  
|
 |
| April 2005 |
Feature |
|
Adding new dimensions to handwriting recognition
|
|||||||||||||||||||
Dr Malome Khomo |
Pictured to the right is a handwriting recognition device, the brainchild of system's scientist, Dr Malome Khomo. The system is built on a straight forward idea, the formulation of a new computer text input interface using three-dimensional (3D) handwriting information, to improve the recognition hit rate over existing stochastic recognition methods.
This is a third version of Malome's design featuring improved acoustics for its ultrasonic sensors. Its centerpiece is a cue card that has a letter inscribed on it like a stencil. To use it, one places the cue card onto the device, and uses the transducer-tipped stylus to trace the outline on the cue card stencil. The device immediately recognises the written character, and a sound chip loudly enunciates the identified character in the appropriate language. The prototype also has a visual display output and other demonstration I/O interfaces. Malome's hope is that this device can ultimately supplant the electronic keyboard in a culturally neutral way.
Dr Khomo found that most documented recognition methods began with 2D X-Y handwriting traces, performed their translation onto a standard reference coordinate frame, and then made a maximum likelihood selection of the character that was assumed to have been inscribed in the 2D trace.
The most frequently cited technique in the patent literature was the Hidden Markov stochastic process model. Such stochastic methods assume the existence of a stationary process that generates the signal of interest, and a stochastic error that obscures the signal. To Dr Khomo this statistical technique did not lend itself very well to rapid on-line handwriting recognition (in real-time). It was an inherently cumbersome way of categorizing handwriting that included all the peculiarities of individual style, factoring in things like size and slant which should not feature among topological invariants.
As it turned out, there existed a body of knowledge from classical differential geometry that essentially explained the topological recognition of spatial paths some two centuries ago. So there were no real ground-breaking discoveries. Rather it brought new perspective to classic knowledge. The names Frennet-Serret, and Euler, come to mind.
In particular, the motion of the stylus in the kinematic unit speed frame of reference depends on a few 'tensorial' attributes that include direction of the path, its curvature and torsion. Together, these can be categorised into discreet torsion-free orbital paths connected by finite-torsion ligatures. In other words, each handstroke has a sequence of inferred rotations.
If this sounds bizarre, consider how you effect writing. You begin above the writing surface, attack down to it, make markings 'in-the-ink', and then lift up to retract at the end of the stroke. Those three actions complete a circular (as in connected) path, however 'flatly' you may prefer to think of them. Hence the orbital paths.
So Dr Khomo's deterministic handwriting recognition idea boiled down to classifying the shape of letters into a minimal set of stylus tip orbits that project the desired written image. In this realization additional scientific names such as Pauli and Cartan informed on the projection (gauges) of quantized orbital paths. The patent for this idea is pending.
The initial product is an educational toy, named ANDIKO, the "learn to write" device. This was the centerpiece of the invention that was showcased at the March 2005 SASOL SciFest in Grahamstown by Embizeni Innovation Support Centre of Durban, as sponsored by the Godisa Project. In its turn, the Godisa Project is sponsored by South Africa's Department of Trade and Industry, and Department of Science and Technology, and the European Union.
The invention uses ultrasonic position tracking in the prototype construction. That design choice added inadvertent problems which new commercialization projects can ill afford. These included acoustic fidelity, and the more subtle signal detection jitter. Jitter arises from the limited dynamic pressure sensitivity of the piezoelectric crystals used in transducers and sensors.
Because the method is geometrical, it can be applied to any writing system, not just the Latin alphabet. In fact the patent specification cited the Tamil writing system, one of the richer geometrical character sets, as an example of the character images that can be classified in the recognition method. So this device is truly global, and it should be implemented equally well for Greek, Hebrew, Arabic, Amharic, Osmani, Indo, and Hun based scripts.
The obvious users of this product are educators and learners. Especially early learners, as they begin learning to write. There are also the occupational health professionals who perform remedial training, and who have expressed an interest in a variant of the device that can be used as a corrective handwriting teaching aid. The Professional variant will use a USB tether that is presently used only for development of the embedded electronics application. The professional variant will also be accompanied with control software to be run on a host workstation PC. Typically, the professional occupational health therapist will tailor the method to encourage recommended hand stroke sequences in their remedial training work with learners. Finally, there is also a possibility of embedding the device in a stuffed animal for preschooler's to role-play with as handwriting teachers.
Because of its real-time responsiveness, this device and recognition technique can potentially be deployed in sign language communication. It would be interesting to see how existing sign languages apply the device. Furthermore, the device provides new opportunity for phonetically mapping speech. Because of its culturally neutral reliance on geometry, a noble objective would be to come up with a universal geometric mapping of phonetics so that it can be used in active cross-cultural sign communication. Dr Khomo, is pleased to support this kind of research that can, at a time of the rapid loss of exotic cultures, help capture cultural folklore into writing, as-is.
Ironically, the Professional variant is nearest to market simply because it resembles the prototype development system that interacts with a personal computer via a high-speed USB interface. Dr Khomo can presently fill small custom orders for such a system since the control software must be tailored to the professional's requirements, and the electronics are assembled by hand. (See end of article for contact details)
Next in availability is the Learner Device which still requires one more round of redesign to allow for volume production at lower cost in order to reach the educational market. This product is one year away from the market.
Following the success of the Learner product, a truly low-cost retail for pre-schoolers should become viable in the foreseeable future.
More information:
Malome Khomo holds a PhD. in Systems Science from City University, and was born of South African parents in Tanzania.
Email address: mkhomo@dslextreme.com
Copyright 2002, Science in Africa, Science magazine for Africa CC. All Rights Reserved
