Machine Vision Using Cellphone Camera: A Comparison of deep networks for classifying three challenging denominations of Indian Coins
This is an incremental improvement for automated coin recognition systems, addressing a specific challenge in currency handling.
The paper tackled the problem of classifying three visually similar Indian coin denominations using digital images from cellphone cameras, achieving 97% accuracy with two deep neural network models.
Indian currency coins come in a variety of denominations. Off all the varieties Rs.1, RS.2, and Rs.5 have similar diameters. Majority of the coin styles in market circulation for denominations of Rs.1 and Rs.2 coins are nearly the same except for numerals on its reverse side. If a coin is resting on its obverse side, the correct denomination is not distinguishable by humans. Therefore, it was hypothesized that a digital image of a coin resting on its either size could be classified into its correct denomination by training a deep neural network model. The digital images were generated by using cheap cell phone cameras. To find the most suitable deep neural network architecture, four were selected based on the preliminary analysis carried out for comparison. The results confirm that two of the four deep neural network models can classify the correct denomination from either side of a coin with an accuracy of 97%.