Concurrent Detection of Known Defects and Out-of-Distribution Instances in Building Inspections: Advancements in Deep Classification

Early detection of structural damages is crucial for maintaining a good health level in buildings preventing further deterioration and irreversible harm. Current structural inspections are predominantly carried out by industry professionals, often aided by drone technology to facilitate faster assessments and access hard-to-reach areas. To analyze the data gathered from visual inspections, machine learning-based approaches are commonly employed. While these existing machine learning classifiers excel in identifying the defects they have been trained on, they face limitations in recognizing unseen or unknown data classes.In this study, we propose a novel approach to enhancing the capability of civil structural defect classifiers to classify damages in buildings based on known damage types while also identifying unknown damage categories by incorporating GAN assistance. The proposed GAN-assisted classifier demonstrates significant improvements in detecting Out-Of-Distribution (OOD) data, showing a remarkable enhancement of 15% to 35% compared to the conventional defect classifier. Additionally, we further improved the OOD detection metrics by augmenting the GAN-assisted classifier’s training set with GAN-synthesized data, resulting in an impressive increase of approximately 35-40% in performance. Despite these advancements, the classifier maintains a high classification accuracy of around 96% for known distribution data. The developed approach showcases promising potential for implementation within drones, enabling automated damage detection and classification in buildings, a domain where publicly available data is typically limited.