DTU-Wind Turbine Blade Drone Inspection Images

Drones can be used to monitor the surface condition of wind turbine blades. When transmitting image data to servers via 4G/5G technology, defect detection models often have to deal with ultra high-resolution images.

Annotations

This repository contains two annotation folders for an openly available drone captured dataset of wind turbine blades. The images can be found here https://data.mendeley.com/datasets/hd96prn3nc/2.

Under Annotations folder, the dataset is divided into three parts that is train, test and val with a ratio of 70:15:15.

The file names are self-explanatorty, train-1024-s refers to images that is sliced into square size of 1024 pixels. Test set has two files that is test-HR which represents the original images and test1024-s represents the sliced set.

Re-annotation

Annotation process is often tedious in nature, with many challenging decisions to make. For instance, some defects may occupy a large area of blade surface causing bounding boxes to be unwieldy large. Meanwhile, long and narrow components of the WTB can result in extreme aspect ratios, i.e. boxes with amuch larger horizontal side than the vertical side and vice versa, as exemplified in below Figure. Annotators could either choose to keep the entire defect intact or break the defect into smaller boxes. Such decisions are not without consequences; small boxes are notoriously difficult to detect, as many findings have acknowledged while boxes with extremely large or small aspect ratios faced data scarcity for training purposes. Therefore, it is important to explore and determine optimal aspect ratios by assessing their class-wise performances on benchmark datasets.

Re-annotation sub-folders

Under this folder two sub-folders are placed with the name D2 and D3. These folders contain the re-considered annotations of the Crack category with extreme aspect ratio. For details see the paper.

Extreme Aspect Ratio

Citation

If you are using these annotations in your research or considering it as base for your annotation process please cite:

@article{gohar2023slice,
  title={Slice-Aided Defect Detection in Ultra High-Resolution Wind Turbine Blade Images},
  author={Gohar, Imad and Halimi, Abderrahim and See, John and Yew, Weng Kean and Yang, Cong},
  journal={Machines},
  volume={11},
  number={10},
  pages={953},
  year={2023},
  publisher={MDPI}
}

@inproceedings{gohar2023automatic,
  title={Automatic Defect Detection in Wind Turbine Blade Images: Model Benchmarks and Re-Annotations},
  author={Gohar, Imad and See, John and Halimi, Abderrahim and Yew, Weng Kean},
  booktitle={2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW)},
  pages={290--295},
  year={2023},
  organization={IEEE}
}