Deep Learning for Concrete Surface Crack Detection: Sliding‑Window Inference, Non‑Maximum Suppression, and Confidence Heatmaps

Automatic crack detection on concrete surfaces is important for infrastructure inspection because manual assessment is time-consuming, subject to observer variability, and difficult to apply consistently across large and heterogeneous inspection areas. This study evaluated an integrated deep learning workflow for practical concrete surface crack detection under realistic inspection conditions. The objective was to examine whether a patch-based classification workflow using a pretrained ResNet-50 model could support crack presence detection and approximate spatial localization on full-resolution concrete surface images. The dataset combined field-acquired images and four public datasets, totaling 135,711 images, including 55,559 cracked and 80,152 non-cracked samples. Image pre-processing, data augmentation, transfer learning, batch normalization, dropout regularization, and class balancing were applied during model de- velopment. Model performance was evaluated on 41,119 independent test images using accuracy, precision, recall, and F1-score. The workflow achieved an accuracy of 81.98%, with a macro precision of 0.8176, a macro recall of 0.8275, and a macro F1-score of 0.8179. In addition to patch-level classification, the workflow generated spatially interpretable outputs through sliding-window aggregation, suppression of overlapping detections, confidence heatmaps, and overlay visualization. The visual inference results showed that crack-prone regions could still be localized under both ideal and non-ideal image conditions, including varying illumination and surface blemishes. These findings indicate that the evaluated workflow is useful for practical crack-presence detection and approximate spatial localization on concrete surfaces. The study contributes a structured evaluation of an integrated inspection workflow that may support more consistent and interpretable visual assessment in civil engineering applications under diverse field conditions and realistic infrastructure inspection scenarios.