The line of thought which starts off with segmentation of a page of text into lines, then into characters by checking that the set of points is separated by white pixels from other sets, and then tries to recognise the characters by looking at their shape, is liable to several sorts of criticism. Contemplation of Fig.2.3, the postcode digits, suggests that the shape based methods might work well with printed characters where there is a high degree of consistency and where, except when noise levels get very high, characters are isolatable by finding boundaries. Up to a point this is correct, as some experimenting with the exercises will show you. On the other hand, for hand-written characters, there is something unsatisfactory about this whole approach. As a person who writes with a pen, I am conscious that I make strokes with more or less care and attention; that the changes of direction are pretty important, but that once I have embarked upon generating a stroke, the thickness is of minor significance. I can easily tell by looking at Fig.2.3 that the horizontal stroke belongs with the second character, i.e. is part of a /5/ and not of a /7/. I have no trouble about the lower end of the /5/ intruding into the /6/, nor the /2/ intersecting the horizontal stroke of the /5/.
Some sort of syntactic processing is going on here, whereby the image is decomposed by the eye not into pixels and then into characters in one step; there is an intermediate entity, the stroke, made up out of pixels, and itself a component part of a character. The eye has no difficulty joining the two components of the /5/; it expects a horizontal stroke because of the other two strokes making up the /5/. If the horizontal stroke were moved two centimetres to the left, it would have been a /3/, but it has to be one or the other. And so the arrangement of some of the strokes leads to hypotheses about the remaining strokes, hypotheses which are confirmed and lead to identifications of other elements.
Attempts to articulate this view of optical pattern recognition have tended into a quagmire. The Artificial Intelligentsia have tried (in such programs as HEARSAY, from Carnegie Mellon University) to articulate the notion of competing hypotheses resolved by some higher level adjudicator, but their methods of representing hypotheses have been mathematically naive and have had only limited success. To be fair, the problem is far from trivial.
What is known of the physiology of the the lower level processing in the mammalian eye leads us to the notion of a hierarchy of levels, proceeding from essentially pixels, retinal cells triggered by rhodopsin release, to edges or blobs or motion. It is known that a layer away from the retinal cells, there are cells which respond to edges in specific orientations. It is not too hard to believe that a little further back again there are cells which respond to strokes, sequences of edges. And so on, until, conceivably, there are cells responding to digits.
Whether this continues to the level of the mythical `grandmother neuron' which fires when and only when you see your grandmother, is a matter into which we shall not enquire. What is plausible from neurophysiology is that some aggregation process builds `molecules' out of `atoms', and then supermolecules out of molecules. In the same way, one might say, as letters aggregate to words, and so on. By analogy with the structure of language, whereby a sentence is considered to be derived from more primitive entities by means of rewrite rules, we talk of a grammar and of syntax of the components.
The diagram in Fig.2.9 is, more or less, a parse of a sentence in terms of a phrase structure grammar, showing how each symbol (like `sentence', Noun phrase') at any level from the top down, is rewritten to a string of symbols at the level below it, terminating in symbols which are words of the English language. Actually, we could go one level further down into a string of letters; we don't do this because meddling grammarians a few centuries ago regularised the spelling of words, so there is only one way to spell a word, which makes the rewrite rather uninteresting. Around Shakespeare's day, when an English Gentleman spelled the way he fancied and took no nonsense from grammarians, there was a choice. Winston Churchill was notorious for being atavistic in this and other respects, and if your spelling is not all it should be, console yourself with the thought that you have an eminent rôle model.
The question `how can one decompose objects like Fig.2.3 into strokes and then characters?' leads us into interesting places which will be explored more thoroughly when we get to Syntactic Pattern Recognition. The motivation for doing so is worth reflecting on; strokes are produced by people and recognised by people as comprising characters. This is true not just in European scripts of course; Chinese characters and Japanese characters as well as Arabic and the many other scripts are usually decomposed in this way. Cuneiform is not, but you don't see a lot of that about these days. The idea that it might be good policy to unravel the problem of reading characters by studying the processes at work in human recognition methods produces interesting results upon which I shall expand later.
