Wind Turbine Surface Damage Detection Using YOLOv7 with Slicing Aided Hyper Inference (SAHI)

This paper expands on the publicly available dataset of wind turbine surface damage and uses this dataset to finetune all P5 models for both YOLOv5 and YOLOv7 object detection frameworks. YOLOv7 outperformed YOLOv5, with the YOLOv7x model achieving the highest recall score and best F1-confidence. This model was therefore selected for inference on both images and video of wind turbines. Slicing Aided Hyper Inference (SAHI) has also been used to improve detection capability for smaller instances of damage. The model was further evaluated on a dataset collected from a scaled model of a wind turbine, with hand drawn damages. Lastly, this dataset was used for inference using SAHI, which showed slight improvement for detecting damage instances. More accurate results were observed when evaluating the model on real damage examples compared to simulated damage.