Annotated Image Dataset for Shielded Metal Arc Welding (SMAW) Surface Defect Detection
Description
This dataset contains 448 annotated, high-resolution images of Shielded Metal Arc Welding (SMAW) joints on SS45 medium-carbon steel, developed to support automated visual identification of three common surface-level weld defects: spatter, slag inclusion, and undercut. Images were acquired using a 50-megapixel smartphone camera mounted on a tripod at a fixed 15 cm distance, perpendicular to the weld joint, under uniform ambient lighting. An initial set of 71 high-resolution images (3456×3456 px) was reviewed and verified by a certified welding inspector, who identified and marked the location of each visible defect. Defects were subsequently annotated using polygon and bounding-box tools on the Roboflow platform, then auto-oriented, resized to 640×640 px, and tiled into a 5×5 grid, producing 448 processed image samples with COCO-format JSON annotations. The processed dataset was split 80:20 into training (358 images) and testing (90 images) subsets. The training subset was augmented (flipping, rotation, shear, grayscale conversion, colour jitter, Gaussian blur) to 3,580 images, and further partitioned into five folds for cross-validation (approximately 2,864 training / 716 validation images per fold). Folder structure: - train/ — training images and COCO annotations - test/ — held-out test images and COCO annotations (not augmented) - val/ — validation images and COCO annotations This dataset supports research in object detection, image segmentation, and automated non-destructive weld inspection. It is the companion dataset to the article: Mustajib et al., "Deep learning-based detection transformer (DETR) for defect identification in shielded metal arc welding plate joints," Welding International, 2026, doi: 10.1080/09507116.2026.2708882.
Files
Steps to reproduce
1. Perform SMAW welding on SS45 steel plates to produce specimens with visible surface defects (spatter, slag inclusion, undercut). 2. Capture images using a 50 MP smartphone camera on a tripod, positioned perpendicular to the weld joint at a fixed 15 cm distance, under uniform ambient lighting. 3. Have a certified welding inspector review each image and mark visible defect locations. 4. Annotate defects with polygon/bounding-box tools on Roboflow, assigning one of three class labels: spatter, slag inclusion, undercut. 5. Auto-orient and resize images to 640×640 px; tile each source image into a 5×5 grid (25 tiles) to increase sample count. 6. Split the processed images 80:20 into training and testing subsets. 7. Augment the training subset (horizontal/vertical flip, 90°/random rotation, shear, grayscale, colour jitter, Gaussian blur). 8. Partition the augmented training set into five folds for cross-validation.
Institutions
- Trunojoyo UniversityEast Java, Bangkalan
- INTI International UniversityNegeri Sembilan, Kampung Baharu Nilai
- King Fahd University of Petroleum and MineralsEastern Province, Dhahran