Optimizing deep learning-driven computer vision for civil infrastructure defect Identification: Challenges and strategies

Optimizing deep learning-driven computer vision for civil infrastructure defect Identification: Challenges and strategies

WoS Article

With advancements in Internet of Things (IoT) technologies and deep learning, Structural Health Monitoring (SHM) is progressing towards long-distance and intelligent applications. To promote the widespread adoption of deep learning in vision-based SHM, this survey compiles optimization strategies derived from deep learning models specifically designed for defect identification in civil infrastructure-an area that has not been comprehensively explored. First, a concise overview of fundamental deep learning models for vision-based defect identification is provided. Next, optimization methods are categorized into three main groups, each addressing a distinct challenge encountered in practical vision-based SHM: optimizing considering defect-specific characteristics, lightweight design for real-time identification, and enhance robustness under complex environmental conditions. These methods are further classified systematically based on their similar features. This survey offers researchers a deeper understanding of the challenges in vision-based defect identification and assists them in selecting appropriate optimization techniques to address these challenges. Ultimately, it aims to enhance the effective deployment of deep learning models in vision-based SHM by improving accuracy, enabling real-time operation, and facilitating automated defect identification.

With advancements in Internet of Things (IoT) technologies and deep learning, Structural Health Monitoring (SHM) is progressing towards long-distance and intelligent applications. To promote the widespread adoption of deep learning in vision-based SHM, this survey compiles optimization strategies derived from deep learning models specifically designed for defect identification in civil infrastructure-an area that has not been comprehensively explored. First, a concise overview of fundamental deep learning models for vision-based defect identification is provided. Next, optimization methods are categorized into three main groups, each addressing a distinct challenge encountered in practical vision-based SHM: optimizing considering defect-specific characteristics, lightweight design for real-time identification, and enhance robustness under complex environmental conditions. These methods are further classified systematically based on their similar features. This survey offers researchers a deeper understanding of the challenges in vision-based defect identification and assists them in selecting appropriate optimization techniques to address these challenges. Ultimately, it aims to enhance the effective deployment of deep learning models in vision-based SHM by improving accuracy, enabling real-time operation, and facilitating automated defect identification.

Deep learning, Multiscale defect, Defect morphology, Lightweight, Robustness, Generalization, Complex background

Deep learning, Multiscale defect, Defect morphology, Lightweight, Robustness, Generalization, Complex background

FU, R.; HUANG, Z.; CAO, M.; NOVÁK, D.; XIE, C.; HUANG, J.

22.10.2025

Engineering Applications of Artificial Intelligence

Part B

158

United Kingdom of Great Britain and Northern Ireland

1

29

29

https://www.sciencedirect.com/science/article/pii/S0952197625015234?via%3Dihub