A curated catalog of openly available photo datasets showing real-world welds in industrial settings — suitable for labeling pixel-level segmentation masks for defect detection, bead localization, and quality inspection models.
The largest fully open datasets — no registration, no access requests.
Industrial weld seam inspection images with object detection annotations: COL, EXPOSURE HOLE, IO, PLASMA, UNWLD. Pre-trained YOLOv11 and YOLO-NAS models available. Created December 2024.
640×480 images in 5 classes (background + 4 weld seam geometries) with 10,000 pixel-wise semantic segmentation labels in color-coded masks. Designed for robotic abrasive belt grinding verification.
Real MAG robotic welding images at 640×480 and 2048×1080. COCO JSON bounding boxes for 4 defect classes: pores, deposits, discontinuities, stains. Published IEEE Access 2024 with k-fold splits. Cleanest download path of any dataset here.
Well-documented datasets with existing segmentation masks or detection labels ready to use.
Weld surfaces with YOLO bounding boxes for 3 classes (bad weld, good weld, defect). A companion instance segmentation version with polygon masks is also available. Referenced in PLOS ONE SH-DETR paper.
448×448 color images with binary pixel masks for weld bead segmentation. Straight-bead welds only; masks are coarse/sparse. Referenced in the INWELD paper (MDPI 2025).
Real-world manufacturing and construction weld defect photos on steel structures. Used to benchmark YOLO v5/v8/v11/v12, Faster R-CNN, and SSD. Good diversity of environments.
Focused collections that add weld-type diversity or high-quality annotation references.
Curated from published papers — conventional welding and metal deposition additive manufacturing. Hand-annotated pixel-wise masks with source spreadsheet. University of Michigan / Purdue.
Bounding boxes for Defect, Welding Line, and Workpiece — the class structure separating weld line from workpiece is a useful reference for structuring region annotations. 580+ downloads.
Manually produced welds photographed from above with YOLO bounding box annotations. Published in IET Collaborative Intelligent Manufacturing (2024). Benchmarked YOLOv5–v8.
Key fields at a glance — task type, image count, and download link. All fully open, no registration required.
| # | Dataset | Images | Task type | Weld process | License | Link |
|---|---|---|---|---|---|---|
| 01 | Roboflow Weld Seam Insp. | 9,377 | Detection | Industrial seam | Public | Roboflow |
| 02 | Mendeley Weld Seam Geom. | 4,000 | Segmentation | Industrial | CC BY 4.0 | Mendeley |
| 03 | LoHi-WELD | 3,022 | Detection | MAG robotic | Open access | GitHub |
| 04 | Welding Defect OD | ~2,154 | DetectionSegmentation | General | Public | Kaggle |
| 05 | Welding Seams | 1,394 | Segmentation | Straight bead | Public | Kaggle |
| 06 | Weld Defects | 1,134 | Detection | Steel / construction | Public | Kaggle |
| 07 | Weld Bead Literature | ~300 | Segmentation | Various | CC BY 4.0 | Mendeley |
| 08 | Roboflow FYP | 408 | Detection | General | CC BY 4.0 | Roboflow |
| 09 | Szőlősi et al. | Small | Detection | Manual | GPL-3.0 | GitHub |