JPH06274636A - System and method for identification and three dimensional identification system - Google Patents

Abstract

PURPOSE:To provide the system and method for quantatively deciding an optimum sensing position and the number of these positions in the case of identifying an object to be identified. CONSTITUTION:This system is provided with an image data input device 3 for the identifying object, first pattern learning/identifying means 11 for performing learning so as to close to a desired value, noise data generating means 12 for applying noise data to this means 11, identifying characteristic arithmetic means 13 for calculating an identifying characteristic from the output of the means 11, unit area identifying characteristic storage part 41 for storing this identifying characteristic, synthetic area constituting means 20 for selecting two unit areas from this identifying characteristic and synthesizing the identifying characteristics of these areas, second pattern learning/identifying means 21, noise data generating means 22, identifying characteristic arithmetic means 23, identifying characteristic evaluating means 24 for deciding the degree of satisfaction of the condition of this identifying characteristic, and output device 5 for receiving the output of the means 21 and outputting it to the outside.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an identification system and method, and a stereo identification system.

[0002]

2. Description of the Related Art An identification system identifies an identification target as one of a plurality of identification species. The identification target includes, for example, a planar object such as a bill and a three-dimensional object such as a component. First, a bill identifying system will be described as an example of a system for identifying a planar object.

In recent years, there have been attempts to use a neural network (hereinafter referred to as NN) for an identification system. Since the NN can learn a certain teacher pattern and output the degree of approximation to the teacher pattern when another pattern is input, its application to various recognition systems has been tried.

An example of a conventional identification system using an NN is described in "1992, IEEJ Transactions, Vol. 112, No. 2, 2".
49-258 ". The banknote discrimination system described in this paper discriminates banknote types, front and back, left and right postures, that is, 12 types, based on the density characteristics of the banknotes. It is composed of a plurality of sensors (four locations) fixedly arranged, and an NN for identifying a banknote type from characteristics such as a change in the density of the banknote obtained by the sensor. Using a basic banknote, a density pattern or the like is given to the NN input layer for learning, data of the banknote to be identified is input, and the identification result is output from the NN output layer. FIG. 13 shows a position where the density change is read and an example of the density change.

There are four sensing positions S1, S2, S3
And S4. The density data of the entire banknote is not detected because the number of data to be input is too large and the memory usage amount and the identification processing time become too large to be suitable for practical use. For example, even if the whole banknote is divided into 1 mm meshes and used as density data, 150 × 75 = 1
It requires about 1,250 input data, which is N
Practical application is impossible with the current data processing technology using N. It is also considered that one of the reasons is that the sensing device is empirically set to about four places even in the conventional identification device that does not use the NN.

Items required for the bill identifying system are: (1) It has a means for extracting features of each bill.
In other words, it should be possible to extract the "individuality" in which changes in the concentration of each of the 12 types are clear. (2) Resistant to sensor mounting misalignment, banknote misalignment and dirt. (3) It is possible to identify in a minimum required processing time while ensuring identification accuracy. And so on.

[0007] These matters are similarly required not only for bill recognition but also for other planar recognition systems.

However, in the conventional identification system, the determination of the sensing position is based on the empirical rule of the skilled person.
It was not possible to quantitatively determine the optimal sensing position and its number for.

Further, the sensing position and the number determined by the expert are for an identification system different from the NN, and are means for quantitatively determining the optimum one when the NN is used as the identification means. Did not exist.

Next, a system for identifying a three-dimensional object will be described.

When stereoscopic identification is performed, there are mainly the following two positions. One is the position to construct a three-dimensional model inside the computer based on the measurement value of the distance sensor and the image data, etc. Is
The candidate for identification is known in advance and is in a position to identify one candidate from the candidates.

In parts identification on a production line, the type of identification target is known in advance, and the identification system generally identifies only one type of identification target part.

This patent relates to an identification system for identifying one type of identified solid from a plurality of types of three-dimensional shape candidates.

As the identification procedure of the stereoscopic identification system, (i) the stereoscopic feature amount data is acquired. For example, the unevenness data of the three-dimensional surface is acquired by a sensor, the ridge line or curved surface data of the three-dimensional is acquired by an image processing device, and the like.

(Ii) Based on the characteristic amount, the identified solid is identified as one of the candidates.

It can be divided into two procedures.

An example of a conventional stereoscopic identification system is "19.
1991, Japan Society of Mechanical Engineers, Vol. 57, No. 535, pp. 172-179. In the stereoscopic identification system described in this paper, as the stereoscopic feature amount data, the stereoscopic unevenness data obtained by scanning the laser rangefinder in a circle around the binary image centroid of the stereoscopic image is used as an identification method. It uses fuzzy theory.

Further, a method of acquiring a stereoscopic feature amount by image processing is well known in the art.

However, in the conventional stereoscopic identification system, much time is required for the acquisition process of the feature amount data, for example, the measurement process by the laser range finder, the calculation process for extracting the ridge line from the image data, and the like. Moreover, much time is required for the identification processing based on the feature amount.

In some cases, the objects to be identified may have many solids with a high degree of similarity, or conversely, the degree of similarity may be low and a lot of processing is not required for the identification. Since the process for identification is uniform regardless of the sex, there was a problem of excessive or insufficient processing (indiscernibility).

[0021]

As described above, in the conventional system for identifying a planar object, the optimum sensing positions and the number thereof cannot be quantitatively determined, and a stereoscopic object is not detected. In the identification system, there is a problem that a lot of time must be spent for processing regardless of the similarity of the identification target.

The present invention has been made in view of the above circumstances, and an identification system and a method therefor capable of quantitatively determining an optimum sensing position for realizing a high identification accuracy and a minimum required processing time, and a minimum. It is an object of the present invention to provide a stereoscopic identification system that realizes the stereoscopic feature amount acquisition processing, high identification accuracy, and minimum necessary processing time.

[0023]

The identification system of the present invention is a system for identifying an identification target as one of a plurality of identification species, and is a first identification target divided into a plurality of unit areas. First pattern learning / identifying means for performing an output for identifying a corresponding identifying species given an identification pattern, and learning so that the output approaches a desired value, and the first pattern learning / identifying means after learning. A first noise data in which noise is included in the first identification pattern,
Noise data generating means, and first discriminating characteristic calculating means for obtaining a first discriminating characteristic for each unit area from the output of the first pattern learning / discriminating means given the first noise data. One or more of excellent identification characteristics among the unit area identification characteristic storage means for storing the first identification characteristic for each unit area and the first identification characteristics stored in the unit area identification characteristic storage means Unit area is selected, and a second identification pattern as an identification target is given to each of the combined unit areas combined by the combined area forming unit that combines the identification characteristics of the selected unit area and the combined area forming unit. The second pattern learning / identifying means for performing an output for identifying the corresponding identifying species and learning so that the output approaches a desired value, and the second pattern learning / identifying means after learning for the second pattern learning / identifying means. Identification of A second noise data generating means for providing a second noise data moistened with noise turn,
From the output of the second pattern learning / discriminating means given the second noise data, the
Second discrimination characteristic calculating means for obtaining the discrimination characteristic of, and discrimination characteristic evaluating means for judging whether or not the second discrimination characteristic obtained by the second discrimination characteristic calculating means satisfies a desired condition, At the time when the identification characteristic evaluation means determines that the second identification characteristic satisfies a desired condition,
And an output unit for outputting the output of the second pattern learning / identification unit related to the combined unit region selected and combined by the combined region forming unit to the outside.

Further, the identification method of the present invention is a method of making the identification target similar to one of a plurality of identification species, and the first identification pattern of the identification object divided into a plurality of unit areas is A step of giving an output for identifying the corresponding discriminant species to the pattern learning / identifying means No. 1 and learning so that this output approaches a desired value; and the first pattern learning / identifying means after learning. , Providing first noise data including noise in the first identification pattern,
Determining a first discriminating characteristic for each unit area from the output of the first pattern learning / discriminating means given the first noise data; and the first discriminating characteristic for each of the obtained unit areas. Among them, a step of selecting at least two unit areas having excellent identification characteristics and synthesizing the identification characteristics of the selected unit areas, and a second identification pattern to be identified for each synthesized composite unit area A step of giving the first pattern learning / discriminating means an output for identifying the corresponding discriminant species, and learning so that the output approaches a desired value; and the second pattern learning / learning after learning.
A step of giving the identification means second noise data in which the second identification pattern includes noise; and a synthesis from the output of the second pattern learning / identification means provided with the second noise data. The step of obtaining the second identification characteristic for each unit area, the step of determining whether or not the obtained second identification characteristic satisfies a desired condition, and the second identification characteristic satisfies the desired condition. The step of outputting the output of the second pattern learning / identifying means relating to the composition unit area at the time when it is determined that the condition is satisfied to the outside.

Furthermore, the stereoscopic identification system of the present invention is a stereoscopic identification system for identifying a solid as one of a plurality of identification species, and an image capturing device for capturing a solid from a predetermined direction, and the image capturing device. One or a plurality of illuminating means arranged on the circumference around the center, means for arranging and controlling the illuminating means in a predetermined direction with respect to the solid, and means for illuminating the illuminating means simultaneously or one by one. Image data storage means for storing image data obtained by the image capturing device;
A stereoscopic feature amount acquiring unit having a stereoscopic feature amount configuring unit that generates a required stereoscopic feature amount from the image data, and a stereoscopic feature amount acquired by the stereoscopic feature amount acquiring unit are divided into unit areas. A first pattern learning / identifying means for giving an identification pattern, performing an output for identifying a corresponding identification species, and learning so that the output approaches a desired value; and a means for obtaining an identification characteristic for each unit area Of the unit areas, a combination area forming unit for selecting and combining a plurality of unit areas having good identification characteristics, and a second identification pattern to be identified to the combination area combined by the combination area forming unit. Then, a second pattern learning / discriminating means for performing an output for identifying the corresponding discriminating species and learning so that the output approaches a desired value, a means for obtaining the discriminating characteristic of the combined area, and The discriminant characteristic evaluation unit that determines whether or not the discriminant characteristic of the synthetic region satisfies a predetermined condition, and the synthetic region configuration at the time when the discriminant characteristic evaluation unit determines that the discriminant characteristic evaluation unit satisfies the predetermined condition. Output means for receiving the output of the second pattern learning / identifying means relating to the combined unit area selected and combined by the means and outputting the output to the outside.

[0026]

The first pattern learning / identifying means outputs the first identifying pattern for each unit area to be identified and identifies the corresponding identifying species, so that the output approaches the desired value. After learning, the first noise data generating means gives the first noise data to the first pattern learning / identifying means after learning. From the output of the first pattern learning / discrimination means given the first noise data, the first discrimination characteristic computing means obtains the first discrimination characteristic for each unit area, and the unit area discrimination characteristic storage means Remember. The combination area forming means selects at least two unit areas having excellent characteristics among the first identification characteristics, and combines the selected unit areas as second pattern learning / identification means. Second
The pattern learning / identification means of (1) gives an output for identifying the corresponding identification species given the identification pattern regarding the composite area, and learns so that this output approaches a desired value. The second noise data generating means gives the second noise data to the second pattern learning / identifying means after learning, and the second noise learning data is outputted from the output of the second pattern learning / identifying means. , The second identification characteristic calculation means obtains the second identification characteristic of the combined area. The discriminating characteristic evaluating means judges whether or not the second discriminating characteristic satisfies a desired condition,
When it is determined that the desired condition is satisfied, the output of the second pattern learning / identifying means relating to the combined area selected and combined by the combined area forming means is provided to the output means and output to the outside. As a result, a region having excellent identification characteristics is quantitatively selected from the plurality of unit regions.

Further, by using such an identification system,
The identification method of the present invention can be implemented.

Further, according to the three-dimensional identification system of the present invention, the illumination means is turned on simultaneously or one by one, the image photographing device photographs the three-dimensional object from a predetermined direction, and the obtained image data is stored in the image data storage means. Then, the stereoscopic feature quantity constructing means constructs a required image feature quantity from this image data. next,
The first pattern learning / identifying means divides the stereoscopic feature amount thus obtained into unit areas, and outputs an identification pattern of the unit area to identify the corresponding identification species, and this output is predetermined. Learn to approach the value of. Further, the means for obtaining the identification characteristic for each unit area obtains the identification characteristic for each unit area, and the combining area composing means combines a plurality of unit areas having good identification characteristics. The second pattern learning / discriminating means outputs the discriminant corresponding to the discriminant species to which the second discriminant pattern to be discriminated is given to the synthesized region so that the output approaches a desired value. learn. The means for obtaining the identification characteristic of the combined area obtains the identification characteristic of the combined area, the means for obtaining the identification characteristic obtains this identification characteristic, and the identification characteristic evaluation means determines whether or not the identification characteristic satisfies a desired predetermined condition. When it is determined that the above condition is satisfied, the output means is given the output of the second pattern learning / identification means and outputs it to the outside. As a result, the minimum area required for three-dimensional identification is quantitatively obtained, and reliable identification is possible with the minimum processing amount and processing time.

[0029]

An embodiment of the present invention will be described with reference to the drawings.

FIG. 1 shows the configuration of a system for identifying a planar identification target according to the first embodiment of the present invention. In this embodiment, an image data input device 3 such as an image sensor and a unit area data processing device 1 that operates under program control.
And a composite area data processing device 2, a storage device 4 for storing information, and an output device 5 for transmitting a display and an identification result to the next process.

The unit area data processing device 1 includes a first pattern learning / identifying means 11 and a noise data generating means 12 for giving banknote data including noise to the first pattern learning / identifying means 11 after learning. The discrimination characteristic calculating means 13 for calculating the discrimination characteristic of the unit area at that time by using the output of the first pattern learning / identifying means 11 is provided.

The storage device 4 includes a unit area identification characteristic storage unit 41 for storing the identification characteristics of the unit area calculated by the identification characteristic calculation means 13.

The first pattern learning / identifying means 11 has a three-layer back propagation N as shown in FIG. 4, for example.
You may comprise N system.

The composite area data processing device 2 comprises a composite area composing means 20 for forming a new NN identification means based on the data stored in the unit area identification characteristic storage unit 41, and a second pattern learning / combining area for the composite area. Discrimination means 21 and discrimination characteristic calculation means 23 that computes the discrimination characteristics of the combined area at that time using the outputs of the second pattern learning / identification means 21 and whether or not the discrimination characteristics satisfy the required accuracy. And a discrimination characteristic evaluation means 24 for judging whether or not.

The storage device 4 also includes a synthetic area identifying data storage section 42 for storing the identification characteristics of the synthetic area calculated by the second pattern learning / identifying means 21 and the NN parameters obtained as a result of learning. .

The second pattern learning / identifying means 21 has a three-layer back propagation N as shown in FIG. 5, for example.
You may comprise N system. The first and second pattern learning / identifying means 11 and 21 function as pattern learning means as they are after the pattern learning is completed. Here, the output when data is given to the NN input layer and an example of the error are shown in Table 1 below.

[0037]

[Table 1] This table 1 shows an output example when the data of the used thousand yen table right is given to the learned NN. "Output layer 1"
Corresponds to the right side of the thousand yen table, the output from this "output layer 1" is a value close to "1", and the other output layers are
The output is close to "0". In general, the output of the NN is not completely "1" or "0", but some error is left. If the used bills have wrinkles or stains, or if the identification system has poor characteristics, this error will increase. Stable identification is required for any of the three types of banknotes to be identified × 4 attitudes = 12 types of banknotes.

There are several possible definitions of the discrimination characteristic to be obtained by using the output result from the NN, but here, 1 / (sum of errors) is used. That is, banknote data including noise such as 12 kinds of used banknotes is given to the learned NN, and the reciprocal of the sum of the errors at that time is used.
Therefore, the larger the discrimination characteristic, the smaller the discrimination error.

Next, an identification method according to the second embodiment of the present invention will be described in which identification is performed according to the procedure shown in FIG. 3 using the identification system according to the first embodiment shown in FIG.

First, the unit area data processing will be described.

The data given from the image data input device 3 is the image data of the entire area or wide range of the bill.
For the sake of explanation, it is assumed that this data is horizontal 128 × vertical 32 data, and the unit area is vertical 32 line data. That is, there are 128 unit areas. Codes of I = 1 to 128 are attached to these 128 unit areas.

First, I = 2nd line is selected (step 1
01), the first pattern learning / identifying means 11 composed of NNs performs NN learning on this line (step 102). That is, 32 pieces of image data of I = 2nd line regarding 3 denominations × 4 attitudes = 12 kinds of banknotes are given to the input layer (for example, p1 to p32 in FIG. 4),
Learn. Since the input data used at this time is the reference banknote data, it is, for example, data of an unused new banknote. Next, the noise data generating means 12 gives the following three types of noise data to the input layer to the NN for which learning has been completed (step 103).

I = I-1, I, I + 1 (= 1, 2, 3)
The image data of the line The discrimination output error due to this is e1,
Assuming e2 and e3, of course, e2 is usually small,
e1 and e3 become large. In addition, the identification characteristic calculation means
In, for example, E = (e1 + e2 + e3) is calculated (step 104).

When the above processing for the I = 2nd line is completed, the error E is stored in the unit area identification characteristic storage unit 41. Then, the same processing is performed for the I = third line and thereafter, and the unit area data processing is completed at the I = 127th line.

I = 2 to 12 obtained by the above processing
FIG. 6 shows an example of the discrimination characteristic for 7 lines. In this graph, the horizontal axis represents the line positions 2 to 127, and the vertical axis represents the reciprocal 1 / E of the error in each line. The larger the value 1 / E on the vertical axis, the better the discrimination characteristic.

Next, the composite area data processing will be described.

The composite area forming means 20 selects a plurality of lines from the lines having excellent unit area identification characteristics based on the data stored in the unit area identification characteristic storage unit 41, and newly selects N lines.
Configure N (step 202). For example, by selecting four lines L1 to L4 having better characteristics from FIG.
An NN having input layers corresponding to the number of input layers, that is, 32 × 4 = 128 input layers, is configured as shown in FIG. Then, the second pattern learning / identifying means 21 performs learning (step 203), and noise-containing data is input from the noise data generating means 22 to the learned NN (step 204). The noise data at this time may be data of adjacent lines as in unit area data processing,
The data of used banknotes prepared may be used. Using the output of the NN to which the noise data has been input, the discrimination characteristic computing means 23 computes the discrimination characteristic in which the discrimination errors are integrated (step 205).

Whether or not this discriminating characteristic satisfies the required characteristic is evaluated by the discriminating characteristic evaluating means 24 (step 206). If it is satisfied, the processing is ended, and if it is not satisfied, the number of unit areas used. Is increased (step 207), a large-scale NN is configured, and the same processing is performed. In the above description, the number N of combined lines in the combined area data processing is N.
Starts from N = 4, but generally starts the process from N = 1 (step 201) and ends when the required discrimination characteristic is satisfied.

By the way, in the unit area data processing device 1, the image data including the adjacent areas from the noise data generating means 12, that is, the image data of I = I-1, I, I + 1 line is subjected to the first pattern learning / identification. Given to the means 11, its meaning will be described.

Image data of I = 1st line and I = I-
When the image data of the 1st and I + 1th lines are close to each other, the sum E of identification errors is small, and when these image data are different, E is large. This means that the stability with respect to the relative position error between the sensor and the bill can be evaluated by the size of the error E. If E is small, it is determined that the line I is the sensing position in the direction of the adjacent area. Position error, such as a bill setting error and a sensor assembly error.

However, if the image data between the adjacent areas are simply approximated, the sum E of the identification errors does not necessarily become small. For example, in a pattern in which there is no change in density such as a simple white background, it can be said that the image data of the adjacent areas are close to each other, but the distinguishability is also poor. The layer error increases.

For these reasons, the fact that the above-mentioned error E is small means both the strength against the positional error between the sensor and the bill and the discrimination characteristic for extracting the density change of each bill as a clear individual. Will be done. Furthermore, the noise data is not limited to the data in the adjacent area, and may be any data that takes into account the error unique to the identification system. Therefore,
For example, the data obtained by changing the relative position between the sensor and the bill by an appropriate amount may be used as noise data.

Here, the lines (L1 to L4) having a good unit area identification characteristic are selected by the composition area composing means 20. The quality of the unit area identification characteristic has a positive correlation with the quality of the identification characteristic after composition. It has been confirmed experimentally.
For example, in FIG. 6, a line L having a poor unit area identification characteristic
When the combination of 5 to L8 is performed, the identification error becomes more than twice as large as that of the combination of L1 to L4.

FIG. 7 shows the experimentally obtained result of the relationship between the increase in the number of lines to be synthesized and the discrimination characteristic. The horizontal axis indicates the number of combined lines, the vertical axis indicates 1 / E, the solid line L12 indicates the case where noise data of adjacent lines is given, and the broken line L11 indicates the case of actual banknote identification. As shown in the figure, it can be seen that the identification characteristic improves as the number of combined lines increases. However, in the bill identifying system illustrated here, when the number of combined lines is 4 or more, the identifying characteristic is saturated. Therefore, in order to minimize the processing time, it is considered appropriate to set the number of combined lines to four.

As described above, the parameters of the identification system are decided and the ordinary bills are discriminated. However, as shown in FIG. 2, the bill discriminating system after the parameters are decided is only the minimum elements necessary for the discrimination. Constitute. That is,
An image data input device 3 having line-shaped image sensors arranged at several required sensing positions, and a data processing device 51 having a final second pattern identification means 21 and a composite area identification data storage section 42. , And an output device 5 for transferring data to a process after display or identification. This new configuration is effective in the case where the type of identification target does not change except when issuing a newly designed bill, such as bill identification.

The identification system of the present invention is not limited to bill identification but can be widely applied to other identifications. For example, the invention can be applied to a fish species identification system, a component identification system in a manufacturing line, and the like, and a similar identification system can be configured by using concentration data regarding an identification target. In these cases, since the type of the identification target changes frequently, it is more appropriate to perform the identification processing with the configuration shown in FIG. 1 rather than configuring the identification system with the minimum elements as shown in FIG. Is considered to be easy.
Further, in addition to the line shape shown in the present embodiment, various shapes of the unit area are possible, such as an arc shape, a concentric shape, a radial shape, or a scatter shape, depending on the purpose of the identification system.

Next, a stereo identification system according to a third embodiment of the present invention will be described with reference to the drawings. First, the stereoscopic feature amount acquiring means included in the system will be described.

FIG. 8 shows the configuration of the stereoscopic feature amount acquisition means.

The three-dimensional feature amount acquisition means in this embodiment is
It is composed of an image capturing device 61 such as a CCD camera and four illumination devices 65a to 65d. Image capturing device 6
No. 1 is arranged so that it can be photographed with the substantially central portion (center of gravity) of the identified solid 64 as the image center. Further, the illumination devices 65a to 65d are arranged at equal intervals, that is, at a 90-degree sense on the circumference of the image capturing device 61, and illuminate the identified solid 64. Further, the number of lighting devices 65a to 65d is 1.
It can be lit individually or simultaneously. Further, the image data acquired by the image capturing device 61 is stored by the image data storage device means 62, and the stereoscopic feature amount forming means 63 is stored.
By this, a configuration suitable for the identification process is formed.

Next, a method for acquiring the stereoscopic feature amount of the identified solid 64 by the stereoscopic feature amount acquisition means will be described.

Now, six types of three-dimensional identification shown in FIG. 9 will be taken as an example. The six types of solids have a square planar shape, that is, an outer shape captured from above, and the type of the solid cannot be identified only from the outer shape, and the direction of the solid cannot be specified. Generally speaking, even if it is called stereoscopic identification, if it is possible to identify one type from the external shape, it is not particularly necessary to obtain the stereoscopic feature amount, and the identification can be performed by a conventional technique such as image binarization. A feature amount other than the outer shape is required, and the problem of the stereoscopic identification system is to obtain and identify the feature amount.

Then, in the six types of three-dimensional identification shown in FIG. 9, it is necessary to identify the direction as well as the three-dimensional identification. Therefore, in this example, one kind is identified from among the 6 kinds × 4 directions = 24 identification kinds.

From the six types of solids shown in FIG. 9, for example, FIG.
A method of acquiring the stereoscopic feature amount shown in (c) will be described.

First, when the illumination 65a is turned on, a shadow image as shown in FIG. 10A is obtained on the image capturing device 6.
Similarly, when the illuminations 65b, 65c, and 65d are turned on, shaded images as shown in FIGS. 10B, 10C, and 10D are obtained, respectively. It should be noted that although shadows also occur outside the solid, this is not taken into consideration in this example.

Since the image data is the density data in each pixel, the obtained shadow image has the shape of the density distribution in the square.

Further, here, the three-dimensional shaded images (a), (b), (c) and (d) shown in FIG. 9 (c) were obtained, and a set of these images is shown in FIG. 9 (c)-( a), (b), (c),
It will be referred to as (d).

However, since the three-dimensional shape shown here has a square outer shape, the direction is also designated as one of the four directions. 9 (c)-(b), (c) as a set of shadow images when the planar placement is changed at 90 degree intervals,
(D), (a), FIG. 9 (c)-(c), (d),
(A), (b), FIG. 9 (c)-(d), (a),
(B) and (c) are obtained.

When these operations are performed for all six types of solids, a total of 24 sets of shadow image data can be obtained.

Then, the three-dimensional feature amount constructing means D collects the four pieces of shadow image data of each set. For example, Fig. 4-
As shown in (a), FIG. 9 (c)-(a), (b),
The image data of (c) and (d) are vertically combined, and this is 1
Each image data. That is, the image data shown in FIG. 11 has a corresponding relationship with the three-dimensional directions (a), (b), (c), and (d) of FIG. 9C. Since the same processing is performed for each of the 24 sets in total, 24 pieces of image data can be obtained.

The stereoscopic identification process is a process of acquiring 24 pieces of image data by the above procedure and identifying the image density data obtained for the identified solid as one of the 24 pieces of image data.

It should be noted that the time required for acquiring the stereoscopic feature amount shown here is such that one illuminating device per identified stereoscopic object is turned on four times, four screens are photographed by the image photographing device, and the image data is stored in the memory. It takes a very short time because it only needs to store and synthesize.

The three-dimensional feature amount can be obtained as described above.

Next, the stereoscopic identification system according to this embodiment using the acquired stereoscopic feature amount will be described. FIG. 12 shows the configuration of the identification system.

In the first and second embodiments described above, the feature amount distribution such as the density distribution on the plane is used as an example of bill identification, but in the present embodiment, the stereo identification is used to identify the density data of the image. Since it is converted into, the completely same identification method can be applied.

That is, for example, with respect to the image data as shown in FIG. 11A, this is divided into a plurality of linear unit areas, and a density change pattern of each linear area is obtained. FIG. 11B shows the density change pattern of the linear region at the position of arrow A in FIG. Using these patterns, the first pattern learning / identifying means learns the identification patterns of all the unit areas, and the same processing as that described in the first and second embodiments is performed.

Next, a stereoscopic feature amount acquisition means in the stereoscopic identification system according to the fourth embodiment of the present invention will be described. This means is characterized in that the distance pattern to the appropriate area of the solid is measured from a fixed direction, and the measured distance pattern is used as the solid feature amount.

The system configuration corresponds to the configuration shown in FIG. 8 in which a laser rangefinder is used instead of the image pickup device 61 including a CCD camera and the like. The three-dimensional feature amount acquisition means in the fourth embodiment is
A distance measuring unit such as a laser range finder, a distance data storage unit, and a stereoscopic feature amount forming unit are provided. The distance measuring means scans about the central portion of the identified solid 64, so that the identified solid 64 in a certain area is scanned.
It is possible to acquire the unevenness information of.

By the way, the data format of the unevenness information acquired here is exactly the same as the image data format. That is, image data is a set of density data in a grid-shaped coordinate system called pixels or pixels, and uneven data is also a set of distance data in a grid-shaped finite number of coordinate systems. is there. Of course, it may be not a grid pattern but another segmented form, but it can be considered as the same data format in the sense of a concentration or distance data set at positions that are finely segmented in two dimensions.

Therefore, by using the distance data acquired by the distance measuring means as the stereoscopic feature amount, the stereoscopic identification can be performed in the same procedure as that of the third embodiment.

As described above, in the stereoscopic identification systems according to the third and fourth embodiments, the stereoscopic feature quantities such as density and distance data are acquired to quantitatively determine the minimum attention area required for identification. It can be obtained, and reliable identification can be performed with the minimum processing amount and processing time. As the three-dimensional feature amount, various forms and methods, or data conversion as preprocessing can be considered.

By the way, an identification experiment of 6 types of shapes × 4 directions = 24 types shown in FIG. 9 was carried out by the above-mentioned system. The result shows that the unit area, which is comparatively gentle and has a distinct shape difference, has good discrimination characteristics.

[0082]

As described above, according to the identification system and the method of the present invention, it is possible to quantitatively select the area having the best identification characteristic from the unit areas of the identification target data, and therefore the minimum It is possible to identify with high accuracy in the data processing time.

According to the three-dimensional identification system and method of the present invention, the three-dimensional characteristic amount acquisition means acquires the three-dimensional characteristic amount such as density and distance data.
It is possible to quantitatively determine the minimum area required for identification, and it is possible to reliably identify with the minimum processing amount and processing time.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an identification system according to a first exemplary embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration after parameters are determined in the identification system.

FIG. 3 is a flowchart showing a procedure of an identification method according to a second embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a neural network system shown as a specific example of a first identification pattern learning / identifying means in the identification system according to the first exemplary embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a neural network system shown as a specific example of a second identification pattern learning / identifying means in the same identification system.

FIG. 6 is an explanatory diagram showing an example of identification characteristics with respect to a line position when a banknote is identified using the identification system.

FIG. 7 is an explanatory diagram showing an example of identification characteristics with respect to the number of lines when a bill is identified using the identification system.

FIG. 8 is a perspective view showing a configuration of a stereoscopic feature amount acquisition means in a stereoscopic identification system according to a third embodiment of the present invention.

FIG. 9 is a perspective view showing six types of solids that can be identified using the system.

FIG. 10 is an explanatory diagram showing a shadow image formed by illuminating the solid body shown in FIG. 9C.

FIG. 11 is an explanatory diagram showing a combination of the shadow images shown in FIG. 10 in the vertical direction and a change in density on the line indicated by arrow A.

FIG. 12 is a block diagram showing a configuration of a stereoscopic identification system according to a third embodiment of the present invention.

FIG. 13 is an explanatory diagram showing an example of an identification line set on a bill that can be identified by using the identification system according to the embodiment of the present invention, and a density change on the identification line.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 unit area | region data processing apparatus 2 synthetic | combination area | region data processing apparatus 3 image data input apparatus 4 memory | storage device 5 output apparatus 11 1st pattern learning / discrimination means 12 noise data generation means 13 discrimination | determination characteristic calculation means 20 synthetic | combination area | region composition means 21 2nd Pattern learning / identification means 22 noise data generation means 23 identification characteristic calculation means 24 identification characteristic evaluation means 41 unit area identification characteristic storage section 42 synthetic area identification data storage section 51 data processing apparatus 61 image photographing apparatus 62 image data storage section 63 Stereoscopic feature amount configuration means 64 Identification solids 65a to 65d Illumination 71 Stereoscopic feature amount acquisition means

Claims (8)

[Claims] 1. An identification system for identifying an identification target as one of a plurality of identification species, wherein a first identification pattern of the identification target divided into a plurality of unit areas is given and identified to a corresponding identification species. And a first pattern learning / identifying means for learning such that the output approaches a desired value, and the first pattern learning / identifying means after learning is provided with noise in the first identification pattern. From the output of the first noise data generation means for giving the included first noise data and the output of the first pattern learning / identification means given the first noise data, the first identification is made for each unit area. First identification characteristic calculation means for obtaining characteristics, unit area identification characteristic storage means for storing the first identification characteristics for each unit area, and the first identification characteristic stored in the unit area identification characteristic storage means No From the above, one or more unit areas having excellent identification characteristics are selected, and the combination area forming means for combining the identification characteristics of the selected unit areas, and the combination area combined by the combination area forming means Second pattern learning / discriminating means for performing an output for identifying a corresponding discriminative species given a second discriminating pattern, and learning so that this output approaches a desired value; and the second pattern after learning. Second noise data generation means for giving learning / identification means second noise data in which noise is included in the second identification pattern, and second pattern learning given the second noise data / From the output of the identification means, the second for each synthesis unit area
Second discrimination characteristic calculating means for obtaining the discrimination characteristic of, and discrimination characteristic evaluating means for judging whether or not the second discrimination characteristic obtained by the second discrimination characteristic calculating means satisfies a desired condition, The second pattern learning / identification means relating to the composition unit area selected and composed by the composition area composition means at the time point when the identification characteristic evaluation means determines that the second identification characteristic satisfies a desired condition. And an output means for outputting the output of the output to the outside, and an identification system. 2. The first noise data generating means includes one including at least a part of characteristic data of a unit area adjacent to a unit area when learning an identification pattern for each unit area. The identification system according to claim 1, wherein the identification system is generated as data. 3. The first noise data generating means generates, as the first noise data, data including at least a part of characteristic data in the vicinity of a unit area when learning an identification pattern for each unit area. Claim 1 characterized by the above.
Identification system described. 4. An identification method for identifying an identification target as one of a plurality of identification species, wherein the first identification pattern of the identification target divided into a plurality of unit areas is used as first pattern learning / identification means. A step of giving an output to identify the corresponding discriminant species, and learning so that this output approaches a desired value; and Providing first noise data containing noise, and obtaining a first identification characteristic for each unit area from the output of the first pattern learning / identifying means provided with the first noise data. A step of selecting at least two unit areas having excellent identification characteristics from the first identification characteristics obtained for each unit area and synthesizing the identification characteristics of the selected unit areas; The second discrimination pattern to be discriminated is given to the first pattern learning / discrimination means for each of the synthesized unit areas thus generated, and the output for discriminating the corresponding discriminant species is performed, so that this output approaches the desired value. A step of learning; a step of giving the second pattern learning / identifying means after learning second noise data in which the second identification pattern includes noise; and a step of giving the second noise data. From the output of the second pattern learning / identifying means, the second
Determining the discriminating characteristic, determining whether the obtained second discriminating characteristic satisfies a desired condition, and determining that the second discriminating characteristic satisfies the desired condition. A step of outputting the output of the second pattern learning / identification means regarding the composition unit area at the time point to the outside. 5. The first noise data includes at least a part of characteristic data of a unit area adjacent to the unit area when learning an identification pattern for each unit area. Identification method described. 6. The identification method according to claim 4, wherein the first noise data includes at least a part of characteristic data in the vicinity of a unit area when learning an identification pattern for each unit area. . 7. A stereoscopic identification system for identifying a solid as one of a plurality of identification species, an image capturing device for capturing a solid from a predetermined direction, and one image capturing device on the circumference of the image capturing device. Alternatively, a plurality of illuminating means, a means for arranging and controlling the illuminating means in a predetermined direction with respect to a solid body, a means for lighting the illuminating means simultaneously or one by one, and image data obtained by the image capturing device Image data storage means for storing, and three-dimensional feature amount constructing means for generating a required three-dimensional feature amount from the image data, and three-dimensional feature amount obtained by the three-dimensional feature amount acquiring means as a unit The first pattern learning / learning is performed by dividing into regions and given an identification pattern of a unit region, and performing an output for identifying to a corresponding identification species, and learning so that this output approaches a desired value.
Identification means, means for obtaining identification characteristics for each unit area, combination area composition means for selecting and combining a plurality of unit areas having good identification characteristics from the unit areas, and composition area composition means for composition. A second pattern learning / identifying means that is provided with a second identification pattern to be identified in the combined area, performs an output for identifying a corresponding identification species, and learns so that this output approaches a desired value; A means for obtaining a discrimination characteristic of the composite area, a discrimination characteristic evaluation means for judging whether or not the obtained discrimination characteristic of the composite area satisfies a predetermined condition, and the discrimination characteristic evaluation means satisfies a predetermined condition. When it is determined that the combination unit area is selected, the second pattern learning / related to the combination unit area selected and combined by the combination area forming unit is performed.
A stereoscopic identification system, comprising: an output unit that receives the output of the identification unit and outputs the output to the outside. 8. A stereo identification system for identifying a solid as one of a plurality of identification species, the distance obtained by scanning an arbitrary area of the solid from a predetermined direction and detecting the distance to this area. A stereoscopic feature amount acquisition unit that uses a pattern as a stereoscopic feature amount, and a stereoscopic feature amount acquired by the stereoscopic feature amount acquisition unit is divided into unit areas, and an identification pattern of the unit area is given to identify the corresponding identification species Output the first pattern learning / learning so that the output approaches the desired value
Identification means, means for obtaining identification characteristics for each unit area, combination area composition means for selecting and combining a plurality of unit areas having good identification characteristics from the unit areas, and composition area composition means for composition. A second pattern learning / identifying means that is provided with a second identification pattern to be identified in the combined area, performs an output for identifying a corresponding identification species, and learns so that this output approaches a desired value; A means for obtaining a discrimination characteristic of the composite area, a discrimination characteristic evaluation means for judging whether or not the obtained discrimination characteristic of the composite area satisfies a predetermined condition, and the discrimination characteristic evaluation means satisfies a predetermined condition. When it is determined that the combination unit area is selected, the second pattern learning / related to the combination unit area selected and combined by the combination area forming unit is performed.
A stereoscopic identification system, comprising: an output unit that receives the output of the identification unit and outputs the output to the outside.