Abstract
Concrete structures form the backbone of modern civil infrastructures, yet they are inherently susceptible to cracking over time due to mechanical stress, environmental exposure, and aging. Undetected cracks can escalate into catastrophic structural failures, making timely inspection critical. Traditional manual inspection methods are labor‐intensive, subjective, costly, and unsuitable for large‐scale deployment. Unlike prior surveys that treat vision‐based crack detection in isolation, this paper positions automated crack detection as one component of a broader vibration‐aware, multisensor structural health monitoring (SHM) ecosystem, in keeping with the scope of this journal. We present a critically synthesized review of automated concrete crack detection, spanning classical image processing and machine learning, convolutional neural networks, object detection, semantic segmentation, Vision Transformers (ViTs), generative adversarial networks, Explainable Artificial Intelligence (XAI), and, distinctively, vibration‐vision sensor fusion, Digital Twin (DT) and physics‐informed frameworks that connect surface cracking to internal structural state, emerging vision‐language and foundation models (SAM, Grounding DINO, CLIP, and GPT‐assisted inspection), and depth‐aware 3D crack characterization via stereo vision, structured light, and LiDAR. Rather than only tabulating reported metrics, each methodological family is critically appraised along robustness, computational cost, practical applicability, and scalability, and every method‐specific section closes with an explicit synthesis of strengths, weaknesses, and open failure modes. We further address two dimensions largely absent from earlier reviews: the severity of a crack is shown to be inseparable from its structural and environmental context (a small crack in a seismically active or dynamically loaded structure is not equivalent to a similar crack in a quiescent one) and the environmental operating condition (above‐water, underwater, shadowed, or variably illuminated) fundamentally changes which detection strategy is viable. We have found that CNN‐based transfer learning methods routinely exceed 99% classification accuracy on curated benchmarks, YOLO‐based detectors provide real‐time localization with competitive mAP scores, ViTs and hybrid architectures offer superior global feature modeling, and XAI tools such as Grad‐CAM and SHAP enhance interpretability required for regulatory compliance, but all of these figures should be read against the robustness, cost, and generalization caveats we discuss explicitly. A dedicated comparison against existing review papers clarifies the specific gaps this survey closes, and quantitative bibliometric figures summarize publication trends, architectural evolution, dataset popularity, and accuracy‐versus‐year patterns across the literature. This survey concludes with a substantially expanded, author‐articulated research agenda covering domain generalization, sensor fusion, DT‐enabled predictive maintenance, and regulation‐compliant explainability for next‐generation automated structural inspection systems.