High-Resolution Images for Coating Defect Detection in Wind Turbine Structures

CoatingDet is a dataset of 5,416 high-resolution RGB images of coating surfaces on wind turbine tower components, acquired in production, with bounding-box annotations in YOLO format for defect localization and classification. This version supersedes the initial release. The first version described 2,697 images and a three-label `class_labels.txt`; it has been replaced by the complete 5,416-image, two-class release described in the accompanying paper. Annotation scheme (two classes). Each image has a YOLO `.txt` file with one line per instance (`class cx cy w h`, normalized): - `0` — defect: critical coating defects requiring rework, comprising inclusions, pinholes, scratches and contaminants. - `1` — particle: benign surface irregularities (dust) that do not require repair. The release contains 2,526 boxes: 768 defect and 1,758 particle. The finer-grained defect subtype is retained per image in the `label` field of `metadata.csv`, not per box. Class names are declared in `class_labels.yaml`. Structure. Full-resolution images only; tiles are generated locally by the `tile_creator.py` script in the code repository. ``` metadata.csv 5,416 rows, semicolon-separated class_labels.yaml the two-class scheme train/raw_data/{images,labels}/ 1,154 images (1,128 train + 26 val) test_1/raw_data/{images,labels}/ 138 images Low-Warm test_2/raw_data/{images,labels}/ 349 images Low-Cold test_3/raw_data/{images,labels}/ 3,775 images High-Cold ``` Image geometries. Three, reflecting successive camera configurations: 5472 × 3648 (537 images, 2023 sessions, alternative camera body) 5320 × 3032 (1,104 images, 2024 sessions and March 2025, Basler ace 2 landscape) 3032 × 5320 (3,775 images, May 2025, same sensor in portrait on the robotic platform) Note for users combining subsets. Test 1 is the only subset acquired under warm lighting balance and contains particle instances exclusively.