Attention-enhanced deep learning framework for automated concrete crack depth prediction using infrared thermography

Traditional manual methods for measuring concrete crack depth are inefficient, time-consuming, and heavily reliant on operator experience, often resulting in inconsistent and subjective outcomes. Moreover, most existing studies on crack characterization primarily emphasize surface-level parameters such as crack length, width, and area. The crack depth, a key indicator of structural integrity and residual load-bearing capacity, remains insufficiently addressed. To bridge this gap, this study proposes an automated crack depth prediction framework that integrates infrared thermography (IRT) with an enhanced SE-ResNet-18 deep learning model. Concrete beam specimens with precisely calibrated crack depths were fabricated under controlled laboratory conditions, and corresponding thermal images were acquired to establish a robust training dataset. By embedding a squeeze-and-excitation (SE) attention mechanism into the conventional ResNet-18 architecture, the model’s capacity to capture and emphasize salient thermal features was significantly improved, resulting in more accurate and stable depth predictions. Experimental results demonstrate that the proposed SE-ResNet-18 achieves 93.77% accuracy within a ±1 mm tolerance, outperforming the baseline ResNet-18 network by a substantial margin. This solution is fully automated in its predictive analysis and non-contact in its sensing modality. It shows strong potential for practical implementation in real-world structural health monitoring and provides a foundation for future research on field-scale applications and model generalisation under varying environmental conditions.