JPH1083473A - Method for recognizing paper money - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for precisely and quickly identifying the kind of paper without being influenced by stain, strain or the like. SOLUTION: The picture data of a recognized paper money is first fetched using a light having plural different wavelengths (S2). Then, it is roughly judged whether the recognized paper money is a true note or a false note from this picture data (S3). Then, an identification error is predicted from the picture data of the recognized paper money thus judged to be a true note, and at the same time, it is precisely judged whether the identified paper money is a true note or a false note using this identification error (S4). Then, highly accurate judgment is performed for the picture data thus judged to be a true note using a neural net (S5).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for identifying paper sheets such as banknotes and securities, and more particularly to a method for identifying highly accurate and high-speed judgments by suppressing the influence of various stains on the identified banknotes. About possible identification methods.

[0002]

2. Description of the Related Art Japanese Patent Laid-Open No.
-215293. In this publication, a bill is divided into a plurality of zones, detection data for each zone is compared with reference data predetermined for each zone, and the bill is identified based on a comparison result in each zone. In the bill discriminating method, a plurality of bills are set in accordance with the front and back, the orientation, and the displacement at the time of discrimination, and the data of each zone are totaled for one bill, and a standard value is determined by a ratio value to the total value. It is stored as pattern data, a total value of the detection data is obtained, and a ratio value to the total value is calculated as detection pattern data, and whether or not the detection pattern data is within an allowable value range of the reference pattern data is determined. Is determined, and the absolute value of the difference between the reference pattern data and the detected pattern data is calculated for each of the zones, and the total is calculated. The total value banknote identification method and performing to the bill validator determines whether less than the allowable value is disclosed.

[0003]

However, in the above-mentioned conventional technology, there is a problem that a fake bill is erroneously recognized as a genuine bill because variations in identification due to dirt, distortion and other reasons are allowed. Yes, this causes a reduction in identification accuracy.

The present invention has been made to solve the problems of the conventional method, and a method for accurately and quickly identifying bills without being affected by dirt, distortion and the like. The purpose is to provide.

[0005]

According to the present invention, there is provided a method for determining a genuine bill or a fake bill from image data of a bill to be identified, which is captured by using a plurality of light beams having different wavelengths. A precise determination step of predicting an identification error from the image data of the bill to be identified determined as a genuine bill in the step and determining whether the bill is genuine or fake using the identification error; And a high-precision determination step of performing a high-precision determination on the image data determined to be using a neural network.

[0006]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a perspective view of a bill discriminating method according to an embodiment of the present invention;

FIG. 1 is a flowchart showing the basic algorithm of the present embodiment. In the figure, when a program is started by inserting a bill (step S1), multi-value density data by an optical sensor of the bill is captured (step S2).

A part of the captured density data is subjected to matching with three denominations and reference waveforms in four directions, and the denomination and the input direction are determined using the matching degree obtained as a result. .

Using the determination result thus obtained, the grayscale data is converted into data in the front direction and in the erect direction.
Next, a rough determination process in step S3 is performed. This process irradiates the identified bill with, for example, infrared light as the second wavelength light and red light as the first wavelength light to obtain image data, respectively, and calculates the difference or ratio of the transmittance of these image data. This is a process of judging a genuine bill or a fake bill by comparing these calculated values with a predetermined threshold value.

The bill to be identified, which has been roughly determined to be a counterfeit note in step S3 through the above processing, enters a fine determination processing routine in step S4. In this routine, a transport deviation correction process, a level matching process, an irregular component extraction process, an irregular component prediction process, a prediction error calculation process, a prediction error threshold value determination process, and the like are performed.

Using this result, the authenticity of the bill to be identified which was previously determined to be genuine is determined. If it is determined to be a counterfeit note, this is determined. If it is determined to be a genuine note, the process proceeds to a high-precision determination processing routine in step S5. In this routine, identification processing using a neural network (hereinafter abbreviated as NN) is performed.

Next, details of the flowchart of FIG. 1 will be described. The determination processing for three denominations and four directions is performed, for example, as follows. That is, as shown in the flowchart of FIG. 2 and the conceptual diagram of FIG. 3, when the bill to be identified is inserted into the bill validator (not shown) in step S20, LE is input.
The image sensor (line sensor) comprising a light-emitting diode (D) that emits light of the first and second wavelengths selected from red and infrared colors and the like and a light-receiving element converts the image data of the bill surface to the vertical axis. Is obtained in the form of two waveform data in which is the transmittance (sensor value) and the horizontal axis is the position information (point) (see FIG. 3A).

Since the number of data points used in the calculation in the denomination / direction determination is only used for the denomination and direction determination, it does not need to be all the points of the banknote.
The number of points required for the direction determination may be used.

In step S21, of the obtained waveform data (banknote data) based on the second wavelength light,
Three denominations and four directions previously obtained from the genuine bill by the image sensor (FIG. 3 shows front right A, front left B, back right C and back left D in the case of one denomination and four directions) (Step S22), and the minimum required sum of squares of the differences for each point is calculated (FIG. 3).
b).

Next, the obtained sums of squared differences (matching degrees) in each direction are compared, and in step S23, the direction of the direction-specific reference waveform data having the minimum value is determined as the bill insertion direction (FIG. 3c). reference).

In the rough determination process in step S3, the difference or ratio of transmittance between the image data obtained using the second wavelength light and the image data obtained using the second wavelength light is calculated by the image sensor. Then, the calculated value is compared with a threshold value obtained in advance, and a genuine bill or a fake bill is roughly determined.

The fineness determination process in step S4 is roughly divided into a process for correcting the transport deviation of the banknote to be identified and a process for authenticity identification. The transport shift processing is performed in step S4 as shown in FIG.
In step 1, the shift width K (for calculation) is set. This shift width K is a value between a minimum shift width value at which a transport shift can occur and a maximum shift width, and starts from the minimum value.

Then, in step S42, data is generated by shifting the input signal of the bill conveyed by K, which is set. Next, in step S43, data at a plurality of target locations is extracted from the data shifted by K generated in step S42.

In step S44, a total absolute value of the difference between the extracted data and the corresponding data of the reference waveform is calculated. If the total value obtained in step S45 is the minimum, the value of K at that time is temporarily recorded as the calculated deviation width.

The above operation is repeated from the minimum value to the maximum value of K. By doing so, every time the value of K changes between the minimum value and the maximum value, the cumulative value of the difference from the reference waveform is obtained, and the minimum cumulative value of K is updated each time. Go.

Then, by setting the finally remaining value of K as the target conveyance deviation width, the conveyance deviation value can be accurately obtained. If more accuracy is required, instead of calculating the absolute value of the difference in step S44, the square of the difference is calculated (step S441).
It is also possible to determine the minimum deviation value as the target deviation width.

The extraction of data at a plurality of locations in step S43 is performed by a method as shown in FIG. That is, in step S431, a basic representative waveform is created from a banknote (closed note) free from dirt and tear.

In step S432, a displacement representative value K is set and a displaced basic representative waveform is created in the same manner as in the previous calculation of the displacement displacement width. In step S433, a difference value (absolute value) for each position between the shifted basic representative waveform and the original basic representative waveform is recorded.

In the above operation, the value of K is changed from the minimum value to the maximum value, and the difference value is sequentially recorded. In step S434, the minimum value of the difference value at each position is calculated. Finally, a plurality of points are selected in ascending order of the smallest difference value obtained in step S435, and this point is set as a plurality of points from which data should be extracted.

A conceptual diagram of the waveform in the middle is shown in the right half of FIG.
The left half shows a conceptual diagram of the steps for selecting a plurality of locations. After the transport deviation correcting process, the level of the captured image of the banknote to be identified is adjusted, and the process proceeds to a process of reducing the identification error due to an irregular component such as a dirt component from the image.

This processing is performed using, for example, an auto (AR) regression model. That is, as shown in FIGS. 6 to 8, the present processing method is largely divided into a pre-processing and a processing at the time of bill insertion, and FIGS. 7 and 8 are flowcharts respectively.

The pre-process is a process of creating a fluctuation component estimating model learning of sensor input data of a banknote. As shown in FIG. 7, first, in step S151, a plurality of genuine bills (new bills) are sensed. Then, sensor signals of the transmittance and density of images and characters on the bill are obtained, and a reference waveform (for example, an average data waveform) as reference data is obtained from each sensor signal.

Next, in step S152, using the reference waveform, fluctuation components such as dirt and distortion are extracted from the genuine note data for each genuine note. Then, the data of the fluctuation component extracted in step S153 is regarded as a periodic time series signal, and an equation as an autoregressive model is used.

[0029]

(Equation 1)

The coefficients a1, a2,..., Ap of the dirt expression at a certain time represented by the following equation are obtained. The learning data in this case is comparable to the time-series signal data of the dirt component when several banknotes are arranged and input.

The periodic time-series signal of the dirt component created in this way is learned by an auto-regression analysis method in step S4, and as a result of the learning, an estimation model of a variation component of dirt for one banknote is created.

After the pre-processing is performed in this manner, the flow proceeds to a processing at the time of insertion for making a true / false determination when a bill is actually inserted. In the bill insertion process, first, a sensor signal from the bill inserted in step S251 is input.

Next, in step S252, the difference between the signal input and the reference waveform is calculated, and the
The distortion component is extracted. In step S253, based on the data of the dirt and distortion components obtained in step S252, dirt estimation using the estimation model is performed by an auto-regression model technique, and a predicted value is calculated.

Here, the estimation method will be described.
Since the estimation model has been obtained by the pre-processing, the dirt component predicted from the estimation model of the autoregressive analysis is calculated from the equation 1 and the input dirt component data of the banknote (step S253).

A prediction error is calculated in step 254 from the stain waveform as a fluctuation component of the input bill obtained in this way and the fluctuation component waveform of the estimation model. If it is within the range of the error, it is determined to be a genuine bill, and otherwise, it is determined to be a fake bill (step S255).

Of the banknotes subjected to the prediction error processing as described above, those that are determined to be counterfeit are determined to be counterfeit, and those that are determined to be genuine are further subjected to the high-precision determination processing in step S5. Proceed to.

The high-precision judgment processing is roughly divided into a pre-processing and a bill insertion processing. In the pre-processing, waveform data of several points are prepared from a plurality of genuine bills and fake bills prepared in advance. The waveforms of a genuine note and a counterfeit note are learned as inputs of the neural network. The outputs of this neural network are true: 1 and false: 2.

In the processing at the time of bill insertion, the waveform data of each point of the bill to be identified is input to the input of the learned neural network obtained in advance, and the higher value of the obtained output is used as the value of the bill to be identified. Is determined as the determination result.

[0039]

As described above, the present invention discriminates a bill to be identified, which has been determined to be true in the rough determination and the fine determination, into a true bill and a fake bill by high-precision determination using a neural network. The effect of accurately identifying counterfeit notes with high difficulty can be expected.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an embodiment of a bill discriminating method according to the present invention.

FIG. 2 is a flowchart of a denomination / direction determination process.

FIG. 3 is a conceptual diagram of a denomination / direction determination process.

FIG. 4 is a flowchart illustrating a conveyance deviation width calculation method.

FIG. 5 is a flowchart illustrating a method of selecting data at a plurality of locations from input data.

FIG. 6 is a conceptual diagram of a prediction process using an irregular component such as a stain.

FIG. 7 is a flowchart of a pre-processing.

FIG. 8 is a flowchart of a bill insertion process.

FIG. 9 is a conceptual diagram illustrating a high-precision determination processing method using a neural network.

Claims (1)

[Claims] 1. A rough judgment step of judging a genuine bill and a fake bill from image data of a banknote to be identified taken by using light of a plurality of different wavelengths, and the genuine bill is determined in the rough judgment step. Precise discrimination from the image data of the bill to be discriminated and a genuine bill using this discrimination error, a fine judgment step of judging a counterfeit note, A bill discriminating method comprising a high-precision determining step of performing a high-precision determination using a neural network.