YOLO-BFID: Real-Time Building Facade Inspection by Using an Advanced Artificial Vision Model Integrated with Drone Technology

This study presented the You Only Look Once (YOLO) for building façade inspection with drones (YOLO-BFID), an enhanced version of YOLOv11 designed to improve small detection and operate efficiently across challenging real-world scenarios. The model compactness was increased using a spatial-channel decoupled downsampling module and a cross-stage partial compact inverted block module to minimize model parameters without compromising accuracy. An efficient intersection over union loss function (EIoU) was used to train the YOLO-BFID model and facilitate the independent optimization of object width and height, leading to more rapid convergence and more accurate segmentation masks. Experiments were subsequently conducted using a data set containing approximately 2,000 images collected across various locations in Taiwan depicting tile peeling, defective tile sealing, and concrete spalling. The YOLO-BFID model exhibited a 3.8% and 1.9% higher segmentation mask accuracy relative to the YOLOv11_small and YOLOv11_medium models, respectively. Additionally, the YOLO-BFID model achieved a 2.3% higher mean average precision than that of both YOLOv11 models at an intersection over union threshold of 50% (mAP50=74.5%). The YOLO-BFID model also outperformed other baseline models (Mask R-CNN and SOLOv2) in precision (by 13.6% and 14.0%, respectively). In real-world tests, the YOLO-BFID model had a processing speed of 27.1 frames per second, 27% greater than that of YOLOv11_medium. These results demonstrate the feasibility of YOLO-BFID for use in real-time drone-based building façade inspections.