Enhanced Small-Scale Crack Detection in High-Resolution Imagery Using Object Detection Model and Slicing Aided Hyper Inference (SAHI)

Unmanned Aerial Vehicles (UAVs) and high-resolution images are becoming more important for Structural Health Monitoring (SHM) to observe anomalies like cracks in concrete. But cutting-edge object identification methods like YOLOv8 typically have trouble finding extremely thin cracks in large images since they have to down-sample them aggressively during standard inference. This study presents a pipeline that combines YOLOv8 with Slicing Aided Hyper Inference (SAHI) to solve the problem of detecting small objects in civil infrastructure. The study compares the usual full-image inference method with the new sliced inference method. Experimental findings demonstrate that the SAHI-integrated pipeline drastically improves the ability to find weak, fine-grained, and hairline fractures compared to traditional inference. Finally, the sliced images are merged into continuous crack structures using the morphological dilation method, thereby combining multiple overlapping bounding boxes together to form polygonal detection boxes.