KSG-YOLO: Application of YOLO-Based Detection with Knowledge Distillation and Structured Pruning in Green Assembly Building Scenarios

Green assembly building emphasizes low-carbon, standardized construction, where reliable on-site perception is essential for safety supervision and material management under clutter, occlusion, and the prevalence of small targets. To reconcile accuracy with the stringent efficiency demands of edge deployment in such scenarios, we develop KSG-YOLO, a YOLO-based compression pipeline tailored to construction-site object detection. First, a high-capacity teacher distills a compact student to transfer both classification and localization cues, thereby preserving small-object recognition without increasing runtime cost. Then, structured channel pruning guided by L1/L2 importance—with random pruning as a stochastic baseline—is applied and followed by short fine-tuning to recover performance under explicit parameter/FLOPs budgets. Moreover, we conduct a comprehensive evaluation on representative construction datasets, including SODA, MOCS, and CIS, under consistent training and validation protocols. Finally, experiments show that the proposed pipeline attains markedly higher efficiency while maintaining competitive, state-of-the-art detection performance across key categories relevant to green assembly practice. On SODA dataset, our compressed model reduces computation from 171.1 GFLOPs to 12.41 GFLOPs (-92.7%) and parameters from 30.21M to 7.7M (-74.5%), while maintaining accuracy (mAP50 from 0.8637 to 0.8652, mAP50-95 from 0.5760 to 0.5796). These findings provide an actionable route toward resource-aware vision systems for sustainable construction and offer a solid empirical reference for subsequent research in this domain.