Every year, the spring freeze-thaw cycle leads to significant structural damage to critical infrastructure. Visible cracks in roads, bridges and buildings are the first signs and, if left undetected, can present serious dangers to the public.
To increase crack detection speed and accuracy, researchers at Concordia have developed a new method that merges drone camera technology with image-based artificial intelligence. Their segment-any-crack (SAC) model builds on existing AI systems for crack detection while requiring far less retraining.
The approach, described in the Journal of Computing in Civil Engineering, achieved higher accuracy than fully retrained models , according to standard AI metrics, and only adjusted less than 0.05 per cent of the system's parameters, saving a large amount of computational power. It also demonstrated strong performance on entirely new datasets, suggesting it can adapt well to different environments without additional training.
Po-Han Chen, centre, with Ghodsiyeh Rostami, left, and Mahdi Hosseini: "Automating the inspection process can be hugely beneficial in territories like Quebec and Canada as a whole, which have vast territories but limited manpower."Fine-tuning instead of retraining
While AI-based inspection tools are faster than manual methods, they typically require large, carefully labelled datasets to be trained effectively. Adapting them to new environments, such as different materials, lighting conditions or surface textures, can be slow and computationally expensive. In some cases, retraining the full model can even overwrite useful knowledge learned earlier.
To address this, the researchers focused on modifying only a small but important part of the model known as normalization layers. These components help stabilize how data moves through the network and can be adjusted to better match new types of images. By fine-tuning these layers, they were able to improve the system without retraining the entire model.
"Instead of going to a site, taking a large number of photos and videos to examine in post-analysis and identifying cracks using conventional algorithms, our system is working to automate the entire process," says Po-Han Chen, a professor in the Department of Building, Civil and Environmental Engineering and study co-author.
"Automating the inspection process can be hugely beneficial in territories like Quebec and Canada as a whole, which have vast territories but limited manpower."
A transferable tool
The team applied their approach to a large, pre-trained AI system designed for image segmentation, which identifies and outlines objects at the pixel level. They then tested their method across multiple crack detection models using a large and diverse dataset of more than 30,000 images, representing a wide range of materials and real-world conditions.
The results show that selective fine-tuning can outperform traditional approaches while using far fewer computational resources. By reducing training time and demands, the method lowers costs and makes large-scale infrastructure monitoring more feasible.
The technology is also more scalable than previous models, allowing it to be applied across a wider range of inspection scenarios.
PhD candidate Ghodsiyeh Rostami and Mahdi Hosseini, an assistant professor in the Department of Computer Science and Software Engineering, co-authored this paper.
This study was funded by the Natural Sciences and Engineering Research Council of Canada.
Read the cited paper: "Segment Any Crack: Deep Semantic Segmentation Adaptation for Crack Detection"
