ISI Journals

Concrete remains the cornerstone of modern civil infrastructure, yet its vulnerability to crack formation represents one of the most persistent threats to structural durability and public safety. Effective detection and monitoring of cracks in concrete structures are critical for preventing catastrophic failures, reducing maintenance costs, and extending service lifespans. Traditional inspection methods, reliant on visual evaluation or sensor-based techniques, often suffer from limitations in accuracy, scalability, and timeliness. With the advent of deep learning, computer vision-based approaches have emerged as a transformative solution to automate crack detection with high precision. Among these, the Swin U-Net architecture—a hierarchical transformer-based model integrating the strengths of convolutional neural networks and self-attention mechanisms—has gained prominence due to its ability to capture both global contextual dependencies and fine-grained local features. This study investigates the optimization of crack detection in concrete infrastructure through attention-based Swin U-Net frameworks. The integration of attention mechanisms enhances the model’s capacity to discriminate between cracks and surface noise, thereby addressing challenges such as varying lighting conditions, background complexity, and multi-scale crack patterns. Additionally, the hierarchical window-based self-attention in Swin Transformers enables computational efficiency while maintaining robustness in large-scale imagery typical of civil engineering applications. The research highlights how these architectures outperform conventional CNN-based models, offering improved accuracy, recall, and intersection-over-union scores across diverse datasets of concrete images. Furthermore, the implications of adopting such architectures extend beyond academic research to practical structural monitoring systems. By embedding attention-based Swin U-Net models into digital twins of civil infrastructure, stakeholders can achieve real-time crack detection and predictive maintenance strategies, significantly improving safety outcomes. The paper also addresses computational considerations, model scalability, and integration challenges with existing monitoring platforms. The findings underscore the transformative potential of attention-driven architectures in advancing concrete safety and reliability. This research provides both a theoretical foundation and practical roadmap for deploying Swin U-Net in structural monitoring pipelines, positioning the approach as a next-generation solution for civil engineering.