HFE-YOLO: Hybrid Feature Enhancement with Multi-Attention Mechanisms for Construction Site Object Detection

Construction sites require integrated monitoring of equipment and structural safety. This study systematically compares four feature enhancement mechanisms at the YOLOv11n backbone–neck transition: HFE-YOLO (hybrid multi-attention), FPN-YOLO (feature pyramid), C2F-YOLO (cross-stage partial), and Identity-YOLO (baseline). Evaluation utilized two datasets with contrasting class distributions: ConstructSight (eight equipment classes, severe imbalance) and SafeGuard (five safety classes, balanced distribution). All models were trained for 200 epochs using identical configurations to ensure controlled comparison. On the imbalanced CS dataset, HFE-YOLO achieves superior performance (95.0% mAP@50, 82.6% mAP@50–95), followed by FPN-YOLO (94.8%, 82.4%), Identity-YOLO (92.5%, 74.4%), and C2F-YOLO (92.4%, 72.1%). On the balanced SG dataset, performance differences compress substantially: HFE-YOLO (96.8%, 79.4%), C2F-YOLO (96.6%, 78.2%), Identity-YOLO (96.3%, 78.1%), and FPN-YOLO (96.1%, 76.1%). HFE-YOLO provides 8.2 percentage points mAP@50–95 improvement over the baseline on imbalanced data versus 1.3 percentage points on balanced data. Enhancement mechanism effectiveness varies substantially between dataset distributions, with sophisticated mechanisms providing greater benefits for imbalanced scenarios. These findings offer insights for architecture selection based on dataset distribution characteristics.