output_unipen

1. Resampling. A typical allograph, i.e. a letter of the alphabet in a particular shape, contains 3-7 strokes, where a stroke is defined as the collection of (x,y,z) coordinates between two subsequent minima in the velocity-signal. On the average, each stroke can be represented by 5 data points. To be able to compare allographs containing a different number of (x,y,z) coordinates, each raw signal is spatially and temporally resampled to 30 samples.

   Figure 1. Original allograph <A>

   Figure 2. Resampled allograph <A>

2. Size normalization. Handwriting is extremely variable. Different writers do not only produce different shapes for the same character, but also shapes of different size. This may even hold for characters produced by one writer. The first is called between-writer variance, the latter within-writer variability. To be able to compare two resampled allographs, they are normalized to unit-size via:
3. Feature generation. After these two steps, which can be considered as preprocessing steps, a resampled and normalized signal of 30 (x,y,z) coordinates results. From this signal, features can be generated. At the moment, we are using image moment invariants and angular shape features.

output_image

Situation A	Situation B	Situation C
Situation D	Situation E	Situation F

---