Gravel Particle Shape Classification from Half-Particle Point Clouds using a Dynamic Graph Edge Convolution Neural Network

Abstract

Obtaining the three-dimensional (3D) shape of gravel particles is essential for calculating their roundness and sphericity. However, cost-effective, and rapid non-penetrating 3D imaging technologies, such as 3D laser scanners, stereophotography, and structured light techniques, only capture the geometric shape of the upper half particles (2.5D particles), unable to penetrate a particle to acquire the shape of the lower half. Current algorithms cannot accurately classify real 3D particles using easily available and low-cost 2.5D particles. To address this issue, this study aims to develop a dynamic graph edge convolution neural network (DGECNN) based on deep learning, utilizing 2.5D point clouds to characterize and classify the roundness and sphericity of 3D particles. The dataset comprises 4200 2.5D point clouds labeled into 12 roundness-sphericity categories based on corresponding complete 3D particle characterizations. Experimental results demonstrate that with a sampling point of 1200 and a batch size of 64, the training time is relatively shorter, and the automatic classification accuracy reaches a peak of 90.76%. Finally, compared to the traditional 3D CG method, the DGECNN classification is equally applicable to sand-size particles and exhibits significant advantages in roundness-sphericity, volume, surface area, and convex hull volume.