GoKawiil"The identification of a symbol is based on a data-dependent sequence of tests. At each step in the decision making process there are several potential identifications. Some of these are eliminated by testing the key features of the track. The particular test applied at any step depends on the set of possible identifications at that step, and on those characteristics of the track which have already been examined. The decision-making scheme thus has a tree structure. Its original design was based on an examination of the handwriting of four users. The author changes its structure, to accommodate additional symbol variations, as he acquires more experience."
The recognizer begins by generating a list of potential characters based upon the first four cardinal directions of the stroke. Strokes which cannot, or would not, normally be drawn using these first four directions are immediately eliminated from consideration. Sometimes, the first four directions provide enough information to precisely identify a character. For example, a stroke which moves down, right, and then stops ▼ ▶ can be immediately identified as an L . Similarly, the only possible match for a stroke beginning with the four directions
▶ ◀ ▶ ◀ is the character 3 . Usually, the first four directions will not lead to an immediate identification, and we will be left with a list of potential characters. For example, a stroke beginning with ▶ ▼ ◀ ▲ might end up as a O , 2 or 3 . Click on the images below to see how each stroke is drawn. Notice that all three begin in the exact same way.
We can resolve these sorts of ambiguities by testing other aspects of the stroke description. For example, we can identify a 2 if the stroke ends in the bottom right corner of the symbol. A three can be identified by checking to see if the stroke ends in the bottom left corner of the symbol. We can identify the O by noticing that the start and end positions both lie close to one another at the top of the symbol. In practice, the decision-making process is implemented by writing a large cascade of if statements. There is no “magic” in this process, and the decision tree must be handcoded by a diligent and perceptive programmer who is willing to improve and modify the decision tree as they encounter new and unexpected ways of writing certain characters.
The recognizer I'm presenting in this essay is capable of identifying uppercase letters and numbers which are drawn in a single stroke. In other words, each character must be drawn in its entirety without lifting the pen from the writing surface. Groner’s original recognizer was capable of identifying multi-stroke characters which allowed the user to input text in a completely natural way. Groner’s aim was to leverage existing competencies of the user, to make it unnecessary for the user to adapt their writing style to fit the needs of a software system. In fact, Alan Kay claims that one of the main motivations for the system was to obviate the need for typing instruction. Because my aim is far less ambitious — to explain the basic methodologies employed Groner — I’ve decided to implement a simpler, single-stroke variant of the recognizer. Multi-stroke identification isn’t that much more complex than single-stroke identification, but it is a bit more complicated. The approach is clearly spelled out in Groner’s original memo if you are interested in implementing it yourself.
My constraint that all characters must be written in a single stroke means that the user must adapt their natural handwriting to some degree. I have written a recognizer which expects that the user will input letters and numbers roughly following the vocabulary described below. The F and X in particular, require a bit of unnatural mental contortion from the user. Click on a stroke to see an animation of it being drawn.
While it’s necessary that a general shape be followed when drawing a given character, Groner’s method is robust enough to allow for a great deal of variation and sloppiness when drawing the characters. For example, all of the following strokes will be identified as the number 3 .
... continue reading