hahahang2/blade-defect-detection

Blade Defect Detection Dataset (notch / crack / ablation)

A YOLO-format object detection dataset for blade surface defect detection, covering three defect types: notch, crack, and ablation.

Dataset Structure

├── anomaly.yaml   # Ultralytics YOLO dataset config (relative paths)
├── train/         # images + labels
├── val/           # images + labels
└── test/          # images + labels

• —anomaly.yaml uses relative paths — training can be launched directly from this folder.
• —Labels follow the standard Ultralytics YOLO format: class cx cy w h (normalized).
• —The test split also contains defect-free (negative) images with intentionally empty label files, so the false-positive rate on normal blades can be evaluated.

Classes

Usage

# Ultralytics YOLO
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.train(data="anomaly.yaml", epochs=100, imgsz=1024)

Notes

• —Images are of varying resolution; no size normalization was applied.
• —The dataset is small-scale; consider data augmentation during training and expect some variance in metrics across runs.
• —Val and test metrics are not directly comparable, since the test split includes negative samples.

License / Usage Terms

This dataset is released for non-commercial academic research and educational purposes only. By downloading or using it you agree to the following:

• —Academic research / personal study: free to use, please cite or credit the author.
• —Any commercial use, redistribution, or incorporation into other datasets or products: requires prior written consent from the author.

For permission requests, please contact the dataset author via the HuggingFace repository page (Community tab) or the contact information on the author's profile.

All rights not explicitly granted here are reserved by the author.