.png){kind=icon}
.png){kind=icon}
Return
Defects inspection system for building facades using drones and deep learning method
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.129715.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Regular inspection and maintenance of building facades are essential for preserving structural integrity and aesthetic quality, especially in aging urban high-rises. While drone-based visual inspection powered by artificial intelligence (AI) offers benefits in speed, safety, and scalability, existing methods are typically limited to single defect types or uniform facade categories due to the challenges of detecting multi-scale defects in complex, heterogeneous environments. This study introduces an automated multiclass defects inspection system for building facades by integrating drone technology, an AI-driven segmentation platform, and automatic report generation. Central to the system is a segmentation AI model capable of detecting multiclass defects with orders-of-magnitude differences in scale across diverse facade backgrounds. To handle the pixel imbalance of defects ranging from fine cracks to large spalling and glass breakage, the model is built upon EfficientUNet++, trained on a carefully curated dataset and optimized using adjustable batch sizes and active learning rates to improve multi-scale feature learning and mitigate overfitting. Evaluations on validation and out-of-sample datasets demonstrate that the proposed model achieves superior performance across all defect classes. Real-world drone experiments further confirm the model’s practical applicability, with high recall rates in detecting spalling, water seepage, cracks, and glass breakage across different types of facades. This work pioneers a robust, scalable, and efficient AI-based framework for automated multiclass facade defect inspection, providing actionable information for engineers and supporting urban infrastructure maintenance.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W