As more and more of our daily transactions become electronic, the prospect of counting more than a few coins, or even small bills, begins to feel all the more tedious. For someone who has trouble seeing at all or a traveler unfamiliar with the local currency, counting money quickly and accurately may seem all the more daunting.
The coin and bill detection algorithm solves the problem of computing the total value of U.S. currency displayed in a given image by solving a set of detection problems – namely, to detect the number of each possible bill and each possible coin in the image – and then aggregates over the results. Although both classes of objects are semantically similar, their visual properties and relative expected quantities recommend a distinct approach for each. Dollar bills reflect light relatively uniformly and have a diverse combination of distinctive patterns that identify each.
This suggests that a feature-based approach should be effective. Since such an approach requires exhaustively comparing candidate images from a database, there should be relatively few objects to match in a given image for reasonable runtime on computeconstrained systems. This is a reasonable assumption for bill counting because bills occur in much smaller amounts than coins.
They are also physically much larger than coins – there is roughly an order of magnitude of difference in area between a bill and the largest coin – so for a fixed image size, one could capture ten or more times as many coins as dollars.