Q. Why did one of your five dollar bills succeed and one of them fail?

A. They are actually the same bill! The image that succeeded is merely a cropped version of the one that failed. The bill that was not recognized either had some dirt and grime around the edges or had a visible shadow as a result of the scanning process. The corner detector saw the edge, and interpreted the edge as the top of the image on the bill. Therefore, when the image was automatically cropped and resized in Matlab, it did not actually line up with any of the samples. It was apparently close enough to both the one dollar bill and the five dollar bill to be considered a possible match, and matched the one dollar bill slightly more than the five dollar bill. To make our program read this bill successfully, we manually cropped the edge of the five to where only the image of the bill was visible.

Q. Could you fool this program with a black-and-white copy of a bill?

A. Probably not. Although we didn't really test this (as making copies of money is a federal offense), we would surmise that a black-and-white copy would fail. The reason is, since the copy would effectively be in grayscale, the red, green, and blue values would be approximately equal for any pixel in the bill. In an actual bill, the green values would be much higher in certain areas. This would create a lower correlation for a copied bill. It would take some experimentation with our threshold value to be sure of throwing out copies, however.

Q. Could you fool this program with a color copy of a bill?

A. Undoubtedly. But then you'd be breaking the law, and the Secret Service would have a few words to say to you. Actually, there is a good chance of catching these if your money scanning hardware is powerful enough, and your samples contain enough details. Color copies lack the detail of the original money and have many distortions. The problem is that the more detail you use, the longer it takes to scan the bill.

Q. Why did you look across the entire bill, rather than looking for key features like the number in the corner or the face?

A. There are several reasons for this. The first is that while our scanner probably will not pick out a counterfeit bill made with good equipment, it should not be a complete idiot either. In other words, it needs to be able to determine whether an entire bill is being verified, or if it has only been presented with a small bit of one.

The second reason is that every denomination has different features, enough so that no one algorithm could easily distinguish between the different varieties. The new $20, $50, and $100 have drastically different features from the older $1, $2, $5, and $10. Being able to find the key features also brings us back to the problem of shifting the money. We would waste so much time looking around the bill for something by which we could evaluate it, that we could have found the dot product of the whole thing and be finished.

Q. Could you add more bills to your sample library, and scan across those?

A. Yes! Not only can we do that, it's trivial. To do this, first find a good clean sample bill that you wish to include within the library. Scan the bill, and size it to about 600x250. While the program is more than capable of accepting any size image, this is about the same size as the other bills within our inventory.

Once you have done that, save the image as a JPEG, and load it into Matlab using imread. Use the findedge.m script to locate its corners. Next, edit the inp2.m script, adding a line to load the new image as well as a line containing the corners found above. Use the rest of the file as a model. Finally, add a line to the whatami.m script that scans across the new bill. It can go anywhere in the tree of if statements. Once again, use the rest of the file as a model.

The nice thing about this technique is that it is very easy to add in bills such as the old $20, $50, and $100. It's also possible to add foreign currency, such as money from Canada, Mexico, England, etc. The trade off is that the more money you add, the longer it takes to determine its denomination. Therefore, a scanner that is capable of determining the variety of any world currency would take a prohibitively large amount of computational time.