Feature line detection of noisy triangulated CSGbased objects using deep learning

Abstract

Feature lines such as sharp edges are the main characteristic lines of a surface. These lines are suitable as a basis for surface reconstruction and reverse engineering [1]. A supervised deep learning approach based on graph convolutional networks on estimating local feature lines will be introduced in the following. We test this deep learning architecture on two provided data sets of which one covers sharp feature lines and the other arbitrary feature lines based on unnoisy meshed constructive solid geometry (CSG). Furthermore, we use a data balancing strategy by classifying different feature line types. We then compare the selected architecture with classical machine learning models. Finally, we show the detection of these lines on noisy and deformed meshes.

Nodal Cosine Sine Material Interpolation in Multi Objective Topology Optimization With the Global Criteria Method for Linear Elasto Static, Heat Transfer, Potential Flow and Binary Cross Entropy Sharpening

“…In topology optimization, a clear black-white pattern is required (Kang and Wang, 2012) so that the results can be directly interpreted. For binary decisions, the binary cross-entropy loss (Denk et al, 2019;Mannor et al, 2005) and polynomial loss function with…”

Euclidian Distance Transformation, Main Axis Rotation and Noisy Dilitation Supported Cross-Section Classification With Convolutional Neural Networks

“…The classification of cross-section types is required in reverse engineering of the beam-like shapes. In this article, a convolutional neuronal network for the classification of different cross-sections is shown, which can be downloaded at (Denk, 2021). A data set using polygon and image cross-section is created for eight common cross-section types such as I, U, or T sections.…”

“…Geometry deep learning is one of the recent challenges, especially for 3D objects (Ahmed et al, 2019;Bronstein et al, 2017;Cao et al, 2020). Due to the rapid development in deep learning, these models can perform classification and shape recognition tasks for different kinds of 3D geometries (Ahmed et al, 2019;Bronstein et al, 2017;Cao et al, 2020;Denk et al, 2019). Especially the use of convolutional neuronal networks (CNN) allows a high degree of parallelization.…”

“…While deep learning methods applied on Euclidean data structures such as 3D images (voxels) or 2D images (pixels) are largely established, current research fields summarized in (Ahmed et al, 2019;Bronstein et al, 2017;Cao et al, 2020) use CNN for non-Euclidean data such as graphs, polygonized surfaces, and point clouds. Properties such as shift-invariance and a global parametrization in the image can miss in non-Euclidian shapes (Bronstein et al, 2017), so that processing on non-Euclidian data is quite challenging (Bronstein et al, 2017;Denk et al, 2019). In our contribution, we address the topic of geometric learning for polygon cross-section.…”

Subdivision Surface Mid-Surface Reconstruction of Topology Optimization Results and Thin-Walled Shapes Using Surface Skeletons

“…For further research, the local thickness and the change of the skeleton (curvature, normal vectors) should also be considered segmenting the mid-surface. Additionally, the surface control grid can be sharpened by detecting sharp edges using a feature line detection method similar to (Denk et al, 2019).…”