CNN-Based YOLOv12 for Damage Assessment in Residential Roofs

Roof maintenance is crucial for preserving buildings, shielding them from adverse weather conditions, and extending their lifespan. Manual inspections face significant challenges due to their repetitive nature, susceptibility to errors, and reliance on specialized labor. These limitations highlight the need for automated methods for regular evaluations, especially considering the risks associated with working at heights and the additional costs due to human error. To address these gaps, this study proposes an innovative model that leverages Unmanned Aerial Systems (UASs) for swift image capture, combined with Artificial Intelligence algorithms for automated analysis. Still unexplored in residential roof inspections, the YOLOv12 architecture was used to train and test a model with 1,333 images captured by UASs across 167 residential buildings in Salvador, Bahia, Brazil. The results showed a mAP50 of 86.2% for open trapdoors, 84.4% for extra tiles, and 69.3% for dirty gutters. Precision and recall for identifying open trapdoors were 94.8% and 85.7%, respectively; for extra tiles, 80.0% and 81.1%; and for dirty gutters, 77.6% and 55.1%. Furthermore, the algorithm was implemented in a user-accessible application via Streamlit, enhancing preventive management of urban buildings with greater efficiency and safety. Future studies will include other roof types, newanomaly classes, and the integration of results into a building maintenance management system.