Stroke features used in the NICI/VHS handwriting recognizer

Figure 1 displays the features used in our stroke-based character recognition module.

_____________________________________________________________________________________

Figure 1. The features used to characterize a stroke and its vicinity.

(And, of course, this picture was generated on a pen computer under Windows for Pen Computing, using Paintbrush, and the IBM ThinkPad 360 PE with color LCD).

_____________________________________________________________________________________

The 14 stroke features
0 Yb Vertical starting point with respect to the baseline
1 Ye Vertical end point
2 Phi0 Angles of spatially equidistant subvectors
3 Phi1 (c'ued)
4 Phi2 (c'ued)
5 Phi3 (c'ued)
6 Phi4 (c'ued)
7 Phi-1 Two angles from the end of the previous stroke
8 Phi-2 (c'ed)
9 Phi+1 Two angles from the beginning of the next stroke
10 Phi+2 (c'ed)
11 a loop area, zero if no loop
12 p percentage pen-up time, zero if all ink
13 l length of the stroke trajectory

The 14 stroke features
0	Yb	Vertical starting point with respect to the baseline
1	Ye	Vertical end point
2	Phi0	Angles of spatially equidistant subvectors
3	Phi1	(c'ued)
4	Phi2	(c'ued)
5	Phi3	(c'ued)
6	Phi4	(c'ued)
7	Phi-1	Two angles from the end of the previous stroke
8	Phi-2	(c'ed)
9	Phi+1	Two angles from the beginning of the next stroke
10	Phi+2	(c'ed)
11	a	loop area, zero if no loop
12	p	percentage pen-up time, zero if all ink
13	l	length of the stroke trajectory

Notes

These features gradually evolved from a number of experiments and trials, over several years. Occasionally we try to improve (e.g. by taking more angles or taking normalized XY coordinates) but, basically, the 14-dimensional feature vector captures enough stroke-shape information. The heterogeneous features are scaled such that their range is comparable. Part of the vector is Cartesian (feature 0,1,11,13), but the majority of features is angular (feature 2-10). The directions in handwriting are much more stable than the sizes, over several replications of a character (Teulings & Schomaker, 1993).

The trick to use information of the predecessor and successor strokes was taken from examples in (early) speech recognition, where sliding time-windows with a width of three 10 ms frames were used as a vocal feature vector. As a result, due to the overlap between successive strokes, this scheme is more robust to spurious stroke concatenation than sometimes is suggested in the literature. Also, if bending points (changes from clockwise to anti-clockwise movement or vice versa) are missed in the velocity-based segmentation, the overlapping parts of the feature vector will still capture enough shape information.

Please refer to our Publications when using anything from the shown material.

_____________________________________________________________________________________

to the "NICI stroke-based recognizer of on-line handwriting" page

_____________________________________________________________________________________

Other interesting material:

Handwriting Recognition and Document Analysis Conferences
Pen & Mobile Computing
NICI Handwriting Recognition Group home page
UNIPEN tools
Handwriting-related Java demos

_____________________________________________________________________________________

since 2/Oct/1996