Construction dust detection in existing building renovation under limited annotation: a semi-supervised framework based on improved YOLOv11
Abstract
This study develops a data-efficient method for construction-dust detection in existing-building renovation, where limited annotations, blurred boundaries, irregular morphology, variable illumination and complex backgrounds hinder reliable visual monitoring. A lightweight YOLOv11n-DFD detector is developed by integrating DualConv, Dynamic Feature Fusion and Dynamic Tanh. A simplified single-model, single-round semi-supervised strategy is then used to exploit unlabeled construction-site images through confidence-filtered pseudo-labeling. Across three independent runs, the complete framework achieved a mean mAP@0.5 of 0.969 ± 0.002, representing an improvement of 0.063 over the supervised YOLOv11n baseline. It achieved mean mAP@0.5 comparable to that of an EMA-based teacher–student method while using a simpler training pipeline and producing a larger pseudo-labeled pool. The lightweight detector and simplified training procedure provide a potential vision-based tool for construction-dust monitoring, supporting environmental and safety management with the potential to reduce reliance on extensive manual annotation. This study provides a task-oriented integration of lightweight detection and semi-supervised learning for construction-dust monitoring in existing-building renovation. It demonstrates a practical approach to incorporating additional unlabeled training samples through pseudo-labeling when reliable annotations are limited.