A scale-adaptive Mask R-CNN strategy for foreground particle segmentation and geometrical analysis of granular aggregates

Abstract

Instance segmentation algorithms play an increasingly crucial role in evaluating the geometrical characteristics of granular particles in geotechnical engineering. However, the accurate segmentation of the foreground particles of densely packed granular aggregates from multi-scale images remains challenging due to the time-consuming, laborious manual annotation process. This study proposes a scale-adaptive mask regional convolutional neural network (Mask R-CNN) strategy method for the automatic segmentation and geometrical analysis of granular aggregates in multi-scale images with less manual labeling. This new Mask R-CNN strategy involves an iterative process of automatic annotation of large-scale images and updating the preliminary model initially trained by small-scale images. The large-scale images are initially decomposed into a series of small-scale blocks. Subsequently, the preliminary model automatically identifies and annotates foreground particles within these small blocks. The segmentation results of the small blocks are then integrated into the original size, serving as annotations for the large-scale images. Finally, a multi-scale model can be obtained by retraining the preliminary model with annotated large-scale images and repeating the decomposition and integration process. Experiments based on densely packed ballast particles and cobble particles were conducted to validate the effectiveness of the proposed strategy. The results indicate that the trained multi-scale model exhibits satisfactory performance in segmenting foreground particles from densely packed granular aggregates, achieving a detection rate of 90.66% and an error rate of 18.81%. The proposed framework provides a feasible and effective tool for onsite inspection and the morphology analysis of densely packed granular aggregates.