AI-powered Siamese Networks for banknote quality control
A Banca d'Italia working paper applies Siamese Networks, an AI neural network model, to automate banknote quality control. This approach achieves high accuracy in detecting defects, even when full enumeration of defect types is unfeasible.
AI spots elusive banknote flaws
Banknote production is a highly articulated process, requiring comprehensive quality control measures to ensure integrity and security.
While many stages are automated, final assessment relies on manual visual inspection by specialized operators.
This complex task involves identifying various printing defects, which is crucial for determining batch conformity.
The assessment is time-consuming and prone to subjectivity, as enumerating all possible defect occurrences (type, shape, severity, location) is unfeasible.
This work applies Siamese Networks, a class of artificial neural networks, to support human experts in this critical area.
The AI tool, developed using few-shot learning, aims to spot potential defects on banknote images, extending previous research by widening defect classes and analyzing both front and back sides of banknotes.
It also includes a visualization strategy to support defect localization by human experts.
Training AI on diverse defect data
Quality assessment is crucial for reducing waste in banknote production.
Traditional machine learning struggles with comprehensive defect datasets, as representing all possible defects is often impractical.
Few-shot learning, or k-shot learning, offers a viable solution by mimicking human ability to associate unfamiliar entities based on similarity.
This approach, often relying on Siamese networks, is well-suited for defect detection with highly variable data.
The study's dataset comprises 446 images, including both fit (not defective) and unfit (defective) banknotes.
Defects include missing ink, misaligned printing, ink smears, and physical damage, categorized by severity (severe, mild, light).
The training strategy uses images annotated as fit or unfit, with the unfit class encompassing a diverse and non-exhaustive range of defects, allowing the model to classify unseen defect types based on limited examples.
AI's opaque logic, human's clear vision
This study successfully applies advanced AI to a complex, real-world problem where traditional methods struggle due to the sheer variability of potential defects.
While the model achieves high accuracy, its inherent 'black box' nature necessitates an image segmentation-based explainability approach for transparency.
This highlights the critical need for AI systems in sensitive applications to not only perform well but also provide clear, interpretable insights for human oversight and trust.