Summary of OCR Measurement Methods

So far in this chapter we have focussed on the recognition of characters, principally on printed characters, and we have been exclusively concerned with getting from a point set in the plane to a vector of real numbers which describes that set. The first difficulty is the problem of segmenting an image of text into separate characters, and I remarked that this approach, in which the character is regarded as the pixel set of interest, cannot be expected to work well for cursive hand-written text, suggesting that the methodology needs attention. We then examined three methods of turning the point set into a real vector. The first was to put the character into a box with a fixed number of pixels in a rectangular array (after some transformation), and then raster scan the array. This method was not recommended. The second was to apply a family of masks to the set and extract information about intersections between the mask and the pixel set. And the third was a family of methods which included taking a Fourier series expansion as a special case, but is more commonly described as the use of moments. This entails computing inner products in the function space ${\cal L}_2(D^2)$, and hence projecting down onto a basis for the set of square integrable functions defined on the unit disk. Choosing polynomials in the radius and trigonometric functions in the angle and orthogonalising gives us the Zernike moments.

In deciding between different methods, we were much concerned with the problems of having the representation invariant with respect to the kinds of transformations which occur in practice with characters in text- shifting, scaling and the `deck transformation' were mentioned, as was rotational transforms. Making the system robust, or invariant under low levels of noise, can also be thought of as part of the same framework of looking at the problem.

It was observed that these methods could be applied to just the boundary set with some possible saving in computation but at the risk of being more vulnerable to noise.

Specialised methods arising from boundary tracing exist, such as fitting functions to the border of a character and counting convexities, curvatures, or other quantities.

We concluded by observing that the possibility of going through intervening steps, so that a character should more usefully be regarded as being built up out of strokes, and strokes being built up possibly out of something else, giving a hierarchy of structures and substructures, was attractive as a generic model for human information processing. Again, there were promises to keep.