It is common practice in manufacturing steel to stamp an identification code on each bar of the stuff. This is done with a metal stamp and a hammer or equivalent. The result may be several digits punched into one end of a bar of steel, and it would be nice in some applications to be able to read these digits automatically by pointing a camera at the bars as they roll by. Variants of this problem occur throughout industry; sometimes the only action taken is to make stock entries, sometimes other automation responses are appropriate. Sometimes, the stamping process is itself automated and needs to be verified.
The image now is a grey scale one, some typical images are provided as TIF files on the accompanying disk. I am grateful to Chris de Silva and Gek Lim of CIIPS, The University of Western Australia, for image files and also many programs used in image processing.
The first thing to do with a greyscale image of a character, if at all possible, is to threshold it back to a binary image. Inspection of an image will, sometimes, allow you to decide if this
is feasible. In the case of the stamped characters, thresholding to a binary image does not work well here, since
information relevant to the recognition is lost. Ideally one would have a thresholding process which adapted the threshold to the local environment rather than one which merely set a value and stuck with it. This sort of consideration leads us to the theory of filtering, which is properly part of an image processing course rather than a pattern recognition book, and some pointers to the literature are provided in the bibliography at chapter's end. Removing `textural noise' can be done to some extent, but it still leaves us normally with a grey scale image.
In Fig.2.11 we indicate two images, of a
letter E in terms of the function from 
to the
brightness value in 
, one is a binary image
and the other a greyscale image. It is useful
to look
at images as functions defined in some region
of the plane in much of what follows.
