Effectively segmenting firmly bonded composite NHCDW particles remains an unresolved challenge, as conventional sorting methods fail to separate these complex materials. To address this issue, this study proposes an intelligent approach for identifying and segmenting bonded NHCDW gravel (mortar-brick) using advanced You Only Look Once deep learning models. The models were trained for 100 epochs on two independently annotated gravel datasets: one containing individual gravel classes and another comprising composite mortar-brick gravel. Results show YOLOv11 outperforming YOLOv8 across both datasets, achieving box and mask mAP50 scores of 0.846/0.811 and 0.965/0.965, respectively, compared to YOLOv8’s 0.835/0.815 and 0.964/0.963. Corresponding F1-scores of 0.824/0.793 on the first and 0.950/0.965 on the second dataset further highlight YOLOv11’s superior precision, recall, and accurate surface fractions. This research is conducted within the framework of the EU Horizon MOBICCON-PRO project, which highlights the potential of artificial intelligence to enhance CDW recycling in circular economy strategies. • Deep learning enables reliable, automated NHCDW composite gravel analysis for reuse. • Composite (mortar-brick) gravel accurately separated through instance segmentation. • YOLOv11 outperforms YOLOv8 in precision, recall, and mAP for gravel detection. • Real-world validation confirms robust surface area fraction quantification. • Framework paves the way for future volumetric analysis of NHCDW for recycling.
Safi et al. (Fri,) studied this question.