A computational method for real-time roof defect segmentation in robotic inspection

Roof inspections are crucial but perilous, necessitating safer and more cost-effective solutions. While robots offer promising solutions to reduce fall risks, robotic vision systems face efficiency limitations due to computational constraints and scarce specialized data. This study presents real-time roof defect segmentation network (RRD-SegNet), a deep learning framework optimized for mobile robotic platforms. The architecture features a mobile-efficient backbone network for lightweight processing, a defect-specific feature extraction module for improved accuracy, and a regressive detection and classification head for precise defect localization. Trained on the multi-type roof defect segmentation dataset of 1350 annotated images across six defect categories, RRD-SegNet integrates with a roof damage identification module for real-time tracking. The system surpasses state-of-the-art models with 85.2% precision and 76.8% recall while requiring minimal computational resources. Field testing confirms its effectiveness with F1-scores of 0.720–0.945 across defect types at processing speeds of 1.62 ms/frame. This work advances automated inspection in civil engineering by enabling efficient, safe, and accurate roof assessments via mobile robotic platforms.