Siamese Networks for AI-powered automated banknote quality control

Banca d'Italia today publishes 'Siamese Networks for AI-powered automated banknote quality control', the new issue of the series 'Markets, infrastructures, payment systems'.

We apply Siamese Networks, a class of artificial intelligence (AI) neural network models, to the field of banknote production. The criticality of quality control in banknote production requires human oversight throughout the process. We rely on few-shot learning to develop an AI-powered tool that could support quality control by highly trained human experts at the end of the production pipeline, by spotting potential defects on banknote images. We show our approach succeeds in such a complex application, where the enumeration of all possible defect occurrences - in terms of type, shape, severity class and location - is unfeasible. Our proposal achieves high accuracy and proves especially reliable in ensuring that no defects occur, i.e. classifying a banknote as fit. As is the case with most neural network models, our tool is accurate but inherently difficult to read in its underlying logics. To address this potential issue while enhancing transparency, we complement our proposal with an image segmentation-based explainability approach to support the application of our tool.