JPH103543A - Method for collating signature - Google Patents

Abstract

PROBLEM TO BE SOLVED: To attain accurate and real time collation by inputting the data of a horizontal direction component in the writing speed of a signature to be collated to an FIR filter constructed using horizontal and vertical direction data in the writing speed of a reference signature and collating the inputted data, based on output data from the FIR filter. SOLUTION: The data of horizontal and vertical direction components in the writing speed of a signature are inputted to a digitizer by a pen. The self- correlation of the horizontal and vertical direction components in the writing speed outputted from the digitizer is found out, the influence of a deviation on a write starting position which is generated in each writing is removed, the information volume of the self-correlation is compressed and a discrete cosine conversion coefficient is found out by executing discrete cosine conversion to speed up the collation. The FIR filter is constituted so as to have the discrete cosine conversion coefficient as an I/O characteristic and the signature is collated by using the FIR filter.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for collating a signature made by an unknown writer with a previously recorded reference signature.

[0002]

2. Description of the Related Art Hitherto, a technique of extracting characteristics of information using a computer and collating it with a unique pattern registered in advance, that is, a technique of so-called pattern recognition has been advanced, and is also used for signature collation.

[0003] Such signature verification means uses only handwriting as an information source without using information on the writing process, and further includes, for example, pen coordinates,
A document using information on a writing process such as writing pressure is presented.

[0004]

However, most of the conventional signature verification methods are not accurate enough or difficult to verify in real time.

The present invention has been made in view of such circumstances, and it is needless to say that an accurate verification can be performed, and a signature verification method capable of performing verification in real time is provided.

[0006]

In order to solve the above-mentioned problems, a signature verification method according to the present invention comprises the steps of: writing horizontal component data of a writing speed in a signature to be verified; It is characterized in that it is input to an FIR filter constructed using directional data, and collation is performed based on the output data.

[0007]

Next, a preferred embodiment of the present invention will be described.

In the embodiment of the present invention, the data of the horizontal and vertical components of the writing speed for the signature is inputted by a digitizer (or tablet) using a pen.

Then, the autocorrelation function of the horizontal component and the vertical component of the writing speed output from the digitizer is obtained, and the influence of the shift of the writing start position that occurs every time writing is performed is eliminated. Discrete cosine transform (hereinafter, referred to as “DCT”) is performed to compress the amount of information of the function to speed up the matching, and a discrete cosine transform coefficient (hereinafter, referred to as “DCT coefficient”) is obtained.
This DCT coefficient was used as data for collation.

More specifically, the data output from the digitizer is a time-series pattern composed of coordinate data in the horizontal and vertical directions of the pen and up / down information of the pen, and the horizontal and vertical components are the digitizer. It is indicated by the coordinate values above.

On the other hand, the pen up / down information outputs a data value 0 at the moment when the pen tip touches the digitizer, and subsequently, a state in which the pen touches (down state), that is, data 1 in the writing process. In the state where the pen tip is not in contact with the digitizer (up state), nothing is output including the components in the horizontal and vertical directions.

Here, a sequence of sample values until the pen up / down information 0 is output from the digitizer and immediately before the next 0 is output is called a stroke.

As described above, the signature data from the digitizer is represented by a horizontal component x (n), a vertical component y (n), and pen up / down information p (n). However,
n indicates a sample number of the nth day on the digitizer sampled at the sampling interval T.

In general, the size of a signature differs every time it is written. Therefore, processing is performed using the following equation,
Normalization is performed so that the maximum amplitude of the horizontal component x (t) and the vertical component y (t) of each signature stroke is 1.

[0015]

(Equation 1)

Furthermore, the writing time for writing a signature is generally different every time it is written, and the writing time is normalized for each stroke as follows.

First, a continuous time function obtained by interpolating the sample points for each stroke is defined as follows.

[0018]

(Equation 2)

Then, the number of samples per stroke is normalized to the sampling interval as in the following equation.

[0020]

(Equation 3)

Next, each of the horizontal component (hereinafter referred to as "x (n)") and the vertical component (hereinafter referred to as "y (n)") of the signature is normalized for each stroke. Writing speeds V _x (n) and V _y (n) are defined as follows.

[0022]

(Equation 4)

FIGS. 2 and 3 show examples of the signature data and the writing speed of the actual signature "Kubo" shown in FIG.

[0024] Further, each time a signature is written, there is a shift in the writing start position. Therefore, in order to remove the influence of the positional shift, the writing speeds V _x (n) and V _y (n) are calculated by the following equations. Calculate the correlation function.

[0025]

(Equation 5)

Next, the maximum value of the autocorrelation function obtained by the following equation is normalized to 1.

[0027]

(Equation 6)

FIG. 4 shows an example of the autocorrelation function, FIG. 5 shows several examples of the writing speed of a signature written by the same writer, and FIG.
5 shows the autocorrelation function of the writing speed shown in FIG.

Further, in order to perform the comparison quickly, it is desirable to reduce the feature amount to be processed. For this purpose, a low-order DCT is performed by the following equation to obtain a DCT coefficient.

[0030]

(Equation 7)

FIG. 7 shows the DCT coefficients for the data shown in FIG.

In the embodiment of the present invention, the characteristics (DCT coefficients) of the horizontal component and the vertical component of the writing speed are extracted as follows.

Next, the DCT coefficients Ψ _vx (n), Ψ
A signature matching algorithm using an FIR filter expressing a writing speed feature using _vy (n) will be described.

Here, a filter expressing the characteristics of the writing speed is defined as a writing speed filter (hereinafter referred to as "WVF").

It is assumed that this WVF can be approximated by a suitable order FIR filter.

Here, the input and output of the WVF are obtained by dividing the DCT coefficient Ψ of the autocorrelation function of the horizontal and vertical components of the signature data divided for each stroke.
Consider the case where _vx (n), _Ψvy (n).

Next, a method of determining the impaired response of the FIR filter approximating WVF will be described.

FIG. 8 shows an approximate model of a WVF FIR filter.

Here, the output Ψ _vy (n) of the _SIR stroke FIR filter is given by the following equation.

[0040]

(Equation 8)

At this time, the impulse response of the FIR filter is expressed by the following equation.

[0042]

(Equation 9)

Further, the impulse response of the FIR filter is determined so that E shown in the following equation becomes minimum.

[0044]

(Equation 10)

That is, the optimum impulse response is given by the following equation.

[0046]

[Equation 11]

Next, FIG. 9 shows a signature collation algorithm using an FIR filter expressing the writing speed characteristic obtained as described above.

The DCT obtained from the writing speed of the reference signature
The coefficients are Ψ _vx (n) and Ψ _vy (n). It is also assumed that an FIR filter having the writing speed characteristic of the reference signature data has already been obtained.

[0049]

Next, an embodiment of the present invention will be described.

In this embodiment, as shown in FIG. 10, a digitizer (tablet) [transfer rate;
9,600b 5 data / s baud rate;
it / s] on a personal computer (NECPC-9801BX2, CP
U; 80486SX-20M) and signature-verified using the algorithm shown in FIG.

It should be noted that at the time of implementation, the following Table 1 was randomly extracted from handwritten Kanji signatures obtained by inputting data of 10 signatures per person from the tablet once a week for about one year. The signature data shown in FIG. 11 was used.

[0052]

[Table 1]

In the present embodiment, the training signature data is extracted at random from the signature data of the person and the fake signature data, and 20 pieces of data are extracted for each signature data of the person in the signature data for training. From the autocorrelation function of velocity, D
CT coefficients Ψ _vx (n) and Ψ _vy (n) were determined.

Further, their DCT coefficients Ψ _vx (n), Ψ
FI configured to have _vy (n) as input / output characteristics
After the impulse response of the R filter was determined for each signature data, the average of those impulse responses was used as the WVF impulse response of the reference signature data.

Furthermore, the sample points of the data for each stroke of the signature are normalized to 30, and the approximation 2 of the FIR filter is performed.
The approximate order of the FIR filter was determined so that the power error was 1/10 ² or less.

The signature of the person (true signature) on the result of the signature verification under the above conditions is replaced by the signature of another person (false signature).
FIG. 12 shows the first-type erroneous recognition rate in which a false signature is erroneously recognized and the second erroneous recognition rate in which a false signature is erroneously recognized as a true signature.

Table 2 shows the first-class misrecognition rate and the second-class misrecognition rate when the threshold value is appropriately determined.

[0058]

[Table 2]

According to the above results, the erroneous recognition rate is extremely low, and the autocorrelation function of the writing speed of the signature data is subjected to discrete cosine transform to sufficiently express the writing characteristics of the signature data in which the number of feature parameters is reduced. It turns out that you can.

In the present embodiment, the collation time per signature data (including the characteristic display for each stroke) is 20 to
It was about 30 seconds, and it was also found that if only the collation result was output, the writer could be sufficiently identified in real time even with the simple system used in this embodiment.

[Brief description of the drawings]

FIG. 1 is an example of an actual signature.

FIG. 2 is a relationship diagram showing data of a horizontal component and a vertical component of a signature.

FIG. 3 is a relationship diagram showing velocity data of a horizontal component and a vertical component of a signature.

FIG. 4 is an autocorrelation diagram.

FIG. 5 is a relationship diagram illustrating an example of a signature writing speed.

6 is an autocorrelation diagram of the writing speed shown in FIG.

FIG. 7 is a relationship diagram showing DCT coefficients for the data shown in FIG.

FIG. 8 is an explanatory diagram showing an approximate model of a WVF FIR filter.

FIG. 9 is a process chart showing a signature collation algorithm.

FIG. 10 is a schematic view showing an apparatus according to an embodiment of the present invention.

FIG. 11 shows signature data of a formal signature and a fake signature in the embodiment of the present invention.

FIG. 12 is a relationship diagram showing an erroneous recognition rate in signature verification.

 ────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Kimio Tanaka 8-7-15 Seijo, Setagaya-ku, Tokyo

Claims (1)

[Claims] 1. A horizontal component data of a writing speed of a signature to be verified is input to an FIR filter constructed using horizontal and vertical data of a writing speed of a reference signature, and based on output data thereof. Signature verification method characterized by performing verification by using a signature.