Comparative Analysis of Deep Learning Models for Structural Defect Segmentation in Bridges

• Published in: Proceedings of IEEE Southeastcon pp. 166 - 170
• Main Authors: O'Neal, Kiara, Hu, Da
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.03.2025
• Subjects: Accuracy; Benchmark testing; Bridge infrastructure; Bridges; Corrosion; Deep learning; Defect detection; Feature extraction; Maintenance; Safety; Semantic segmentation
• DOI: 10.1109/SoutheastCon56624.2025.10971462

Automated structural defect detection is essential for ensuring the safety and maintenance of civil infrastructure, particularly in bridges where defects such as cracks, spalling, and corrosion can compromise structural integrity. This paper presents a comparative study of three semantic segmentation models-U-Net, Feature Pyramid Network (FPN), and DeepLabv3+-for detecting and classifying structural defects in bridge imagery. Each model was evaluated using two encoder architectures, EfficientNet B3 and MobileOne S4, to assess the impact of different feature extraction strategies on segmentation accuracy. The experiments were conducted using the DACL benchmark dataset, which includes a wide range of defect classes. FPN paired with EfficientNet B3 demonstrated the highest mean accuracy across most defect categories, outperforming the other combinations, particularly in detecting common defects such as cracks and graffiti. However, certain defect types, such as hollow areas and cavities, presented challenges for all models. These results highlight the effectiveness of deep learning models in automated defect detection, while also identifying areas where further refinement is needed to improve performance in more complex defect scenarios.