A SYSTEMATIC REVIEW OF DEEP LEARNING TECHNIQUES FOR CRACK DETECTION AND STRUCTURAL DAMAGE ASSESSMENT

Abstract

Structural health monitoring has become increasingly critical for ensuring the safety and longevity of civil infrastructure, yet traditional manual inspection methods remain time-consuming, subjective, and often hazardous. Deep learning techniques have recently emerged as powerful tools for automated crack detection and damage assessment, but the rapidly expanding literature in this domain presents challenges for researchers seeking to understand prevailing trends, comparative performance, and remaining gaps. This systematic review therefore aims to synthesize and critically evaluate the state of the art in deep learning approaches for crack detection and structural damage assessment, with particular focus on architectural innovations, multi-modal data integration, and deployment feasibility. We conducted a structured search and rigorous screening of peer-reviewed publications, then extracted and analyzed key findings related to model performance, data strategies, and application contexts. Our analysis reveals that convolutional neural networks, particularly those with encoder-decoder and attention mechanisms, consistently achieve high accuracy for pixel-level crack segmentation in standard image datasets. We further observe that hybrid frameworks combining deep learning with complementary sensors, such as ground-penetrating radar or acoustic emission, significantly improve detection under occluded or noisy conditions. However, critical challenges persist: data scarcity and class imbalance remain inadequately addressed across most studies, and few works demonstrate real-time capability in field deployment. We also find that domain adaptation techniques, although promising, have been applied predominantly to laboratory settings rather than to extreme events like earthquakes or fires. Based on these synthesized findings, we propose a set of best practices for model selection, data augmentation, and validation protocols, and we identify several high-priority directions for future research, including unsupervised learning for scarce damage scenarios and lightweight architectures for embedded systems. This review provides a comprehensive roadmap for practitioners and researchers advancing automated structural damage assessment