Unveiling Building Façade Deterioration: A Drone-Powered Deep Learning Approach for Seamless Tile Peeling Detection

The accurate segmentation of tile peeling on building facades holds considerable significance for effective building maintenance, particularly in regions like Taiwan, where tiles are the predominant facade protection. This research introduces YOLOM, a novel deep-learning-based segmentation model designed to address this challenge. YOLOM harnesses the capabilities of You Only Look Once version 7 (YOLOv7) and incorporates the BlendMask-based segmentation technique, further augmented by the Efficient Layer Aggregation Network (ELAN) to enhance feature discrimination and extraction capabilities specifically tailored for scenarios involving tile peeling. Employing a dataset comprising 400 images featuring 758 instances of peeling and 525 instances of sealed tiles observed during on-site surveys of public buildings, YOLOM exhibits outstanding segmentation performance. It outperforms the Resnet-BlendMask50 FPN with improvements of 7.1% of mean average percentage (mAP) and 0.4% of the average precision (AP) at the intersection over union (IoU) of 50%. Remarkably, YOLOM consistently surpasses other models, showcasing a 19.5% and 2.2% lead in AP for small and large objects, respectively. In a noteworthy advancement, YOLOM seamlessly integrates with drone technology, enhancing its capabilities for aerial surveying of building facades. This integrated approach proves invaluable for building maintenance teams, enabling proactive and cost-effective interventions. The study introduces a distinctive framework seamlessly integrating cutting-edge backbone and neck modules, particularly emphasizing the ALAN. The innovative YOLOM model establishes a new standard in artificial intelligence (AI) techniques for building maintenance, contributing significantly to academic discussions surrounding AI-enhanced image segmentation.