A statistical mixture approach to automatic handwriting recognition

Claim 2

In a handwriting recognition system, the combination comprising: means for sampling known handwriting inputs from at least one writer; means for providing a parameter vector representation of said handwriting inputs; means for providing a spliced vector representation of said handwriting inputs; and means for providing a feature vector representation of said handwriting inputs.

Claim 3

The combination claimed in claim 2, including: means for performing clustering in a feature vector space to provide prototype distributions in said feature vector space; means for performing Gaussian modeling in said feature vector space; means for computing mixture coefficients to evaluate the relative contribution of each prototype distribution to the current sample handwriting inputs.

Claim 4

The combination claimed in claim 2 or 3, including: means for sampling an unknown handwriting input from the same or another writer; means for providing a feature vector representation of said unknown handwriting input; means for estimating the likelihood that the feature vector representation corresponds to each character in the underlying alphabet, utilizing said chirographic prototype distributions and said mixture coefficients; said combination preferably further comprising: means for performing mixture decoding in said feature vector space to produce recognized handwriting; means for utilizing multiple feature vector spaces for producing a more robust recognition.

Claim 5

In a handwriting recognition system, the combination comprising: an electronic tablet which can be written on with a stylus, and on which writing can be displayed; means for sampling known characters comprising handwriting inputs from at least one writer, in response to said one writer writing on said tablet with said stylus; means for providing a feature vector representation in chirographic space of the sampled known characters; means for providing chirographic prototype distributions and mixture coefficients in response to the provision of the feature vector representation in chirographic space of the sampled known characters; means for sampling unknown characters comprising handwriting inputs to be recognized from the same or other writers, in response to said writers writing on said tablet with said stylus; means for providing a feature vector representation in said chirographic space of the sampled unknown characters; means for estimating the likelihood that at least one character of the underlying alphabet is a candidate character for the recognition of each unknown character, based on a probabilistic comparison of said chirographic prototype distributions, appropriately weighted by said mixture coefficients, and said parameter vector representation in chirographic space of the sampled unknown characters, including means for making a list of candidate characters; and means for performing a probabilistic analysis of said list of candidate characters to recognize the sampled unknown characters comprising the handwriting inputs to be recognized; said combination preferably further comprising: means for displaying the recognized characters on said electronic tablet and/or means for providing a library of probable combinations of at least characters occurring in handwriting; and wherein said means for performing a probabilistic analysis includes means for making a probabilistic comparison of said list of candidate characters and said library of probable combinations of at least characters occurring in handwriting, to recognize the sampled unknown characters comprising the handwriting inputs to be recognized.

Claim 6

In a handwriting recognition system, the combination comprising: an electronic tablet which can be written on with a stylus, and on which writing can be displayed; means for sampling handwriting inputs from at least one writer, in response to said one writer writing on said tablet with said stylus; means for providing a parameter vector representation of said handwriting inputs; means for providing a spliced vector representation of said handwriting inputs; means for computing a covariance matrix of said spliced vector representation; means for calculating eigenvalues and eigenvectors associated with the computed covariance matrix; means for applying a linear transformation along the direction of the leading eigenvectors, this providing a feature vector representative of said handwriting inputs; said combination preferably further comprising: means for performing clustering in a feature vector space to provide prototype distributions in said feature vector space; means for performing Gaussian modeling in said feature vector space; means for computing mixture coefficients to evaluate the relative contribution of each prototype distribution to the current sample handwriting inputs and/or means for recognizing unknown handwriting in response to a probabilistic comparison of the unknown handwriting in chirographic space with said chirographic prototypes, as combined in accordance with the mixture coefficients.

Claim 7

In a handwriting recognition system, a method comprising the steps of: sampling handwriting inputs from at least one writer; signal processing said inputs to extract both local and global feature elements; windowing said inputs to detect broad feature events; rotating and projecting said inputs to eliminate redundancy; performing clustering in a feature vector space to provide prototype distributions in said feature vector space; performing Gaussian modeling in said feature vector space; computing mixture coefficients to evaluate to relative contribution of each prototype distribution to the current sample handwriting inputs; and performing mixture decoding in said feature vector space to produce recognized handwriting.

Claim 8

In a handwriting recognition system, a method comprising the steps of: sampling known handwriting inputs from at least one writer; providing a parameter vector representation of said handwriting inputs; providing a spliced vector representation of said handwriting inputs; and providing a feature vector representation of said handwriting inputs; said method preferably further comprising the steps of: performing clustering in a feature vector space to provide prototype distributions in said feature vector space; performing Gaussian modeling in said feature vector space; computing mixture coefficients to evaluate the relative contribution of each prototype distribution to the current sample handwriting inputs and/or sampling an unknown handwriting input from the same or another writer; providing a feature vector representation of said unknown handwriting input; estimating the likelihood that the feature vector representation corresponds to each character in the underlying alphabet, utilizing said chirographic prototype distributions and said mixture coefficients and/or performing mixture decoding in said feature vector space to produce recognized handwriting; utilizing multiple feature vector spaces for producing a more robust recognition.

Claim 9

In a handwriting recognition system, a method comprising the steps of: sampling known characters comprising handwriting inputs from at least one writer, in response to said one writer, in response to said one writer writing on an electronic tablet with a stylus; providing a parameter vector representation in chirographic space of the sampled know characters; providing chirographic prototypes in response to the provision of the parameter vector representation in chirographic space of the sampled known characters; sampling unknown characters comprising handwriting inputs to be recognized from said at least one writer, in response to said one writer writing on said tablet with said stylus; providing a parameter vector representation in said chirographic space of the sampled unknown characters; estimating the likelihood that at least one of said chirographic prototypes is a candidate character for being recognized as an unknown character, based on a probabilistic comparison of said chirographic prototypes, and said parameter vector representation in chirographic space of the sampled unknown characters, including making a list of candidate characters; and performing a probabilistic analysis of said list of candidate characters to recognize the sampled unknown characters comprising the handwriting inputs to be recognized; said method preferably further comprising the steps of: displaying the recognized characters on said electronic tablet and/or providing a library of probable combinations of at least characters occurring in handwriting; and wherein performing a probabilistic analysis includes making a probabilistic comparison of said list of candidate characters and said library of probable combinations of at least characters occurring in handwriting to recognize the sampled unknown characters comprising the handwriting inputs to be recognized.

Claim 10

In a handwriting recognition system, a method comprising the steps of: sampling handwriting inputs from at least one writer, in response to said one writer writing on an electronic tablet with a stylus; computing a covariance matrix of the sampled handwriting; calculating eigenvectors associated with the computed covariance matrix; applying a linear transformation along the direction of the eigenvector; and producing chirographic prototypes of the sampled handwriting in response to the application of the linear transformation; said method preferably further comprising the steps of: providing mixture coefficients indicative of how to combine the chirographic prototypes for handwriting recognition and/or recognizing unknown handwriting in response to comparing the unknown handwriting to the chirographic prototypes as combined in accordance with the mixture coefficients.