Automated quality control of recycled aggregates via deep learning: A unified framework for instance segmentation and mass estimation

Abstract

The large-scale use of recycled aggregates (RA) in high-grade construction applications is currently hindered by the high variability of their physical properties. Current quality control relies on manual sorting, which is labor-intensive and limits scalability. This study presents RAMSES (Recycled Aggregates Mass estimation and Segmentation), an automated framework based on deep learning, designed to bridge the gap between high-speed production and rigorous material characterization. A central contribution of this work is the introduction of a large-scale, publicly available dataset comprising 90,000 labeled and batch-weighed aggregate instances. This extensive dataset supports a strong statistical robustness across diverse RA compositions and serves as a benchmark for automated waste characterization. Using this dataset, RAMSES performs simultaneous instance segmentation and direct mass estimation from 2D images. By integrating a dual-branch architecture, the model effectively decouples morphological features from instance-dependent density factors. The framework achieves high precision in particle identification (mean Average Precision mAP@[0.5:0.95] = 0.84, mAP@0.5 = 0.91) and a 0.3% relative error in total mass prediction, which meets industrial requirements for batch monitoring. By providing a scalable alternative to manual inspection, this approach improves the consistency of RA-based concrete mixes, directly supporting the transition to a circular construction economy.

Keywords: circular economy, construction and demolition waste, recycled aggregates, artificial intelligence, deep learning, convolutional neural networks, instance segmentation, mass prediction