JPH07146936A - Pattern recognizing method by hierarchical neural network - Google Patents

Abstract

PURPOSE:To eliminate the need for relearning including a pattern which can not be recognized and to improve the capability of recognition by storing the pattern which can not be recognized, and accumulating and utilizing coupling weight obtained by additional learning for pattern recognition. CONSTITUTION:In an evaluation process 107, the minimum pattern error between the output of a recognition process 106 and a target value used at the time of learning pattern learning for used finding coupling weight is found and when no pair of the minimum pattern error and target value is stored, a message showing that pattern recognition can not be performed is displayed on a display 7. In an additional pattern storing process 108, the pattern which can not be recognized is stored for subsequent additional learning, and the number of patterns which can not be recognized is counted; when the counted value reaches a prescribed value, those patterns are learnt together additionally and a discrimination symbol is added as to the coupling weight between respective obtained layer units and stored in a learning pattern storage device 6, and utilized for pattern recognition.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to input images, graphics,
The present invention relates to a pattern recognition method using a hierarchical neural network for classifying, identifying, judging, or recognizing patterns such as voices, and in particular, makes it possible to appropriately deal with recognition processing for patterns that are significantly different from learned patterns. The present invention relates to a pattern recognition method using a hierarchical neural network that enables effective processing by appropriately selecting a weight between units used for a neural network.

[0002]

2. Description of the Related Art Various types of patterns such as images, figures, and sounds
A layered neural network used to recognize
First, the learning pattern and the hierarchical neural network
The target value of the output of each unit in the
(Called) is input as a set and learning is performed. Study
In practice, the number of learning patterns is equal to or more than
Enter groups in sequence or randomly and add to each learning pattern
In contrast, each unit in the output layer of the hierarchical neural network
So that the actual output of the knit approaches each target value
Exist in the input layer, middle layer, and output layer respectively.
Correct the coupling load between units. Typical of that modification
The error backpropagation method is well known, but the error backpropagation method is often used.
Since it is explained in detail in various documents and is well known, here
Omit it. In learning, the pattern to be recognized
Learning is performed for all of the
Completion is determined by error evaluation as described below.
It For each learning pattern of each unit in the output layer
The square of the difference between the actual output value and its target value is calculated for all output layers.
Add for units, divide by the number of units, and add
The square root of the normalized value is called the pattern error of this pattern.
Bu That is, the units in the output layer for the pattern p
the target value of k is t_k ^P, The actual output of unit k is y _k ^P
And the pattern error ε (p) is

[0003]

[Equation 1]

It is defined as However, K is the number of units in the output layer. The average value of all pattern errors will be called the average pattern error. Average pattern error ave
Is

[0005]

[Equation 2]

It is represented by Here, P is the number of learning patterns. This average pattern error for each learning
If ave is calculated and it is less than the allowable value, learning is completed. In the case of recognizing the learning pattern by using the connection weight at the time of completion of learning, it is possible to recognize the pattern with a precision less than the maximum pattern error among the pattern errors. Even a pattern that has not been used for learning can be approximately recognized within this maximum pattern error if it is similar to the learning pattern.

[0007]

The conventional problem of pattern recognition by a hierarchical neural network is to recognize a new pattern other than a pattern greatly different from a learning pattern or a pattern given for learning. This is the case. Recognition of these patterns using the learned connection weights is generally unsuccessful. In order to enable recognition of this kind of pattern as well, it is generally performed to add these as new patterns to the learning pattern and to perform learning again as a learning pattern. In this case, re-learning requires a longer learning time because the number of sets of learning patterns increases, and the connection weight obtained by the learning is reset and the learning is restarted from the beginning.

The object of the present invention is to solve the above-mentioned problems of the prior art, and to appropriately process even if the input pattern is very different from the pattern used at the time of learning and cannot be identified or judged. It is to provide a pattern recognition method using a hierarchical neural network.

[0009]

In order to achieve the above object, the invention claimed in the present application is as follows. (1) A method of inputting patterns such as images, figures, and voices to a hierarchical neural network and classifying, identifying, and judging these (hereinafter referred to as recognition) includes learning and recognition steps, and a learning step Then, input the pattern to be learned and the target value, add the identification symbol to the learned connection weight between each layer unit of the neural network obtained by learning about these and store it, and store it in the recognition step. The specified connection weight is specified, the recognition process is performed on 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 to obtain the used connection weight Based on and, the minimum value of each pattern error calculated by using each connection weight is temporarily stored, and all the specified connection The minimum pattern error minimum value is obtained from the above pattern error minimum values obtained by using the overlap, and if the minimum pattern error value is equal to or less than the preset allowable value, it is used for learning about this pattern error minimum value. A pattern recognition method using a hierarchical neural network, which outputs a pattern or a target value thereof as a final recognition result, and outputs a message as unrecognizable unless it is less than an allowable value.

(2) A method of inputting patterns such as images, figures, and voices into a hierarchical neural network and recognizing these patterns includes learning, recognition, and evaluation steps, and learning patterns and goals in the learning step. By inputting a value, the identification symbol is added to the learned connection weights between the units of each layer of the neural network obtained by learning about them, and they are sequentially stored and accumulated. In the recognition step, among the accumulated connection weights, From the output values obtained as a result of the above processing and the used coupling weight, the recognition processing is performed for the pattern to be recognized using the specified coupling weight. The minimum of the pattern error with the target value used at the time of learning pattern learning is found, and this pattern error is less than or equal to the preset 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 selected from the stored pattern error minimum values obtained by using all the specified connection weights. , And output it as the optimal recognition result,
A pattern recognition method using a hierarchical neural network, wherein an unrecognizable message is issued when no set of minimum pattern error and target value is stored.

(3) A method of inputting patterns such as images, figures, and voices to a hierarchical neural network and recognizing these patterns is provided with initial learning, recognition, evaluation, and additional learning steps. In the learning step, the learning pattern and the target value are input, and the learning symbols between the respective layer units of the neural network obtained by learning of the learning pattern are added with an identification symbol and stored and accumulated. From the accumulated connection weights, specify the one to be used, recognize the pattern to be recognized using the specified connection weight, and in the evaluation step the output obtained in the above recognition step and the connection weight used. The minimum of the pattern error with the target value used at the time of learning the learning pattern is found. If it is less than the specified allowable value, the learning pattern, the target value, and the minimum pattern error are temporarily stored, and the minimum value of the pattern error minimum values obtained using all the specified connection weights is used. A pair of pattern error and target value is obtained and output as an optimum recognition result.If no pair of minimum pattern error and target value is stored, a pattern recognition not possible message is output and the recognition is performed. When the impossible patterns are stored and the number of stored patterns reaches a predetermined number, these patterns are collectively subjected to additional learning in the additional learning step, and the obtained connection weight between each layer unit is added with an identification symbol and accumulated. Pattern recognition method using a hierarchical neural network.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an overall 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 is a hierarchical neural network processing device having a central function of the present invention. Two
Is a pattern input device. 3 is a device for storing the learning pattern, 4 is a device for storing the target value of the output of the hierarchical neural network, 5 is a device for storing the learned connection weight, and 6 is a device for storing the additional learning pattern. is there. A display 7, a keyboard and a mouse 8 are connected to the hierarchical neural network processing device 1. To explain the operation and function of the hierarchical neural network device, referring to FIG. 1 while showing the operation screens of FIGS. 2, 3, 4 and 5 displayed on the display 7,
The details will be described below.

When the apparatus is activated, the control and screen display routine 100 shown in FIG. 1 causes the display 7 to display an operation screen as shown in FIG. "Initial learning", "recognition"
Alternatively, when one of the "additional learning" is clicked with the mouse, the neural network device enters the selected mode. When "initial learning" is selected, the control and screen display routine 100 of FIG. 1 displays a menu as shown in FIG. 3 for operating instructions and setting of the initial learning mode.
When you click the "Parameter Setting" menu with the mouse, a submenu surrounded by a chain line appears. The operator sets the configuration of the hierarchical neural network, the number of learning patterns, the learning parameters, and the like by using the keyboard and the mouse 8 while viewing this submenu. These set values are fetched by the parameter setting routine 101. After completing the settings, click "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 will be displayed. When the pattern number is designated by "pattern designation", the corresponding learning pattern is input from the pattern input device 2 by the pattern setting routine 102 of FIG. 1 and displayed on the display 7. The "target value setting" submenu allows the operator to set a target value for the displayed pattern. The “Target value setting” submenu is shown in Figure 1.
The target value setting routine 103 is started, a pattern number or a symbol indicating that it is a set corresponding to the display pattern is added to the set target value, and the display pattern is stored in the learning pattern storage device 3 and set. The desired value is stored in the desired value storage device 4. This operation is repeated for the number of patterns set in the "parameter setting", and the "setting of learning pattern and target value" is ended by clicking "setting complete". After this, select "learn". When “learning” is selected, the control / screen display routine 100 starts a learning routine 105. The learning routine 105 sequentially takes in a learning pattern and a target value thereof, which form a set 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 reduction state of the total error is displayed on the display 7. When learning is completed, “learning completed” is displayed on the display 7. If the learning is not successful or if the learning speed is desired to be improved, it is possible to select "stop learning" and return to "parameter setting" again to change the parameters and re-learn. When learning is completed,
Select "Coupling load identification symbol setting". This is because an identification symbol is added to the learned coupling weights and it is effectively used for subsequent recognition. If you click the "Identification symbol list" submenu, a list of identification symbols that have been set up to now will appear on the screen.The operator can refer to this list and identify the learned coupling load with the "Identification symbol setting" menu. The symbol is set using the keyboard 8. When "setting complete" is clicked, the learned connection weight is stored in the learned connection weight storage device 5 with the set identification symbol added, 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 is shifted to the recognition mode and the menu shown in FIG. 4 is displayed. The “load identification symbol designation” is for determining the learned binding weight used in the recognition routine 106 by designating the identification symbol of the learned binding weight stored in the learned binding weight storage device 5. When "Load identification symbol designation" is selected, "Identification symbol list" surrounded by the one-dot chain line, "Batch designation", "Individual designation", "Range designation",
A submenu such as "Specify logical expression" appears. The “identification symbol list” displays a list of identification symbols that have been set so far in the “coupling load identification symbol setting” in the initial learning mode. By referring to this list, the binding weights used for recognition are specified as follows. "Batch designation" is selected when designating all registered identification symbols. "Individual designation" is selected when individually designating an arbitrary number of identification symbols.
"Range designation" is selected when designating an identification symbol in a range when the identification symbol is registered as a continuous natural number.
"Logical expression designation" is used when the identification symbol is attached by a combination of numbers and keywords, and is designated by a logical expression that combines logical sums and logical products. The operator clicks on the appropriate submenu above to set the discriminant for the learned coupling weights. Click "Designation Complete" to complete "Identification Symbol Designation".

Next, "recognition" is selected. Control and image
The surface display routine 100 selects the “recognition” to select the recognition process.
A command for executing the recognition operation is issued to the physical routine 106.
Emit. That is, the finger from the learned combined load storage device 5
Take in the trained connection weights with a defined identifier,
Next, input the recognition pattern from the pattern input device 2.
Instruct them to recognize. In the following, it operates in recognition mode.
Recognition routine 106 and evaluation routine 107 and addition
Since the learning pattern processing routine 108 is related to each other,
This will be described with reference to the flowchart shown in FIG. First, the figure
No. 6 601 of the combined load identification code designation is “identification code” in FIG.
Corresponding to "designation". Here, for example, identify with "range specification"
If you specify the symbol ID = 1 to M, this is the identification symbol.
M learned connection weights WT (ID
= 1 to M) to recognize the pattern to be recognized.
Will be done. First recognition pattern number in 602
N _PSet = 1. In 603, the maximum specified in 601
The first identification code ID = 1 is set. Learned in 604
Only the combined load WT (ID = 1) has been learned.
The recognition processing routine 106 is read from the unit 5. At 605
Is the pattern for the first recognition set in 602
(N_P= 1) is input from the pattern input device 2 and recognition processing is performed.
It is read into the routine 106. First in the output calculation of 606
Pattern (N_P= 1) learned weight WT (ID =
The output y of the neural network using 1)_k ^Np(N
_P= 1), (k = 1 to K) are calculated. The above is Fig. 1
Is a process of the recognition processing routine 106. From the following
This is the processing of the price processing routine 107. In the evaluation of 607
Eyes input when obtaining the learned weight WT (ID = 1)
Threshold t_k ^P((K = 1 to K), (learning pattern p = 1
~ P)) in this neural network output y_k ^Np
((N_P= 1), (k = 1 to K)) is the closest
Evaluate if it exists The method has already been explained
The following pattern error ε (p) which has been clarified is used.

[0017]

[Equation 3]

In the evaluation processing routine 107, y _k ^{Np in the} above equation
Y _k ^Np ((N _P = 1), obtained by the output calculation of 606,
(K = 1 to K)), the target value t _k ^P of the unit k (k = 1 to K) for the learning patterns p = 1 to ^P ((k
= K to K), (learning pattern p = 1 to P)), the pattern error ε (p) is calculated. P thus obtained
The pattern Pmin which is the minimum ε (p) among the individual pattern errors ε (p) is obtained. That is, Min {ε (p) | p
Then, a pattern Pmin for which = 1 to P} is obtained. If the minimum pattern error ε (Pmin) of the pattern Pmin is less than or equal to the specified value, the target value t _k paired with the pattern Pmin.
^P (p = Pmin) and its minimum pattern error ε (Pmi
n) is temporarily stored as a recognition result in 609. If there is no minimum pattern error ε (Pmin) below the specified value,
At 608, the next coupling load identification symbol is set, and 604,
Similar processing is performed at 605, 606, and 607. At 610, it is determined whether all the learned connection weights set at 601 have been used for the first recognition pattern (N _p = 1), and if there are still unused connection weights, then 6
At 08, the next learned connection weight is designated and the same processing is repeated. After using all the specified learned connection weights, in 611, the target value t _p ^k (p = Pmin) temporarily stored as the recognition result in 609 and the minimum pattern error ε.
Check how many sets (Pmin) are stored. If only one set of the target value and the minimum pattern error is stored, this is output as the final recognition result to the display 7 via the control and screen display routine 100 at 613. This corresponds to the "recognition result" submenu of "recognition" in FIG.
If you click the "Confirmation" submenu, you will proceed to the recognition of the next pattern. When the stored target value and minimum pattern error are two or more sets, the smallest combination 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. If you click the "Confirmation" submenu, you will proceed to the recognition of the next pattern. If no set of the target value and the minimum pattern error is stored, the pattern cannot be recognized, and the sub-menu “Unrecognized message” and “Additional learning pattern storage” of “recognition” in FIG. ”, An alarm“ unrecognizable ”is issued at 616. The above is the evaluation processing routine 107 of FIG.
Processing.

From the following, additional pattern storage processing routine 1
The process moves to 08. Additional pattern storage processing routine 10
At 8, the unrecognizable pattern is stored for additional learning thereafter. The number of unrecognizable patterns stored in 617 is counted, and if the count value is less than the specified value N in 618, the unrecognizable pattern is stored in 619 and the next pattern is recognized in 614. When the count value reaches the specified value N in 618, the "additional learning message" is output in 620 and the recognition mode is ended, as shown in the "additional learning message" submenu of "recognition" in FIG. The above is the additional learning pattern storage processing routine 10 of FIG.
It is the process of 8. Such processing is performed by using the recognition pattern (N
_P = 1 to NP) are sequentially performed. To end the recognition, click "recognition completed" in FIG. 4 with the mouse to return the screen to FIG. Here, it is also possible to select and execute "initial learning" or "recognition" again. FIG. 4 at 620
When the “additional learning message” submenu of “recognition” is output, the menu of FIG. 5 appears when “additional learning” of FIG. 2 is selected. In “Pattern number setting”,
The number of N “unrecognizable” patterns stored in the additional learning pattern storage device 6 is set as the additional learning pattern. When "Transfer pattern" is clicked, the number of "unrecognizable" patterns set in "Set number of patterns" 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. . The respective menus of "parameter setting", "learning pattern and target value setting", "learning", and "coupling weight identification symbol setting" are the same as those in the initial learning mode, and therefore description thereof will be omitted.

[0020]

When a pattern such as an image, a figure, or a voice is recognized by a hierarchical neural network device, if the recognition is impossible, the unrecognizable pattern is stored for additional learning and added later as a learning pattern. Since it is possible to perform learning and accumulate the connection weights obtained by the additional learning sequentially and use them for subsequent pattern recognition, it is possible to redo learning including unrecognizable patterns, which is conventionally performed, from the beginning. Further, it becomes unnecessary, and the accumulated coupling load can be arbitrarily and appropriately designated and utilized for pattern recognition, so that the recognition ability can be improved.

[Brief description of drawings]

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

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

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

FIG. 5 is a screen display diagram showing a menu of operations, settings, processing results, etc. 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]

1 ... Hierarchical neural network processing device, 2 ... Pattern input device, 3 ... Learning pattern storage device, 4 ... Target value storage device, 5 ... Learned connection weight storage device, 6 ... Additional learning pattern storage device, 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.

Claims (3)

[Claims] 1. A hierarchical neural network for images,
Input patterns such as figures and voices to classify them.
The recognition method for identifying and determining includes learning and recognition steps. In the learning step, a pattern to be learned and a target value are input, and learning connection between each layer unit of the neural network obtained by learning about these. An identification symbol is added to the load to store and accumulate it. In the recognition step, the accumulated bond weight is specified, and the recognition process is performed for the pattern to be recognized using the specified bond weight. Based on the output value and the target value in the learning process to obtain the used connection weight, the minimum value of each pattern error obtained by calculation using each connection weight is temporarily stored, The minimum pattern error minimum value is obtained from the above minimum pattern error values obtained using all the specified connection weights, and the minimum pattern error value is If less beforehand scheduled tolerance, learning about this pattern minimum error pattern,
Alternatively, a pattern recognition method using a hierarchical neural network is characterized in that the target value is output as a final recognition result, and if the target value is not less than the allowable value, the recognition is not possible and a message is issued. 2. An image is applied to a hierarchical neural network,
The method of inputting patterns such as figures and voices and recognizing them includes learning, recognition, and evaluation steps, and inputs learning patterns and target values in the learning step to obtain them by learning. Neural network Learning codes between units of each layer are added with an identification symbol and sequentially stored and accumulated. In the recognition process, the one to be used is specified from the accumulated connection weights, and the specified connection weight is specified. The recognition process for the pattern to be recognized is performed by using, and in the evaluation step, the output value obtained as a result of the above process and the pattern error of the target value used at the time of learning pattern learning for obtaining the used coupling weight Find the smallest of these, and if this pattern error is less than or equal to the preset allowable value, set the learning pattern, target value, and minimum pattern error. The set of the minimum pattern error and the target value is obtained from among the stored minimum values of the pattern error obtained by using all the specified connection weights, and this is set as the optimum recognition result. A pattern recognition method using a hierarchical neural network, which outputs and outputs an unrecognizable message when no set of minimum pattern error and target value is stored. 3. An image is applied to the hierarchical neural network,
The method of inputting patterns such as figures and voices and recognizing them includes the steps of initial learning, recognition, evaluation, and additional learning. In the initial learning and additional learning steps, the learning pattern and the target value are input. , The neural network obtained by learning about the learning pattern is stored with the identification symbol added to the learned connection weight between each layer unit, and in the recognition step, the one to be used is specified from the accumulated connection weights. Then, the pattern to be recognized is recognized using the specified connection weight, and in the evaluation step, the output obtained in the above recognition step and the target value used during learning pattern learning for obtaining the connection weight used If the pattern error is less than or equal to the preset tolerance, the learning pattern and the target value are calculated. And the minimum pattern error are temporarily stored, and the set of the minimum pattern error and the target value is obtained from the minimum values of the above pattern errors obtained by using all the specified connection weights, and the optimum recognition result is obtained. When 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 the number of stored unrecognizable patterns reaches a predetermined number, the addition is made. A pattern recognition method using a hierarchical neural network, characterized in that additional learning is collectively performed on these patterns in a learning process, and an identification symbol is added to the obtained connection weights between the layer units to be accumulated.