A compared approach on how deep learning may support reverse engineering for tolerance inspection

Reverse Engineering (RE) may help tolerance inspection during production by digitalization of analyzed components and their comparison with design requirements. RE techniques are already applied for geometrical and tolerance shape control. Plastic injection molding is one of the fields where it may be applied, in particular for die set-up of multi-cavities, since no severe accuracy is required for the acquisition system. In this field, RE techniques integrated with Computer-Aided tools for tolerancing and inspection may contribute to the so-called “Smart Manufacturing”. Their integration with PLM and suppliers' incoming components may set the information necessary to evaluate each component and die. Intensive application of shape digitalization has to front several issues: accuracy of data acquisition hardware and software; automation of experimental and post-processing steps; update of industrial protocol and workers knowledge among others. Concerning post-processing automation, many advantages arise from computer vision, considering that it is based on the same concepts developed in a RE post-processing (detection, segmentation and classification). Recently, deep learning has been applied to classify point clouds, considering object and/or feature recognition. This can be made in two ways: with a 3D voxel grid, increasing regularity, before feeding data to a deep net architecture; or acting directly on point cloud. Literature data demonstrate high accuracy according to net training quality. In this paper, a preliminary study about CNN for 3D points segmentation is provided. Their characteristics have been compared to an automatic approach that has been already implemented by the authors in the past. VoxNet and PointNet architectures have been compared according to the specific task of feature recognition for tolerance inspection and some investigations on test cases are discussed to understand their performance.