Siamese neural networks for detecting banknote printing defects

The production of banknotes is a complex process, composed of different printing steps, in which various kinds of defects can be generated which, if not adequately monitored, can lead to production waste, significantly impacting productivity and costs. This paper proposes a novel approach for identifying defects during banknote production using ‘one-shot learning’ methods. These methods rely on a small number of observations to train a Siamese neural network to reproduce the similarities between pairs of samples. The network can then identify defects in new banknote images by comparing them to benchmark samples. The proposed approach allows the correct identification of some specific defects on banknotes, even with limited training data, laying the foundation for the development of a solution for the recognition and intelligent classification of defects on banknotes.