<p>Vision-based roof inspection plays a critical role in ensuring the functionality, safety, and durability of building systems. However, transferring pre-trained models to real-world inspection tasks poses substantial challenges, particularly because crack-induced leakages resulting from material deterioration are often subtle and small in scale, making them difficult to identify through standard image recognition techniques. To address this practical challenge, we develop an efficient and cost-effective anomaly detection framework that leverages an Attention-Enhanced Deep Neural Network to predict potential percolation using Red–Green–Blue images captured by Unmanned Aerial Vehicles. Specifically, the backbone network is improved through Attention-Enhanced Feature Extraction modules that integrate both spatial and channel attention, enabling more fine-grained defect feature extraction from feature maps. Together with the focus layer, the attention-enhanced module preserves more sensitive and informative content while reducing overall information loss. Furthermore, we construct a Mixed Loss Function based on a more focused Intersection over Union loss to strengthen the framework’s anomaly detection capability by emphasizing different regression samples. Both comparative and ablation experiments conducted on a real roof inspection task demonstrate that the proposed framework achieves superior detection accuracy, efficiency, and robustness. For practitioners, this work not only offers a proactive solution for preventing building failures but also encourages broader adoption of advanced vision-based anomaly detection tools across diverse engineering domains.</p>