A Contour-Preserving Based Graph Convolutional Networks for Mesh Simplification

Triangular meshes serve as the core carrier in the fields of 3D digital twins and industrial simulations. Their face count often reaches millions or even hundreds of millions, posing challenges such as high time complexity, large memory consumption, and low parallelization for traditional simplification algorithms. To address the need for efficient simplification of large-scale meshes, this paper proposes a lightweight framework based on graph convolutional networks (GCN) and region growing. Through multi-granularity geometric feature extraction and partitioned parallel simplification, our method makes large-scale mesh simplification speed approach that of small-scale meshes. Unlike traditional simplification methods, our approach dynamically constructs simplification clusters with geometric consistency through region growth, followed by learning optimal simplified vertices using GCN. Experiments show that under the condition of maintaining an average Chamfer distance error below 0.2, our method achieves mesh simplification of a million-face model in just 27 seconds, accelerating over the latest learningbased method SubdivNet by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5 0 \%}$</tex>.