The quality of calibration will depend on the size of the object relative to the size of the image, the (nonlinear) optical distortion due to the lens, and whether or not the object to be measured appears in the same location within the image for each inspection.
If you want to calibrate an entire 2D image, one of the best and most common methods is to use an array of dots. You could create such an image by taking multiple snapshots of your coin in different locations in the image, following which you could use image arithmetic to combine the multiple saved images into a single calibration image. (It'd be an odd method, but it could work.) In order for this to work you would have to be able to locate the coin or calibration object using a precise XY stage.
When possible, use a NIST-traceable calibration target for gauging applications. Calibrating a system using a "golden sample" is possible, and such a calibration may be adequate for your application, but it's generally a better idea to calibrate your system using a target traceable to national standards.
