Separation of sand and gravel particles in volume images using a random forest

Geometric features of sand and gravel particles contained in building materials significantly influence the macroscopic properties of the final structure. Investigating volume images of these particles acquired by micro-computed tomography is of interest, as all features are then directly derivable. However, particles touch and must be separated in the 3D image before being analyzed. For multi-sized and strongly non-spherical particles, there is no standard solution for this task available. We first separate the particles by the established morphological procedure of applying the watershed transform marked by the minima of the inverted distance map, here with the notorious over-segmentation alleviated by pre-flooding. Subsequently, remaining errors – particles split into several fragments – are detected by a random forest and automatically united. Training on one to two 3D images, only, we achieve accuracies rendering the currently still necessary manual correction obsolete. More precisely, less than 3% of the particles of natural origin including crushed ones remain incorrectly segmented. Even more remarkable is the result for recycled materials, where the final rate of wrongly segmented particles does not exceed 6%.