model residual clustering. For an initial set of correspondences containing outliers, the proposed method conducts sampling in the model space derived from the initial set of correspondences, computes the residual matrix for each point pair under the sampling models, and then refines the inlier set by employing density peak clustering (DPC). Finally, model estimation and consistency verification are performed on the refined inlier set to obtain stable and accurate matching results. Classical and novel image matching methods are compared on VGG, Kusvod2, and Mag datasets. The results demonstrate that the proposed method shows its superiority in indicators such as RMSE and Acc, thus validating its adaptability and effectiveness in complex scenarios.