Drone Footage Wind Turbine Surface Damage Detection
Ashley A. Foster, Oscar Best, Mario Gianni, Asiya Khan, Keri Collins, Sanjay Sharma
Abstract
In this work a new publicly available dataset of wind turbine surface damage images is presented. Moreover, a comparison between ResNet-101 Faster R-CNN and YOLOv5 for Wind Turbine Surface Damage Detection is analysed and performance of these models on drone footage with active turbines is also discussed. Results show that YOLOv5 outperforms ResNet-101 Faster R-CNN in predicting the bounding box coordinates of the damaged surfaces of the wind turbines. However, unlike YOLOv5, ResNet-101 Faster R-CNN estimates an entire instance of damage as a single prediction.
Topics & Concepts
DroneTurbineResidual neural networkBounding overwatchComputer scienceWind powerMarine engineeringEnvironmental scienceMinimum bounding boxWind speedWork (physics)Remote sensingArtificial intelligenceDeep learningAerospace engineeringMeteorologyGeologyEngineeringElectrical engineeringPhysicsMechanical engineeringImage (mathematics)BiologyGeneticsAdvanced Neural Network ApplicationsRemote Sensing and LiDAR ApplicationsIndustrial Vision Systems and Defect Detection