Real-time instance segmentation and quantification method for concrete dam cracks based on RCIS-YOLO

To address the high complexity of dam crack monitoring workflows involving detection followed by segmentation and the limitations of unmanned aerial vehicle applications, this article first proposes reparameterized crack instance segmentation YOLO (RCIS-YOLO). This model uses YOLOv8s-seg as a baseline, optimizing the network backbone and neck through a reparameterized structure and lightweight modules, while introducing a hybrid loss function to resolve the imbalance in crack foreground data. Subsequently, we integrate monocular camera, laser rangefinder, RCIS-YOLO, and Gaussian imaging model to construct an instance segmentation-based crack monitoring framework suitable for confined spaces. Experiments demonstrate that RCIS-YOLO achieves higher accuracy and lighter weight compared to similar models, while its integrated detection-segmentation capability significantly reduces monitoring complexity. The proposed crack width quantification method achieves a relative error of 3.62% on eight real-world cracks, meeting engineering requirements. Finally, an intelligent crack monitoring application was developed by integrating these techniques.