Pattern recognition of quartz sand particles with PointConv network

Zhiren Zhu, Jianfeng Wang, Mengmeng Wu

Research output: Journal article publicationJournal articleAcademic researchpeer-review

2 Citations (Scopus)

Abstract

Particle kinematics plays a significant role in the mechanical response of granular soils. Accurate particle identification and tracking holds the key to the investigation of particle kinematics during the soil lab test. This paper introduces a novel pattern recognition method for identifying and tracking intact Leighton Buzzard sand (LBS) particles in a miniature triaxial sample with a neural network, called the PointConv network. Firstly, the image processing techniques are applied on the 2D slices to reconstruct the realistic morphology of LBS particles. Secondly, 100 LBS particles are randomly selected as the objects of particle tracking and are handled by an operation of sampling and grouping to prepare the training and testing datasets for the PointConv network. A set of Gaussian noise is generated and injected into the particle point subsets to increase the robustness of the training model. Next, the PointConv network is implemented to learn morphological features of the LBS particles in multiple dimensions and successfully recognized and tracked all LBS particles. Finally, the discussions about the effects of various parameters of the model on the prediction results are made. The merits of the proposed method are highlighted by comprehensively comparing with several existing particle tracking methods.

Original languageEnglish
Article number105061
JournalComputers and Geotechnics
Volume153
DOIs
Publication statusPublished - Jan 2023
Externally publishedYes

Keywords

  • Deep learning
  • LBS sands
  • Particle recognition and tracking
  • Point cloud
  • PointConv neural network

ASJC Scopus subject areas

  • Geotechnical Engineering and Engineering Geology
  • Computer Science Applications

Fingerprint

Dive into the research topics of 'Pattern recognition of quartz sand particles with PointConv network'. Together they form a unique fingerprint.

Cite this