JPH07220087A - Neuro identification/loss separation device for sheet paper or the like by random masking system - Google Patents

Abstract

PURPOSE:To provide a neuro identification/loss separation device for sheet paper or the like by a random masking system which reduces pattern image data measured with a sensor by using plural column masks and performs the pattern recognition of the sheet paper or the like in a small scale neural network. CONSTITUTION:A slab value (image representative value) reduced by the column mask remains unchanged by the disolaction of a carrier pattern image due to the certain degree of bias of the sheet paper or the like. and it is inputted to a separation arithmetic part (neural network) 152, and a separation arithmetic value at every decision pattern of the pattern image is calculated by neuro weight adjusted at the optimum value in decision pattern classification in advance. The pattern image can be judged according to the maximum value of the separation arithmetic value. In this way, it is possible to make the constitution of the neural network into a small scale and also, to make a controller into the same scale.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device capable of recognizing patterns of paper sheets such as banknotes and gift certificates by a neural network, and recognizing a damage level due to stains and tears of the paper sheets. By randomly covering the paper sheet mask with a column mask that is vertically long in the paper sheet transport direction, the pattern image used for identification is reduced and efficient small-scale separation calculation is performed. The present invention relates to a sheet neuro-identification / correction-separation device by a random mask method for determining and calculating a loss level.

[0002]

2. Description of the Related Art Conventionally, a neural network for recognizing patterns such as currency and characters is as follows.
There is one as shown in FIG. The recognition device of FIG. 27 is
An input image 3 is obtained by reading the characters or the like of the identification target 1 with the image sensor 2. In the case of the input image 3 in FIG. 27, since it is divided into eight vertically and horizontally, D (i.j) (where i = 1 to 8 and j = 1
8 to 64) as input information of 64 pixels. And
Each of the 64 pixel data is directly input to the separation calculation unit 5. At this time, the number of neuro elements in the input layer of the separation calculation unit 5 is required to be 64 corresponding to each pixel. Then, a pattern is recognized using a neural network based on this input information.

On the other hand, the conventional separation of normal and defective sheets is broken,
A technique for ranking and separating perforations, stains, etc., and finally integrating these results to determine the damage level of paper sheets. In particular, the basis of the separation technology is to set a threshold value based on the information obtained from each sensor, and mainly to separate information with low damage level and information with high damage level, and select the level. Was left to the user. In the neural network, the sigmoid function is approximated by a simple line graph. This has a large error, and there is a problem that an accurate neuro output value cannot be obtained depending on the input. Then, the values of the sigmoid function are discretely referred to as a table at the time of recognition, and when a detailed sigmoid output value is required, linear interpolation calculation is performed and used. Further, learning of the weight of the neural network is not separated from the product, so there is a problem in terms of security. That is, there was a risk that the learning method and learning data would be leaked to a third party. For this reason, the banknotes of the country in which the product is used were procured from the field, the banknotes were actually flowed and learned, and the learning results were stored in the product, but this took a lot of time and money. .

[0004]

However, in the conventional recognition apparatus as described above, since the number of pieces of information to be input to the separation calculation unit 5 is the same as the number of pixels of the input image 3, the neuro element of the input layer also has pixels. You will need as many as you have. Therefore, it is necessary to provide an input layer having a large number of neuro elements, and correspondingly, it is necessary to provide a large number of neuro elements in the hidden layer as well. In the end, pattern recognition is required unless a very large-scale separation operation unit 5 is used. Couldn't do.

Although edge extraction and binarization can be considered as image processing, it is necessary to increase the number of pixels and measure the banknote image finely in order to perform the processing technique for extracting the characteristics of the banknote, and the processing speed. It was also difficult in terms of.
Further, in recent years, a neural network that models the information transmission of a biological neural network has excellent characteristics such as information interpolation, learning algorithm self-organization, and parallel processing, and is suitable for various pattern recognition. In particular, it is said that the self-organization of the algorithm by learning can reduce the task of searching the characteristic parameters of banknotes by the conventional experience, and the banknote validator using the neural network is being developed in recent years. In particular, it has been desired to construct a neural network of a suitable scale for discriminating a plurality of denominations having the same bill width, bill length and hue such as US dollar bills.

By the way, the sensitivity of a person intervenes in the normal separation of paper sheets, and the separation level differs depending on the user of the apparatus. Therefore, it is necessary to set several damage levels for the paper sheets. Further, since the damage level is detected separately for each of the above-mentioned items, the final judgment obtained by integrating these may not necessarily match the actual damage level of the actual item.
For these reasons, there has been a demand for a device that can set the damage level for each user and at any time so as to meet small demands. Further, in the neuro-arithmetic device, it is impossible to make the computer calculate the sigmoid function itself due to the processing time constraint, so the above-mentioned approximate value was used, but this has a large error. Yes, and in some cases there was a risk of making an erroneous decision. Further, there has been a demand for the advent of a device that can easily reset the damage level of a paper sheet that meets the customer's request, for example, at a sales office or the like in charge of maintenance of the user near the user.

The present invention has been made under the circumstances as described above, and an object of the present invention is to ensure even if the number of input information (hereinafter referred to as slab number) input to the input layer of the separation calculation unit is reduced. In addition to realizing the pattern recognition, the separation calculation unit can be downsized, and the same size, the same hue, and the similar pattern can be distinguished between denominations such as US dollar bills by imitating the neural network of organisms. An object of the present invention is to provide a neural discriminating device for paper sheets by a random mask method realized by a neural network and a random mask covering a part of an image. Further, the object of the present invention is to enable the setting of the damage level of the paper sheets to meet the detailed requirements for each user, and there is no problem in terms of security. An object of the present invention is to provide a damage / separation device for paper sheets, in which the damage level can be easily reset.

[0008]

SUMMARY OF THE INVENTION The present invention relates to a neuro-identifying device for paper sheets by a random mask method,
The above object of the present invention is stored in the image input means for inputting an optical pattern image of a paper sheet to be identified, the storage means for storing the image data read by the image input means, and the storage means. Image extracting means for extracting the optical pattern image of the paper sheet from the image data, and a plurality of types of columns that are randomly parallel to the transport direction of the paper sheet with respect to the optical pattern image extracted by the image extracting means. A pre-processing unit for applying a mask and converting into a total sum (slab value) of non-masked pixels in each of the optical pattern images, and inputting the slab value, and preliminarily adjusting the neuro patterns by the neuro weights optimally adjusted to the judgment pattern classification. A separation calculation unit of a neural network that calculates a separation calculation value for each judgment pattern of an image, and a judgment pattern having the maximum value among the separation calculation values It is accomplished by comprising a determination unit of the neural network for outputting a decision value of the optical pattern image of the serial paper sheet.

The present invention relates to a neuro-loss / separation device for paper sheets using a random mask method, and an object of the present invention is to provide an image input means for inputting an optical pattern image of a paper sheet to be identified, and the image. Image data storage means for storing the image data read by the input means, image extraction means for extracting the optical pattern image of the paper sheet from the image data stored in the image data storage means, and the image extraction means A plurality of kinds of column masks are applied to the optical pattern image extracted by the method in a random and left-right symmetry parallel to the transport direction of the paper sheet, and converted into slab values of non-mask pixels in each optical pattern image. A preprocessing unit and the slab value are input, and the weight is divided for each fitness level of the optical pattern image by a neuro weight that is optimally adjusted in advance for the fitness separation. A separation calculation unit of a neural network that calculates a calculation value, and a judgment unit of a neural network that outputs a fitness level having a maximum value or a minimum value among the separation calculation values as a determination fitness level of the paper sheet. Achieve by having.

[0010]

In the present invention, the pattern image data optically input by the sensor is used in order to efficiently perform the pattern recognition and the damage separation of the paper sheet (in particular, the 7-denomination of the US dollar bill) by the neural network having the learning function. Is reduced using a plurality of column masks to obtain a plurality of image representative values (slab values). Image data is divided into a number of strips,
A plurality of column regions are masked. The image representative value reduced by such a column mask is invariable to a slight skew of the pattern image, is input to the neuro-operation unit (neural network), and is adjusted by the neuro-weights optimally adjusted for the judgment pattern classification in advance. The separation calculation value is calculated for each determination pattern of the pattern image. The pattern image is determined according to the maximum value of the separation calculation values. As a result, the structure of the neural network becomes small and the control device becomes small. Also, without increasing the scale of the neural network circuit, by connecting a small one in a cascade, or by exchanging weighting functions in the same circuit and inputting the same slab value, the neural network (cascade processing) separates multiple times. Paper sheets can be identified by calculation. This allows
As the structure of the neural network becomes smaller,
The control device itself is also small.

In the present invention, an image image of the target paper sheet is input, the image image is cut out from the image data stored in the storage means, and the paper sheet is conveyed with respect to the cut out optical pattern image. A mask is applied in a direction parallel to the direction in a random and symmetrical manner, a slab value is obtained through the mask, and a neural network separation operation is performed based on this slab value to determine the correctness / separation degree. .
By using symmetrical mask patterns in this way, it is possible to reduce the number of types of patterns to be discriminated and to speed up the process of separating the damage. Also, the sigmoid function e
Since the part of xp (-x) is approximated by a linear function, the error can be reduced and the risk of misjudgment is eliminated.
In addition, learning can be done by a neuro learning calculation unit by an external personal computer (neuro engine), the main unit and the external personal computer are connected by an interface means, image data or a slab value is sent to the external personal computer, and the resulting neuro weight is obtained. Since it can be downloaded from an external personal computer and the result is written to the rewritable ROM of the neuro operation unit attached to the main unit, the neuro weight can be easily updated from the outside and the two can be separated. Therefore, the safety of the device can be secured. When the data passed from the device to the external personal computer is image data, the mask pattern itself can be simulated, and the neuro operation having a wide range of application can be performed. Further, when the data passed from the device to the external personal computer is a slab value, the slab value calculated by the actual machine is used, so that the weighting coefficient can be calculated with data closer to the actual machine.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In the case of performing pattern recognition and damage separation of banknotes (especially US dollars) with a neural network, conventionally, information obtained from a sensor of an identification / separation device is directly input to the neural network. For this reason, in the neural network, the weight of information transmission must be given to all of the obtained input information, and the scale thereof is also large. Therefore, in the present invention, a plurality of column masks for reducing the input information are provided in the preprocessing unit, and the image representative value (slab value) is obtained from the image covered with each column mask. The column mask has a large number of rectangular small sections, and has a structure in which a plurality of rectangular regions in the same direction as the banknote transport direction are covered.

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows an external configuration of the present invention, in which a banknote identification / correctness separation device 100 is provided with a panel section 120 on the right front side of the bank, and a banknote (US dollar banknote) 10 to be identified on the upper part. 1 for aligning and storing
01 is provided. The banknotes 10 stored in the hopper 101 are stacked one by one from the impellers 102 to 103, banknotes of a desired denomination to be selected are held in the first stacker 104, and banknotes whose denomination has been determined are stored in the second stacker 105. The rejected banknotes are suspended and retained in the reject stacker 106. Further, a guide plate 107 that can be opened and closed is arranged above the first stacker 104, and the second stacker 1
A guide plate 108 that can be opened and closed is arranged above 05. This mechanism will be described later.

The panel unit 120 is provided with a power switch 121, a start / stop button 122 for instructing the start / stop of the identification / separation operation in a toggle manner, and a clear button 123 for instructing to clear the display. Is provided. Below that, a display unit 130 for displaying the counted number (amount of money) and the like is provided, and further, a guidance display unit 124 for displaying a failed portion and contents of the bill identifying / separating apparatus 100 is provided.
Then, a mode switch 125 that sequentially selects and switches the identification mode, the learning mode, and the test mode, a batch switch 126 that sets a batch mode for stopping counting when the designated number of sheets is reached, and the number of stopped sheets is increased each time the button is pressed ( When the denomination denomination other than the first identified bill is identified by pressing the number designation switch 127 and the start / stop button 122 for setting the designated number (for example, every 10), reject as a different denomination. Different denomination reject setting buttons 128 are provided in a column. Further, at the lowermost portion, an unfit note level setting switch 12 for setting a level for detecting dirt of the bill and rejecting it as an unfit note.
9A and a tape length setting switch 129B for setting the tape length when detecting a banknote with a tape attached are provided.

FIG. 2 shows an internal mechanism of the bill discriminating / damage separating device 100. The bill 1 stored in the hopper 101 is shown in FIG.
0s are fed one by one by the feeding roller 110, the direction of which is changed by the roller 111 via the conveyance path P1 and later conveyed to the roller 112 via the conveyance path P2. A passage switching member 113 for switching to the conveyance path P3 or P8 is provided at the conveyance exit portion of the roller 112, and the banknotes sent to the conveyance path P8 pass through the conveyance path P9 and the reject stacker 1
Pending at 06. The banknotes sent to the conveyance path P3 by the passage switching member 113 are further sent to the rollers 114, and the passage switching member 115 provided at the conveyance exit portion of the rollers 114.
Is switched to the transport path P4 or P6 and transported. The bills sent to the transport path P4 are sent to the transport path P5 via the rollers 116, and thereafter, are held in the first stacker 104 via the blade members provided around the impeller 102 and sent to the transport path P6. The bills thus sent are sent to the conveying path P7 via the rollers 117, and thereafter are held in the second stacker 105 via the blade members provided around the impeller 103. The banknotes stored in the first stacker 104 are taken out by opening the guide plate 107, and the banknotes stored in the second stacker 105 are similarly guided.
It is taken out by opening 8. And
A line sensor 11 for reading an optical pattern image of a banknote being conveyed is integrally provided with the light emitting unit 12 on the conveying path P1. Further, a rotary encoder 13 that outputs a sample pulse SP that determines the timing of taking in the output data of the line sensor 11 is connected to the transport feeding means such as the roller 111. Further, although not shown, sensors are provided on each of the transport paths P1 to P9 to detect passage of bills, and the first stacker 104, the second stacker 105, and the reject stacker hold and take out bills. A sensor is provided to detect the.

The details of the display unit 130 are as shown in FIG. 3. The uppermost row is a quantity display column 131 for displaying the counted number or amount, and the middle row is a batch display column 132 for displaying the number of batch processing stopped. Is provided, and a mode indicator lamp 133 for lighting and displaying the mode selected by the mode switch 125 is provided at the bottom. In the present invention, the bills 10 stored in the hopper 101 are not limited to the front side and the back side, but the banknotes 10 to 10 shown in (A) and (B) of FIG. It defines the D direction. That is, in the present invention, each denomination of the US dollar bill can be identified / separated in any of the directions A to D.

FIG. 5 is a block diagram showing the internal construction of the bill discriminating / impairment separating device 100. Light from the light emitting section 12 is reflected on the surface of the bill 10 on the conveying path P1 and inputted to the line sensor 11. The read signal of the line sensor 11 is input to the image frame memory 141 via the sensor control unit 14 and stored as a frame image as shown by 141A in FIG. The sensor control unit 14 generates a clock signal and a control signal for reading the data from the line sensor 11, converts the read analog signal into 8-bit to 12-bit digital data by an AD converter, and outputs the bill image and its image. The image data is written in the image frame memory 141 which stores the image data of the peripheral portion. Neuro operation unit 150 is a DS
P (Digital Signal Processo)
r) 154, RAM 155, ROM 156, rewritable R
It is composed of the OM 157 and is built in the apparatus 100 as a separate unit. The data in the image frame memory 141 is shown in FIG.
141B, only the banknote portion is cut out, and the cut-out image data becomes the processing target. Neuro operation unit 1
The identification result DR of 50 is input to the identification device control unit 160 including a CPU, a ROM, a RAM, and the like. The neuro operation unit 150 functions as an identification determination / damage determination operation processing unit, the RAM 155 provides a work area when the DSP 154 executes processing, and the ROM 156 includes a DSP.
A neuro program executed by 154 is stored. The rewritable ROM 157 stores neuro weighting factors downloaded from an external personal computer (neuro learning engine) 170, parameters defining the mask shape, and the like, and the rewritable ROM 1 is used during normal operation.
The contents of 57 are not updated. When the DSP 154 actually operates, the contents of the ROM 156 and the rewritable ROM 15 are rewritable.
All the contents of 7 are copied to the RAM 155 to execute the processing, and in order to speed up the processing speed of the DSP 154, the data is transferred from the ROM 156 having a slow access time to the RAM 155. An external personal computer 170 is further interfaced (I / F) to the banknote identification / property separation device 100.
Serial communication I / F171
Is an external personal computer 170 and a banknote identification / property separation device 10
The command and its response are exchanged with 0, and the parallel communication I / F 172 is used to send the image data or the calculated slab value from the bill validator / damage separator 100 to the external personal computer 170, and a large amount of data is used. It uses parallel communication to transfer. Identification device control unit 16
0 controls the illuminance of the light emitting unit 12, controls the passage switching members 113 and 115, and controls the display of the display unit 130 and the guidance display unit 124. Further, setting input means such as a start / stop button 122 is connected to the identification device control section 160 to control the drive of the conveying and feeding means such as the feeding roller 110, and further detection signals from various sensors are sent. It is supposed to be entered.

FIG. 7 is a block diagram showing a detailed configuration example of the neuro calculation unit 150 of the present invention.
Is measured by an image sensor 11 such as a CCD, and image processing is appropriately performed to obtain a frame image 141A in the image frame memory 141. In the frame image 141A, only the banknote portion is cut out by the neuro calculation unit 150, and 141B
The banknote image is specified as follows. Next, the preprocessing unit 1
At 51, the column masks 41 to 4n are applied to the bill image data 141B to obtain the sum total values (slab values) SB1 to SBn of the pixel data of the unmasked portion. The slab values SB1 to SBn are input to the separation calculation unit 152 of the neural network, and the separation calculation value SP is calculated by the neuro weight optimally adjusted in advance for the pattern classification of the bill for determination. The separation calculation value SP is input to the determination unit 153, and the pattern having the maximum value in the separation calculation value SP is output as the pattern image of the bill 10 of the object. The column masks 41 to 4n in the pre-processing unit 151 are elongated strip-shaped column masks that are randomly covered by using uniform random numbers in parallel to the transport direction, and each cover a different small section.

Here, the pretreatment unit 1 using a column mask
51 will be described. In the present invention, the column masks 41 to 4
The reason for using n is as follows. As shown in FIG. 8, in the binary image of “0” and “1” on the 8 × 8 matrix, the slab value (the number of “1” of the binary image) which is the sum of the pixel values as the feature amount of the image. 8A, “14” is obtained as a value characterizing the character “E” in FIG.
In the same figure (B), "1" is set as a value characterizing the character "H".
2 ”is obtained. Therefore, it is possible to separate“ E ”and“ H ”by using the slab value. However, there are cases where the slab value is the same even in images having different patterns. In A), the slab value of the character “F” is “10”, but the slab value of the character “K” in FIG. 7B is also “10”, which makes them inseparable. This can be solved by introducing a column mask in which a specific rod-shaped region corresponding to the pixel of the input image is covered as shown in FIG.
When the image of FIG. 9 (A) is covered with the mask of FIG. 9, the image shown in FIG. 10 (A) is obtained, and the slab value in this case is “8”. On the other hand, when the image of FIG. 9B is covered with the mask of FIG. 10C, the image shown in FIG. 10B is obtained, and the slab value is “9”. By introducing a column mask in this way,
"F" and "K" can also be separated. Therefore, in order to separate various images in this manner, the positions of a plurality of rectangular-shaped covered portions in the mask as shown in FIG. 10C are selected in advance. In this case, the probability of producing slab values that can separate different images with only one such mask is extremely small. However, different slab value sequences can be obtained in the same image by using different column masks as described above. Often, one of the slab value sequences differs between images, and by using various column masks, it is possible to stochastically enhance the separation capability between images. The use of the plurality of different column masks described above has the following physical meaning. In other words, when observing a three-dimensional object while changing the viewpoint from another direction, different information can be obtained even for the same target. Similarly, using various column masks changes the viewpoint in the two-dimensional plane to observe the image, and as described above, it is possible to generate different information even for the same image. In this case, the input image is covered with various different column masks in the pre-processing, and the sum of the pixels not covered is the slab values SB1 to SB.
n, which corresponds one-to-one with the neuro element in the input layer. Further, the unit value of the output layer corresponds to the determination pattern (or the damage level).

FIG. 11 shows the effect of column masks in the vertical direction such as the masks 41 to 4n, which are the same for the lower moving input image of FIG. 11A and the upper moving input image of FIG. It is possible to obtain a stable slab value of "5", and it is possible to obtain a slab value that does not change even if the input image vertically shifts. Note that the banknotes are conveyed in the lateral direction, and it is extremely rare that the banknotes are conveyed while being held at a right angle with respect to the conveyance path during the conveyance, and they are usually conveyed to the left or the right in advance to some extent. , So-called skew conveyance occurs.
Since the column mask also has a shorter long side length than the row mask (horizontal-direction bar-shaped mask) even with respect to the skew, the pixel area that changes is small and it is not easily affected by it. In the case of the embodiment, a skew angle of up to ± 8 ° is allowed.

Next, at the time of recognition, the number of kinds of column masks selected in advance, suitability for the contents, and points to be noted in the selection will be described. In the preprocessing unit 151 in the neuro operation unit of the external personal computer 170, the column mask types 41 to 4n and the covered areas in the masks 41 to 4n are considered as parameters. However, external PC 17
The learning algorithm used by the neuro-operation unit of 0 uses the back-propagation method given by the following equation 1.

[Equation 1] Further, the weight is corrected every time each pattern is presented. The convergence judgment is performed when the sum of squares of the difference between the output layer value obtained for each pattern and the teacher value is equal to or less than the convergence judgment error.
Alternatively, the number of presentations reaches the maximum number of presentations. Here, the number of presentations means one time when the teacher is presented to all patterns of seven denominations from USD1 $ to USD100 (three levels from the level 1 to the level 3 for the damage level). Define as. The learning data sequentially presents up to 7 denominations from US $ 1 $ to US $ 100 to the neural network. Further, the discrimination rate ES given by the following equation is used as the evaluation standard of the cognitive ability.

[Equation 2] ES = number of correctly recognized events / evaluated number of all events × 100

The learning situation and recognition ability are examined according to the type of mask, that is, the number of slabs. First, the column mask used in the present invention is created as follows. That is, a uniform random number of 8 which is the same as the number of columns of the matrix of the input image is generated in the interval [−1, 1], and the column region of the number corresponding to the negative random number value is covered therein. Various column masks are created by changing the initial value of the uniform random numbers by such a method. The number of slabs examined here is 6, 2, 3, 4, 5, 6, 7.
On the street. FIG. 12 shows the learning state of the neural network until the number of presentations reaches 30,000 for these six slab numbers. The horizontal axis represents the number of times teacher data is presented, and the vertical axis represents the squared error. Figure 1
From 2, it is apparent that when the number of slabs is "2", learning does not converge and pattern separation is impossible. The number of slabs is "3"
In the case of, the learning of the neural network shows a convergence tendency to some extent, but the error curve is oscillatory. In this case, the learning was further continued up to 60,000 times, but the squared error did not become 1.0 or less. It is shown from FIG. 12 that the learning converges when the number of slabs is “4” or more. Therefore, it is sufficient to set the number of slabs to 4 or more in the input layer of the separation calculation unit 152. Compared to the conventional input pixel 64 (= 8 × 8) input to the input layer, the number of neuro elements can be extremely reduced, and as a result, the separation calculation unit 152 can be simplified.

Next, the learning situation and the recognition ability will be described based on the coverage area of the column mask. First, the change of the covered area is performed by the fluctuation range of the random number. That is, the range of random numbers [-
1, 1] as a reference, in the increasing direction of the coverage area,
Random number range is [-2, 1], [-3, 1], [-4, 1]
And the width of the random number in the decreasing direction of the coverage area is [-1, 1],
Random numbers are generated as [-1, 3] and [-1, 4].
However, the number of slabs is "4". FIG. 13 shows the learning situation when the number of presentations reaches 30,000 when the width of the random number is changed. The horizontal axis and the vertical axis are the same as in FIG. It is clear from FIG. 13 that the learning does not depend so much on the masked area. It was also found that the recognition ability does not depend much on the masked area. Therefore, the coverage area of each column mask should be selected randomly without any special care.
Consider the type of mask.

The number of slabs can be reduced by using the column mask as described above, even for a light and shade pattern such as a bill. Therefore, when the banknotes are identified, it is possible to surely identify them if there are 16 types of masks, and it is sufficient to input 16 or more as the number of slabs to the input layer of the separation calculation unit 152. Various simulation experiments were conducted on specific column masks. That is, since one pixel has a width of 1 mm and a length of 4 mm,
The width of the column mask was increased from 1 mm to 10 mm in increments of 1 mm, and such a column mask was randomly taken to determine the most efficient column width and its position by experiment. In the embodiment, as these optimum values, width 6 mm × length 66
It was decided to prepare 16 mm column masks.
FIG. 14 is an example of mask information of a column mask. Mask patterns 1, 2, ... 9, 0, ... 5, 6 show mask positions when a banknote is divided into 16 in the horizontal direction, and 0 is The mask is not applied, and 1 indicates that the mask is applied. Unit number NO. In the case of 1, the slab value which is not masked at all is input and the unit number NO. 2-NO. At 16, a slab value on which the mask is randomly applied is input.

In the conventional case, the number of neuro elements can be extremely reduced as compared with the case where the input pixel 6480 (= 216 × 30) is input to the input layer in the configuration of FIG. Therefore, the separation calculation unit can be simplified. Note that the learning speed is not limited to the case where the number of slabs is 16, but the learning speed is reduced even if the number of slabs is somewhat decreased, but the identification is possible, and more reliable identification is possible if the number of slabs is increased.

Next, a description will be given of the separation calculation section for inputting the information preprocessed by the preprocessing section of the neuro calculation section of the external personal computer 170 to perform the separation calculation. The separation operation unit 152 having a hierarchical structure is roughly composed of three layers of an input layer, a hidden layer, and an output layer. As shown in FIG. 15, the input layer is provided with neuro elements so as to correspond one-to-one to each mask type from the pre-processing unit, and the slab value pre-processed by each mask type (mask processed The total number of subsequent pixels) is input to the corresponding neuro element. The hidden layer is composed of at least one neuron layer, and plays a role of separating and calculating the information of the input layer and transmitting it to the output layer. As the number of hidden layers increases, it becomes possible to perform the calculation separately for each pattern invariantly with respect to the fluctuation of information of each neuro element of the input layer. Neuro elements are provided in the output layer so as to correspond one-to-one to the categories to be identified. Then, the output unit values by the weighting factors between the neuro elements completed by learning are calculated for several output units. The maximum value of these plural output unit values (taking a value between 0 and 1) (about 0.99 in the output unit of the denomination of the normal inspection bill) and the quasi-maximum value (the second denomination candidate) And less than 0.2). Next, it is judged whether or not the maximum value is larger than the threshold value 1 (usually 0.6), and when it is smaller, it is excluded from the learning data. Then, it is determined whether or not (maximum value-quasi-maximum value) is larger than the threshold 2 (normally 0.4), and when (maximum value-quasi-maximum value) is small, it is excluded, and when it is large, the maximum value is set. The pattern of the unit included is determined to be the judgment pattern of the evaluation bill. Here, the reason why the maximum value and the quasi-maximum value are checked is to prevent misidentification between denominations.

In the pattern recognition using the neural network, the performance evaluation has so far been performed by the discrimination rate ES which is a statistical probability, and only the maximum value of the output value of the output unit of the neural network is obtained. Is the subject of evaluation. However, in such an evaluation index, even if the output value other than the maximum value is considerably close to the maximum value, the evaluation result does not reflect the direct influence. For example, when two types of data are identified, the result of the differentiation rate of the neural network is 100%, but in this case, the difference between the maximum output value of a certain output layer and the output value of the other output layer in the neural network is There are cases where it is very large and cases where there is not much difference. From the detection of the degree of separation, it can be easily understood that the larger the difference is, the better the discrimination is. Generally, in the products on the market to which the neural network is applied, it is required that not only the maximum value of the output unit but also the output values other than the maximum value are sufficiently apart from the maximum value. This is because the market demands that the output result be as close as possible to the output result of normal data even when noise is mixed in the input data due to external influences. In other words, a product that gives outputs at a sufficiently distant distance is desired rather than a product in which the maximum output value of the output unit and other output values are close to each other. Therefore, thresholds are set for the maximum value of the output values of the output unit and the difference between the maximum value and the output values of the other units in order to check whether the neuro judgment is sufficient or not.

For example, when identifying a US dollar bill of 1 $ to 100 $, the neuro elements in the output layer are sequentially arranged from 1 $ to 1 $ from the top.
Corresponding to the denomination of 100 $, the corresponding neuro element outputs the value closest to 1, that is, the maximum value, and the other neuro elements output the value close to 0 at the time of identification. When identifying the 1 $ bill, the highest neuro element in the output layer outputs the maximum value, and it is determined as the determination pattern.

Similarly, in the above-mentioned bill identification, the neuro elements in the output layer are sequentially arranged from the top in the order of 1 $, 2 $, 5 $, 10 $,
20 neuron elements of 7 denominations × 4 directions correspond one-to-one, such as 20 $ ... 100 $. The mechanism that connects the neuro elements from the input layer to the output layer and transfers signals is called a synapse. The synapse stores the strength of coupling between the neuro elements as a weighting function. One neuro element receives a signal from a plurality of neuro elements in the preceding layer through the synapse, and multiplies the weight of the synapse that has passed through it as an input value. The neuro element sums the input values when it receives the signals from all the neuro elements to which it is coupled. When the total value exceeds the threshold value set in advance in the neuro element, the neuro element "fires", sends an output signal to the neuro element in the next subsequent layer, and repeats this process to output information from the output layer.
The weight of each of these synapses is determined in advance by learning by the back propagation method, corresponding to the identification target. The method is performed by the above-mentioned equation 1.

Next, the judging section will be described. The determination unit inputs the information output from the output layer of the separation calculation unit,
The maximum value is discriminated from among them, and it is discriminated that it is the cadegory corresponding to the neuro element having the maximum value. For example, when the neuro elements of the output layer correspond to the cadegories in the A direction of 1 $ bills, the 1 $ bill B direction, and the 1 $ bill C direction from the top, the highest neuron outputs the maximum value. If so, it is determined to be “1 $ A direction”. External personal computer 17
The configuration of the neuro computation unit of 0 is the same as the configuration of the neuro computation unit 150 of the bill discriminating / correctness separation device 100 except for the learning function. Since the former requires a differential value of the sigmoid function, the approximation formula Cannot be used, but the latter is different in that a value obtained by approximating the exp (-x) function by a linear function y = ax + b is used in order to secure the calculation speed. That is, the built-in neuro calculation unit 150 does not have a learning function.

An operation example of bill identification in such a configuration will be described with reference to the flowchart of FIG. Hopper 1
When the banknote 10 is placed on 01 and then the start / stop button 122 is pressed, the identification device control section 160 drives the conveying and feeding means, and the banknote 1 is fed by the feeding roller 110.
0s are fed one by one to the transport path P1. The bills 10 that have been fed out are irradiated with the light from the light emitting unit 12, and the reflected light is C
It is input to the line sensor 11 such as a CD, and its read signal is input to the sensor control unit 14. Sensor control unit 14
Is the sample pulse SP from the rotary encoder 13.
Is input, the sweep is started at every predetermined clock of the mechanical clock, the pixel output for one line at a pitch of 1 mm in the horizontal direction is AD-converted at an interval of 4 mm in the transport direction, and the converted digital numerical value is stored in the image frame memory. By writing to 141, the optical image of the banknote 10 can be input to the storage means (steps S1 to S5). The image frame memory 141 of this example has 256 mm in the horizontal direction for each banknote.
× A length of 128 mm is prepared. Since the size of the bill is 156 mm × 66 mm in the case of a US dollar bill, the bill cutting unit is set to the image frame memory in order to avoid setting the mask on the non-billing portion when masking in the next step. The banknote edge in the image frame developed in 141 is extracted (step S6), the data of the banknote portion is specified and cut out (step S7), and the banknote portion is image-processed to obtain an input image. Since the optical system is a reflection type sensor, the presence or absence of the banknote medium is determined by using a threshold value that is a predetermined multiple of the AD value of the darkest portion of the portion without the medium, and the bright portion is taken out. In the case of using the transmissive sensor, a place darker than the threshold value that is set oppositely is the bill portion.

The image data cut out by the bill cutting unit is processed by the neuro calculation unit 150, the preprocessing unit 151 reads the mask information (step S8), and creates a slab value based on the mask information (step S9). ). Steps S8 and S9 until the number of slab values reaches the number of input units
The processing of (1) is repeated (step S10), and upon arrival, the separation calculation section 152 calculates the output unit value by the weighting coefficient between the neuro elements completed by learning (step S1).
1). Then, the maximum value of output unit value (takes a value between 0 and 1) (about 0.99 in the output unit of the denomination of the normal inspection bill) and the quasi-maximum value (0 in the second denomination candidate). .2 or less) is extracted (step S12). next,
It is judged whether or not the maximum value is larger than the threshold value 1 (normally 0.6) (step S13), and when it is smaller, it is excluded from the learning data (step S15). Then, it is determined whether (maximum value-quasi-maximum value) is larger than the threshold value 2 (normally 0.4) (step S14), and when (maximum value-quasi-maximum value) is small, it is excluded (step S15). If it is larger, the unit pattern having the maximum value is determined as the evaluation bill determination pattern (step S16). Judgment unit 153
Inputs the information output from the output layer of the separation calculation unit 152, discriminates the maximum value (may be set to the minimum value) from the information, and discriminates it from the category corresponding to the neuro element having the maximum value. To do. Then, the identification result DR is input to the identification device control unit 160. During the identification operation, the bills are conveyed through the conveyance paths P1 and P2, and the identification result D input by the identification device control unit 160 is reached by the time the rollers 112 reach the rollers 112.
The passage switching members 113 and 115 are driven based on R.
That is, when the bill is identified as a bill of the desired denomination, the passage switching member 113 is rotated to the roller 112 side to retract from the conveyance path P3, and the passage switching member 115 is projected onto the conveyance path P6. The P3, P4, and P5 are conveyed and fed by the impeller 102 to the first stacker 10
Reserved to 4. Further, when it is identified as a banknote of an undesired denomination, the passage switching member 113 is rotated to the roller 112 side to retract from the conveyance path P3, and the passage switching member 115 is rotated to the roller 114 side to retract from the conveyance path P6.
The banknotes are transported through the transport paths P3, P6, P7, are dispensed by the banknote dispensing unit 103, and are held in the second stacker 105. When the denomination cannot be identified, the passage switching member 113 is projected to the conveyance passage P3, the bills are conveyed along the conveyance passages P8 and P9, and are held in the reject stacker 106. When the identification device control unit 160 receives the banknote hold detection signal from the sensor, the identification device control unit 160 displays the number and the amount of the bills on the display unit 130. If the determination of the denomination has not been completed, the process returns to step S11 and the above-described processing is repeated (step S17).

Next, an embodiment in which the neural network is connected in a cascade will be described. FIG. 17
[Fig. 3] is a diagram showing the concept of connecting neural networks in a cascade. First, as the first step, 16 units of input layers, 16 units of hidden layers, and 8 units of output layers are set. Each of the eight units in the output layer is 1 $ ~
Seven pieces are assigned to the A direction (defining four directions of ABCD including front and back) of a 100 $ bill, and the other is assigned to the eighth output unit which is the remaining one unit. As a result, the A direction of 7 denominations out of 28 patterns (7 denominations × 4 directions) is separated, and the rest is 21 patterns. Then the first
If there is the eighth output of the stage neural network, the same slab value as that input to the input section of the first stage neural network is input to the second stage neural network. In the eight output layers of the output of the second stage, the B direction of the 1 $ to 100 $ banknotes is 1 to 7 respectively.
The second output is assigned. Here, B patterns of 7 denominations of 21 patterns are separated. At this stage, the rest is 14 patterns.

Next, when there is the eighth output of the second-stage neural network, the same slab value input to the input section of the first-stage neural network is input to the third-stage neural network. input. Each of the eight output layers of the output of the third stage has 1 $ to 100 $.
The 1st to 7th outputs are assigned to the C direction of the banknote. Here, out of 14 patterns, the C directions of 7 denominations are separated. At this stage, the rest is 7 patterns. Next, when there is the eighth output of the third-stage neural network, the same slab value as that input to the input section of the first-stage neural network is input to the fourth-stage neural network. In the eight output layers of the output of the fourth stage, the D direction of 1 $ to 100 $ bills is 1 to
The seventh output is assigned. Here, of 21 patterns, the D direction of 7 denominations is separated. If there is an eighth output at this stage, it means that the bill is not a dollar bill.

In the above, the neural network is physically cascade-connected, but the same neural network is used as shown in FIG. It is also possible to sequentially reload the data from the device control section 160 for each of the above steps and perform the processing in cascade. That is, the weighting coefficient value of the neural network is forcibly rewritten from the outside to the target weighting coefficient value for the separation calculation immediately before the separation calculation is performed on the part. As a result, the separation operation can be performed by using the same slab value for the cascade without increasing the number of physical hardware.

The output from the determination unit 153 of the neuro calculation unit 150 is read and determined by the CPU of the identification device control unit 160, and if the denomination set by the setting input means is sent to the stacker 104, otherwise it is passed. Switching member 1
13 is controlled to convey the bill to the reject stacker 105. Then, the display unit 130 is updated. Also,
The identification device control unit 160 counts the number of fed-out bills from the feeding roller 110 by a bill passage sensor (not shown) provided immediately after the feeding unit 110, and when the setting input means designates the counted number. Stops the conveying means when the counting of the designated number of bills of the designated denomination is completed. In addition, although the pattern recognition of the US dollar bill is described in the above-described embodiment, the pattern recognition can be performed by the bill character reader of another country or the OCR, and the securities reader such as a check. Of course, other pattern reading devices are also possible. In addition, although the line sensor is used to acquire the optical pattern image in the above description, an area sensor can also be used.

By the way, in the normal separation of the paper sheets, the conveying direction of the paper sheets is not limited. Therefore, in identifying the denomination of a paper sheet, it is necessary to use the denomination, its front and back, and the upside-down inverted shape as identification patterns, and the neuro element of the output layer of the neuro-operation unit 150 also corresponds to this. However, it is necessary to perform the loss separation regardless of whether the head is inverted and the number of neuro elements in the output layer needs to be reduced. Therefore, the mask processing that is not related to the inverted head is realized.

That is, in the normal / separation separation of paper sheets according to the present invention, as shown in FIG. 19, the masks M1 and M2 are set symmetrically with the central axis of the paper sheet 1 as a symmetrical axis, and the number of unmasked pixels is set. to add. As a result, the added value becomes an invariable amount regardless of the image pattern and is upside down. 20 and 21 are diagrams for explaining a method of creating a slab value that is irrelevant to the upright and inverted positions, and it can be seen that the slab values with the central axis being symmetrical are the same in the upright and inverted transportations. It In order to reduce the number of output neuro elements of the neural network, it is necessary to match the number of judgment patterns with the number of output neuro elements. Therefore, the center axis of the image of the paper sheet in the image data is calculated, and the center axis is symmetrical. Set the mask so that the part not masked appears in. This allows
As shown in FIGS. 20 and 21, a slab value irrelevant to the upside-down of the paper sheet is obtained. Therefore, for example, by learning only the image in the upright conveyance direction as data, information irrelevant to the upright inversion can be externally provided to the neuro calculation unit 150 by downloading. The number of neuro elements in the output layer of the neuro calculation unit 150 corresponds to the fitness level. That is, neuro elements are provided in the output layer so as to correspond one-to-one to the fitness level to be determined. Then, the weighting factor between the neuro elements completed by the learning by the external personal computer 170 is stored in the rewritable ROM 157 by downloading, and the output unit value is calculated for several output units by the weighting factor.

In the correctness separation, the degree of normality of the paper sheet is determined within the denomination specified after the denomination determination on the premise of the neuro denomination identification. FIG. 22 shows a state in which the denomination discrimination neuro is followed by the fitness separation neuron to perform the fitness separation, and in FIG. 23, the mask image processing unit 15 receives image image data.
After setting the mask in 8, the neuro decision unit 159 composed of the input layer, the hidden layer, and the output layer separates the genuine note and the unfit note (unfit note 1 to unfit note 3). That is, the pixels effective for separation of the fitness are extracted from the paper sheet by various masks, and the smoothing information (sum) of the extracted pixel group is used as the neuro determination unit 1.
Input to 59 input layers. The neuro determination unit 159 has a hierarchical structure, for example, one input layer and a plurality of hidden layers,
It also has one output layer. Each neuro element of the output layer corresponds to a damage level of a paper sheet. For example, in the simplest case, there are two neuro elements, one is a neuro element indicating a genuine note, and the other is It is a neuro element showing an unfit note. In addition to this, the unfit notes are divided into levels according to the degree of damage, and the level is set to, for example, a normal note, a small unfit note level, an unfit note level, and a large unfit note level, and outputs four neuro elements. It can also be set in layers. The unfit note level is a comprehensive expression of the degree of dirt, holes, and tears. Regarding the neuro learning that enables separation of unfit notes, several kinds of paper sheet data to be judged as genuine notes as teacher data,
Also, several types of paper data to be determined are prepared as unfit notes, and learning is performed using the backpropagation method. The masking method is to search the coordinates of the symmetric center line of the cut-out banknote image image, apply some kinds of masks shown in FIG. 23 based on the coordinates, and sum the pixel values of the unmasked image, that is, Calculate the slab value. Since the mask is constructed symmetrically, the slab values of the upright and inverted images are the same as shown in FIGS. 20 and 21, and different slab values can be obtained for the front and back images.

Further, FIG. 24A shows the sigmoid function y.
= 1 / {1 + exp (-x)}, and the function exp (-x) in a part of the section of FIG.
By approximating with the linear function y = ax + b shown in (B), the processing speed can be increased. The values of x shown in the figure and the values of a and b of the linear function are tabulated and stored in the ROM 156. The DSP 154 uses this table copied to the RAM 155 as described above,
Exp (-X) is approximately calculated by y = ax + b. Table 1 below shows the functions of a, b and x.

[Table 1] Since the neural operation unit 150 uses the differential value of the sigmoid function for learning, it is not possible to use an approximate expression. After the data from the sensor is expanded in the image frame memory 141, the stored data in the memory is transferred to the neuro calculation unit 150, the data from the sensor is expanded in the image frame memory 141, and the DSP 154 is operated to obtain the slag value. Are processed by the identification / property separation device 100,
The slab value is sent to the neuro calculation unit 150.

Next, with reference to the flow charts of FIGS. 25 and 26, the operation of separating the fitness will be described. First, the RAM 15 is used as a parameter for determining the neural network configuration.
5 (step S20), and then the neuro weights suitable for the neural network configuration are stored in RAM1.
55 is downloaded (step S21). Then, the banknote data is read from the sensor control unit 14 (step S22), the edge of the banknote is detected (step S23) as in the case of sheet identification, and the banknote image is extracted from the image frame (step S24). After that, various masks are set on the bill image (step S25), and slab values by various masks are created (step S2).
6) The slab value is input to each unit of the input layer (step S27).

Each unit of the hidden layer is multiplied by the output value of the input layer and the weight to obtain the sum (step S30), and the sum of the units is input by the approximate sigmoid function (linear function) described in FIG. The output value of the layer unit is obtained (step S31). Then, for each unit of the output layer, the output value of the hidden layer is multiplied by the weight to obtain the total sum Xi (step S32), and the total sum of the units is input to the approximate sigmoid function (linear function) to output the unit of the output layer. The output value is Oi. Then, check the maximum output value (step S
34), and rejects in the case of NG (step S3
7) In the case of OK, the maximum value and the quasi-maximum value of the output value are further checked (step S35), in the case of NG, it rejects, and in the case of OK, the pattern of the unit having the maximum value is used as the judgment pattern. (Step S36). Although the fitness level of the paper sheet is determined based on the fitness level using the determination pattern, the determination result DR is input from the neuro calculation unit 150 to the identification device control unit 160, and operates as a separation device control unit here. In the above, banknotes were explained, but a symmetrical random mask method can be used for the separating unit used for sorting devices such as beer tickets, gift certificates, etc. At this time, since there is no pattern on the back side, only the front side image Only process data.

[0043]

As described above, according to the neuro-banknote discriminating / damage separating device by the random mask method of the present invention, the image representative value (slab value) is obtained by using the column mask parallel to the lateral direction of the paper sheet. Therefore, the structure of the neural network is small and the control device is also small.
Further, even if the paper sheet is displaced, the pixel value is not affected, so that the paper sheet can be reliably identified. Also, by connecting a neural network in a cascade, even in the case of a large number of category classifications, categories (7 denominations of dollar bills, front and back,
The transport direction) can be determined.

According to the right / left symmetric random mask type damage separation device, a column mask parallel to the lateral direction of the sheet is symmetrically applied to obtain the image representative value (slab value). Less affected by misalignment of leaves,
Moreover, the configuration of the neural network is small and the control device is also small without distinguishing whether it is upright or inverted. Separate masks are prepared for the front and back sides,
The same bilaterally symmetrical mask pattern can be used for the A-direction conveyed bill and the B-direction conveyed bill, and similarly, a common mask pattern for the C-direction and the D-direction can be used. The calculation of the natural logarithm part of the sigmoid function which requires complicated operation is performed by y = ax
Since the natural logarithm approximation of the sigmoid function is approximated by the linear line of + b, the recognition calculation process becomes faster.
As a result, an accurate approximate value can be obtained, and the reliability of the determination result by the neuron can be improved. Further, the number of sheets to be processed per time can be increased, and a large amount of processing can be performed.

By having an interface connected to an external neuro-calculation unit, it is possible to easily learn the neural network when it is desired to change the identification medium or change the fitness level, and the neuro weight of the result is calculated. The contents of the rewritable storage means of the identification device can be downloaded and updated, and additional learning can be facilitated. Therefore, it is possible to finely adjust the damage level using the actual product or to newly register at the business office where the user is located.

[Brief description of drawings]

FIG. 1 is a perspective view showing an external configuration example of a neuro-identification / stainlessness separation device for paper sheets by a random mask method according to the present invention.

FIG. 2 is a diagram showing an internal mechanism of a neuro-identification / stainlessness separation device for paper sheets by a random mask method according to the present invention.

FIG. 3 is a diagram showing details of a display unit of a neuro-identification / damage-separation device for paper sheets by a random mask method according to the present invention.

FIG. 4 is a diagram for explaining banknote data and a collection direction.

FIG. 5 is a block diagram showing an internal configuration of a sheet neuro-identification / damage-separation device according to a random mask method of the present invention.

FIG. 6 is a diagram for explaining banknote data and a slab value creation method.

FIG. 7 is a block diagram showing a main part configuration of a neuro-banknote discriminating apparatus using a random mask method according to the present invention.

FIG. 8 is a diagram for explaining different slab values for different patterns used in the present invention.

FIG. 9 is a diagram for explaining different slab value generation for different patterns used in the present invention.

FIG. 10 is a diagram for explaining a column mask used in the present invention.

FIG. 11 is a diagram for explaining a column mask effect with respect to movement of an image in a vertical direction.

FIG. 12 is a diagram showing learning situations for various slab numbers.

FIG. 13 is a diagram showing a learning situation with respect to various covering areas of a mask.

FIG. 14 is a diagram showing an example of encoded mask information used in the present invention.

FIG. 15 is a block diagram showing the details of the main part configuration of the neuro-identification / damage-separation device for paper sheets by the random mask method according to the present invention.

FIG. 16 is a flow chart showing an operation example of the sheet neuro-identification / damage-separation device by the random mask method according to the present invention.

FIG. 17 is an explanatory diagram of a case where the neural network of the sheet neuro-identification / loss detection device by the random mask method according to the present invention is connected to a cascade.

FIG. 18 is an explanatory diagram in the case of rewriting the weighting factor of the neural network of the sheet neuro-identification / damage separation apparatus by the random mask method according to the present invention.

FIG. 19 is a diagram showing a state of a symmetrical mask.

FIG. 20 is a diagram showing an example of slab values during upright conveyance.

FIG. 21 is a diagram showing an example of a slab value during inverted conveyance.

FIG. 22 is a diagram showing a relationship between a denomination identification neuron and a fitness separation neuron.

FIG. 23 is a diagram showing a neuron calculation unit in the fitness separation.

FIG. 24 is a diagram for explaining approximation of a sigmoid function.

FIG. 25 is a part of a flow chart showing an example of the operation for separating the fitness.

FIG. 26 is a part of a flow chart showing an example of the operation of the normality separation.

FIG. 27 is a block diagram showing a configuration example of a conventional pattern recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 banknote (identification target) 2 image sensor 3 input image 4 preprocessing section 5 separation calculation section 10 banknote (US dollar banknote) 11 line sensor 12 light emitting section 14 sensor control section 141 image frame memory 150 neuro processing section 151 preprocessing section 152 Separation calculation unit 153 Judgment unit 158 Mask processing unit 159 Neuro judgment unit 160 Identification device control unit 170 External personal computer (neuro learning engine)

[Procedure amendment]

[Submission date] January 25, 1995

[Procedure Amendment 1]

[Document name to be corrected] Drawing

[Correction target item name] Fig. 25

[Correction method] Change

[Correction content]

FIG. 25

[Procedure Amendment 2]

[Document name to be corrected] Drawing

[Correction target item name]

[Correction method] Change

[Correction content]

FIG. 26

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kengo Terada 1-3-1, Shimodeno, Himeji-shi, Hyogo Prefecture Glory Industry Co., Ltd.

Claims (8)

[Claims] 1. An image input unit for inputting an optical pattern image of a paper sheet to be identified, a storage unit for storing image data read by the image input unit, and image data stored in the storage unit. Image extraction means for extracting the optical pattern image of the paper sheet from, and a plurality of column masks at random in parallel to the transport direction of the paper sheet with respect to the optical pattern image extracted by the image extraction means. A pre-processing unit for converting the sum of the non-masked pixels in each of the optical pattern images (slab value), and inputting the slab value, and preliminarily adjusting the neuro weights of the optical pattern images optimally adjusted for the determination pattern classification. A separation calculation unit of a neural network for calculating a separation calculation value for each judgment pattern, and a judgment pattern having a maximum value or a minimum value among the separation calculation values A neural network discriminating device by a random mask method, comprising: a neural network discriminating unit that discriminates and outputs as an optical pattern image of the sheet. 2. The neuro-identifying device for paper sheets by a random mask method according to claim 1, wherein the neural networks are cascade-connected in accordance with categories of the paper sheets to be classified. 3. The neuro-identifying device for paper sheets by the random mask method according to claim 1, wherein the neuro weights in the separation calculation unit are sequentially rewritten according to the categories to be classified of the paper sheets. 4. An image input unit for inputting an optical pattern image of a sheet to be identified, an image data storage unit for storing image data read by the image input unit, and an image data storage unit for storing the image data. Image extraction means for extracting the optical pattern image of the paper sheet from the image data,
A plurality of types of column masks are randomly and bilaterally symmetric parallel to the transport direction of the paper sheet with respect to the optical pattern image extracted by the image extracting unit, and a slab of non-masked pixels in each optical pattern image. A pre-processing unit for converting into a value and a separation operation of a neural network which inputs the slab value and calculates a separation operation value for each fitness level of the optical pattern image by a neuro weight optimally adjusted in advance for the fitness separation. Unit and a judgment unit of a neural network for outputting a damage level having a maximum value or a minimum value among the separation calculation values as a judgment damage level of the paper sheet. Neuro-property separation device for paper sheets by. 5. The sigmoid function f (x) = 1 / {1 + exp (−) of the nonlinear characteristic of the neural network.
x)} exp (-x) term is linearly approximated by a linear function y = ax + b. 6. A built-in neuro-operation unit located inside the apparatus and formed by the pre-processing unit, the separation operation unit and the determination unit,
An electrically programmable non-volatile storage means for storing the neuro weight, a pre-processing section capable of changing and setting the mask pattern, and a neural network composed of a separation calculation section and a determination section for learning the neuro Interface means for exchanging data with an external neuro-operation unit provided outside for calculating weights, and data of the optical pattern image of the paper sheet stored in the image data storage means by the interface means Alternatively, the slab value calculated by the built-in neuro arithmetic unit is uploaded to the neuro arithmetic unit, and the neuro weight from the external neuro arithmetic unit is downloaded by the interface unit to update the non-volatile storage unit. The random mask method according to claim 4, further comprising a control means. Neuro positive loss separation apparatus of the sheets. 7. The neuro-loss / separation device for paper sheets by a random mask method according to claim 6, wherein the data sent to the external neuro-operation unit through the interface means is image data. 8. The neuro-loss / separation device for paper sheets by a random mask method according to claim 6, wherein the data sent to the external neuro-operation unit through the interface means is a slab value.