Intelligent construction and demolition waste detection in complex environments using an improved CLRD-YOLO model with knowledge distillation

Rapid global urbanization and construction industry growth have significantly increased construction waste generation, posing substantial environmental and resource challenges. To address the limitations of traditional detection methods for complex, multi-label construction and demolition waste (CDW), this study introduces several key technical innovations. First, we developed the Multi-label Mixed Construction and Demolition Waste dataset (MM-CDWaste), accurately reflecting typical complex waste distributions on construction sites. Building upon this dataset, we propose a novel CLRD-YOLO-based detection model featuring a multiscale-context-spatial fusion framework, including Context and Spatial Feature Calibration Network (CSFCN) with multi-level cascaded pyramid pooling and pixel-level contextual similarity aggregation. The model also integrates Large Separable Kernel Attention (LSKA) and Robust Feature Down-sampling (RFD) strategies, improving adaptability and detection accuracy in complex environments while reducing redundant information. Additionally, for the first time in CDW detection, a knowledge distillation strategy is applied, utilizing YOLOv8s as the teacher model to refine and transfer high-level feature knowledge to optimize CLRD-YOLO. The system-level design further demonstrates the potential integration of CLRD-YOLO with optical inspection, gripping devices, and hardware control for automated CDW sorting. Experimental results show that CLRD-YOLO achieves an mAP of 92.4% on MM-CDWaste, representing a 3.6% improvement over baseline methods, highlighting its competitive accuracy and potential for lightweight deployment in complex CDW detection scenarios.