JPH05324837A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To provide the pattern recognizing device which performs pattern recognition for paper money, etc., by a small-scale neural network without reference to an up-down or right-left position shift by reducing pattern image data measured by a sensor by using plural rod type masks. CONSTITUTION:This pattern recognizing device is provided with the sensor 2 which measures a pattern image of an object 1, a preprocessing part 4 which obtains plural image representative values (slab value) for the object 1 by the rod type masks 41-4n, a separate arithmetic part (neural network) 5 which calculates separate arithmetic values every decision patterns by using a weight coefficient, and a decision part 6 which decides the pattern image of the object 1 from the separate arithmetic values. The weight coefficient is previously learnt with a teacher by the neural network, so the object 1 can securely be recognized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for recognizing patterns of coins (coins, banknotes), characters, etc. by a neural network, and in particular, by covering with a rod-shaped mask, it is possible to keep the positional deviation of the pattern image unchanged. In addition, the present invention relates to a pattern recognition device that reduces the size of pattern image data and efficiently performs separation calculation.

[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.
The 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. 12, 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
The 64 pieces of pixel data are directly input to the separation calculation unit 5. At this time, 64 neuro elements corresponding to each pixel are required as the neuro elements in the input layer of the separation calculation unit 5. Then, a pattern is recognized using a neural network based on this input information.

[0003]

However, in such a conventional recognition device, 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 the number of pixels. Will need the same number as. Therefore,
It is necessary to provide an input layer having a large number of neuro elements, and correspondingly, a large number of neuro elements must be provided in the hidden layer as well, and eventually pattern recognition is performed unless a very large-scale separation operation unit 5 is used. I didn't get it.

The present invention has been made under the circumstances as described above, and an object of the present invention is to reduce the number of input information (hereinafter referred to as slab number) input to the input layer of the separation calculation unit. Make sure you can recognize the pattern,
It is an object of the present invention to provide a pattern recognition device that can realize a reduction in size of a separation calculation unit and can reliably recognize a positional deviation.

[0005]

DISCLOSURE OF THE INVENTION The present invention relates to a pattern recognition device, and an object of the present invention is to provide a sensor for optically or magnetically measuring a pattern image of an object to be recognized and a large number of rectangular shapes. With a plurality of rod-shaped masks that cover the rod-shaped area of the rectangular small partition with a small section, a pre-processing unit that converts the pattern image data measured by the sensor into a plurality of image representative values, and the plurality of An image representative value is input in parallel, and a separation calculation unit that calculates a separation calculation value for each judgment pattern of the pattern image by a weight optimally adjusted in advance for the judgment pattern classification, and a maximum value among the separation calculation values or This is achieved by including a determination unit that determines and outputs the determination pattern having the minimum value as the pattern image of the object.

[0006]

According to the present invention, pattern recognition of currency or the like is performed by a neural network having a learning function, and pattern image data optically or magnetically measured by a sensor is reduced by using a plurality of rod-shaped masks. Obtain the image representative value. The rod-shaped mask has a large number of rectangular small sections, and a plurality of rod-shaped regions of the rectangular small sections are covered.
The image representative value reduced by such a rod-shaped mask is invariable to the positional deviation of the pattern image and is input to the separation calculation unit (neural network), and the pattern image is adjusted by the weight optimally adjusted for the judgment pattern classification in advance. The separation calculation value is calculated for each of the determination patterns of. The pattern image is determined according to the maximum value (or the minimum value) of the separation calculation values. As a result, the structure of the neural network becomes small and the control device becomes small.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In the present invention, when pattern recognition of currency or the like is performed by a neural network, information obtained from a sensor of a recognition device has been directly input to the neural network. Therefore, 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. For this reason,
In the present invention, a plurality of rod-shaped masks for reducing the input information are provided in the preprocessing unit, and the image representative value (slab value) is obtained from each rod-shaped mask. The rod-shaped mask has a large number of rectangular small sections, and has a structure in which a plurality of rod-shaped regions are covered in the horizontal direction or several vertical directions.

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of a currency identifying device of the present invention. A banknote 1 to be recognized is measured by an image sensor 2 composed of a CCD or the like, and an image is appropriately processed to obtain an input image 3. The input image 3 is the mask 4 in the preprocessing unit 4.
1 to 4n, the sum values from the masks 41 to 4n are calculated, and the image representative values SB1 to SBn are obtained. The image representative values SB1 to SBn are input to the separation calculation unit 5 of the neural network, and the separation calculation value SP is calculated by the weight optimally adjusted in advance for the pattern classification of the judgment bill. The separation calculation value SP is input to the determination unit 6, and the pattern having the maximum value (or the minimum value) in the separation calculation value SP is output as the pattern image of the bill 1 of the object. The masks 41 to 4n in the pre-processing unit 4 have a plurality of rectangular sub-sections, and the bar-shaped areas in the vertical or horizontal direction of the sub-sections are masks of bar-shaped areas randomly coated with, for example, uniform random numbers. Yes, each is coated for a different rod-shaped area.

First, the pretreatment section 4 using a rod-shaped mask will be described. The reason why the rod-shaped masks 41 to 4n are used in the present invention is as follows. As shown in FIG. 2, in a binary image of “0” and “1” on an 8 × 8 matrix, a slab value that is the sum of pixel values as a feature amount of the image (in the case of FIG. If the number of 1 ") is used, FIG.
"A" is obtained as a value characterizing the character "E" in (A), and "12" is obtained as a value characterizing the character "H" in FIG. Therefore, "E" and "H" can be separated by using the slab value. However, there are cases where the slab values are the same even in images having different patterns. For example, in FIG. 3A, the slab value of the character "F" is "10", but the slab value of the character "K" in FIG. 3B is also "10", and the characters cannot be separated. Such a problem can be solved by introducing a rod-shaped mask in which a specific rod-shaped region corresponding to a pixel of the input image as shown in FIG. 4C is covered. Figure 4
When the image in FIG. 3A is covered with the mask in FIG.
In this case, the slab value is "8". On the other hand, when the image of FIG. 3 (B) is covered with the mask of FIG. 4 (C), the image shown in FIG. 4 (B) is obtained, and the slab value is “9”. By introducing a rod-shaped 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 rod-shaped covered portions in the mask as shown in FIG. 4C 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 masks as described above. It is often the case that one of the slab value sequences differs between images, and by using various rod-shaped masks, it is possible to stochastically enhance the separation capability between images. The use of the plurality of different rod-shaped masks described above has the following physical meaning. That is, when observing a three-dimensional object while changing the viewpoint from another direction, different information can be obtained even for the same target. Similarly, the use of various masks changes the viewpoint in the two-dimensional plane to observe the image, and as described above, different information can be generated even in the same image. In this case, the input image is covered by various different rod-shaped masks in the pre-processing, and the sum of the pixels not covered becomes the corresponding slab values SB1 to SBn, which correspond to the neuro elements of the input layer one to one. Further, the unit value of the output layer corresponds to the judgment pattern.

FIG. 5 shows the effect of vertical bar masks such as the masks 41 to 4n, which are the same for the downward movement input image of FIG. 5A and the upward movement input image of FIG. It is possible to obtain a slab value of "5", and it is possible to obtain a slab value that does not change even if the input image is vertically displaced in the vertical direction. Figure 6 shows the effect of a horizontal bar mask,
The same slab value “6” can be obtained for the left-moving input image of FIG. 11A and the right-moving input image of FIG.
It is possible to obtain an invariable slab value even if the input image is horizontally shifted from side to side. FIG. 7 shows a vertical bar mask 4
It shows an example of a currency identification device provided with a preprocessing unit 4A having A1 to 4An and a preprocessing unit 4B having horizontal bar masks 4B1 to 4Bn.

Next, at the time of recognition, the number of types of the rod-shaped masks selected in advance, suitability for the contents, and points to be noted in the selection will be described. In the preprocessing unit 4 (4A, 4B), the mask types 41 to 4n and the masks 41 to 4n (4A
1 to 4An, 4B1 to 4Bn) are considered as parameters. Here, a parameter search is performed using 12 binary images of alphabets A to L drawn on an 8 × 8 matrix to examine the recognition ability of the present invention. However, the learning algorithm of the neural network uses the backpropagation method given by the following equation 1.

[Equation 1] In addition, 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 smaller than the convergence judgment error.
Alternatively, the number of presentations reaches the maximum number of presentations. Here, the number of presentations is defined as one when the teacher is presented in all the patterns A to L. The learning data sequentially presents A to L 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 the recognition ability are examined according to the type of mask, that is, the number of slabs. First, the rod-shaped mask used in the present invention is prepared as follows. That is, a uniform random number of 8 which is the same as the number of rows or columns of the matrix of the input image is [-
, 1], and covers the row or column region of the number corresponding to the negative random number value therein. Various rod-shaped masks are created by changing the initial value of uniform random numbers by such a method. The number of slabs examined here is 2, 3,
There are 6 ways of 4, 5, 6, 7. FIG. 8 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. As shown in FIG. 8, when the number of slabs is “2”, learning does not converge and pattern separation is impossible. When the number of slabs is "3", the learning of the neural network shows a convergence tendency to some extent, but the error curve is oscillating. In this case, continue learning, and
The operation was repeated up to 10,000 times, but the squared error did not fall below 1.0. If the number of slabs is "4" or more,
It was shown that the learning converged. Further, in the case where the alphabets A to L are recognized by using this weight, even in the unit value corresponding to the determination pattern of the output layer, the recognition ability is almost the same when the number of slabs is "4" or more. It was found that it is possible to obtain an output value sufficient for pattern recognition. Therefore, in order to recognize the letters of the alphabet (A to L), it is possible to recognize if the mask type is 4 or more, and it is sufficient for the input layer of the separation calculation unit 5 to have 4 or more slabs. Conventional input pixel 64
Compared with the case where (= 8 × 8) is input to the input layer, the number of neuro elements can be extremely reduced, and the separation calculation unit 5 can be simplified accordingly. When the number of characters to be recognized is increased, the number of masks may be increased accordingly.

Next, the learning situation and the recognition ability will be described based on the masked area. First, the coverage area is changed by the fluctuation range of random numbers. That is, the width of the random number [-1, 1]
, The random number widths are [−2,1], [−3,1], and [−4,1] in the increasing direction of the covered area, and the random number width is [−, 1] in the decreasing direction of the covering area. 1, 1], [-1,
Random numbers are generated as 3] and [-1, 4]. However,
The number of slabs is "4". FIG. 9 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. Figure 9
From this, it is clear that learning does not depend much on the masked area. It was also found that the recognition ability does not depend much on the masked area. Therefore, it suffices to select the covering region of each rod-shaped mask at random to some extent and pay attention to the mask type without paying much attention to the covering region of each rod-shaped mask.

The number of slabs can be reduced by using the mask as described above not only in the case of characters but also in a light and shade pattern such as a bill. The effectiveness of the present invention is verified using the banknote data that is a grayscale image of 0 to 255.
However, the banknote data handled here is time-series data collected from a sensor layout as shown in FIG. 10, and is empirically representative of the characteristics of banknotes. The dotted rectangles in the figure indicate the areas where individual sensors sense by transporting banknotes. Here, the following two cases will be performed. One is the case of inputting time series data of banknotes (hereinafter referred to as time series data) to the present invention, and the other is Fourier transform of the time series data by the following equation 3 for each sensor, and the power spectrum ( Hereinafter, a Fourier power spectrum) is input to the present invention.

[Equation 3] The Fourier transform of f (k) is F (n), that is,
F (n) = A (n) + jB (n). Here, the reason why the Fourier power spectrum of the time series data is adopted is based on the following reason. That is, even in the bill validator handled in the experiment, bills are identified at 8 sheets / sec or more, and the transport speed is 600 mm / sec or more. The data collected by this identification device contains various errors due to the conveyance of banknotes, and it is expected that the weight of the errors due to sample deviation will be particularly large. Therefore, this error can be removed by using the Fourier power spectrum of the time-series data, which leads to improvement of the discrimination ability. In the configuration of the present invention, the number of slabs is 16 both when using time series data and when using its Fourier power spectrum. Therefore, the number of units in the input layer of the neural network is 16, and the number of units in the hidden layer is the same as that of the input layer. The unit value of the output layer corresponds to the judgment pattern, and the number is 12
Is. This is a true banknote that does not include counterfeit banknotes as the target of identification, and for each of the three denominations of 10,000 yen, 5,000 yen, and 1,000 yen, upright, upside down, upside down, upside down It means that it is classified into four conveyance directions. Note that the number of slabs when the time series data and the Fourier power spectrum are used is determined in consideration of the convergence of learning and the number of units by various experiments. FIG. 18 shows the result of bill identification performed under these conditions. However, in the figure, the FFT data means the Fourier power spectrum. As is apparent from FIG. 11, when time series data and its Fourier power spectrum were applied to the present invention, learning could be converged until the square error was 0.01 in both data. Here, the horizontal axis and the vertical axis in FIG. 11 represent the number of times the teacher has presented and the squared error, respectively. Further, when 10 pieces of banknote data each having a total of 12 patterns in 3 denominations and 4 conveyance directions different from the learning data were identified using these weights, the discrimination rate ES was 100% in all cases.

Therefore, when the banknotes are identified, if there are 16 kinds of masks, they can be identified with certainty, and the separation calculation unit 5
It is sufficient for the input layer to have 16 or more slabs.
In the conventional case, the input pixel 128 (= 4
Compared with the case where x32) is input to the input layer, the number of neuro elements can be extremely reduced, and the separation calculation unit can be simplified accordingly. The number of slabs mentioned above is 1
Not only in the case of 6, the learning speed decreases even if the number of slabs is slightly reduced, but the identification is possible, and more reliable identification is possible if the number of slabs is increased. Moreover, the number of slabs (the number of masks) can be increased according to the number of foreign banknotes as well as the identification of the three denominations of 10,000 yen, 5,000 yen and 1,000 yen × 4 conveying directions. You can easily increase it.

Next, a description will be given of the separation operation unit 5 which inputs the information preprocessed by the preprocessing unit 4 and performs the separation operation. The separation operation unit 5 having a hierarchical structure is roughly divided into an input layer, a hidden layer, and an output layer. The input layer is
Neuro elements are provided so as to correspond to the mask types from the pre-processing unit 4 in a one-to-one manner, and the pre-processed slab value (the number of pixels after mask processing (total sum) is supported by each mask type. The hidden layer is made up of at least one neuron layer, and plays a role of separating the information of the input layer and transmitting it to the output layer. The more hidden layers, the more the input layer It is possible to separate and operate each pattern invariantly with respect to changes in the information of each neuro element of the output layer.The output layer is provided with neuro elements so as to correspond one-to-one to the category to be identified. For example, when recognizing the letters of the alphabets A to Z, the neuro elements in the output layer correspond to A to Z in order from the top, and the corresponding neuro element has a value closest to 1. (Maximum value) and other neuro elements output values close to 0. When recognizing the letter A, the highest neuro element in the output layer outputs the maximum value. In this case, the neuro elements of the output layer are in order from the top, 10,000 yen (front right direction), 10,000 yen (front left direction), 10,000 yen (back right direction), 10,000 yen (back left direction), 5,000 yen (front There are 12 neuro elements in a one-to-one correspondence, such as right direction), ………, 1,000 yen (back left direction), and connect the neuro elements from this input layer to the output layer to output signals. The transfer mechanism is called a synapse.The synapse stores the strength of coupling between neuro elements by a weighting function.One neuro element receives a signal from a plurality of neuro elements in the preceding layer through the synapse and passes through it. Input value by multiplying the weight of the synapse To.
The neuro element sums the input values when it receives signals from all the neuro elements to which it is coupled. When the total value exceeds the threshold value set in advance for the neuro element, the neuro element "fires", sends an output signal to the neuro element of the next subsequent layer, and repeats this process to output information from the output layer. The weight of each synapse 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 discrimination section 6 will be described. The discriminator 6 inputs the information output from the output layer of the separation calculator 5, discriminates the maximum value (may be the minimum value) from the information, and corresponds to the neuro element having the maximum value. Distinguish from Cadegory. For example, if the neuro elements in the output layer correspond to the caddies from A to Z from the top, and if the highest neuro element outputs the maximum value, "A"
It is determined that For convenience of explanation, the patterns of characters and banknotes have been mainly described, but it goes without saying that the effect surface pattern is also possible and other patterns are also possible.

[0018]

As described above, according to the pattern recognition apparatus of the present invention, the necessary identification algorithm can be self-formed by the learning function using the neural network. Further, according to the present invention, since the image representative value (slab value) is obtained by using the horizontal and vertical bar masks in the preprocessing unit, the neural network has a small configuration and the control device has a small scale. Become. Further, since the neural network includes the non-linear element, there is an advantage that the recognition performance is less likely to deteriorate with respect to variations in the image data in the light and shade direction. Furthermore, the rod-shaped mask can realize pattern recognition using a neural network that is invariant to positional deviation.
In order to consider the recognition ability according to the present invention, a comparison is made with the conventional method using a binary image of alphabets from "A" to "L". The conventional algorithm has a three-layer structure as in the present invention. However, the number of units in the input layer is 64 (= 8 × 8), which is the same as the number of input pixels. Also,
The number of units in the hidden layer is 32, and the number of units in the output layer is 12, which is the same as the determination pattern. However, in the conventional algorithm, the number of hidden layer units was determined by various experiments in consideration of the recognition ability and the number of units. In the present invention and the conventional algorithm, the number of presentations when learning is performed until the squared error reaches 0.01 is 61,983 for the former.
And the latter was 978 times. However, for the former, ε = 0.01, α = 0.9, β = −0.1, and for the latter, ε = 0.1, α = 0.9, β = −.
Learning was performed with 0.1. Moreover, when the data used for learning was recognized again by using the above-mentioned weights by both algorithms, the discrimination rate ES was 100% in all cases. Here, the number of weights of the present invention is 8 × 8 + 8 × 12 =
160, while the number of weights in the conventional algorithm was 64 × 32 + 32 × 12 = 2,432. From these, it was revealed that although the present invention takes time to converge, the scale of the neural network can be reduced without impairing the recognition ability.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a pattern recognition device according to the present invention.

FIG. 2 is a diagram for explaining a slab value used in the present invention.

FIG. 3 is a diagram for explaining a slab value used in the present invention.

FIG. 4 is a diagram for explaining a rod-shaped mask used in the present invention.

FIG. 5 is a diagram for explaining a rod-shaped mask in the vertical direction.

FIG. 6 is a diagram for explaining a horizontal rod-shaped mask.

FIG. 7 is a block diagram showing another example of the currency identifying device according to the present invention.

FIG. 8 is a diagram showing learning situations for various slab numbers.

FIG. 9 is a diagram showing a learning situation for various covering areas of a mask.

FIG. 10 is a diagram for explaining banknote data and a collection method.

FIG. 11 is a diagram showing a learning situation of a Nutel network according to the present invention and a conventional method when time series data of banknotes and Fourier power spectrum data are used.

FIG. 12 is a block diagram showing a configuration example of a conventional pattern recognition device.

[Explanation of symbols]

 1 banknote (identification target) 2 image sensor 3 input image 4, 4A, 4B pre-processing unit 5 separation calculation unit 6 determination unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI Technical display location G07D 5/00 9340-3E 7/00 H 9340-3E

Claims (3)

[Claims] 1. A sensor for optically or magnetically measuring a pattern image of an object to be recognized, a plurality of rectangular small sections, and a plurality of rod-shaped masks covering a rod-shaped region of the rectangular small sections. By using a pre-processing unit that converts the pattern image data measured by the sensor into a plurality of image representative values, and the plurality of image representative values are input in parallel, the weights optimally adjusted to the determination pattern classification in advance, A separation calculation unit that calculates a separation calculation value for each judgment pattern of the pattern image, and a judgment unit that judges and outputs a judgment pattern having a maximum value or a minimum value among the separation calculation values as a pattern image of the object. A pattern recognition device characterized by being provided. 2. The pattern recognition device according to claim 1, wherein the object is a bill or a coin. 3. The pattern recognition device according to claim 1, further comprising a first rod-shaped mask in which the rod-shaped region is horizontal, and a second rod-shaped mask in which the rod-shaped region is vertical.