Image analysis – Applications – Personnel identification
Reexamination Certificate
1999-04-21
2003-07-01
Mehta, Bhavesh M. (Department: 2625)
Image analysis
Applications
Personnel identification
C382S116000, C382S160000, C382S187000, C382S209000
Reexamination Certificate
active
06587577
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates generally to identity verification, and specifically to on-line determination of whether an identifying sample, such as a signature, belongs to the same owner as a plurality of reference samples.
BACKGROUND OF THE INVENTION
A large amount of work has been performed regarding identity verification. In general the problem attempted to be solved in identity verification is: given a plurality of identifying reference samples and a test sample, does the test sample belong to the owner of the reference samples? The samples may be signatures, voice samples, face pictures, etc.
Due to the complexity of the human handwriting and the difference between pluralities of authentic signatures made by a given person, signature verification is very complex. The success of a solution is generally classified according to two variables: the percentage of forgeries which were identified as authentic (false positive), and the percentage of authentic signatures identified as forgeries (false negative).
Various features of signatures are used in signature verification. These features include the geometrical shape of the signature, the speed at which the signature was signed, the number of times the pen was lifted, the pressure of the pen on a pad on which the signature was signed, etc. None of these features are generally distinctive enough so that all signatures of a single individual will have the same value of the feature. Therefore, a plurality of reference signatures are taken from the individual, and the test signature is verified against the plurality of reference signatures.
“Progress in Automatic signature Verification” edited by R. Plamondon, which is incorporated herein by reference, describes various methods of signature verification. A first method includes generating an average signature from the reference signatures and comparing the test signature with the average signature. Such a method is also described for example, in U.S. Pat. No. 4,040,010, which is incorporated herein by reference. In this method, data descriptive of the reference signatures is lost due to the averaging.
In a second method, described, for example, in U.S. Pat. No. 4,724,542, which is incorporated herein by reference, the test signature is compared to each of the plurality of samples and is considered authentic if it is close enough to one of the samples. In this method a large number of reference signatures are required in order to achieve suitable results.
U.S. Pat. No. 5,111,512, which is incorporated herein by reference, describes a method in which a first reference signature is chosen from a plurality of samples, and a second reference signature is constructed by averaging the samples. Some features of the test signature are compared to the first reference signature, while others are compared to the second reference signature. However, this method still does not overcome the problem of information loss due to the averaging.
U.S. Pat. No. 5,680,470, which is incorporated herein by reference, describes a method in which the features of reference signatures are stored in an associative memory. A test signature is compared to the reference signatures using the associative memory. However, the features of the reference signatures may be so diverse as not to allow the associative memory to generate internal rules governing the differences between true and false signatures. For better results, it is desirable to more closely define for the associative memory the points of comparison between the reference signatures and the test signature.
SUMMARY OF THE INVENTION
It is an object of some aspects of the present invention to provide methods for determination whether a given test sample belongs to the same person as a plurality of reference samples.
It is another object of some aspects of the present invention to provide methods for reliable determination of whether a given test sample belongs to a given person using a relatively low number of reference samples.
It is another object of some aspects of the present invention to provide a relatively simple method and apparatus for determining whether a signature is authentic.
It is another object of some aspects of the present invention to provide methods and apparatus for fast determination of whether a signature is authentic.
In preferred embodiments of the present invention, a processor receives a plurality of reference samples and a test sample, for example, samples of a person's signature, and calculates a correlation value between each pair of the samples for each of one or more features of the samples which are identified by or input to the processor. For each of the one or more features, a matrix of the correlation values is formed, and the processor evaluates the regularity of the matrix. The test sample is labeled as authentic if the matrix has a predetermined measure of regularity. In other words, the test sample is identified as authentic if the regularity of its correlation with the reference samples is similar to the regularity of the correlation of the reference samples with one another. Thus, unlike methods of signature identification known in the art, the method used in preferred embodiments of the present invention identifies the test sample as authentic not if it is closely correlated with one of the reference samples, but rather if its degree of correlation (and deviation from full correlation) with the reference samples is comparable to the degree of correlation between the reference samples.
In some preferred embodiments of the present invention, the regularity of the matrix is evaluated by forming a set of points in an n-Euclidean space (n being the number of reference samples) which is representative of the matrix, and evaluating the regularity of the points of the set. Preferably, the points of the set are generated such that the distances between the points are the correlation values of the matrix. Preferably, the set includes n points representative of the reference samples and one test point representative of the test sample.
In some preferred embodiments of the present invention, the regularity of the points of the set is evaluated by evaluating the regularity of a test polyhedral formed of the points. Preferably, in order to establish authenticity of the test sample, the test polyhedral must have a shape resembling a reference polyhedral formed from the points corresponding to the reference samples. Preferably, one or more parameters representative of the relation between the test point and the test polyhedral are calculated. The one or more parameters are preferably normalized by dividing them by an average of the values of the parameters calculated for each of the reference points corresponding to the reference samples. The normalized parameters of each set of points representing a particular feature are used to determine whether the test sample is authentic.
In some preferred embodiments of the present invention, the one or more parameters comprise two parameters: a height of the polyhedral, and a variance of the test point relative to the polyhedral.
In some preferred embodiments of the present invention, the determination whether the sample is authentic is performed by providing the one or more parameters of each of the one or more features to a suitably trained neural network. Preferably, the neural network comprises a Boltzmann Perceptron classifier, although other networks may be used. Alternatively or additionally, other classifiers may be used instead of neural networks, such as a decision tree (for example, an ID3-type tree), as is known in the art.
In some preferred embodiments of the present invention, the samples comprise signatures of a human individual. Preferably, the features of the signatures comprise features indicative of the correlation between two signatures. Therefore, the correlation values of the matrix are given by the values of the features. Preferably, the features include an absolute speed mismatch, an average speed mismatch, a uniformity of point-
Darby & Darby
Desire Gregory
International Business Machines - Corporation
Mehta Bhavesh M.
LandOfFree
On-line signature verification does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with On-line signature verification, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and On-line signature verification will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3054267