Bank of Italy publishes research on AI banknote quality control tool with high accuracy in identifying defect-free notes

The Bank of Italy has published a research paper in its Markets, Infrastructures, Payment Systems series on using Siamese neural networks and few-shot learning to support final banknote quality checks. The paper examines an AI tool that could help highly trained human inspectors spot potential printing defects in banknote images, including in cases where defect types, shapes, severity classes and locations cannot be exhaustively listed in advance. It reports high accuracy, with particularly strong reliability in classifying banknotes as fit, and adds an explainability component to make the model’s output more transparent to human users. The analysis uses 446 scanned banknote images covering both front and back sides, 21 defect types, and severe, mild and light defect classes alongside fit notes. Performance improved when the model was trained on all defect severities rather than only severe and mild defects. On the reported test sets, the front-side model trained on all defects achieved precision of 1.0, recall of 0.926 and specificity of 1.0, while the back-side model reached 0.95, 0.633 and 0.963 respectively. The paper also proposes a segmentation-based visualization method to highlight image regions likely to be driving the score, while noting that this approach does not fully mirror the network’s internal logic. The paper says future experimental work and additional samples could help validate performance further and refine classification thresholds for production use. It also points to scope for a model-specific explainability method tailored to Siamese networks.