A Bibliometric Analysis and Systematic Review of Image Recognition for Intelligent Damage Detection in Engineering Structures

Abstract

Structural health monitoring and regular damage inspection are critical to ensure the operational safety of civil infrastructure and reduce life-cycle maintenance costs, while traditional manual inspection suffers from low efficiency, high subjectivity, and occupational safety risks for inspectors in hard-to-reach areas. Although existing reviews have explored image-based damage detection, most focus on single damage types or individual infrastructure categories, with few providing quantitative bibliometric mapping of the whole field. This study combines bibliometric analysis and systematic review to trace the development trajectory, identify unresolved technical bottlenecks and industry–academia gaps, and provide a structured reference for researchers and engineering practitioners. Following PRISMA guidelines, 171 peer-reviewed publications from the Web of Science Core Collection (2009–2025) were included after two rounds of screening (initial retrieval: 892 records). CiteSpace and VOSviewer were jointly used to analyze publication trends, institutional cooperation networks, and emerging research hotspots, followed by a systematic review of technical evolution and engineering applications. Results show that annual publications have maintained a growth rate of over 40% since 2019, with China (54.4%) and the United States (22.2%) as the core global contributors; 89.5% of research outputs come from universities and research institutes, while enterprise participation accounts for only 8.3%, indicating a clear technology translation gap. Technically, the field has evolved from traditional digital image processing to deep learning paradigms (CNN, YOLO, U-Net, GAN, Transformer), integrated with UAV platforms and 3D reconstruction to achieve both intelligent damage identification and 3D quantitative assessment. Key bottlenecks include scarcity of high-quality multi-class annotated datasets, poor model robustness in complex field environments, insufficient pixel-to-engineering scale conversion accuracy, and low model interpretability. Future directions include multimodal sensor fusion, unsupervised domain adaptation for real-world generalization, lightweight edge-deployable detection models, strengthened industry–academia collaboration, and explainable artificial intelligence to accelerate technology deployment in engineering practice.