riawelc-rg

RIAWELC-RG: A Radiograph-Level Split of RIAWELC for Weld Defect Classification RIAWELC-RG is a radiograph-level reorganization of the RIAWELC dataset (Totino et al., 2023), a public collection of radiographic weld defect patches spanning four classes: crack, lack of penetration, porosity, and no defect. The original RIAWELC release distributes patches into training/validation/testing subsets at the patch level. This allows patches extracted from the same source radiograph to appear in more than one subset, which permits models to memorize radiograph-specific characteristics rather than learn to generalize — 2,443 files in the original test set were found to be byte-identical duplicates of training files (100% of the original test set). RIAWELC-RG corrects this by assigning every patch from a given source radiograph to a single subset (training, validation, or testing), based on the radiograph ID embedded in each filename. The resulting split: Training: 18 radiographs, 15,120 patches Validation: 6 radiographs, 3,357 patches Testing: 5 radiographs, 3,487 patches Total: 29 source radiographs, 21,964 patches, 4 classes. One test radiograph, bam5, is an anomalously named radiograph included as a deliberate domain-shift case study; it is discussed in detail in the accompanying paper. Related resources: Paper: "Revisiting RIAWELC: Radiograph-Level Auditing Reveals Data Leakage and Class-Selective Domain Shift in Weld Defect Classification" (citation to be added upon publication) Code: https://github.com/adamantoi84/riawelc-rg Original dataset: Totino, B., Spagnolo, F., & Perri, S. (2023). RIAWELC: A novel dataset of radiographic images for automatic weld defects classification. International Journal of Electrical and Computer Engineering Research, 3(1), 13–17. This release is independent of, and should not be confused with, the concurrent RIAWELC-RS resplit by Thfmn. License: CC-BY 4.0