Improved Dexel Representation: A 3-D CNN Geometry Descriptor for Manufacturing CAD
Xingyu Fu, Dheeraj Peddireddy, Vaneet Aggarwal, Martin Byung‐Guk Jun
Abstract
In this article, we present a novel 3-D descriptor, improved dexel representation (IDR), which assists to input holistic information from an engineering computer-aided design (CAD) model to convolutional neural network (CNN) based manufacturing applications. The IDR carries the model’s position, size, and surface information, which not only provides high resolution to small-scale local (machining) features, but also has the potential to reconstruct the original CAD model. Data conversion algorithms between IDR and other CAD models (mesh and NURBS model) are efficient. CNNs with IDR input can largely improve the prediction accuracy compared to other 3-D descriptors, which reaches 98.8% in the modified machining-process-identifier dataset and 100% on the FeatureNet style 3-class dataset. IDR benefits both the manufacturing industry and other CAD-related deep learning applications in engineering fields.