LAWS@IU.AI.SRI.COM (Ken Laws) (02/19/88)
Much of the MIT vision literature deals with data smoothing and interpolation by fitting mathematical "thin plates" through the image data. The data I get is usually too smooth already, which may be why the human vision system introduces the Mach effect. The question is, once you have smooth data (e.g., if it were given to you initially) what are you going to do with it? Threshold it? Detect edges? Segment it? Match it to templates? To generic models? Take Fourier transforms? Moment invariants? Count concavities relative to the convex hull? The vision literature in graphics tends to consider only binary data, ignoring the gray levels that high-quality scanners pick up. There are shrink/expand techniques for smoothing and many papers on how to characterize approximations to straight lines and arcs on a digital grid. You should check out the IEEE book list, particularly the pattern recognition conferences and related books such as "Machine Recognition of Patterns" and "Computer Text Recognition and Error Correction". There is a very old book called "Optical Character Recognition" that still has some good info on recognition by moments and some examples of just how bad scanned characters can be. -- Ken -------