MVIT_metal_datasets

Provided by Machine Vision and Industrial Testing Laborator (MVIT Lab).

📌 Overview

We release two metal surface defect datasets with instance-level pixel annotations: Casting Billet and Steel Pipe. Additionally, we provide a Medium and Heavy Plate Surface Defect Dataset annotated in YOLO format, along with an unlabeled bridge damage dataset.

🗃️ Datasets

1. Casting Billet Dataset

• Images: 1,060 (780 defective)
• Resolution: 96×106 to 3,228×492
• Defect Types:
  • Scratch
  • Weld slag
  • Cutting opening
  • Water slag mark
  • Slag skin
  • Longitudinal crack

2. Steel Pipe Dataset

• Images: 1,227 (554 defective)
• Resolution: 728×544 (fixed)
• Defect Types:
  • Warp
  • External fold
  • Wrinkle
  • Scratch

3. Medium and Heavy Plate Surface Defect Dataset

• Images: 680 (480 defective)
• Resolution: 256×256 (fixed)
• Defect Types:
  • Inclusions (In)(120 samples)
  • Blocky Scale (Bs)(120 samples)
  • Striated Scale (Ss)(120 samples)
  • Foreign Object Embedding (Foe)(120 samples)

4. Bridge Damage Dataset

• Images: 781 (489 defective) + 21 Sets of Evolutionary Defects
• Resolution: 47x51 to 1,782 x 1,457
• Defect Types:
  • Blistering
  • Corrosion
  • crack
  • Discoloration
  • rust
  • scratch
  • crack-Corrosion
  • crack-rust

✏️ Annotation Process

1. AI Pre-segmentation
   Leverage SAM's predictive interface to perform batch automatic segmentation, generating initial masks based on the provided bounding box annotations and images.

2. Expert Refinement
   1). Identification of Suboptimal Segmentation:
   Review the initial masks to identify suboptimal segmentation results through human assessment.

   2). Interactive Refinement:
   For suboptimal results, use SAM's interactive segmentation by iteratively adding:

   • Positive sample points to guide the identification of the target region.
   • Negative sample points to exclude interference regions.
     Continuously update the segmentation results in real-time until the desired accuracy is achieved.

   3). Post-processing:

   • Perform threshold-based segmentation using optimal thresholds for the specific dataset.
   • Apply morphological operations, including opening and closing, to smooth edges, eliminate noise, fill holes, and perform other enhancements.

🖼️ Samples

📥 Download

• Download Link(baiduyun) | Alternative links(google drive)
• Bridge Damage Dataset: Download Link(baiduyun)

📜 Citation

@article{liu2025advancing,
  title={Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset},
  author={Liu, Chuni and Li, Hongjie and Du, Jiaqi and Hou, Yangyang and Sun, Qian and Jin, Lei and Xu, Ke},
  journal={arXiv preprint arXiv:2509.18919},
  year={2025}
}

@article{li2025few,
  title={A Few-Shot Steel Surface Defect Generation Method Based on Diffusion Models},
  author={Li, Hongjie and Liu, Yang and Liu, Chuni and Pang, Hongxuan and Xu, Ke},
  journal={Sensors},
  volume={25},
  number={10},
  pages={3038},
  year={2025},
  publisher={MDPI}
}

📧 Contact

For dataset inquiries or collaboration opportunities: 📧 xuke@ustb.edu.cn 📧 chuniliu@xs.ustb.edu.cn

---

Maintained by MVIT Lab @ Collaborative Innovation Center of Steel Technology, University of Science and Technology Beijing