Estimating Manufacturing Errors of Industrial Products with Neural Optimization-based Non-rigid Registration

Abstract

When inspecting an industrial product, it is important to detect the manufacturing errors on its surface because they may affect the performance of the product. In this paper, we propose a non-rigid registration method that can detect the manufacturing errors on the surface of each product component separately. Given the product’s point cloud obtained by 3D scanning and the meshes of the components generated from CAD models, we first perform template matching to extract the point clouds that match with each component from the product’s point cloud, and then measure the surface deformations on each component by performing non-rigid registration from the component’s mesh to the matched point cloud. The efficiency of our method is tested using the CT-scanned data of a 3D-printed model. Experiments demonstrate that our method can deal with large and small manufacturing errors on the product’s surface.