Comparative Evaluation of YOLOv11n Neck-Level Modifications for Precast Component and PPE Object Detection in Construction Environments

Construction monitoring systems can benefit from automated detection of both structural components and personal protective equipment (PPE). While previous studies focus on single-task applications using YOLOv5-YOLOv10 architectures, this study presents a systematic comparative evaluation of four neck-level architectural modifications within the YOLOv11n framework: Depthwise separable convolutions (DSC) for computational efficiency, multi-scale dilated convolutions (MSDC) for expanded receptive fields, feature refinement interfaces (FRI) for learnable feature adaptation, and dual attention mechanisms (DAM) for enhanced feature discriminability. Controlled experiments were conducted on precast components (3771 images, 6 classes) and PPE (5201 images, 3 classes) datasets. DAM-YOLO achieved the highest performance with mAP@50 of 0.972 (precast) and 0.968 (PPE), while the performance range across all variants spanned 0.942–0.972 (precast) and 0.936–0.968 (PPE). All variants demonstrated robust detection capabilities with mAP@50 ≥ 0.936 and mAP@50–95 spanning 0.760–0.825. The narrow performance differences across variants indicate that different neck-level modifications yield comparable detection accuracy, providing empirical evidence to support architecture selection within these two evaluated construction object detection scenarios.