A Comprehensive Framework for Automated Facade Defect Evaluation Using Deep Learning

Evaluation of facade degradation is crucial to determining the need for further examination and maintenance, safeguarding the structural health of the building. Traditional evaluation methods rely on visual inspection and subjective judgment by surveyors, which requires working at heights and manual detection and severity assessment. This process is timeconsuming, poses safety risks, and may result in potential errors. To address these challenges, unmanned vehicles have been deployed for building inspection tasks, saving time and minimizing safety concerns. Additionally, diverse deep learning techniques have been utilized to automate the visual evaluation process and reduce subjective errors to classify, detect, and segment defects. However, previous research has primarily focused on improving image processing accuracy without incorporating the entire evaluation process into industry evaluation standards. In this paper, we propose a framework for automated building facade defect evaluation that can be applied to both unmanned and manned data collection systems. Our approach employs the deep neural network for defect segmentation. The trained model accurately recognizes individual defects and extracts their properties, such as width, length, and area using morphological operations. Following ISO standards, evaluation results are automatically obtained within defined effective evaluation areas. Furthermore, defect information and evaluation results are registered to a model generated with 3D reconstruction techniques, providing a valuable reference for experts in formulating maintenance plans. Finally, experiments are conducted on a tall building to verify the effectiveness of our proposed method.