Thousands of high-resolution factory images reveal how real-world lighting, surface variation, and subtle defects challenge automated inspection of wind turbine coatings.
Paper: High-Resolution Images for Coating Defect Detection in Wind Turbine Structures. Image Credit: Rafal Rutkowski / Shutterstock
A recent study in the journal Scientific Data presents CoatingDet, a high-resolution image dataset for detecting coating defects in wind turbine manufacturing. The researchers collected 5,416 RGB images of coated wind tower components and added detailed defect annotations to them. The dataset provides a useful foundation for developing computer vision tools for faster and more consistent industrial coating inspection.
Addressing Challenges in Wind Tower Coating Inspection
Wind turbine towers use protective coatings to reduce corrosion and extend structural service life. Small coating defects can compromise the protective layer on wind turbine towers. Common defects include pinholes, inclusions, scratches, and contaminants. Detecting these defects early helps manufacturers identify areas that require repair and maintain the coating's long-term performance.
Quality control for wind tower coatings relies heavily on manual inspection. Inspectors must examine large, curved surfaces that can exceed 2,000 m² per tower. This process can take considerable time. Fatigue and differences in individual judgment can also affect inspection results. Computer vision and deep learning could make coating inspection faster and more consistent.
However, these technologies require large, well-annotated datasets that represent real manufacturing conditions. Existing industrial datasets often focus on metal surfaces, composite materials, or controlled laboratory environments. The researchers addressed this gap by developing CoatingDet. The dataset contains high-resolution images of wind tower coating surfaces and includes both critical defects and small particles that may resemble defects.
Capturing Coating Defects Under Real Production Conditions
The researchers collected images at GRI Renewable Industries’ wind tower manufacturing facilities in Galicia, Spain. They initially used a mobile computer vision system that moved around tower sections while maintaining a consistent distance from the surface. The system was initially operated manually before being integrated into an autonomous robotic platform used during the 2025 acquisition campaign. Most images were captured using a high-resolution industrial camera with a Sony IMX542 sensor. It captured images at 5,320 × 3,032 pixels through a 25 mm focal-length lens. A custom tangential illumination system enhanced the visibility of subtle surface features, while positioning the light close to the coating reduced the influence of ambient light changes.
Image collection took place over approximately three years. They recorded defective areas identified during quality-control inspections and captured non-defective surfaces during inspection downtime. After each acquisition session, the team reviewed the images and removed blurred images and annotation errors.
The dataset includes five surface categories. Four categories cover critical coating defects, including inclusions, pinholes, scratches, and contaminants. The fifth category captures minor surface particles, such as dust and dirt. In the released object-detection annotations, the four critical defects were grouped into a single defect class, and particles were treated as a separate class. The released YOLO bounding-box annotations therefore use two classes, while the finer defect categories are retained at the image level in the accompanying metadata. This classification helps the detection models distinguish defects that may require repair from minor surface particles.
Deep Learning Models Identify Defects Across Different Conditions
CoatingDet contains 5,416 images, including 699 images with critical defects. From these images, the researchers created 2,385 image tiles with 2,461 annotated instances. The annotations included 773 critical defects and 1,688 particles. The dataset also separates the images into training, validation, and test sets based on acquisition sessions and surface conditions. The team created bounding-box annotations using the VGG Image Annotator (VIA). A second expert independently reviewed 150 images containing 250 annotated boxes. The two experts agreed on the classification of defects and particles in 88% of the cases.
For technical validation, the researchers tested two object-detection models. YOLOv11n is a lightweight model designed for fast detection. Real-Time Detection Transformer (RT-DETR) uses transformer-based attention to capture broader image context. The experiments were intended to validate the dataset’s usability rather than establish which architecture performed best. Both models used 512-pixel input images and the Ultralytics framework. The original high-resolution images were divided into non-overlapping 512 × 512 pixel tiles. They retained tiles that contained sufficiently large portions of annotated surface features, as defined by predefined area thresholds.
The models achieved [email protected] values ranging from 0.70 to 0.86 across most validation and test conditions. YOLOv11n achieved 0.856 on the validation set, while RT-DETR achieved 0.812. On the largest operational test set, RT-DETR achieved 0.833, compared with 0.700 for YOLOv11n. The study also showed that lighting conditions can strongly influence detection performance. Test 1, the only warm-lighting subset and one containing particles but no critical defects, produced a particularly large decline for RT-DETR. Applying white-balance correction substantially improved performance, indicating that color normalization or training across a wider range of lighting conditions may be important for reliable automated coating inspection.
Towards More Reliable Automated Coating Inspection
CoatingDet provides a practical resource for developing artificial intelligence systems for automated inspection of wind turbine coatings. Its main strength is the use of high-resolution images collected under realistic manufacturing conditions. The dataset captures large, curved, and reflective coating surfaces, helping researchers develop detection models that can better address challenges found in industrial environments.
The study also identifies areas that need further improvement. Particle instances outnumber critical defect instances, creating a class imbalance that can lead to more false alarms. The researchers found that this imbalance primarily led to an increase in unnecessary repair calls rather than to missed critical damage. Future work could expand the dataset with more examples of less common defects and a wider range of lighting and surface conditions. Adding pixel-level segmentation masks could also support additional tasks, including anomaly detection. The available preprocessing tools and metadata can support reproducible research and allow researchers to compare different detection approaches.
Overall, CoatingDet connects high-resolution imaging with deep learning for industrial quality control. It could support the development of faster, more consistent inspection systems and contribute to more efficient wind turbine manufacturing.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.