Efficient identification of 2D/3D particle morphology parameters based on a dual-stage neural network approach
Ping Chen, Ziwei Zhao, Ziyi Liu, Wei Hua Yao, Chaomei Liu, Bo Wang · Nondestructive Testing And Evaluation · 2025
The macroscopic mechanical behaviour of geotechnical materials in underground space resources is significantly influenced by particle morphology parameters. Traditional two-dimensional (2D) image analysis methods are unable to accurately reflect the three-dimensional (3D) geometric characteristics of particles and suffer from issues such as low efficiency and insufficient automation. To address these problems, this paper proposes a rapid identification method for 2D/3D size parameters of particles based on a dual-stage neural network model. Firstly, particle models are generated and rendered using computer graphics software to construct a large-scale annotated dataset. Subsequently, a two-stage neural network model is designed: the first stage extracts 2D size parameters of particles through image segmentation and identification, while the second stage constructs a 3D parameter recognition network based on multi-view images to achieve precise measurement of the three-dimensional geometric features of particles. The experimental results demonstrate that this method exhibits high recognition accuracy and efficiency. Specifically, the dual-stage network model (e.g. Mobile-UNet + MobileViT) excels in 2D parameter recognition, with MAE reaching 1.67 pixels, MSE reaching 5.38 pixels, and R2 reaching 98.51%. In 3D parameter recognition, the model with bidirectional input significantly improves recognition accuracy. Additionally, the lightweight network design maintains high accuracy while reducing the number of parameters, showing potential for deployment on edge computing devices.