A Multi-Source Feature Registration Network Based on Differential Noise Suppression

Cross-source point cloud registration is a key task in computer vision and the 3D reconstruction, aiming to accurately align point cloud data from different sensors. Due to the different working mechanism of the sensor itself, the obtained point cloud data has significant differences in density, noise, and geometric features, which brings challenges to registration. The existing point cloud registration algorithms are only verified on simulation data sets, but their performance is not satisfactory on real cross-source data sets. To solve this problem, a multi-source feature registration network based on differential noise suppression is proposed. Differential noise suppression point cloud registration network (DSPR) combines a multi-layer MLP network with a multi-head self-attention mechanism to learn the characteristics of structure distribution at different points. The transformation matrix is solved by matching the features through finite difference and iterative loop. The core component is the differential noise suppression encoder, which can effectively suppress heterogeneous noise in point clouds by utilizing feature differences and increasing the structural overlap rate, thereby enhancing the model's understanding of cross-source point cloud data. Experiments on the Multi-Source Satellite Dataset validate the effectiveness of DSPR, and the results indicate that the method significantly improves registration accuracy and stability in cross-source point cloud data registration tasks, demonstrating strong adaptability.