Blade Profile Reconstruction via Multiview Registration With Automatic Overlap Distinction Mechanism and Mixed Weights of Feature-Distance Information

Multiview registration is a key step in achieving accurate reconstruction and measurement of blade profile. However, the complex blade free-form surface often leads to limited overlaps, nonprominent overlap-area features, and inconsistent point density in multiview data, making accurate registration challenging. Moreover, for blade reconstruction, it is essentially a partial-to-partial registration problem and the overlaps extraction process is typically required, which may introduce additional uncertainties and reduce reconstruction reliability. To overcome these challenges, we propose a multiview registration method by integrating an automatic overlap distinction mechanism and an effective mixed weight of feature-distance information. Specifically, by introducing a self-adaptive distance weight and constructing multiple support regions for each point, a multiscale distance-weight covariance descriptor (MDWCD) is designed to reveal the tiny differences between multiview data with less-prominent features and effectively recover their potential correspondences. A robust double bidirectional correspondence between multiview data is established to formulate the objective function of our method, effectively reducing the impact on registration accuracy posed by insufficient overlaps and density inconsistency. An automatic overlap distinction mechanism is proposed to avoid the introduction of additional uncertainties during registration. Here, a probability derived from the established correspondence will be assigned to each point, which can determine whether the point belongs to an inlier or outlier of the overlaps. A dynamic optimization framework based on a mixed weight of feature-distance information is introduced to continuously optimize the established correspondences. It can fully leverage the features and distance information of multiview data to dynamically estimate the optimal transformation in each iteration, achieving accurate registration. Finally, extensive experimental results demonstrate the robustness and accuracy of the proposed method.