Accurately and robustly aligning multiple viewpoints into a complete profile, i.e., 3-D reconstruction, is crucial for the optical measurement of thin-walled, twisted blade profiles. However, existing mainstream reconstruction methods heavily depend on the motion stability and geometric accuracy of the measurement system. To address this issue, we propose a multiview fine registration framework based on self-supervised simultaneous optimization (MFRSSO) to improve the reconstruction accuracy. Specifically, MFRSSO designs a self-supervised loss to simultaneously score the alignment of all viewpoints and proposes a training-as-optimization paradigm to directly optimize global rigid transformations via gradient descent, avoiding challenges, such as cumulative errors, insufficient robustness, and local optima, commonly found in current point cloud registration methods. In addition, the proposed paradigm eliminates the requirement for parameter generalizability, thereby preventing performance degradation caused by domain differences between training and testing datasets inherent in the traditional train-then-test paradigm of machine learning. Experimental results on three representative blades show that MFRSSO consistently outperforms nine state-of-the-art algorithms in both accuracy and robustness, achieving a reconstruction accuracy of 0.001 mm and improving accuracy by over 70%, demonstrating its significant application prospect.

Currently, optical-based method provides a feasible means for blade reconstruction and measurement. However, as the key step, multiview registration is still heavily reliant on the measurement system's motion stability and geometric accuracy. Moreover, the complex and high-reflective freeform surface of blades may result in both insufficient overlaps and less prominent overlap-area features between adjacent views, making accurate multiview data alignment difficult. Thus, we propose a method to realize accurate 3-D reconstruction and measurement of blade profiles based on a scanning system with a laser sensor and a coarse-to-fine registration strategy. First, coarse alignment of the multiview data is achieved by calibrating the system's rotational axis using the blade datum plane feature, which can provide a good initial value for the fine registration and improve the reconstruction efficiency. Then, a fine registration algorithm based on the dynamic encoding of feature-coordinate fusion information is proposed to refine the coarsely aligned data and reduce the effect of system motion error on registration accuracy. Here, the introduction of fusion information can effectively eliminate the redundant correspondences to continuously optimize the matching probability between multiview data in each iteration, improving registration accuracy and efficiency. Finally, experiments on typical blades and comparison with the other fine registration algorithms demonstrate the accuracy and robustness of the proposed method.

Optical measurement methods for blade profiles attract lots of interest in industry. Due to the nature of the thin-walled and twisted spatial freeform surfaces of blades, the measurement accuracy would be significantly affected by the accumulated error associated with the geometric accuracy and motion stability of the developed multiview system. To overcome these issues, this article proposes a deep feature interaction network for fine registration of the multiview data. In our network, we design a two-branch structure to integrate a global and a local feature extraction branch to encode point cloud features. Moreover, we propose a feature interaction module to strengthen information association between two point clouds during feature extraction. Next, an attention mechanism is used to fuse matching information between two matching matrices obtained from the global-based and the local-based features. Experimental results demonstrate the feasibility and good practical application prospect of this method.

Learning an efficient projection to map high-dimensional data into a lower dimensional space is a rather challenging task in the community of pattern recognition and computer vision. Manifold learning is widely applied because it can disclose the intrinsic geometric structure of data. However, it only concerns the geometric structure and may lose its effectiveness in case of corrupted data. To address this challenge, we propose a novel dimensionality reduction method by combining the manifold learning and low-rank sparse representation, termed low-rank sparse preserving projections (LSPP), which can simultaneously preserve the intrinsic geometric structure and learn a robust representation to reduce the negative effects of corruptions. Therefore, LSPP is advantageous to extract robust features. Because the formulated LSPP problem has no closed-form solution, we use the linearized alternating direction method with adaptive penalty and eigen-decomposition to obtain the optimal projection. The convergence of LSPP is proven, and we also analyze its complexity. To validate the effectiveness and robustness of LSPP in feature extraction and dimensionality reduction, we make a critical comparison between LSPP and a series of related dimensionality reduction methods. The experimental results demonstrate the effectiveness of LSPP.