Invariance Class based Surface Reconstruction using CNN Transfer Learning

Geometrical operations are defined in ISO standards on geometrical product specifications and verification (GPS) for obtaining ideal and non-ideal features to represent surfaces on mechanical parts. These geometric features should be categorized by their geometrical properties for further geometry processing. In the context of ISO GPS, surface portions are determined by their kinematic invariance and are classified into planar, cylin-drical, helical, spherical, revolute, prismatic and complex surfaces. In this paper, geometric features based on discrete representation enabled by the Skin Model Shapes paradigm are investigated. An automatic approach for surface reconstruction by invariance class is proposed in this paper based on deep learning architecture to speed up geometry processing. Partitioning operation is implemented to decompose a mechanical part from simulations or measurements into surface portions. Regarding the specificity of the derived point cloud, a view-based strategy is implemented by dimension reduction using point cloud projection. The extracted view-wise features are aggregated into a discriminative global representation for training neural networks. The method is tested on an open dataset and a case study is given for validation. The proposed method enables robust surface reconstruction for reverse engineering with low computational cost in the context of ISO GPS.