Allograph based writer adaptation for handwritten character recognition

1. A system that facilitates analyzing handwriting, comprising:

an interface component that receives at least one handwritten character;

a classifier comprising an allograph neural network trained using allograph data relating to at least one handwriting style and a base neural network trained using non-allograph data;

a combine component that combines outputs from the allograph neural network and the base neural network; and

a personalization component that employs the combined outputs of the classifier to provide handwriting recognition for the at least one handwritten character.

2. The system of claim 1 , further comprising an allograph component that generates the allograph data used to train the allograph neural network.

3. The system of claim 2 , the allograph component automatically generates the allograph data utilizing a clustering technique that groups handwritten characters into clusters, each cluster representing a style of a handwritten character.

4. The system of claim 3 , the clustering technique is a hierarchical agglomerative clustering approach utilizing dynamic time warping as a distance measure.

5. The system of claim 2 , the results of the clustering technique are visualized using at least one of a binary tree or a dissimilarity dendogram.

6. The system of claim 1 , wherein the allograph neural network and base neural network utilize a polynomial feature technique to provide inputs thereto.

7. The system of claim 1 , wherein the allograph neural network is trained to accept feature vectors for handwritten characters as inputs and to map the feature vectors to character allographs.

8. The system of claim 7 , the allograph neural network utilizes at least one of a simple folder, a linear folder, or an allograph folder to map the character allographs to character classes.

9. The system of claim 8 , wherein the at least one of a simple folder, a linear folder, or an allograph folder is trained using gradient descent to map the character allographs to the character classes.

10. The system of claim 7 , wherein the feature vectors are derived for a handwritten character by segmenting the handwritten character and representing each resulting segment in the form of a Chebyshev polynomial.

11. The system of claim 1 , wherein the base neural network is trained to accept feature vectors for handwritten characters as inputs and to map the feature vectors to character classes.

12. The system of claim 1 , the combine component employs at least one of a linear combiner or a linear classifier.

13. The system of claim 1 , the combine component employs a combiner classifier that learns from data.

14. The system of claim 13 , the combiner classifier is a support vector machine.

15. The system of claim 14 , the support vector machine learns to optimally combine the allograph neural network output and the base neural network output utilizing a handwriting sample from a user.

16. The system of claim 1 , the personalization component infers the handwritten character taking in account a deterioration of quality due to fatigue.

17. The system of claim 1 , the allograph data is based at least in part upon at least one of the following: a geographic region, a school district, a language, or a style of writing.

18. A method that facilitates handwriting recognition, comprising:

employing a processor having computer-executable instructions embodied on a computer-readable storage medium to perform the following acts:

generating allograph data;

training a first classifier using the allograph data;

training a second classifier using non-allograph data;

receiving a handwritten character;

deriving a feature vector for the handwritten character;

employing the first classifier to map the feature vector to character allographs and to fold the character allographs into their associated characters;

employing the second classifier to map the feature vector directly to one or more specific characters;

combining outputs from the first and second classifiers; and

employing the combined outputs to provide handwriting recognition for the handwritten character.

19. The method of claim 18 , further comprising:

combining the outputs of the first and second classifier utilizing at least one of a linear combiner, a personalizer, a support vector machine (SVM), or a combiner classifier.

20. A machine-implemented system that facilitates analyzing handwriting, comprising:

means for receiving at least one handwritten character;

means for training a first recognizer using allograph data;

means for training a second recognizer using non-allograph data;

means for combining outputs of the first and second recognizers; and

means for employing the combined outputs to provide handwriting recognition for the at least one handwritten character.