JP3347435B2 - Pattern Recognition Method Using Hierarchical Neural Network - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to input images, figures,
The present invention relates to a pattern recognition method using a hierarchical neural network for classifying, discriminating, judging, or recognizing a pattern of a voice or the like. In particular, it is possible to appropriately cope with a recognition process for a pattern that is significantly different from a learned pattern. The present invention relates to a pattern recognition method using a hierarchical neural network that enables effective processing by appropriately selecting a load between units used in a neural network during processing.

[0002]

2. Description of the Related Art Various putters such as images, figures and sounds.
Is a hierarchical neural network used for
First, the learning pattern and the hierarchical neural network
Target value of the output of each unit in the
) Are input as a set, and learning is performed. Study
In practice, the number of training patterns equals or exceeds
Enter the pairs sequentially or randomly and add them to each learning pattern.
On the other hand, each user in the output layer of the hierarchical neural network
Make sure that the actual output of the knit approaches each target value
Exist in the input layer / intermediate layer and the output layer, respectively.
Correct the coupling load between units. Representative of that modification
As the method, the back propagation method is famous, but the back propagation method is often used.
Are described in detail in many documents
Omitted. In learning, patterns to be recognized
Learning is performed for all pairs of
Completion is determined by an error assessment as described below.
You. For each unit of the output layer for one learning pattern
The square of the difference between the actual output value and its target value is calculated for all output layers.
Add for units and divide by the number of units to get a positive
The square root of the normalized value is called the pattern error of this pattern.
Huh. That is, the unit in the output layer for the pattern p
The target value of k is t_k ^P, The actual output of unit k _k ^P
And the pattern error ε (p)

[0003]

(Equation 1)

[0004] Here, K is the number of units in the output layer. The average value of all pattern errors will be referred to as an average pattern error. Average pattern error ave
Is

[0005]

(Equation 2)

[0006] Here, P is the number of learning patterns. This average pattern error for each learning
The ave is calculated, and if this value falls below the allowable value, the learning is completed. When recognizing a learning pattern using the connection weight at the time of completion of learning, recognition can be performed with an accuracy equal to or less than the maximum pattern error among the pattern errors. Even if a pattern is not used for learning, if it is similar to the learning pattern, it can be recognized approximately within this maximum pattern error.

[0007]

The problems that have conventionally been encountered in the recognition of patterns by a hierarchical neural network are to make it possible to recognize patterns that are significantly different from learning patterns, or new patterns other than those provided for learning. This is the case. Regarding these patterns, recognition is generally not successful even if recognition is performed using learned connection weights. In order to enable recognition of these types of patterns, it is common practice to add them as new patterns to the learning pattern and then perform learning again using the new patterns as learning patterns. In this case, the re-learning is problematic in that the number of sets of learning patterns increases, so that the learning time is further increased, and the connection weight obtained by the learning is reset and the learning is performed again from the beginning.

An object of the present invention is to solve the above-mentioned problems of the prior art, and it is possible to appropriately process even when an input pattern is significantly different from a pattern used at the time of learning and cannot be identified and determined. An object of the present invention is to provide a pattern recognition method using a hierarchical neural network.

[0009]

The invention claimed in this application to achieve the above object is as follows. (1) A method of inputting patterns such as images, figures, and voices into a hierarchical neural network and classifying, identifying, and judging them (hereinafter, referred to as recognition) includes learning and recognition steps. Then, a pattern to be learned and a target value are inputted, and a learning connection weight between each layer unit of the neural network obtained by learning about these is added with an identification symbol, stored and stored. The specified connection weight is specified, the recognition process is performed for the pattern to be recognized using the specified connection weight, the output value obtained as a result, and the target value in the learning process for obtaining the connection weight used. And temporarily store the minimum value of each pattern error obtained by using each of the connection weights based on The minimum value of the pattern error is obtained from the minimum values of the pattern errors obtained using the weights, and if the minimum value of the pattern error is equal to or less than a predetermined allowable value, the learning for the minimum value of the pattern error is performed. A pattern recognition method using a hierarchical neural network, characterized in that a pattern or a target value thereof is output as a final recognition result, and if the value is not smaller than an allowable value, a message is issued that the pattern cannot be recognized.

(2) A method of inputting a pattern such as an image, a figure, and a sound into a hierarchical neural network and recognizing the pattern includes learning, recognition, and evaluation steps. Input the values, add the identification symbol to the learned connection weights between the units of each layer of the neural network obtained by learning about them, store them sequentially, and accumulate them. From the specified, to perform the recognition processing for the pattern to be recognized using the specified coupling weight, in the evaluation step, the output value obtained as a result of the above processing and the coupling weight used to determine the used Find the minimum pattern error from the target value used during learning pattern learning, and determine that this pattern error is less than or equal to a predetermined allowable value. For example, the learning pattern, the target value, and the minimum pattern error are temporarily stored, and the minimum pattern error and the target value are further reduced from the stored pattern error minimum values obtained using all the specified connection weights. And output this as the optimal recognition result.
A pattern recognition method using a hierarchical neural network, characterized in that when no set of a minimum pattern error and a target value is stored, a message indicating that recognition is impossible is issued.

(3) A method of inputting a pattern such as an image, a figure, and a voice into a hierarchical neural network and recognizing the pattern includes initial learning, recognition, evaluation, and additional learning. In the learning step, a learning pattern and a target value are input, and an identification symbol is added to the learned connection weight between the layer units of the neural network obtained by learning about the learning pattern, which is stored and accumulated. From the accumulated connection weights, the one to be used is specified, and the pattern to be recognized is recognized using the specified connection weight. In the evaluation step, the output obtained in the above recognition step and the connection weight used are used. The minimum pattern error from the target value used in learning the pattern is calculated. If it is equal to or less than the specified allowable value, the learning pattern, the target value, and the minimum pattern error are temporarily stored, and the minimum value among the pattern error minimum values obtained using all the specified coupling weights is further reduced. A set of a pattern error and a target value is obtained and output as an optimum recognition result. If no set of the minimum pattern error and the target value is stored, a message indicating that the pattern cannot be recognized is output and When an unacceptable pattern is stored and the number of storages reaches a predetermined number, these patterns are collectively subjected to additional learning in an additional learning step, and an identification symbol is added to the obtained coupling load between the layer units and accumulated. A pattern recognition method using a hierarchical neural network.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a functional block diagram of an apparatus used in a pattern recognition method using a hierarchical neural network according to an embodiment of the present invention. Reference numeral 1 denotes a hierarchical neural network processing device having a central function of the present invention. 2
Is a pattern input device. 3 is a device for storing learning patterns, 4 is a device for storing target values of the output of the hierarchical neural network, 5 is a device for storing learned connection weights, and 6 is a device for storing patterns for additional learning. is there. A display 7, a keyboard and a mouse 8 are connected to the hierarchical neural network processing device 1. In order to explain the operation and function of the hierarchical neural network device, FIG. 1 is used while showing the operation screens of FIG. 2, FIG. 3, FIG. 4, and FIG.
This will be described in detail below.

When the apparatus is started, an operation screen as shown in FIG. 2 is displayed on the display 7 by the control and screen display routine 100 shown in FIG. "Initial learning", "Recognition"
Alternatively, when one of the “additional learning” is clicked with a mouse, the neural network apparatus enters a selected mode. When "initial learning" is selected, a menu for operating instructions and setting of the initial learning mode as shown in FIG. 3 is displayed by the control and screen display routine 100 of FIG.
Clicking the "Parameter setting" menu with the mouse will display a submenu surrounded by a dashed line. The operator sets the configuration of the hierarchical neural network, the number of learning patterns, the learning parameters, and the like using the keyboard and the mouse 8 while viewing this submenu. These setting values are taken in by the parameter setting routine 101. When the setting is completed, click “Setting Complete” to complete the “Parameter Setting” menu.

Next, "learning pattern and target value setting" in FIG. 3 is selected. When this is selected, a submenu is displayed. When a pattern number is designated by "pattern designation", the corresponding learning pattern is inputted from the pattern input device 2 by the pattern setting routine 102 of FIG. The operator sets a target value for the displayed pattern using the “target value setting” submenu. Fig. 1 shows the "Target value setting" submenu.
Is started, and a pattern number or a symbol indicating a set corresponding to the display pattern is added to the set target value, and the display pattern is stored and set in the learning pattern storage device 3. The obtained target value is stored in the target value storage device 4. This operation is repeated the number of times set in the “parameter setting”, and “setting of learning pattern and target value” is completed by clicking “setting completed”. After that, select "learning". When “learning” is selected, the control and screen display routine 100 starts a learning routine 105. The learning routine 105 sequentially takes in a set of learning patterns and their target values from the learning pattern storage device 3 and the target value storage device 4 and performs learning by the error back propagation method. During the learning, the learning process such as the number of times of learning and the state of reduction of the total error is displayed on the display 7. When the learning is completed, “learning completed” is displayed on the display 7. If learning is not successful or if it is desired to improve the learning speed, it is possible to select "stop learning", return to "parameter setting" again, change the parameters, and re-learn. When learning is complete,
Select “Coupling load identification symbol setting”. This is because an identification symbol is added to the learned connection weight and is effectively used for subsequent recognition. When you click the “Identification Symbol List” submenu, a list of identification symbols that have been set appears on the screen. The operator refers to this list and identifies the learned connection weight using the “Identification Symbol Setting” menu. The symbol is set using the keyboard 8. When "setting completed" is clicked, the learned connection weight is added with the set identification symbol and stored in the learned connection weight storage device 5, and the screen returns to FIG. The above is the processing in the initial learning mode.

Next, the recognition mode will be described. When "recognition" is selected in FIG. 2 displayed on the display 7, the mode shifts to the recognition mode and the menu shown in FIG. 4 is displayed. The “designation of load identification symbol” is for determining the learned coupling weight used in the recognition routine 106 by specifying the identification symbol of the learned coupling weight stored in the learned coupling load storage device 5. If you select "Load identification code designation", "Identification symbol list", "Batch designation", "Individual designation", "Range designation",
A submenu such as "Logical expression designation" appears. The “list of identification symbols” displays a list of identification symbols set so far in “setting of coupling weight identification symbols” in the initial learning mode. With reference to this list, the connection weight used for recognition is specified as follows. "Batch designation" is selected when all registered identification symbols are designated. "Individual designation" is selected when an arbitrary number of individual identification symbols are designated.
“Specify range” is selected when the identification symbol is specified as a range when the identification symbol is registered as a continuous natural number.
"Designation of logical formula" is used when specifying an identification symbol by a logical formula in which a logical sum, a logical product, and the like are combined when the identification symbol is given by a combination of numbers and keywords. The operator clicks the appropriate submenu described above to set the learned connection weight identification symbol. Clicking "Specification Complete" completes "Identification Symbol Specification".

Next, "recognition" is selected. Control and painting
The face display routine 100 selects the “recognition” to execute the recognition processing.
Command to execute the recognition operation to the
Emit. That is, the finger from the learned connection load storage device 5
Captures the learned connection weights with defined identification symbols,
Next, a recognition pattern is input from the pattern input device 2.
Give instructions to recognize. Hereinafter, it operates in the recognition mode.
Recognition routine 106 and evaluation routine 107 and additional
Since the learning pattern processing routines 108 are interrelated,
This will be described with reference to the flowchart shown in FIG. First, figure
No. 6 601 designation of the connection load identification code is shown in FIG.
Designation ”. Here, for example, "Specify range"
If the symbol ID = 1 to M is specified, this is the identification symbol
M learned connection weights WT with ID = 1 to M (ID
= 1 to M) to recognize the pattern to be recognized.
Will do it. 602 is the first pattern number for recognition
N _P= 1 is set. In 603, the maximum specified in 601
First identification code ID = 1 is set. Learned in 604
Only the connection weight WT (ID = 1) has been learned.
The recognition processing routine 106 is read from the device 5. At 605
Is the pattern for the first recognition set in 602
(N_P= 1) from the pattern input device 2 for recognition processing
Read into routine 106. First in 606 output calculation
Pattern (N_P= 1) for the learned load WT (ID =
Output y of neural network using 1)_k ^Np(N
_P= 1) and (k = 1 to K). Fig. 1
Of the recognition processing routine 106 of FIG. The following reviews
This is the processing of the price processing routine 107. With a 607 rating
Eyes entered when finding the learned load WT (ID = 1)
Standard value t_k ^P((K = 1 to K), (learning pattern p = 1
~ P)) this neural network output y_k ^Np
((N_P= 1), (k = 1 to K))
Evaluate whether it is present. The method is already explained
The following pattern error ε (p) is used.

[0017]

(Equation 3)

In the evaluation processing routine 107, y _k ^Np
, Y _k ^Np ((N _P = 1) obtained by the output calculation of 606,
(K = 1 to K)), and a target value t _k ^P ((k) of the unit k (k = 1 to K) for the learning patterns p = 1 to ^P
= K to K) and (learning pattern p = 1 to P)) to calculate the pattern error ε (p). P thus obtained
The pattern Pmin which has the minimum ε (p) among the pattern errors ε (p) is obtained. That is, Min ｛ε (p) | p
= 1 to P｝. If the minimum pattern error ε of the pattern Pmin (Pmin) is less than the specified value, the target value t _k which is a pattern Pmin and set
^P (p = Pmin) and its minimum pattern error ε (Pmi
n) is temporarily stored at 609 as a recognition result. If there is no minimum pattern error ε (Pmin) below the specified value,
At 608, the next coupling load identification symbol is set, and at 604,
Similar processing is performed in 605, 606, and 607. It is determined at 610 whether all the learned connection weights set at 601 have been used for the first recognition pattern (N _p = 1), and if an unused learned connection weight still remains, 6
At 08, the next learned connection weight is designated and the same processing is repeated. After using all specified the learned connection weights, the target value has been temporarily stored as the recognition result at 611 in ^{_{609 t p k (p = Pmin}} ) and minimum pattern error ε
It is checked how many (Pmin) are stored. If only one set of the target value and the minimum pattern error is stored, this is output to the display 7 via the control and screen display routine 100 as a final recognition result at 613. This corresponds to the “recognition result” submenu of “recognition” in FIG.
Clicking on the submenu "Recognition confirmation" proceeds to the recognition of the next pattern. If the stored target value and minimum pattern error are two or more pairs, the smallest set of the minimum pattern error and the target value is searched for in 612, and this is output as the optimum recognition result in 613. . This also corresponds to the "recognition result" submenu of "recognition" in FIG. Clicking on the submenu "Recognition confirmation" proceeds to the recognition of the next pattern. If no set of the target value and the minimum pattern error is stored, this pattern cannot be recognized, and the sub menu “recognition unrecognizable message” of “recognition” in FIG. ", An alarm of" unrecognizable "is issued at 616. The above is the evaluation processing routine 107 of FIG.
This is the process.

The additional pattern storage processing routine 1 will be described below.
Move to the processing of 08. Additional pattern storage processing routine 10
In step 8, an unrecognizable pattern is stored for additional learning thereafter. At 617, the number of unrecognizable patterns to be stored is counted. At 618, if the count value is equal to or smaller than the specified value N, the unrecognizable patterns are stored at 619, and the process proceeds to recognition of the next pattern at 614. If the count value reaches the specified value N at 618, as shown in the submenu "additional learning message" of "recognition" in FIG. 4, "additional learning message" is output at 620 and the recognition mode is ended. Up to this point, the additional learning pattern storage processing routine 10 of FIG.
8 is the process. Such processing is performed using the recognition pattern (N
_P = 1 to NP). To end the recognition, click "Recognition Complete" in FIG. 4 with the mouse, and the screen returns to FIG. Here, it is also possible to select and execute “initial learning” or “recognition” again. FIG. 4 at 620
If the "recognition" submenu "additional learning message" is output, selecting "additional learning" in FIG. 2 causes the menu in FIG. 5 to appear. In “Pattern number setting”,
It sets how many of the N “unrecognizable” patterns stored in the additional learning pattern storage device 6 are used as the additional learning patterns. When "pattern transfer" is clicked, the number of "unrecognizable" patterns set in "pattern number setting" is transferred from the additional learning pattern storage device 6 to the learning pattern storage device 3 by the additional learning pattern storage processing routine 108. . Hereinafter, the menus of “parameter setting”, “learning pattern and target value setting”, “learning”, and “coupling load identification symbol setting” are the same as those in the initial learning mode, and thus description thereof will be omitted.

[0020]

According to the present invention, when a pattern such as an image, a figure, and a sound is recognized by the hierarchical neural network device, if the recognition is not possible, the unrecognizable pattern is stored for additional learning and added as a learning pattern later. Learning can be performed, and the connection weights obtained by the additional learning can be sequentially accumulated and used for subsequent pattern recognition.Therefore, it is possible to restart the learning including the unrecognizable pattern, which is conventionally performed, from the beginning. It becomes unnecessary, and the accumulated coupling weight can be arbitrarily and appropriately specified and used for pattern recognition, so that the recognition ability can be improved.

[Brief description of the drawings]

FIG. 1 is an overall view of an embodiment of the present invention.

FIG. 2 is a display diagram of a screen showing a menu of operations, settings, processing results, and the like of the hierarchical neural network device in the present invention.

FIG. 3 is a display diagram of a screen showing a menu of operations, settings, processing results, and the like of the hierarchical neural network device in the present invention.

FIG. 4 is a display diagram of a screen showing a menu of operations, settings, processing results, and the like of the hierarchical neural network device in the present invention.

FIG. 5 is a display diagram of a screen showing a menu of operations, settings, processing results, and the like of the hierarchical neural network device in the present invention.

FIG. 6 is a processing system diagram of recognition, evaluation, and additional learning in the embodiment of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Hierarchical neural network processing apparatus, 2 ... Pattern input apparatus, 3 ... Learning pattern storage apparatus, 4 ... Target value storage apparatus, 5 ... Learned connection load storage apparatus, 6 ... Additional learning pattern storage apparatus, 7 ... Display 8 Keyboard and mouse 100 Control and screen display routine 101 Parameter setting routine 102 Pattern setting routine 103 Target value setting routine 10
5: learning processing routine, 106: recognition processing routine, 1
07 ... Evaluation processing routine, 108 ... Additional learning pattern storage processing routine.

────────────────────────────────────────────────── (5) References JP-A-5-290013 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T ⁷ /00-7/60 G06N 3 / 00-3/10

Claims (3)

(57) [Claims] 1. An image in a hierarchical neural network,
Enter patterns, such as shapes and sounds, classify them,
The recognition method for discriminating and judging includes learning and recognition steps. In the learning step, a pattern to be learned and a target value are input, and learned connections between the respective layer units of the neural network obtained by learning about these are provided. An identification code is added to the load, stored and stored. In the recognition step, the stored connection weight is specified, and a recognition process is performed for a pattern to be recognized using the specified connection weight. Based on the output value obtained and the target value in the learning process for obtaining the used coupling weight, the minimum value of each pattern error obtained by using the respective coupling weight is temporarily stored, From the minimum values of the pattern errors obtained by using all the specified connection weights, a further minimum value of the pattern error is obtained. If less beforehand scheduled tolerance, learning about this pattern minimum error pattern,
Alternatively, the target value is output as the final recognition result, and if the value is not less than the allowable value, a message is issued as recognizable and a message is issued. 2. An image in a hierarchical neural network,
A method of inputting patterns such as figures and voices and recognizing them is provided with learning, recognition, and evaluation steps.The learning pattern and target values in the learning step are input and obtained by learning about them. An identification symbol is added to the learned connection weights between the units of each layer of the neural network, which are sequentially stored and stored. In the recognition process, one of the stored connection weights to be used is specified, and the specified connection weight is designated. In the evaluation step, a pattern error between the output value obtained as a result of the above-described processing and a target value used in learning a learning pattern for obtaining a connection weight used is performed in an evaluation step. Find the smallest one of these, and if this pattern error is less than or equal to a predetermined tolerance, the learning pattern, target value, and minimum pattern error When the time is stored, a set of the minimum pattern error and the target value is further obtained from the stored pattern error minimum values obtained using all the specified coupling weights, and this is set as the optimum recognition result. A pattern recognition method using a hierarchical neural network, comprising: outputting a non-recognizable message when no set of a minimum pattern error and a target value is stored. 3. An image in a hierarchical neural network,
In the method of inputting patterns such as figures and voices and recognizing them, the method includes initial learning, recognition, evaluation, and additional learning steps. In the initial learning and additional learning steps, a learning pattern and a target value are input. In addition, an identification symbol is added to the learned connection weights between the units of the neural network obtained by learning the learning pattern and stored and stored, and in the recognition process, one of the stored connection weights to be used is specified. Then, a pattern to be recognized is recognized using the specified connection weight, and in the evaluation step, the output obtained in the above-described recognition step and the target value used during learning pattern learning for obtaining the used connection weight are used. Find the minimum of the pattern errors in the above, and if this pattern error is less than or equal to a predetermined tolerance, the learning pattern and target value And the minimum pattern error is temporarily stored, and a minimum set of the pattern error and the target value is further obtained from the minimum values of the pattern errors obtained using all the specified coupling weights, and this is determined as the optimum recognition result. If no set of the minimum pattern error and the target value is stored, a message indicating that the pattern cannot be recognized is output. A pattern recognition method using a hierarchical neural network, characterized in that additional learning is performed collectively on these patterns in a learning step, and identification codes are added to the obtained connection weights between the layer units and stored.