JPH05324839A - 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 by reducing pattern image data measured by a sensor by using plural ring-shaped 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 ring-shaped 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 an apparatus for recognizing a pattern of coins or the like by a neural network, and in particular, by covering with a ring-shaped mask, the pattern image data is invariant to rotation of the pattern image and the pattern image data is reduced. The present invention relates to a pattern recognition device that efficiently separates and calculates a pattern.

[0002]

2. Description of the Related Art Conventionally, when discriminating authenticity of coins, denominations, etc., coins are regulated so that they can pass through a predetermined passage or a conveying path at a constant speed, and a predetermined range of coins passing through is optically or magnetically determined. It is designed to be read. Then, by comparing and collating with reference data in a predetermined range for each coin registered in advance, the coin is identified by the degree of collation. For this reason, high-performance identification is difficult unless the coin transport at the time of registration of the reference data and the coin transport at the time of identification are in the same state. As described above, in the conventional coin identification method, it is necessary to regulate the coin delivery state at the time of registration and at the time of identification, which has a drawback that the delivery system becomes complicated.

On the other hand, Rumelhar
t) et al.'s error backpropagation method (Error Back P
Since then, the research on neural networks has been very actively conducted. The BP method is a learning method for a multi-layer feedforward network, and is based on Rosenblatt (Rose
nblatt) perceptron type learning and ADAL
INE (Adaptive Linear Neuro)
It is a general form of the Widrow-Hoff rule (delta rule) for n). In recent years, improvement of the BP method and applied research using it have been particularly active. The most applicable area of the neural network is related to pattern recognition. The reason for this is that the pattern separation ability of the multilayer network is excellent, and it is difficult to recognize the modified pattern by the conventional recognition technology. Reid et al. Use a high-order neural network in the preprocessing unit to form a system that is invariant to displacement, rotation, and scale conversion (enlargement / reduction) (Neural Networks, Vo).
l. 1, P. 689-692 (1989)), various rotations and the like are given to various misalignments, and a combinatorial explosion of the unit occurs, which is not practical. In addition to these, a number of invariant pattern recognition systems using neural networks have been proposed, but they simultaneously realize positional displacement, rotation, scale conversion, and shade invariant properties,
And a system at a practical neural network scale has not been proposed.

[0004]

As another approach to pattern recognition that is invariant to misregistration and rotation, mathematical processing is applied to an input pattern so as to extract a feature amount that is not affected by misregistration and rotation. Then, there is a method of inputting to the identification unit (for example, BP type neural network). As for this approach, Houg is used as a pretreatment.
Although there is a method of using h transform, moment invariance, Fourier Mellin transform, etc., the input image must be binarized in advance. Therefore, in order to handle an image of a gray level of several bits instead of a binary image as a pattern input, a coin is recognized which is considered to require image processing such as boundary extraction as a preprocessing of these mathematical conversions. In this case, if the target coins have different sizes, they can be easily identified. However, coins of almost the same size must be identified by the features (pattern) of the coin surface. The image input by the image sensor is
It is often in a state of being rotated at a random angle from a standard image. To recognize this rotated image, a rotation-invariant pattern recognition structure is required, but no practical method effective against this kind of problem has been proposed so far.

The applicant has proposed a currency identification method and apparatus (Japanese Patent Application No. 3-78513) for identifying currency by a neural network having a learning function. Since the slab value based on the scrambled image is created in the pre-processing unit on the assumption of the above modification, the pre-processing unit requires a large amount of calculation. Further, in the conventional neural network construction method, information obtained from the sensor of the identification device is 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 example, when 32 samples are taken by 4 sensors, 128 sensor information is held, but in the conventional neural network construction method, the 128 information is directly input to the neural network, and the network structure is divided into 3 layers. However, the scale is 128 × 64 + 64
X12 = 8960, and the calculation amount was an obstacle to practical use.

The present invention has been made under the above circumstances, and an object of the present invention is to reduce pattern image data measured by a sensor using a plurality of ring-shaped masks,
An object of the present invention is to provide a pattern recognition device that recognizes patterns of coins and the like that are invariable in rotation with a small-scale neural network.

[0007]

SUMMARY OF THE INVENTION The present invention relates to a pattern recognition device, and the above 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 small sections. With a plurality of ring-shaped masks that cover the ring-shaped region of the 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.

[0008]

According to the present invention, a pattern of coins or the like is recognized 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 ring-shaped masks. The image representative value of is obtained. The ring-shaped mask has a large number of small sections, and the ring-shaped areas of the small sections are randomly covered using, for example, uniform random numbers. The image representative value reduced by such a ring-shaped mask is invariant to the rotation of the pattern image, is input to the separation calculation unit (neural network), and the pattern image is adjusted by the weights 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. This reduces the size of the neural network configuration,
The control device is also small. In addition, since the neural network includes a non-linear element, there is an advantage that the recognition performance does not deteriorate with respect to variations in the grayscale direction of the pattern image data.
It is more practical than conventional pattern matching.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. The present invention is an application of the above-described neural network system. When pattern recognition of coins and the like is performed by such a neural network, conventionally, information obtained from a sensor of a recognition device is directly input to the neural network. It was 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. Therefore, in the present invention, a plurality of ring-shaped masks for reducing the input information are provided in the preprocessing unit, and the image representative value (slab value) is obtained from each ring-shaped mask.

FIG. 1 is a block diagram showing a configuration example of a coin discriminating apparatus of the present invention. A coin 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 r from the xy coordinate system.
The coordinates are converted into the −θ coordinate system and input to the ring-shaped masks 41 to 4n in the pre-processing unit 4, and the ring-shaped masks 41 to 4n.
The sum total value from n is calculated to obtain the image representative values SB1 to SBn.
Is 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 determination coin. 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 coin 1 of the object. The masks 41 to 4n in the pretreatment unit 4 are masks that cover the ring-shaped region, and the masks 41 to 4n are covered with different diameters.

First, the pretreatment section 4 using a ring-shaped mask
Will be described. In the present invention, ring-shaped masks 41 to 4n
The reason for using is as follows. 8 as shown in FIG.
In the binary image of “0” and “1” on the × 8 matrix, the slab value (the number of “1” in the binary image in the example of FIG. 1) which is the sum of pixel values is used as the feature amount of the image. If
In FIG. 2A, "1" is used as a value that characterizes the character "E".
4 "is obtained, and" 12 "is obtained as a value characterizing the character" H "in the same figure (B). Therefore," E "and" H "can be separated by using the slab value. In some cases, images having different patterns have the same slab value.For example, the slab value of the character “F” is “10” in FIG. 3A, but the character “K” in FIG. The slab value of the pattern image is also “10” and cannot be separated.
The y-rectangular coordinate system is converted to the r-θ cylindrical coordinate system, and FIG.
The problem can be solved by introducing a ring-shaped mask in which a portion of a specific ring-shaped region corresponding to the pixel of the input image as shown in (C) is covered. 3 with the mask of FIG.
When the coordinate-converted image of (A) is covered, the image shown in FIG. 4A is obtained, and the slab value in this case is "4". on the other hand,
When the coordinate-converted 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 “3”.
Becomes By introducing the ring-shaped mask in this way, "F" and "K" can be separated, and the slab value thereof is invariant to rotation. Therefore, in order to separate various images in this way, the position of the portion covered by the ring-shaped mask as shown in FIG. 4C is 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 various different masks as described above. It is presumed that it is possible to stochastically increase the separation capability between images by using various masks, because one of the slab value sequences often differs between images. The use of the plurality of different ring-shaped masks 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 ring-shaped masks in the pre-processing, and the sum of the pixels not covered becomes the slab values SB1 to SBn, which correspond to the neuro elements of the input layer one to one.

Next, at the time of recognition, the number of types of the ring-shaped masks selected in advance, suitability for the contents, and points to be noted in the selection will be described. FIG. 5 shows the masks 41 to 4n.
This is for explaining the effect of such a ring-shaped mask, and FIG. 9A shows that the slab value is "6" when the input image is not rotated. And FIG. 5 (B)
Even if the input image is rotated as shown in, the slab value remains "6", and a slab value that is invariant to the rotation of the input image can be obtained.

In the preprocessing section 4, the number of slabs (type of mask) and the covered areas of the masks 41 to 4n are considered as parameters, and the effect of the ring-shaped mask will be considered. 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

Next, the learning situation and the recognition ability will be examined according to the mask type, that is, the number of slabs. First, the ring-shaped mask used in the present invention is prepared as follows. That is, x
A circle having the same size as the input image to be converted from the -y coordinate system to the r-? Coordinate system is divided into four in the radial direction, and this ring-shaped area is arbitrarily covered with a uniform random number or the like. The number of slabs examined here is six, 2, 4, 6, 8, 10, and 12.
In FIG. 6, the number of presentations is 3 for these six slab numbers.
The learning status of the neural network up to 10,000 times is shown. The horizontal axis represents the number of times teacher data is presented, and the vertical axis represents the squared error. As shown in FIG. 6, when the number of slabs is “2”, learning does not converge and pattern separation is impossible. When the number of slabs is "4", the learning of the neural network shows a convergence tendency to some extent, but the error curve is oscillating. In this case, the learning was further continued up to 60,000 times, but the squared error did not become 1.0 or less. The number of slabs is "6"
In the above case, it is shown from FIG. 6 that the learning converges. Also, in the case where the alphabets A to L are recognized using this weight, the recognition ability is almost the same even in the neuro element corresponding to the determination pattern of the output layer when the number of slabs is "6" 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 number of mask types is 6 or more, and it is sufficient to set the number of slabs to 6 in the input layer of the separation calculation unit 5.
Compared to the conventional input pixel 64 (= 8 × 8) input to the input layer, the number of neuro elements can be extremely reduced, and the separation calculation unit 5 can be simplified accordingly. Also, when the number of characters or types to be recognized is increased, the number of masks may be increased accordingly.

Next, the learning situation and the recognition ability will be examined based on the covering area of the mask. 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 “6”. FIG. 14 shows the learning situation when the number of presentations reaches 30,000 when the random number width is changed. The horizontal axis and the vertical axis are the same as in FIG. It is clear from FIG. 7 that 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, it suffices to select the covering region randomly to some extent and consider the mask type without paying much attention to the covering region of each mask.

Next, a description will be given of the separation calculation unit 5 which inputs the information preprocessed by the preprocessing unit 4 and performs the separation calculation. 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 one-to-one to each mask type from the pre-processing unit 4, and correspond to slab values pre-processed by each mask type (total number of pixels after mask processing). Input to the 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. 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 AZ, the neuro elements in the output layer correspond to AZ in order from the top, and the corresponding neuro element outputs a value (maximum value) closest to 1, and The neuro element of outputs a value close to 0. When the character A is recognized, the highest neuro element in the output layer outputs the maximum value. Similarly, in the above bill identification, the neuro elements of the output layer are sequentially arranged from the top in the order of 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 (to the right of the table) ……… 1,000 yen (to the left of the back), 1
The two neuro elements correspond one to one. 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 the coupling between the neuro elements as a weight coefficient. One neuro element receives a signal from a plurality of neuro elements in the preceding layer through a 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 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. Discriminator 6
Inputs the information output from the output layer of the separation calculation unit 5, discriminates the maximum value (may be set to the minimum value) from the information, and discriminates it as the cadegory corresponding to the neuro element having the maximum value. To do. 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 In the above embodiment, the input image is temporarily converted from the xy coordinate system to the r-θ coordinate system for processing, but this coordinate conversion is omitted and the xy coordinate system is directly processed.
Processing in the coordinate system is also possible. Especially in the case of an input image with high resolution, there is little need to convert to the r-θ coordinate system,
A ring-shaped masking process can be performed with the xy coordinate system unchanged.
Further, in the above-described embodiment, the separation calculation unit 5 and the determination unit 6 are configured by separate hardware, but they can be configured by one and the same processing can be performed by a software program.

[0019]

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 ring-shaped mask in the preprocessing unit, the configuration of the neural network becomes small and the control device becomes small. 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. 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, which is the same as the number of input pixels.
(= 8 × 8). The number of hidden layer units is 32.
And the number of units in the output layer is the same as the judgment pattern, 12
And 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 was performed until the squared error became 0.01 was 61,983 times and the latter was 978 times. However, for the former, ε = 0.01 and α =
0.9, β = −0.1, and ε = 0.
Learning was performed with 1, α = 0.9 and β = −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 of the conventional algorithm is 64 × 32 + 32 × 12 =
It was 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 coin identifying 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 ring-shaped mask used in the present invention.

FIG. 5 is a diagram for explaining the effect of a ring-shaped mask.

FIG. 6 is a diagram showing a learning situation for various slab numbers.

FIG. 7 is a diagram showing a learning situation for various covering areas of a mask.

[Explanation of symbols]

 1 coin 2 image sensor 3 input image 4 pre-processing unit 5 separation calculation unit 6 determination unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Internal reference number FI Technical display location G06K 9/66 9289-5L G07D 5/00 9340-3E

Claims (2)

[Claims] 1. A sensor for optically or magnetically measuring a pattern image of an object to be recognized, and a plurality of ring-shaped masks having a large number of small sections and covering a ring-shaped region of the 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 coin.