Mid-Curve Completion Using Convolutional Neural Network

Although of great interest for Finite Element (FE) analysis, the automatic generation of a lower dimensional representation of shapes is not straightforward. One-dimensional (1D) mid-curve model contains entities such as lines, circles and more generally curves representing higher dimension entities like surfaces to be located in the middle of a thin-walled model following its shape. In this paper we present a novel approach to adapt a given medial axis of a polygon in order to generate a dimensionally reduced mid-curve model suitable for FE analysis. The novelty is to use a deep learning completion network [1] which is trained to automatically modify the perturbed regions of the medial axis (end and connection regions). The network takes as input a local image containing the polygon boundary, a mask of the regions to complete and the surrounding valid mid-curves. It returns a predicted mid-curve image which is inserted back into the global mid-curve model.