JPH1196431A - Method for judging unknown pattern sheet paper picture in neuro identification and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To exclude an unknown picture pattern from an object to be discriminated by expressing the picture of the learning pattern of sheet paper with a multi-dimensional Gaussian density function in an identification space, and judging the picture of the paper sheet on a threshold plane. SOLUTION: The picture of cash 11 is processed by a picture collecting part 12 constituted of an image sensor so that a slab value can be obtained, and the cash 11 is relatively identified by a neural network 10 so that a certain pattern can be obtained, and the value of a multi-dimensional Gaussian density function corresponding to this judgment pattern is calculated by an arithmetic means 13 based on the slab value. When the multi-dimensional Gaussian density function is a high value in the same way as the learning pattern of the cash 11, it is judged that the cash evaluated by a comparison judging means 14 is within the range of the pattern of the cash applied at the time of learning. Then, the picture of the paper sheet is expressed with the multi-dimensional Gaussian density function, and judged on the threshold plane so that a learning pattern and a non-learning pattern can be identified for the inputted cash, and the cash of the non-learning pattern is excluded.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neuro discriminator (neural network or neural computer) for recognizing the type of paper sheets having image patterns such as banknotes, checks, gift certificates, and securities. The present invention relates to a method and an apparatus for judging an unknown pattern paper sheet image in neuro discrimination in the case of using an unknown pattern paper by using a multi-dimensional Gaussian density function for image data. The present invention relates to a method and an apparatus for determining a leaf image.

[0002]

2. Description of the Related Art A neuro-identifying machine (neural network or neural network) recognizes or identifies the type of paper sheet on which an image pattern such as a bill, a check, a gift certificate, a security, etc. is printed. In the case of using a computer, if an image pattern of an unlearned (untaught) sheet is input and recognized, it is only necessary to obtain an output by a sigmoid function from the learning result of the neural network. For example, in spite of the difference in the reaction strength, the user is forced to output a result indicating which one of the learned patterns is approximated even if it is forced. For this reason, there is a problem that an accurate result cannot be normally obtained for unlearned paper sheets. That is, the neural network exhibits excellent relative discrimination ability with respect to the learned paper sheets (here, banknotes are taken as an example). However, unlearned banknotes that have not been taught have been eliminated by setting a threshold on the reaction level of the neural network and using template matching in post-processing. Specifically, in the neural network, the given banknotes are identified relative to each other, and the input value is learned by a template or threshold based on the pattern determined by the neural network again in post-processing. It was checked whether or not it was within the range of the pattern of the bills. In this way, the banknotes of the unlearned pattern that were not given in the learning were statistically excluded.

On the other hand, in order to improve the discriminating power, as described in Japanese Patent Application No. 8-70868 filed by the present applicant, image data which is usually considered impossible cannot be used for discrimination. Is included, a trap process (confirmation of validity of input values) is performed so that the corresponding banknote is immediately excluded from recognition and classification. In the conventional trap processing, the range in which each slab value input to the neuro discriminator exists is assumed to be normally distributed, and the outside of the predetermined range is excluded by the trap processing.

FIG. 12 shows a schematic configuration of such a neural network. The separation operation section 12 having a hierarchical structure is roughly divided into an input layer (16 units), a hidden layer (16 units), and an output layer (7 units). ) Consists of three layers.
The input layer is provided with a neuro element so as to correspond to each mask type from the preprocessing unit on a one-to-one basis, and a slab value preprocessed by each mask type (the number of pixels after the mask processing) To the corresponding neuro element. The hidden layer is composed of at least one neuro element layer, and plays a role of separating and calculating information of the input layer and transmitting it to the output layer. As the number of hidden layers increases, the calculation can be performed separately for each of the patterns invariably with respect to the fluctuation of each neuro element of the input layer. In the output layer, neuro elements are provided so as to correspond one-to-one to categories to be identified. Then, the output unit values for several output units are calculated based on the weight coefficients between the neuro elements completed by the learning. The maximum value of the plurality of output unit values (takes a value between 0 and 1) (0.99 rank in the output unit of the normal denomination denomination) and the semi-maximum value (second denomination candidate) Is 0.2 or less), and the maximum value is the threshold TH.
It is determined whether it is larger than 1 (normally 0.6), and if smaller, it is excluded from the learning data. And (maximum value-
It is determined whether the (sub-maximum value) is larger than the threshold TH2 (usually 0.4). If (maximum value-sub-maximum value) is small, it is excluded. If it is large, the pattern of the unit having the maximum value is determined. Is determined as a discrimination pattern of the evaluation bill.

[0005]

According to the above-described method for discriminating by a neural network, the neural network is flexible (the ability to identify dirt, tears, aging bills, and various kinds of bill transporting). Even if noise is added to the bill data, the discriminating ability with the ability to correctly discriminate bills is performed, but the discrimination ability of the target bill is eventually reduced in the post-processing. Was. In other words, in the post-processing in the conventional neural network discrimination, it is checked whether the input value of the discriminated pattern falls within the range of the learning pattern linearly with a template or a threshold. Was. Therefore, although unlearned banknotes having an image pattern different from the image pattern of the learning banknote are excluded, the learning banknote is also excluded when the degree of damage is large and the learning banknote is far from the original image.

Generally, when input information is represented by a vector,
In pattern discrimination for determining whether or not the input data can be regarded as data of a certain pattern, the sum of squared errors between the teaching output (learning) and the model output (discrimination) is minimized. A so-called multilayer perceptron (hierarchical neural network, backpropagation network) for learning by backpropagation is widely used. Although the multi-layer perceptron is very suitable for forming a curved surface that non-linearly separates data of different categories, data far from the center of distribution of data (unlearned data). -), There is a danger of inducing erroneous identification because the output is as if it were any pattern.

FIG. 13 shows the elimination of an unlearned pattern by a conventional template. The hatched area is an exclusion area by a template of a target pattern. As shown in FIG. 14, learning data or unlearned data is distinguished in an area of the specific pattern within the identification plane. In FIG. 15, reference numeral 100 denotes an identification surface by a neural network. This identification surface 100 is an open surface, and has a problem that it is always classified into an unlearned pattern and a learning pattern area somewhere. doing. FIG. 15 shows an identification space formed by the neural network, and shows how each learning pattern is located in the identification space. Paying attention to the specific pattern, FIG. 15 illustrates the slab value in the specific identification plane. In particular, FIG. 15 checks whether a pattern determined by the neural network is a learned pattern or an unlearned pattern, based on a known learning pattern area. Also,
FIG. 13 shows this in detail by focusing on the slab value which is an input value of a specific neural network. FIG. 14 shows how the specific slab value of the inclusion relation of the region in FIG. Is shown. In this figure, the unlearned pattern is located farther away from the target pattern and the learned pattern determined by the neural network data.

[0008] With such a conventional neuro discriminator, Italian neuro (neural network learned with Italian banknote), French neuro (neural network learned with French banknote), Spanish neuro (with Spanish banknote). Learned neural network)
Evaluation banknotes of Italy (8 denominations × 4 directions), France (5
When the rejection rate of the unlearned pattern was actually carried out for the denomination of 4 kinds of denominations, Spain (4 directions of 7 denominations), and blank paper (wrinkled, no wrinkles), the following Table 1 is obtained.

[0009]

[Table 1] As is clear from Table 1, the conventional discrimination device using the template or the trapping process using the threshold value cannot completely eliminate 100% of the banknotes in the unlearned pattern.

To solve the problem which can be said to be fatal in the neural network identification as described above, a statistical method can be considered. Conventionally, there is an example using a statistical test for each variable. Without taking into account the dependencies between the variables, inflexible discrimination has to be performed. However, bills and the like are usually printed from a single plate, and the causes of pattern variations are not always random, and there is often a large correlation between the elements.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a method for reliably discriminating an unknown image pattern such as an unlearned paper sheet in neuro discrimination. An object of the present invention is to provide a method and an apparatus for determining an unknown pattern paper sheet image which can be excluded from a determination target or excluded from a determination result.

[0012]

SUMMARY OF THE INVENTION The present invention relates to a method of judging an image of an unknown pattern in neuro discrimination, and an object of the present invention is to convey a plurality of types of sheets and collect them by a sensor. The slab value obtained by processing the obtained image data with a plurality of masks is input to a neural network to determine the type of the paper sheet. The image of the learning pattern is represented by a multi-dimensional Gaussian density function in the discrimination space, and the image of the paper sheet input to the neuro discriminator at the time of discrimination is represented by the multi-dimensional Gaussian density function and a threshold plane. Is determined by identifying the learning pattern and the unlearned pattern with respect to the input paper sheet, and eliminating the banknote of the unlearned pattern. It is more effectively achieved by expressing the multidimensional Gaussian density function by the following equation 5 of n dimensions.

Further, when the present invention is used as a post-processing of a neural network as a realizing means, the above-mentioned neural network is used.
The multidimensional Gaussian density function may be calculated for the image pattern corresponding to the paper sheet determined by the ral network, and the banknote of the unlearned pattern may be excluded.
Alternatively, when the present invention is used as preprocessing of a neural network, the multidimensional Gaussian density function for each of the image data is calculated using the slab values, and the unlearned pattern is eliminated to perform the learning. Only the pattern may be input to the neural network.

Further, according to the present invention, a slab value obtained by processing a plurality of types of paper sheets by a plurality of masks on image data collected by a sensor is input to a neural network. The object of the present invention is to provide a neuro discriminator for discriminating the type of the paper sheet, and to provide a multi-dimensional image pattern corresponding to the paper sheet determined by the neural network. Computing means for computing a Gaussian density function, and comparison determining means for directly outputting the result of the determination when the value of the multidimensional Gaussian density function computed by the computing means is equal to or greater than a predetermined threshold value. Or calculating means for calculating a multidimensional Gaussian density function for each of the image data with the slab value, and calculating a value of the multidimensional Gaussian density function calculated by the calculating means with a predetermined threshold value that's all Since when the image de - data of the slab values only the news - Rarunettowa - also achieved by having a comparison judgment means for inputting a click.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention can be applied to all paper sheets on which an image pattern is printed.

The multilayer perceptron has hitherto been the most widely used pattern
This is a neural network used for network identification. This method has a very good learning ability for separating a boundary surface between patterns to be identified. However, as described above, for an unknown pattern outside the learning pattern, the conventional identification method is used. Similarly, there is a problem that it cannot always be eliminated by a neural network. Therefore, in the present invention, a probability density function (for example, a multidimensional Gaussian density function) is estimated and used in order to accurately identify a learning pattern and an unknown pattern and to eliminate the unknown pattern from identification targets. Solves the above problem. The present invention focuses on the principal component of each pattern, detects unknown patterns at high speed, and
The identification result of the ral network is made more reliable.

The conventional linear discrimination processing is not multidimensional, but is a discussion of only the inclusion relationship between a learning pattern and an unlearned pattern only in a specific one-dimensional or two-dimensional discrimination plane. (See FIGS. 13 and 15). On the other hand, in the present invention, it is assumed that each pattern of the learning bill forms a subspace composed of a multidimensional Gaussian density function in the discrimination space, and the learning pattern is discriminated by the multidimensional Gaussian density function. Approximate space. Conceptually, as shown in FIG.
Mt. Fuji-shaped solid (subspace by multidimensional Gaussian density function) forms one discrimination pattern in three-dimensional space,
The unlearned pattern (white oval) is located at the far bottom. That is, by approximating the discriminant space by the learning pattern with the multidimensional Gaussian density function, the learning data shows a high value of the multidimensional Gaussian density function, and the unlearned pattern shows the value of the Gaussian density function. The value decreases. Therefore, even if the data of the unlearned banknote is classified into any of the identification areas of the learning pattern by the neural network, the value of the multidimensional Gaussian density function is calculated,
As shown in FIG. 2, the threshold plane 101 makes it possible to clearly distinguish a learned pattern from an unlearned pattern.

As shown in FIG. 3, when the present invention is used as post-processing, an image of a bill 11 is processed by an image collection unit 12 composed of an image sensor or the like to obtain a slab value.
Input to the neural network 10 and the neural network
The bank 10 obtains a judgment pattern in which the banknotes 11 are relatively identified with each other, and the value of the multidimensional Gaussian density function corresponding to this judgment pattern is calculated by the calculating means 13 based on the slab value. The multidimensional Gaussian density function in this case is the learning pattern of the bill 11.
If the value is as high as the value of the banknote, the comparison determining means 14 determines that the evaluated banknote is within the range of the pattern of the banknote given at the time of learning, and determines the result of the discrimination of the neural network 10 as the final discrimination result. And The parameters required for the probability density at this time are obtained in advance from the learning data, and only the patterns determined by the neural network 10 need to be calculated and checked.

When the present invention is used as preprocessing as shown in FIG. 4, the values of the multidimensional Gaussian density function for all the learning patterns obtained by the image collecting unit 22 are calculated based on each slab value. The calculation is performed by the calculating means 23. In this case, if all the values of the multidimensional Gaussian density function are much smaller than those of the learning pattern of the banknote 21, the comparison and judgment means 2
No. 4 judges that the evaluation banknote is not any of the banknotes given to the learning, and at this time, the discrimination in the neural network 20 is not performed, and the evaluation banknote is excluded. That is, it is possible to control the neural network 20 so that the identification operation is not performed.

Next, in order to facilitate understanding of the present invention,
The concept of identification in a two-dimensional space will be described with reference to FIG. Two regions 110 and 120 shown by knitting indicate regions divided by an identification curve (separation surface) 102 obtained by a neural network or the like, and the hidden unit output of the neural network increases as the distance from the separation surface increases. It gradually approaches "0" or "1". The final output of the neural network is a sigmoid function with a weighted sum obtained by multiplying the output of the hidden unit by a constant. Therefore, in the case as shown in FIG.
As the distance from the neural network increases, the neural network becomes "1".
(Or "0") is output, and an output is made that strongly asserts that the pattern is the same. In the case where a feature amount is extracted from a bill and a denomination is determined using the feature amount, it is assumed that the feature amounts of bills of the same type form a pattern. When learning a discrimination model using a plurality of types of banknotes, learning data exists like a nebula in outer space, and the neural network successfully creates a boundary surface between the learning data. However, there may be an area outside of this area where other denomination banknote data (for example, foreign banknotes, old banknotes, etc.) were not taken into account. In the case of FIG. 5, the reliable area is only in the ellipses A and B with respect to the classification obtained from the bill data. A region far away therefrom is a portion corresponding to neither pattern of banknote data or banknote data of another pattern. As described above, when there are various other patterns, it is not possible to presuppose the existence area of the pattern. However, the degree to which the data of a certain pattern is close to the dense area can be represented by a probability density function. As a method of obtaining a probability density function from learning data, in the case of a simple distribution, a Gaussian distribution may be obtained by maximum likelihood estimation, and the Gaussian distribution may be used. This is because if a simple Gaussian density function can be used, the exponent calculation can be omitted, and extremely high-speed determination calculation can be performed.

Incidentally, the problem of identifying the pattern to which the original image belongs from information changed by deterioration, movement or rotation due to various factors, such as bill identification, is common in industrial or industrial applications. In particular, when a large amount of data must be identified at high speed, a clear image that can be identified by simple pattern matching cannot be expected because of the strong restrictions imposed by the transport mechanism and imaging method. There are many. Conventionally, in the denomination of banknotes in an automatic banknote sorting machine, an identification method by a neural network using a mask and optimization of the mask by a genetic algorithm have been performed. That is, first, an image such as a CCD
The image data obtained from the sensor is aggregated into coarse mesh data (pixel values in a rectangular area are added). Even if aggregated data is used, if all mesh data is used, the number of input points to the neural network increases, and redundant data with high correlation is obtained. Generalization to data other than data is reduced. Therefore, the present invention also employs a method in which a plurality of slab values obtained from a mask for extracting points having a high degree of contribution to determination are obtained by using a genetic algorithm, and the obtained values are put into a neural network for determination. (See Japanese Patent Application No. 8-70868).

Here, the operation of recognizing a bill will be described. The denomination and the direction of the bill are identified by using a neural operation of a neural network. Image data from the image sensor of the discriminator is processed by an image processing unit, and parameters used for neuro discrimination are stored in a memory in advance, and necessary data is issued by a command from a CPU or the like. Only the data is read out to the image processing unit and the neural network. The discriminating process in the discriminating machine is performed according to the flow shown in FIG. 6. First, a neuro operation is performed in a neural network (step S30), and discriminating judgment process is performed (step S40). By processing the trap, it is determined whether or not the slab value sequence of the banknote identified this time falls within the range of the density trap indicating the self-recognition area of the pattern (steps S41 and S42). When the loss is identified, a damage neuro operation is performed (step S43), and a damage determination process is performed (step S44).
The trap is a mechanism for excluding the ones that have been learned by the neuro operation, and the discriminator judges whether or not the judgment of itself is correct based on the data of the self-recognizable area for each bill pattern.

In the neuro operation, a matrix operation is performed on the input value (slab value) obtained by the image processing and the weight data read from the memory, and a sigmoid calculation is performed on the image sensor on the matrix operation result ( See FIG. 12). The neuro operation is performed according to the flow shown in FIG. 7. First, the matrix operation from the input layer to the intermediate layer is performed according to the following equation (step S31).

[0024]

X = Wo · slab where Wo is a weight coefficient matrix from the input layer to the intermediate layer, sl
ab is a slab value (created by image processing).

Next, the output value mout of the intermediate layer is expressed by the following equation (2).
(Step S32).

[0026]

Mout = 1 / [1 + exp {(− x + θ) / Tn}] where θ is a threshold and Tn is a temperature gradient of a neural network.

When the output value sequence mout of the intermediate layer is obtained by the calculation of the sigmoid function of the above equation 2, a matrix operation from the intermediate layer to the output layer is performed according to the following equation 3 (step S33).

[0028]

X = W1 · mout where W1 is a weight coefficient matrix from the intermediate layer to the output layer.

After the matrix operation of the above equation (3), a sigmoid calculation is performed by the following equation (4), and the output value sequence n of the output layer is calculated.
out is obtained (step S34).

[0030]

Nout = 1 / [1 + exp {(− x + θ) / Tn}] Here, the neuro operation in the above-described discriminator uses the method proposed in Japanese Patent Application Laid-Open No. Hei 7-22087 by the present applicant. ing.

In the present invention, the learning pattern of a banknote is represented by a multidimensional Gaussian density function as shown in FIG. 1, and the learning pattern and the unlearned pattern are represented by a threshold plane 101 as shown in FIG. And the data to be used is an n-dimensional vector. A multidimensional Gaussian density function is used to determine the degree of membership for each type using the probability density function.
It is assumed that a slab value vector X (one pattern is composed of 50) has an n-dimensional Gaussian density function p (x). This is shown in Equation 5.

[0032]

(Equation 5) Here, it is assumed that m is a mean vector indicating the center of the distribution, M is a covariance matrix, and is a positive definite matrix. The superscript T indicates transposition of a vector.

Whether a pattern is rejected as a pattern of the corresponding type, that is, discrimination between a learned pattern and an unlearned pattern as shown in FIG. 2 is a statistical test. An appropriate threshold constant ε may be determined by the method and may be determined based on whether or not the following equation 6 holds.

[0034]

P (x) <ε Next, a method of actually generating a probability model from bill data will be described. N banknote data collected and processed by the image sensor are used, and the banknote data is set to vector data X = {x (1),..., X (N)}. Then, the likelihood function of the vector data X of the denomination (Equation 7 below)

(Equation 7) The parameter estimated values m 'and M' that maximize are called maximum likelihood estimated values and can be obtained by the following equations 8 and 9, respectively.

[0035]

(Equation 8)

(Equation 9) n-dimensional Gaussian density function p (x) is a threshold constant ε
If it is smaller, it is desired to judge that the data x is a vector that does not belong to this pattern.
The dimensional Gaussian density function p (x) is

(Equation 10) This can be rewritten as the following equation 11 using the constant S. That is, if the vector of a certain evaluation bill data satisfies the following expression 12, it is a learning pattern,
If not, it can be determined that the pattern is unlearned.

[0036]

[Equation 11]

(Equation 12) The rank of the matrix M ′ of the parameter estimation values obtained here (independent degrees of freedom as indices indicating the independence of the elements (small matrices) constituting the matrix) falls to a value smaller than the number of dimensions N. Equation 7 may not be able to be calculated. The reason may be as follows.

(1) The rank of the matrix M 'does not exceed the number N of data no matter what data is obtained from the definition of the matrix M' of parameter estimation values. Therefore, if the number of data is lower than the number of dimensions N of the vector, it is essential that a rank drop occurs.

(2) It is considered that the deterioration due to the displacement during the imaging process is the deterioration that can be described by affine transformation. If the data is used collectively for each area, the deterioration becomes somewhat complicated, but the rank of the data increased by such factors is not so large.

(3) Data elements are linearly dependent. For example, when the sum of the values of all the elements is used as one element of the feature vector as a value indicating the density of the entire image, the rank is obviously lowered by one.

For the reasons (1) to (3) above, if it is easy to obtain many data by repeating experiments, it is necessary to collect as much data as possible to increase the rank. desirable. Therefore, it is necessary to find the subspace formed by the vector data X and redefine the likelihood function on the subspace.

Here, the subspace formed by the vector data X is obtained by diagonalizing the matrix M 'of the parameter estimation values, and the probability density function on the subspace is obtained to obtain the data. We will calculate the degree of belonging. That is, in this method, the number of feature patterns smaller than the number of dimensions N is determined by linear combination of the elements of the pattern vector X.
And a determination is made using the characteristic pattern.

The eigenvalue of the covariance matrix M is defined as λ1 ≧ λ2 ≧
.Gamma..lamda.n, and eigenvectors of length 1 corresponding to these eigenvalues are u1, u2,..., UN. The condition number allowed to select an axis important for expressing this data is set as Tm,

[Mathematical formula-see original document] p = max (i)

Assuming that λi / λ1> 1 / Tm, the relationship between the eigenvalue λi and the eigenvector ui is

[Mathematical formula-see original document] M'.uj = [lambda] j.ujj = 1,. The obtained uj is the j-th data distribution.
It is the main component. Therefore, the (N × p) matrix U = [u1, u
2,..., Up], the following equation 16 holds.

[0043]

M ′ · U = U · ΛpΛp = diag (λ1, λ2,..., Λp) Since U (n × p) is an orthogonal matrix and U ^T U = I, U ^T M ′
U = Λp. Assuming that the axes of the p-dimensional subspace obtained here are y1, y2,..., Yp, the data x is

Equation 17] is converted ^{y = U T (x-m'} ), the probability density function of y is the following Equation 18.

[0044]

(Equation 18) From this, the following equation 19 is obtained.

[0045]

[Equation 19] From the above, the condition of Expression 12 is equivalent to satisfying the following Expression 20. Note that S ′ is a certain constant, and is a value that should be set so as to minimize erroneous identification through experiments.

[0046]

(Equation 20) In the above algorithm, the condition number Tm is determined in advance, and the calculation in FIG. 8 is performed offline in the learning process. That is, first, the parameter estimation values m 'and M' are calculated according to Equations 8 and 9 (Step S1), and the eigenvalues of the parameter estimation values M 'are obtained, and λ1 ≧ λ2 ≧... ≧ λn
.Gtoreq.0 (step S2), and eigenvectors corresponding to the eigenvalues .lambda.i are also determined correspondingly.
un (step S3). Next, p (the number of axes required to represent each component) is determined so as to satisfy the condition number Tm (step S4), and the following equation 21 is determined (step S5).

[0047]

(Equation 21) Then, a constant S 'is determined so as to maximize the discrimination rate by using the learning data, and the threshold value C = -2 log.
S ′ (Step S6).

In the evaluation process, the calculation shown in the flow of FIG. 9 is performed online as a vector for evaluating x.
That is, first, yi when x is an input value (slab value) of the evaluation bill data to the neural network is calculated according to the following equation 22 (step S10).

[0049]

(Equation 22) Then, it is determined whether the following Expression 23 is satisfied (Step S11).

[0050]

(Equation 23) If Equation 23 is satisfied, the banknote of the evaluation data is used as a learning pattern, and this slab value is given to a neural network for identification (Step S12). The banknote of-is made an unknown pattern, and this banknote is eliminated (step S13).

By the way, despite the fact that the printed matter has extremely high similarity, the largest cause of the variation in the data of the bill is a displacement or an inclination caused in a conveying process in which the bill is inserted into the identification device and imaged. It is considered that the deterioration can be expressed mathematically by the affine transformation. Due to these causes, the rank of the data hardly increases. However, since this data is a slab value output in the mask processing at the time of image collection, the rank of the data is slightly increased due to variations. However, other factors such as deterioration, such as dirt, wrinkles, imaging conditions, etc. Considering this, the rank of the main components of the data is quite low.

An example of actual bill discrimination will be described as an example applied to Italian lira bills. Italian lira bills are 1000, 2000, 5000, 10,000, (N) 5
0000, (O) 50,000, (N) 100,000
(O) 100000 total 8 denominations, (N) shows a new bill and (O) shows an old bill. In consideration of the orientation and front and back when the bill is inserted into the discriminating machine, each denomination handles four types of images as separate images. A, B,
C and D are A (table), B (table obtained by rotating A by 180 degrees), C
(Turn over without changing the position of the short side of B), D (C is 180
Degrees). Of the data collected from 80 bills for each denomination, 60 are used for learning data used for learning parameters, and the remaining 20 are used for evaluation. The data were not collected at the same time, but were obtained from 50 and 30 images with slightly different image sensor sensitivities.
Sheets and 20 sheets as learning data, and the remaining 20 sheets as evaluation data
Used as the In addition, banknotes from 16 countries including Italian banknotes were additionally used for evaluation. Table 2 shows the denominations.

[0053]

[Table 2] In the present embodiment, a feature vector is extracted from a bill by the following procedure. That is, the banknote moving at a high speed by the transport mechanism is sampled by the image sensor to sample the density, taken in as image data, the center position is determined, and a rectangular area of a size determined vertically and horizontally is cut out from the center. The rectangular area is a shaded image in which several tens of pixels are arranged vertically and horizontally. This rectangular area is divided into a mesh-like small area consisting of a total of 48, 6 in length and 8 in width. Forty-eight types of mask processing (for optimization of a mask pattern to be applied, refer to Japanese Patent Application No. 8-70868) are performed on the rectangular area composed of each mesh-shaped small area, and each of the small areas not covered by the mask is masked. Slab values are obtained by adding the data of A 49-dimensional feature vector is created from 48 slab values obtained by applying these different 48 types of mask patterns and slab values obtained without applying any mask. In this case, since the scaling according to the size of the bill is not performed, the large-sized bill is data only in the vicinity of the center. The input may be rotated. It is possible to make a soft correction to this rotation, but in the present invention, no correction was made, and it was confirmed whether the deterioration caused by the correction could be absorbed by the present invention.

Here, the purpose is to eliminate different types of data for each denomination of Italian banknotes. Therefore, for the eight types of banknotes, a fraction matrix using learning data for each denomination k is used. Mk was determined and its eigenvalue was determined.
Table 3 shows the eigenvalues of the matrix M 'of the parameter estimation values based on the learning data of 5000 lira.

[0055]

[Table 3] Then, the eigenvalues λi and the eigenvectors ui are arranged in descending order of the eigenvalues, and the following Expression 24 is obtained. FIG. 10 shows the result.

[0056]

(Equation 24) The horizontal axis represents an average vector m indicating the center of the distribution in Equation 24, and indicates how many components from the orthogonal axis having the largest component are included. The vertical axis (logarithmic axis) represents the value of Equation 24. What is displayed here is data in one direction (direction C) on the back of the 5000 lira bill. One curve corresponds to one bill. The solid line is the value corresponding to the learning data, the dotted line is the value corresponding to the evaluation data of the same denomination, and the one-dot chain line is the value corresponding to the bill of another country.

From FIG. 10, the following observation can be made. That is, the learning data is suppressed to a small value, and in the evaluation data, the numerical value of the evaluation data jumps from the vicinity where 40 or more pieces are used. This is because energy in the direction of minute components in the learning data is evaluation data.
It does not match that of the data. For other types of banknotes, the discrepancies are large in all the eigenvector directions, and they can be clearly identified.

Thus, it can be seen that stable results can be obtained when the number of axes is reduced from about 5 to about 40. In the experiment,
The axis was selected with the condition number set to Tm = 1000. Table 4 shows the actual determination results of z (z: a value corresponding to an n-dimensional probability density function) by an Italian banknote.

[0059]

[Table 4] In Table 4, α is evaluation data 2 for the same denomination of Italian banknotes.
The value of the n-dimensional probability density function equivalent value z for 0
The maximum value of z) on the left-hand side of 3, β indicates the minimum value for evaluation data of other denominations of Italian bills and for banknote data of other countries. If α <β, it is shown that if C (threshold) that satisfies α <C (= −2 logS) <β is used, all the evaluation data used here can be correctly excluded. The case where it cannot be excluded is shown in the notes column. However, these are due to the new and old Italian banknotes.

In the above experiment, when the condition number Tm = 1000, when it is determined that the denomination is not an old banknote with four denominations, only one banknote of the same amount and in the same direction cannot be excluded. This occurs because the image patterns are very similar. FIG. 11 shows the value of z for 100,000 lira where (O) misidentification has occurred. The dashed-dotted curves represent values for banknotes of other countries. The solid line at the bottom is the learning data, and the dotted line is the evaluation data of the same denomination.
To the data. The four curves indicated by dashed lines in the middle are the results for a new kind of banknote of the same amount. This intersects with one graph in the old banknote, and it can be seen that erroneous identification occurs depending on the number of axes.
On the other hand, when the number of axes is determined to be relatively small (about 7 to 25), the number of errors is small. This is because the characteristics of the banknote are gathered on a relatively small number of main axes, and if an axis with a low contribution is included, the noise component increases and the judgment is erroneous. Therefore, the condition number Tm = 10
00 is not always appropriate, and a smaller value reduces false identification. However, even if it is determined that the banknotes are erroneously identified, these banknotes that have been erroneously identified are accepted as their regular patterns in the identification of each pattern by the neural network. −
If the identification by the ral network is used together, the identification will be correct.

In the above description, banknotes have been described. However, the present invention can be applied to paper sheets having images such as checks and securities.
The data of genuine and genuine bills are known, but the data of damaged and counterfeit bills are basically unknown, which is effective for determining the degree of damage to paper sheets. In addition, specific personal information is known, but information other than individuals is unknown, so that application to personal authentication technology is effective.

[0062]

As described above, according to the present invention, according to the conventional one-dimensional or two-dimensional input value (slab value),
Since the range check is calculated by a multidimensional Gaussian density function, unlearned patterns of paper sheets can be eliminated while maintaining the discrimination ability (non-linear relative discrimination) of the neural network. Most of the neural network can be calculated in advance using the data of the learning pattern, and the result can be stored in the actual machine as a parameter for each pattern. By simply arranging it in the post-processing section, the effect on the actual machine can be improved. Because the subspace of the learning pattern formed by the multidimensional Gaussian density function is nonlinear,
So-called absolute discrimination, in which only the self is known and non-self is nonlinearly excluded, becomes possible. Also, by arranging it in the pre-processing unit of the neural network identification operation, the effect in the actual machine can be obtained.

[Brief description of the drawings]

FIG. 1 is a diagram schematically illustrating the principle of a neural network identification space according to the present invention.

FIG. 2 is a diagram showing a determination principle of the present invention.

FIG. 3 is a block diagram showing an example of an apparatus for implementing the present invention in post-processing of neuro-identification.

FIG. 4 is a block diagram showing an example of an apparatus for carrying out the present invention in neuro-identification preprocessing.

FIG. 5 is a diagram for explaining a determination principle of neuro identification.

FIG. 6 is a flowchart showing an operation example of neuro bill discrimination.
It is.

FIG. 7 is a flowchart showing an operation example of a neuro operation.

FIG. 8 is a flowchart showing an operation example of an offline learning process of the present invention.

FIG. 9 is a flowchart showing an operation example of an online evaluation process of the present invention.

FIG. 10 is a diagram illustrating an example of evaluation of a 5000 lira bill in Italy.

FIG. 11 is a diagram showing an evaluation example of 100,000 lira bills in Italy.

FIG. 12 is a diagram showing a configuration example of a conventional neural network.

FIG. 13 is a diagram showing an exclusion region of unlearned data with respect to a template of a conventional learning pattern.

FIG. 14 is a diagram illustrating a principle of eliminating an unlearned pattern in the related art.

FIG. 15 is a diagram schematically showing classification of an unlearned pattern by a conventional neural network.

[Explanation of symbols]

10, 20 neural network 11, 21, banknote 12, 22, image collecting unit 13, 23 computing means 14, 24 comparison / determination means 100 discriminating curved surface by neural network 101 threshold plane 102 discriminating curve

Claims (6)

[Claims] 1. A slab value obtained by processing a plurality of types of paper sheets and conveyed by a sensor with image data collected by a sensor and inputting the processed slab value to a neural network. In neuro discrimination for discriminating the type of leaf, an image of the learning pattern of the paper sheet is represented by a multidimensional Gaussian density function in a discrimination space, and input to a neuro discriminator at the time of discrimination. The image of the paper sheet is represented by the multidimensional Gaussian density function and determined by a threshold plane, whereby a learning pattern and an unlearned pattern are identified for the input paper sheet, and the unlearned pattern is identified. A method for judging an image of an unknown-pattern paper sheet in neuro-identification, wherein the paper sheets of the pattern are excluded. 2. The method according to claim 1, wherein said multidimensional Gaussian density function is represented by the following equation (5) in n dimensions. 3. The multidimensional Gaussian density function is calculated for an image pattern corresponding to a paper sheet determined by the neural network to eliminate the unlearned paper sheet. 3. The method according to claim 2, wherein the pattern identification is performed on an unknown pattern paper sheet image in neuro identification. 4. The multidimensional Gaussian density function for each of the image data is calculated with the slab value, and the unlearned pattern is calculated.
3. The method according to claim 2, wherein only the learning pattern is input to the neural network by excluding the learning pattern. 5. A slab value obtained by processing a plurality of types of paper sheets and conveyed by a sensor with respect to image data collected by a sensor is input to a neural network, and the slab value is input to the neural network. A neuro discriminator for discriminating the type of leaves, a calculating means for calculating a multidimensional Gaussian density function for an image pattern corresponding to the paper sheet discriminated by the neural network; And comparing and judging means for directly outputting the judgment result when the value of the multidimensional Gaussian density function calculated by the calculating means is equal to or more than a predetermined threshold value. An unknown pattern paper sheet image determination device. 6. A slab value obtained by processing a plurality of types of paper sheets and conveyed by a sensor with image data collected by a sensor and inputting the slab value to a neural network. A neuro discriminator for discriminating the type of leaves, a calculating means for calculating a multidimensional Gaussian density function for each of the image data with the slab value; and the multidimensional Gaussian calculated by the calculating means. A neuro-discriminating means for inputting only the slab value of the image data to the neural network when the value of the density function becomes equal to or greater than a predetermined threshold value. For determining an unknown pattern paper sheet image in the above.