A CONTROLLED DATASET OF PAIRED VISIBLE AND 365 NM ULTRAVIOLET 500 SAUDI RIYAL BANKNOTE IMAGES WITH SECURITY FEATURE ANNOTATIONS

Abstract

Reusable evaluation of automated secure-document analysis requires datasets that preserve imaging modality, feature semantics, spatial ground truth, and source lineage. This article presents a controlled resource of paired visible-light and 365 nm ultraviolet images of genuine and counterfeit 500 Saudi Riyal banknotes. Reconstruction of the archived YOLO-ready package identified 524 balanced filename-derived source keys (262 genuine and 262 counterfeit), each containing one original visible image and one original ultraviolet image, for 1,048 original modality images. The historical release contains 3,143 image-label pairs after augmentation; one expected visible descendant is absent, but its source key remains paired at the original-image level. A corrected experimental root was rebuilt by assigning source keys before augmentation: 420 keys were used for training, 52 for validation, and 52 for testing. The training set comprised 1 original and 2 images released images per source key and modality, while the validation set and test set consisted of 2,728 images of the original images. There are sixteen visible classes (eight expected features and eight replica classes), and four ultraviolet classes (two expected responses and two counterfeit responses). Valid class identifiers were found and found to be correct at the time of the audit, one-to-one image-label pairing was found and found to be correct at the time of the audit, there was no overlap of any source-key, augmentation locking was found and found to be correct at the time of the audit, and there were no exact-image duplication warnings found for any image with a cross-split except for 443 warnings that were retained for transparent review through perceptual-hash screening. Only annotation-leakability baselines were used with separate YOv8m detectors. On the held-out test set, VISIBLE16 achieved precision 0.9907, recall 0.9977, mAP50 0.9917, and mAP50-95 0.7958; UV4 achieved precision 0.9934, recall 1.0000, mAP50 0.9950, and mAP50-95 0.7829. The resource supports controlled research on localized security-feature detection; it does not establish operational forensic validity, physical-note independence beyond the recoverable source keys, or generalization to other denominations and document families.