Abstract

Coin classification as genuine or counterfeit remains a significant challenge for ensuring the
integrity of monetary systems, particularly with the increasing sophistication of counterfeit
production techniques. Traditional inspection methods often fail to capture the subtle differences
between genuine and counterfeit coins, underscoring the need for more advanced,
automated solutions.
This work proposes a dual-stream deep learning approach that fuses complementary evidence:
a texture-preserving stream operating on original coin images and an edge-focused
stream built from Matched Filter–First Derivative of Gaussian (MF-DOG) edge maps. Each
stream is realized with a ResNet-101 backbone, and their probabilistic outputs are combined
via soft voting to form the final decision.
Under 5-fold cross-validation, the proposed model attains mean accuracies of 94.4% on 2
dollars-head, 96.3% on 2 dollars-tail, 92% on 50 cents-head and tail, 90.5% on 25 centshead,
and 91.4% on 25-cents-tail, with an overall mean accuracy of 92.8%. These results
surpass established CNN baselines (VGG-16, VGG-19, ResNet-18, ResNet-34, and
ResNet-50) evaluated under the same protocol, highlighting the benefit of leveraging both
surface texture and fine structural cues for coin authentication.
A second contribution is a high-resolution Canadian coin dataset covering three denominations
(2-dollars, 50-cents, and 25-cents including both head and tail sides) with verified
genuine and counterfeit samples, collected under controlled imaging conditions. Together,
the proposed methodology and the curated dataset advance practical, scalable solutions for
coin authentication with potential deployment in banking, vending, and currency handling
systems.