Structural health monitoring (SHM) plays a pivotal role in ensuring infrastructure longevity and safety, with crack detection serving as a fundamental task in assessing structural integrity. This study presents a novel image processing framework for efficient and interpretable crack segmentation by integrating squeeze-and-excitation (SE) attention mechanisms into a U-Net architecture, forming the SE-U-Net model. The proposed model enhances feature representation and segmentation accuracy by dynamically recalibrating channelwise feature responses, thereby improving the identification of cracks in structural images. To ensure transparency and interpretability, gradient-weighted class activation mapping (Grad-CAM) and filter visualization techniques are employed, enabling a comprehensive understanding of the model’s decision-making process. The SE-U-Net was trained and evaluated on a publicly available data set, demonstrating superior performance in crack segmentation, with a high dice coefficient of 0.6499 and a mean intersection over union (IoU) of 0.6268. In addition, the model achieved efficient inference speeds of up to 15.7 frames per second (FPS) on standard CPU hardware and maintained high segmentation accuracy with only 0.08 million parameters, highlighting its suitability for deployment on edge devices. The results indicate that the integration of SE blocks significantly enhances segmentation performance and computational efficiency while improving model explainability, a crucial factor for real-world SHM applications. By offering an interpretable deep learning solution, this study contributes to the advancement of artificial intelligence (AI)-driven crack detection, fostering more reliable and efficient structural maintenance strategies.
Zaheer et al. (2026) studied this question.