Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Autonomous Crack Segmentation Based on Segment Anything Model
The infrastructure of Canada, including bridges, roads, and buildings, is aging. It is of paramount importance to evaluate the state of infrastructure to ensure its serviceability and prevent catastrophic failures. One of the most common defects in concrete structures is crack, which develops from the surface of the structure to deeper parts of it. Early detection of cracks is vital for in-time maintenance planning. The traditional methods of inspection are resource-intensive (labor and equipment), and time-consuming, making the procedure expensive and difficult. Therefore, there is a need for fast, automated, and cost-effective inspection methods. Recent advancements in Deep Learning and Computer Vision have resulted in significant progress in defect detection algorithms for visual identification of defects and extracting further data for severity assessment. However, the majority of deep learning-based crack detection models do not demonstrate generalization to conditions of real-world practices as they are bound to limited datasets and relatively small models. Thus, to develop a more robust crack detection model, this study introduces Crack Large Model (CLM) which leverages large-scale semantic segmentation models, specifically, Segment Anything Model (SAM) for crack segmentation. Through transfer learning, SAM is fine-tuned on a diverse and challenging crack dataset including 11k pairs of crack images and masks. The comparison results of CLM with popular segmentation models (DeepLabv3 + and DeepCrack) have shown its promising performance in multi-scale and multi-level semantic segmentation of different types of cracks, which makes it a potential practical method for crack damage assessment in the near future.
Autonomous Crack Segmentation Based on Segment Anything Model
The infrastructure of Canada, including bridges, roads, and buildings, is aging. It is of paramount importance to evaluate the state of infrastructure to ensure its serviceability and prevent catastrophic failures. One of the most common defects in concrete structures is crack, which develops from the surface of the structure to deeper parts of it. Early detection of cracks is vital for in-time maintenance planning. The traditional methods of inspection are resource-intensive (labor and equipment), and time-consuming, making the procedure expensive and difficult. Therefore, there is a need for fast, automated, and cost-effective inspection methods. Recent advancements in Deep Learning and Computer Vision have resulted in significant progress in defect detection algorithms for visual identification of defects and extracting further data for severity assessment. However, the majority of deep learning-based crack detection models do not demonstrate generalization to conditions of real-world practices as they are bound to limited datasets and relatively small models. Thus, to develop a more robust crack detection model, this study introduces Crack Large Model (CLM) which leverages large-scale semantic segmentation models, specifically, Segment Anything Model (SAM) for crack segmentation. Through transfer learning, SAM is fine-tuned on a diverse and challenging crack dataset including 11k pairs of crack images and masks. The comparison results of CLM with popular segmentation models (DeepLabv3 + and DeepCrack) have shown its promising performance in multi-scale and multi-level semantic segmentation of different types of cracks, which makes it a potential practical method for crack damage assessment in the near future.
Autonomous Crack Segmentation Based on Segment Anything Model
Lecture Notes in Civil Engineering
Francis, Adel (Herausgeber:in) / Miresco, Edmond (Herausgeber:in) / Melhado, Silvio (Herausgeber:in) / Rostami, Ghodsiyeh (Autor:in) / Chen, Po-Han (Autor:in) / Wang, Yang (Autor:in)
International Conference on Computing in Civil and Building Engineering ; 2024 ; Montreal, QC, Canada
Advances in Information Technology in Civil and Building Engineering ; Kapitel: 24 ; 283-293
30.03.2025
11 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
PaveSAM – segment anything for pavement distress
| Taylor & Francis Verlag | 2025
A feature fusion method on landslide identification in remote sensing with Segment Anything Model
| Springer Verlag | 2025
| Springer Verlag | 2025