Projection-based transfer learning for point cloud surface classification in reverse engineering and re-manufacturing of mechanical parts: A comparative study

Accurate surface reconstruction from point cloud data is a fundamental challenge in the remanufacturing and reverse engineering of mechanical components, where the preliminary classification of geometric primitives represents a key step toward reliable CAD model generation. Reconstructing mechanical parts from raw point clouds remains particularly demanding due to data heterogeneity, noise, and varying point densities. While state-of-the-art deep learning models such as PointNet and 3D ShapeNets have demonstrated strong performance on generic man-made object categories, their direct applicability to industrial remanufacturing and geometric reconstruction tasks is still limited. To address these limitations, this paper proposes a projection-based transfer learning strategy for the classification of segmented point clouds representing basic geometric primitives. Within an automated reverse engineering framework, the focus is placed on the classification stage, which is essential for associating segmented surface regions with their corresponding primitive types prior to surface fitting and CAD reconstruction. The proposed method relies on 2D projections of segmented point clouds combined with transfer learning using established convolutional neural networks. A comparative evaluation of representative CNN architectures is conducted on a dataset of segmented primitives (plane, cylinder, sphere, cone, and tori), using classification accuracy and confusion matrices as performance metrics. Experimental results demonstrate that projection-based transfer learning provides a robust and computationally efficient solution for primitive surface classification, with certain network–projection combinations offering superior accuracy and robustness. These findings confirm the relevance of lightweight projection-based learning approaches for digital reconstruction and automated inspection tasks in sustainable manufacturing and remanufacturing contexts.