There are some mildly intelligent things one can do to reduce the computational load in getting from an image to a vector of numbers which describes the image. One is to reduce the number of pixels we have to put into computations by ignoring the interior and simply looking at the boundary. This will only work if the noise levels are low, since reducing the number of pixels will probably put up the Signal to Noise ratio. If the boundary is corrupted, it won't matter so much if we have a lot of other pixels used in estimating the shape. On the other hand, there are ways of tracing the boundary and then smoothing it in case the odd pixel has been tacked on or gouged out, and since we can tell each character by looking at its boundary and boundaries have fewer pixels, some amount of boundary smoothing followed by an analysis of the boundary shape may be an attractive option in some cases.
A boundary is a relatively simple set and can be specified by chain code, where we pick a starting pixel and then specify the direction of the next by labelling the adjacent pixels, as in Fig. 2.8.
Either by such a coding or by linescan methods, it is possible to do such things as count convexities. If circumnavigating a letter /E/ for example, the changes of direction can be determined, and much useable information can be extracted. Many variants of this will occur to the ingenious student. Smoothing boundaries is properly part of an image processing course and there are many ways of doing this; an erosion followed by a dilation can accomplish something of this.
Again, much depends on the kind of data with which one is confronted, but hand written characters are harder to process in this way than printed characters; there is less consistency in general. For printed characters, it is possible to segment the boundary, smooth it, identify strokes and generally extract complex relationships. These methods are not generally robust under change of font, still less do they work for handwritten characters.
