Abstract: Aiming at the problem that in industrial defect detection, the collection of point cloud was incomplete due to factors such as component damage and environmental interference, structural discrepancies between incomplete and standard point clouds caused outlier correspondences during registration, resulting in a decrease in registration accuracy, we proposed an improved partial point cloud registration network based on point pyramid and maximal cliques. Firstly, we designed incomplete completion module to generate virtual points, the module extracted local and global  features of point cloud through  multi-resolution dynamic graph convolution, and added a point pyramid fractal prediction network to complete missing regions of the point cloud and compensated for the missing  feature information. Secondly, the maximal cliques algorithm was integrated into the registration module to filter incompatible matching pairs, optimizing the matching matrix and pose estimation strategy. The results of  unknown category partial point cloud registration  on the ModelNet40 dataset show that the average absolute errors of rotation and translation of the proposed network are reduced to  0.650 7 and 0.005 9, respectively, effectively improving registration accuracy and robustness of outlier correspondences. The experiment on the ESB dataset further show that the proposed network provides a reliable partial point cloud registration scheme for industrial defect detection.

Key words: point cloud registration, point pyramid, maximal cliques, dynamic graph convolution