Domain Adaptation Strategies for Cross-Domain Concrete Crack Classification
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The rapid development of modern infrastructure has brought unprecedented challenges to structural health monitoring, particularly in the domain of concrete crack detection. While deep learning has shown promising results in automating this process, its practical implementation faces significant challenges when models are deployed across different materials, environments, and operational conditions. The traditional approach of collecting and annotating extensive datasets for each new deployment scenario is not only resource-intensive but also increasingly problematic due to data privacy concerns and regulatory requirements.
This thesis presents a comprehensive investigation into solving these challenges through two interconnected research directions. Our first approach challenges the conventional wisdom of relying solely on supervised learning by exploring the potential of self-supervised learning in understanding crack features. By leveraging recent advances in vision foundation models, particularly DINOv2, we demonstrate how models can learn robust and transferable representations without the need for extensive labeled data. This research reveals fascinating insights into how self-supervised models develop a more comprehensive understanding of crack characteristics compared to traditional supervised approaches, offering new perspectives on feature learning in structural defect detection.
Building upon these insights, our second research direction addresses the critical challenge of domain adaptation in real-world deployments. We develop an innovative source-free domain adaptation framework that leverages the power of cross-modal understanding through CLIP (Contrastive Language-Image Pre-training). This approach not only eliminates the need for source domain data during adaptation but also introduces novel mechanisms for efficient knowledge transfer across different material types. By incorporating careful consideration of data efficiency and computational resources, our framework provides a practical solution that balances performance with real-world constraints.
The experimental validation of our approaches spans multiple datasets and deployment scenarios, demonstrating significant improvements in both accuracy and efficiency compared to existing methods. More importantly, our research provides a scalable framework for practical implementations, addressing key challenges such as data privacy, computational efficiency, and deployment flexibility. The success of these approaches in laboratory testing and initial field trials suggests promising potential for widespread adoption in automated structural inspection systems.
Please use this identifier to cite or link to this item: