Optical Measurement and 3-D Reconstruction of Blade Profiles With Attention-Guided Deep Point Cloud Registration Network

The optical measurement and precise 3-D reconstruction of the blade are crucial in ensuring its processing quality. However, the complex and highly reflective free-form surfaces of blades may lead to inadequate overlaps and less prominent overlap-area features among multiple scanning data. The spatial sampling variance of optical measurement instruments can result in different point cloud densities. These make it challenging to achieve accurate point cloud registration of multiple scanning data, which is currently a critical step in the 3-D reconstruction of blade profiles. To address these issues and achieve the precise 3-D reconstruction of blade profiles, a learning-based method called attention-guided deep point cloud registration network (AGDnet) is proposed in this article, which mainly consists of a feature extraction backbone module, a cross-attention mechanism module, a self-attention mechanism module, and a transformation estimation module. Specifically, the feature extraction backbone incorporates subsampling to address density variations and noise contamination in point clouds. The cross-attention mechanism is proposed to facilitate critical information exchange between point clouds, while the self-attention mechanism integrates significant global features from individual point clouds to further enhance the feature representation. The transformation estimation module is designed to optimize dynamically the matching matrix, enabling the recovery of the underlying correspondences between point clouds. The experimental results on three representative blades and comparison with six point cloud registration methods prove the feasibility and accuracy of the suggested method.