Image-Based 3D Reconstruction of Granular Grains via Hybrid Algorithm and Level Set with Convolution Kernel

Abstract

This study develops a novel method for reconstructing three dimensional (3D) granular grains from computed tomography (CT) images. Unlike previous studies requiring trial-and-error hyperparameters, the hybrid algorithm introduced here, integrating the random forest (RF) algorithm and enhanced by particle swarm optimization for automatic determination of hyperparameters, is the first to train the model for constituent classification and grain segmentation. In addition, and different from previous manual methods, a convolution kernel is applied to assign an initial level set function inside an individual grain and determine whether to activate the level function for automatically reconstructing 3D grains from a CT image. All results indicate the hybrid algorithm can rapidly search the optimum hyperparameters, providing a more effective way to identify the optimum RF-based model. This model segments grains with an accuracy of 90%, in comparison with a 52% accuracy achieved by the conventional watershed algorithm. The convolution kernel can accurately and automatically identify individual grains, avoiding manual assignment of an initial calculation area and ensuring grains are correctly reconstructed. Overall, the proposed method provides a more intelligent and effective way to reconstruct 3D grains from CT images.