JPS6227886A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To have a learning function at the time of recognition by correcting the order logic of picture information so as to collate with a standard order logic. CONSTITUTION:A memory 6 for compiling a dictionary stores a basic order logic inputted from a console 3, a floppy, etc. Based upon the picture data pretreated by a memory 5, a standard order logic inputted from a console 3 is corrected, the identifying function is established, and dictionaries I and II are compiled. A memory 7 is the one to store the order logic which is a part of the dictionary, the order logic corrected by the memory 6 and the address of the identifying function used in a pair with this are recognized, and a part (dictionary I) of the dictionary at the time of recognization is obtained. A memory 8 is the one to store the identifying function which is an another part of the dictionary, the identifying function established by the memory 6 and the recognizing result identified by using this are stored and the part (dictionary II) of the remaining dictionary is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a pattern recognition device. This paper relates to matching with sequential logic and identification processing using identification functions.

The present invention also relates to a pattern recognition device having a learning function.

The present invention relates to automatic creation of a dictionary when developing OCR.

The present invention particularly relates to a pattern recognition device having a learning function, which includes preprocessing such as pattern size normalization processing, thinning processing, and Cheno coding, major classification processing using nine topological features, and sequential logic. This method consists of intermediate classification based on comparison with other people, and identification processing using a discriminant function.

[Prior art]

BACKGROUND ART Many studies have been conducted on character recognition, and various recognition methods and devices have been put into practical use. However, it has been difficult to develop such things as creating a dictionary, and it has also been difficult to eliminate misreading. However, since it does not have a learning function, similar misreadings may be repeated, resulting in lower recognition rates and lower device usage efficiency, which is one of the reasons why recognition devices are difficult to spread.

In addition, one recognition method, for example, is a method of analyzing the structure of the strokes of a pattern, but this method has the advantage of obtaining a higher recognition rate than other recognition methods such as the similarity method. However, it has the disadvantage that creating a dictionary is difficult and time-consuming. Therefore, automatic creation of a dictionary for the structural analysis method is important in terms of recognition rate, and research on learning functions is required.

Although the known technology and device are not related to the learning function, there are 1
For example, the handwritten character recognition device (H-
8959OCR).

The following processing is performed on the character pattern. That is, a character pattern is converted into a line diagram, expressed as a direction code called a chain code, and identified based on the code by comparison with the order logic of standard patterns in a dictionary. Then, f1 is the number of times the state transition state passes.
1 is provided to improve the recognition rate.

However, depending on the character, the starting point of the chain code cannot be uniquely determined, and there are many things that cannot be expressed using a single sequential logic.For example, it is said that approximately 100 sequential logics are required for handwritten digits. The disadvantage is that it takes a lot of time to create and recognize.

As is clear from the above, the creation of a dictionary with a learning function in character pattern recognition.

Currently, research on learning during recognition is desired.

[Problem that the invention seeks to solve]

As is clear from the above points, an object of the present invention is to eliminate the above-mentioned drawbacks.

Another object of the present invention is to provide a learning function in the development of OCR or the creation of a dictionary during pattern recognition.

Another object of the present invention is to provide a pattern recognition device with a learning function during recognition.

Another object of the present invention is to provide a pattern recognition device which incorporates the idea of sequential logic (automaton) into the stroke structure analysis method and which also performs classification processing using a classification function.

An object of the present invention is to automatically create a dictionary for online or offline character recognition.

The first object of the present invention is to create a Cheno code for dictionary creation based on a Cheno code for pattern recognition.

An object of the present invention is to modify the sequential logic, which is a dictionary, based on the Cheno code, and to learn better sequential logic.

Another object of the present invention is to create a discriminant function whose variable is the number of times each state of sequential logic, which is dictionary data, is passed.

[Means for solving problems]

As a means to solve this problem, for example, the pattern input device shown in the block diagram of FIG.
The dictionary information is preprocessed in the preprocessing working area of , and the dictionary information is corrected and created in the memory 6, and then transferred to the memory 7. The memory 7 stores the sequential logic and the address of the discriminant function used in pair with the sequential logic.

Memory 8 is a part that stores identification functions.

[For production]

In the configuration shown in FIG. 1, the CPU 1 has a memory 4.
Execute the program stored in .

The image information read from the scanner 2 is preprocessed in the memory 5, either manually from the console 3 or from an external memory. The basic sequential logic from the optical disk or floppy disk is modified by the preprocessed data and a discriminant function is created.

〔Example〕

One embodiment of the present invention will be described below.

FIG. 1 is a block diagram of an apparatus for implementing the present invention. 4 is a memory in which a program for carrying out the present invention is stored.

1 is a C'PU, which controls and operates other device parts according to a program stored in the memory 4.

Reference numeral 2 denotes a scanner, which converts image data into information that can be processed by a computer using optical reading means.

Reference numeral 3 denotes a console, which consists of a keyboard and a display, and includes means for interfacing with the CPU, such as data input.

A preprocessing memory 5 is a memory area for storing image data read by the scanner 2 and performing preprocessing such as image processing and coding on the image data.

Reference numeral 6 denotes a memory for creating the dictionary 3, which stores basic sequential logic input from the console 3 or floppy disk. Then, based on the image data preprocessed in the memory 5, the standard sequential logic input from the console 3 (or it may be read out from an external memory such as a floppy disk), and the discriminant function Create dictionaries I, I
Create I.

A memory 7 stores the sequential logic that is part of the dictionary, and stores the sequential logic corrected in the memory 6 and the address of the discriminant function used in pair with it.
).

Reference numeral 8 denotes a memory for storing the discriminant function, which is another part of the dictionary. The discriminant function created in memory 6 and the recognition result discriminated using this are stored, and part of the remaining dictionary for recognition is stored. (Dictionary 11).

[Operation flow diagram]

FIG. 2 is a flowchart showing the operation of the present invention.

The operation will be explained in more detail using FIG.

The flow of operations will be explained below.

In this embodiment, patterns that belong to the same category are used intensively, a dictionary is created for that pattern, and then a dictionary is created for patterns that belong to other categories.

First, the standard sequential logic regarding Bata 12/ belonging to the same category is extracted from the aforementioned console 3 or an external storage unit such as a floppy disk, and is stored in the memory 6 (step 5.
6). Next, in the image input process of step 1, one pattern to be pattern recognized is extracted from the image read by the scanner 2, which is an image league such as a digital copying machine, hand scanner, word processor, facsimile, etc. The preprocessing in step 2 consists of binarization processing, size normalization processing, line thinning processing, and chain scratch coding, and preprocessing to convert image data into encoded data suitable for CPU processing and pattern recognition processing. I do.

The image input processing and pre-processing described above are performed in the working area of the memory 5, and the Che7-encoded image is stored.

[Che 7 code] An example of Che7 encoding is shown in FIG.

21 is an 8-way code, which expresses the relationship between adjacent pixels by 1
It is expressed using chords from 8 to 8. The image 20, which has been thinned so that the line width is one pixel, is subjected to CH7 coding from pixel S to pixel E using an eight-way code 21, resulting in a chain code 22.

Obviously, since the image 20 can be reproduced from the Chen Ken code 22, the Che 7 code 22 stores image information.

The memory 7 stores pattern names corresponding to the data of the previously created sequential logic.

Note that the Che7 code 22 is obtained by orienting adjacent pixels from the start point S to the end point E of the image 20 in FIG. 3 using the direction code 21. For example, picture 2s
The direction from pixel to the next pixel is 3, the next pixel is 2, and so on.

In this embodiment, the creation of a dictionary regarding standard logic and the creation of a dictionary regarding discriminant functions are performed sequentially, and the image input processing in step 1 and the preprocessing in step 2 are executed at the stage of creating dictionaries for both.

In the process determination process of step 3, a flag (discrimination function creation if flag (F)=1, which will be described later) is checked to determine whether it is a sequential logic creation stage or a discriminant function creation stage. If it is the step of creating sequential logic (F=O), proceed to Y. In the sequential logic creation stage, the sequential logic is then modified in step 4.

Modification of the sequential logic in step 4 involves storing the standard sequential logic in step 5, which was input from the console corresponding to the category, into the sequential logic memory 6 in step 6, using actual data, and following the learning rules. , modify practical sequential logic one after another, that is, modify and modify.

[Modification of sequential logic]

The sequential logic modification process will be described in detail below.

When creating sequential logic by learning 1, it is necessary to 1) store necessary information, 2) create rules for the learning method, and for this purpose, in the present invention, sequential logic is created as a sample that ignores noise such as digitization noise. By inputting the standard sequential logic from the console, floppy disk, etc., necessary information is saved, and in particular, the direction representing the characteristics of the stroke of the pattern is defined from the transition between states in the standard sequential logic, and the learning method is thereby improved. It is distinctive in that it allows for the creation of rules.

A transition diagram of the standard sequential logic of pattern "7" is shown in FIG.

In FIG. 4, Al, A2. ...A5 is a state, and the code attached to the arrow gives the transition conditions between states.

For example, if the code is 3 in state A1, state A1
, when the code is 2, it is possible to transition to state A2,
Otherwise, the transition is not possible (i.e., no matching,
) Here, when starting from state S and reaching state iE according to the allowed transitions, the sequential logic matches pattern "7".

An example of a Che 7 code obtained by input processing and preprocessing of an image of the actual pattern "7" is 77E. Since the Chen encoding has been described above, a description thereof will be omitted.

This code can transition from state S to state A5 in the standard state transition diagram shown in FIG. 4 (input from the aforementioned console 3 or external storage such as floppy), but code
Since the transition to is not spoofed, it will not be matched. Please check this! The purpose of modifying the sequential logic in step 4 is to modify the rj:, JII transition.

Definitions of directions are illustrated in FIGS. 5-1 and 5-2.

The forward direction of state B in FIG. 5-1 is determined by the transition conditions for states A and C, that is, the arrangement of direction codes 8 and 7 in FIG. The rotation direction when taking a shortcut from to 7 is the forward direction, and the reverse direction is the reverse direction. Similarly for transition states.

Due to the transition conditions for states C and D, that is, the arrangement of codes 1 and 2, the arrow in FIG. 5-2 is in the forward direction.

[Correction and repair of sequential logic]

FIG. 6 shows a flowchart of the sequential logic correction process.

A detailed explanation of the sequential logic correction process will be given based on FIG. Step 30 is a process of converting the Chen 7 code, and in order to simplify the subsequent process, the Chen Φ code obtained in the preprocessing of Step 2 is converted into a code string representing only changes. Note that the original Chen φ code is also used backward for pattern recognition.

For example, code conversion is Che7 code 533321
11L8766776777677E” becomes 532187676767E through the conversion process.

[Verification process]

and The subsequent processing is performed based on the converted code sequence.

The recorded If order logic is stored in the working memory 6 for modification processing by the storage processing in step 38.
moved to the area. Then, in the matching process in step 31, a code string representing only the changing points of the converted code string is compared with the sequential logic.

Code string 532187676767E and the shape ff in Figure 4
, 15, the modification process after the collation process will be explained using the diagram transfer, or sequential logic, as an example.

First, after code string S, there is code 3.

The code allowed to transition from state S in the fourth interval is 3, and the transition is allowed for code 3, that is, the code was verified in the verification determination at step 32. Then, the state FF, S transitions to the state A1. At the end of the collation process in step 33, it is determined whether the collation process is completed based on whether or not the code is at the last code E of the code string. is transferred to the location of the sequential logic stored in step 6, and the sequential logic is rewritten with the corrected one (step 6-A in FIG. 3).

If the process has not been completed, the process returns to step 31.

Then, for example, following the flow of the above-mentioned story, a process of matching code 3 with the next code, that is, code 2, is performed.

In FIG. 4, a transition is made from state S to state A1 due to code 3, but in state A1, transition is allowed for code 3 and code 2. Therefore, the following code 2 can transition from state mAt to state A2, that is, it verifies,
The state transitions to state A2.

Similar processing is repeated until state A3. A transition is made from A4 to state A5 by code 7 which appears first in the code string.

The code after code 7 is code 6.

In state A5, transition is allowed only for codes 7 and E, and matching is not possible. At this time, rule 1 for correction
- Correction processing is performed according to Rule 5 (described later), and when correction is possible, feedback is sent to the verification processing in step 31. If repair is not possible, the correction process is over.

The sequential logic at the location stored in step 38 during the current modification is ignored. That is, step 6 of the above-mentioned FIG. 2 of the sequential logic stored in the process of step 6
Rewriting as explained in A is not performed.

Here, if the standard state transition diagram in FIG. 4 is modified to the state transition diagram in FIG. 7 according to rules 1 to 5, matching with codes after code 6 in the code string becomes rrl function, and step It is determined that the matching process has ended in step 33, and the correction process for the data is completed. Then, in step 38, the modified modified method sequential logic is stored in the working area of the memory 6.

[Correction Processing, Rules 0 to 4] Hereinafter, Rules 1 to 5 for performing sequential logic modification processing will be described.

Before explaining each rule, let us define the language used here.

The auxiliary state refers to a state added by modification to the standard state transition diagram, and is distinguished from the state in the standard state transition diagram (hereinafter referred to as standard state).

State A6 in FIG. 7 is the auxiliary state.

A jump transition is a transition to a state other than the adjacent state, and the number of jumps is 0 calculated based on the standard state, for example, the fourth
In the figure, the transition from state A1 to state A3 is skipped by 1.
This is called the transition.

Suppose there is an auxiliary state adjacent to state A4,
The transition from the auxiliary state to state A5 is called a transition with jump number 1, and the transition to the adjacent state has jump number O.

As common to rules 1 to 5 below.

[Rule 2] Transitions to rules are not allowed except for transitions to auxiliary states.

In other words, although the state transitions in the directions of the respective arrows shown in FIG. 5-2, the state does not change, for example, from 8 to 1 or from 1 to 8.

[Rule 3] By adding one auxiliary state from the state S (this is called a candy state), if a transition to the standard state can be made with the number of skips, for example, 2 or less, then the auxiliary state is added.

The transition conditions allowed at this time are transition from candy 5B to the auxiliary state, transition from the auxiliary state to the auxiliary state itself, and transition from the auxiliary state to the standard state. An example is shown in FIG. 8, and as is clear from the figure, the state ao in the figure is the auxiliary state.

[Rule 2] When Rule 1 is not satisfied, if a transition from the candy state to the standard state is possible with a transition less than or equal to skip fi2, this transition is added.

An example is shown in FIG. 9, in which a and b are attached, indicating the added transition.

[Rule 3] When the code is in the opposite direction in a state halfway from state ff, S to state g, E, by adding an auxiliary state, transition to the standard state with the number of jumps of 1 or less in the primary code. This transition is added when it is possible to transition from the standard state to the standard state with a jump count of 1 or less, for example. An example is shown in FIG. The one with a attached in the figure is the added transition.

[Rule 4] When Rule 3 is not satisfied, 0 examples of adding transitions from the auxiliary state and standard state to the auxiliary state, transitions from the auxiliary state to the auxiliary state itself, and transitions from the auxiliary state to the standard state are shown in the 11th example. State aQ in the 0 diagram shown in the figure is an auxiliary state. In each judgment process of rules 1 to 4 in steps 34 to 37 in FIG. 6, it is determined whether or not each of the previous rules is satisfied, and if it is satisfied, the sequential logic is modified according to the rule. , feeds back to the matching process. If the rules are not satisfied, judgment processing is performed in the order of rules 1 to 4.

Returning to the overall flowchart of FIG. 2, the steps after the sequential logic correction process in step 4 will be explained.

In step 7, it is determined whether or not the modification of the sequential logic using all the data in the l category has been completed, and if it has not been completed, it belongs to the same category.
Input various next pattern images belonging to the ``category'' and modify the sequential logic.

When all the data in one category has been corrected,
The standard sequential logic stored in the memory 6 by the standard sequential logic processing in step 5 and the sequential logic storage processing in step 6 described above is modified in step 4, but the data is stored in the memory 6 in step 6-A. , is registered in the memory 7 as a dictionary (I) in step IO.

Note that the operation of step 6-A may be performed in step 4. In the determination process of step 8, it is determined whether or not the processing for all categories has been completed. If not, the process returns to step 1. At this time, step 5 mentioned above
Jf:1 quasi-sequential logic processing and the storage processing of the sequential logic in step 6 may be used to store the standard sequential logic of the new category to be modified in the memory 5.

[Discrimination function creation process]

When it is determined that the correction of the sequential logic using all the data of all categories is completed, the answer is "Y" in step 8, the process proceeds to step 9, a flag is set, and step 3 indicates that a discriminant function will be created in the following process. The information is transmitted to the process judgment processing.

As a result, the subsequent processing through step 3 becomes the discriminant function creation processing from step 11 onwards from "NO". Then, the image input processing in step 1 and the preprocessing in step 2 can be used for both modifying the sequential logic and creating the discriminant function.

Next, the discriminant function will be explained.

By using sequential logic in the present invention, the accuracy of pattern identification has become extremely high compared to the conventional technology, but in order to further increase the accuracy, it is necessary to identify extremely similar patterns such as r7J, ``r'', and ``wa''.

Use a discriminant function.

This makes the features of the present invention even clearer.

1. By removing the limit on the number of times each state can be passed through, the number of required sequential logics is greatly reduced.

2) Instead of restricting individual states, we find a function that can express the stroke structure of the relationship between states, and use it as a function for identifying patterns, that is, a discriminant function. It becomes a wide recognition device.

Then, we define a discriminant function whose variable is the number of times it passes through each state in the sequential logic, and calculate the coefficients and threshold values.

By finding a method for extracting features and features, we were able to automatically create a discriminant function.

Next, the principle of automatically creating a discriminant function will be explained.

Consider a case where, in arbitrary sequential logic, each state in the sequential logic represents something that identifies a pattern belonging to a different category compared with the sequential logic (i.e., a discriminant function).

In the simplest case, the number of times the first state is passed x
When n is more than 10 times in the case of pattern Pa, and less than 10 in the case of pattern pb, it is possible to know whether the input image is pattern Pa or pb from the value of xn. Expressing this mathematically, X
n) 10, that is, when Xn-10>0, the input image is the putter 7 P a , Xyl<10, that is, X, -1o<0
(7) When the input image becomes pattern pb, in this case F (Xn) = Xn - IQ is the discriminative function, and pattern 7Pa
The reference value lO that separates pattern pb and pb is called a threshold value.

Generally, it is necessary to use more variables for identification, and when expressed as a linear identification function, equation (1) is obtained.

Here, xn is the number of times An and B m pass through the nth state. 71 is the coefficient G is a function consisting of terms of order higher than the third order of xn (this can be ignored when identifying a simple pattern). S is the two categories. The threshold value (xl) for identifying patterns belonging to is the set of xl, x2) - 11111.
The number of times of passing through the th state is Pan and Pb, respectively.
For any pattern belonging to each category let n, F ((Pan) , S) > 0
(2) F((Pbn), S)<0
(3) Coefficient A n ,
Once B m n and the threshold value S are found, equations (2) and (3)
Using the relationship of the formula, it is possible to identify whether the input image is cover pattern Pa or pattern pb.

[Coefficients An, Bmn] Next, a case will be described in which the coefficients An and Bmn are obtained.

From equations (1) 2 to (3), ΣAn(Pan-Pbn)+ is obtained (below, in order to identify with a simpler function, and to simplify the explanation, the G term will be ignored) Here Pan-Pbn and PamPan-PbmPb
n represents the difference between the pattern Pa and the pattern pb, that is, the characteristic for identification and the characteristic temperature.

In general, if you use something with a clear difference, that is, something with a big difference, it will be easier to notice the difference and make it easier to identify.In order to improve the recognition rate, Pan-Pbn and PamP
When an-PbmPbn is large, the coefficients Am and Bmn need to be large.

From here.

An=C(Pan-Pbn)K
(5) Bmn=C(PamPan−PbmPbn)
K (6) Here, K is an appropriate odd number and C is a positive constant.

Equation (5) is the first and second term on the left side of Equation (4).

Substituting (6), each (Pan-Pbn) K”
1.

(PamPan-PbmPbn)"1, and the purpose is to make the exponent an even number so that the expression is positive even if the value inside the brackets is negative. Since the pattern takes various values, (
Equations 5) and (6) need to be averaged values.

, where 　　 represents the average value.

Next, the threshold value S is determined.

In equations (1)2 to (3), F ((Pan), S) > OFF ((Pbn)
, S), the threshold value S that satisfies equation (9) is approximately equal to =
Bmn (PamPa n-1-P bmPb n)
) (9-1)n, (Pan) in equation (9-1), (
Since Pbn) takes various values depending on the pattern, it is necessary to average it, 5--(ΣAn an+ n+ n ΣΣBmn am an+ bm n )
(10) becomes n.

In equations (7), (8), and (io), C and K can be determined appropriately (for example, C = 1), so (Pan) and (Pbn) are found from pattern Pa and pattern Pb. By doing so, the coefficients An, Bmn and the threshold value S are determined.

In other words, the discriminant function is found.

Although the method for determining the discriminant function for patterns belonging to two categories has been described above, discriminant functions for patterns belonging to two or more categories can be similarly determined.

For example, the patterns belonging to the 3gs category are Pa, P
b and Pc.

First, the discriminant function F for pattern Pa and pattern pb is
! is found, Fl ((Pan) , St) > O(11) Fl (
(Pbn), St)<0 (1
2), the discriminant function F2 is similarly determined for pattern Pa and pattern Pc.

F2 ((Pan) , S2) >O(13)F2 (
(Pcn), S2)<0 (1
4), the discriminant function F3 is similarly determined for pattern Pb and pattern Pc.

F3 ((Pb n) , 33) >0
(15) F3 ((Pcn), S3
) <0 (16).

Using equations (11) and (16), F 1 ((Xn), St) > O and F2 ((Xn),
S2) If >0 (17), the human image is identified as pattern Pa.

Similarly, F z ((Xn), 5 t) <O and F3((Xnl
, S3) >0 (18), the input image is identified as pattern pb.

F2((X, n), S2) < o or F3((Xn),
S3) When <0 (19), the input image is identified as pattern Pc.

FIG. 12 abstractly represents the relationship between the following.

The category spaces of each pattern Pa, Pb, Pc are divided into discriminant functions Fl, F2. It is F3,
Each category space is defined by equations (17) 2 to (19).

For example, if the input image is within the category space of the pattern Pa, the input image is the pattern Pa and a 37+
1 will be given.

Even when the number of categories of patterns matched with the one-order logic is larger, the discriminant function can be found in the same way.

From the above, from equations (7) and (8) for calculating the coefficient An and coefficient Bmn, and equation (10) for calculating the threshold value S,
Equation (1) giving the discriminant function is found. It is possible to identify multiple categories that match any one order logic, but by using this discriminant function and the discriminant conditions of equations (2) and (3), we can identify categories that are effective for discrimination, that is, The features improve recognition rate and processing speed.

[Feature extraction method]

Next, a feature extraction method will be explained.

From equations (1)2 to (2), equation (7), equation (8), and equation (lO).

F ((Pan), 5) F ((Pbn), 5) is obtained.

In equations (20) and (21).

What gives a stable feature is one that is always positive for all patterns Pa and always negative for all patterns pb.

For example, in order for the number of passes through the n-th state to be a stable feature, the following must hold true for all patterns of Pan and Pbn.

This condition may also be true for the sum of multiple terms in equations (20) and (21).

For example, are all patterns of Pan, Pan, Pbm and Pb
If it holds true for n, the linear sum AmXm+A nXn of the number of times of passing through the m-th and n-th states becomes a stable feature.

Generally, in equations (20) and (21), 1
If the linear sum of one term or a plurality of terms has a high probability (90% or more in practice), is positive with respect to pattern Pa, and negative with respect to pattern pb, it can be used as a feature for identification.

In the method created using the sequential logic modification method of the present invention, if there are states that transition with the same code and the number of jumps between states is 1 or less, these states (hereinafter referred to as common states) are passed through. For example, in the state transition diagram of pattern "7" in Fig. 13, the number of times the linear sum is a stable feature is higher than that of a stable feature alone. The number is l, and the state is transitioned by code 3, but in distinguishing between pattern "7" and pattern "wa", the discrimination function can be -7,6(XI+χ1).

Here, xl and χ1 are distinguished by the number of times they pass through state A1 and state a1, respectively. This is because the shape flsAt and the shape 8al share information on the line segment in the stroke.

The next most likely feature is the (X1+Z1)2 term (X1+Z1 corresponds to the distance between state Al and state a1) and (X1+χt)(X2+χ2 ) term (corresponds to the curvature of the product of the number of passages of adjacent states...) Here, x2 is the state IAt for xl and the adjacent state A2
is the number of times it passes through state a2, and χ2 is the number of times it passes through state a2. These are likely to be features because the former is related to the length of the line segment, and the latter is the sum of the products of adjacent lines, which is related to the curvature of the line segment.

Next, the features that make a big difference between the patterns, that is, the features that make them easy to identify, are the coefficients (An) and (
Bmn) with a large absolute value or the average value of the sum of a plurality of coefficients has a large absolute value.

In practice, stable and easily distinguishable features are used as the discriminant function.

[Feature extraction procedure]

A procedure for extracting such features and determining a discriminant function will be explained.

1. Find the number of times each state in the sequential logic is passed through a comparison process with the sequential logic.

2) Find the sum of the number of times the common state is passed.

3. Find the square of the values found in 1 and 2.

4. In each state and common state in sequential logic,
Find the product of things that pass through adjacent states.

5. After completing the matching process for each category,
Find the average value of those obtained in step 4 from l.

6. In equations (7) and (8), K=l, C=1
Then, from what was found in step 5 above, the coefficient (An),
Find (Bmn).

Number the coefficients (An) and (Bmn) obtained in 7.6 in ascending order of their absolute values, for example up to 10.

wear.

In the order of the numbers assigned in 8.7, find the ones that satisfy equations (23) 2 to (25) with a probability of 90% or more, and find, for example, up to about 5 in order of probability.

This is the feature used in the discriminant function.

From the features obtained in 9.8, use the formula (lO) to find the threshold value. Here, the coefficient (An
) and (Bmn) are set to 0 and the threshold value is determined.

10. Perform steps 6 to 9 above for different categories checked against the same sequential logic to obtain a discriminant function.

The process for determining the above-mentioned discriminant function will be further explained based on the flowchart of FIG.

In the flowchart of FIG. 2, the processing after the dictionary search/verification process in step 11 in which the creation of the dictionary related to sequential logic is finished and the creation of the discriminant function is started will be described.

In the dictionary search/verification process in step 11, step 10
The sequential logic is searched from the dictionary related to the sequential logic registered in , and the matching process is performed.The number of times each state is passed through is determined at the stage of the matching process.

In step 12, the matching result is judged, and if there is no matching, search matching is performed against the first-order sequential logic.When matching is performed, the processing is performed to store the matching sequential logic and result in step 13.The details of this process are as follows. A flowchart is shown in FIG.

Step 50 is to perform step 2 of feature extraction to obtain a feature extraction and a discriminant function. For example, in the state transition diagram of pattern "7" in FIG. 13, state A1 and state a1. Condition A
2 and state a2 and state A4 and state a4, the sums passing through each state are determined.

Step 51 is step 3 of extracting features and finding a 1-distinct function.
This process performs the following processing.

Step 52 is step 4 of extracting features and determining a discriminant function.
For example, the state transition diagram of pattern "7" in FIG. 13 is the sum of medium mAt and state a1, state bo, and state A1.
The sum of mat and the sum of state A2 and a2, the sum of state A4 and a4 and mA3, the sum of state A4 and a4 and state A5, and the product of the number of times of passing through each state of state A5 and state a6. is found.

In step 53, the number of times each state is passed and the results of steps 50 to 52 are sorted by sequential logic and category and stored.

The above process is performed by the judgment process in step 14 of FIG.
This will continue until all data have been collected.

After data collection is completed, a discriminant function is created in step 15. A detailed flowchart of this process is shown in FIG.

The process of creating a discriminant function in step 15 is performed for each sequential logic between the categories matched with the sequential logic. For example, if the number of matched categories is one, the process of creating a discriminant function is not performed. In two cases, a discriminant function is created to distinguish between these categories. The matched category is "A".

In the three cases of rBJ and rCJ, "A" and rBJ
, rAJ and "C", and rBJ and "C", respectively. The same process is carried out in the case of 4 or more.

The flowchart for creating the discriminant function shown in FIG. 15 will be described below.

In the average value processing in step 60, the items stored in step 53 are averaged for each sequential logic for each category.Step 61 is a process for calculating coefficients, which are calculated in the average value processing in step 60. From each average value, formula (5) or (6) is calculated between different categories of each sequential logic.
) can be found from the formula.

In the coefficient ranking process of step 62, the coefficients found between different categories of each sequential logic are numbered in descending order of absolute value.

In the feature extraction process in step 63, the coefficients in step 62 are ranked between different categories of each sequential logic according to the numbers assigned in the process, so that each coefficient is between (22)2 and (25).
) that satisfies the conditional expression with a probability of 90% or more, extract at most five items with a high probability and a low rank in step 62 (that is, the absolute value of the coefficient is large), and use this as a feature. shall be.

In the threshold processing in step 64, threshold values are determined according to equation (10) from the features extracted in the feature extraction processing in step 63 between different categories of each sequential logic.

Through the above processing, the constants in the discriminant function can be found, that is, the discriminant function can be created.

The discriminant function created here exists between different categories compared with each of the 5 sequential logics, so step 1 in Figure 2
In registering the discriminant functions in the dictionary in step 6, the sequential logic stored in the memory 7 is made to correspond to the discriminant function that identifies the category matched with each sequential logic. It also stores the results identified by the discriminant function, that is, the recognition results.

Although the dictionary necessary for pattern recognition can be created through the above processing, each of the memories 4 to 8 in FIG. 1 and their mutual relationships will be explained in detail below.

The memory 4 stores programs for performing preprocessing, sequential logic creation processing, discriminant function creation processing, memory management, etc., and these programs are executed by the CPUI.

Memory 5 and memory 6 are work areas for preprocessing and dictionary creation processing, and the memory contents are constantly rewritten.

The memory 7 is divided in advance, and each of the divided sections stores each sequential logic. This simplifies the sequential logic modification process and search process.

[Memory structure of sequential logic]

FIG. 16 shows the storage structure of sequential logic.

In this example, in order to make it easier to modify the sequential logic, we introduce the code for transitioning to a state, the (relative) address at which transition is allowed from one state to the next, and the fixed length for the number of times each state is passed through. Memory capacity (for example, 1 byte) is allocated. In addition, the candy state, the final state, and the sample state in the sample transition diagram are stored at predetermined addresses, and a place is provided between the sample states to add one auxiliary state, making it easy to modify the sequential logic. .

After the final state and code E, a predetermined address for storing the discriminant function is stored, and the sequential logic and the discriminant function are associated with each other.

FIG. 16 is a memory map representing a standard state transition diagram of pattern "7'' of FIG. 4, and FIG. 17 represents a modified state transition diagram of FIG. 7.

The state transition diagram in Figure 7 has been modified by adding an auxiliary state A6 to state A5 in the standard state transition diagram in Figure 4. At this time, state A5 transitions to state A6 by code 6 and a new transition is made. The added state A6 has state A6 with code 6.
A transition to state A5 and a transition to state A5 by code 7 are added.

On the memory map, as you can see from the difference between Figures 16 and 17, there is a 6 (6) at the code (shape B).
A6) is the address of transition destination 6 (A6) at transition condition 6 of 7 (As), and the address of transition destination 6 (As) and 7 (A5) is at transition condition 6 and 7 of 6 (A6). is memorized.

The memory 8 is divided in advance, and each division stores a discriminant function corresponding to sequential logic, a conditional expression using a discriminant function such as equations (20) and (21), and a discrimination result.

[Recognition processing]

A case will be explained in which recognition processing is performed using a dictionary of sequential logic and discriminant functions created using the Bokuto method.

At this time, the configuration of the apparatus remains the same except that a recognition processing program is used instead of the dictionary creation program. FIG. 18 shows a flowchart of the recognition processing.

The operation will be explained below.

The image input process in step 1, the preprocessing in step 2, the sequential logic dictionary search board collation process in step 11, and the collation determination process in step 12 are the same as those performed in the previous dictionary creation.

Step 70 determines whether or not the search in the sequential logic dictionary has been completed, and if the input image does not match all sequential logics, that is, if it cannot be recognized, the process ends;
Resulting in denial of recognition.

In the identification process using the identification function in step 71, the information is obtained at the stage of the matching process. The value of the discrimination function is obtained from the number of times each state in the sequential logic is passed, and the discrimination result is obtained from the above-mentioned conditional expression, and the recognition process is completed.

〔effect〕

As described in detail above, according to the present invention, it is possible to provide a learning function in the development of OCR or in the creation of a dictionary during pattern recognition.

Furthermore, according to the present invention, sequential logic (
It becomes possible to provide a pattern recognition device that incorporates the idea of automata (automaton) and further performs classification processing using a classification function.

Additionally, it has become possible to automatically create a dictionary for online or offline character recognition.

Furthermore, it has become possible to create a chain code for dictionary creation based on a chain code for pattern recognition. By modifying the sequential logic that is a dictionary by starting with the chain code, it became possible to learn better sequential logic. Further, according to the present invention, it has become possible to create a discriminant function whose variable is the number of times each state of sequential logic, which is dictionary data created by code data indicating a direction, is passed through.

[Brief explanation of drawings]

Fig. 1 is an explanatory block diagram of the present invention, Fig. 2 is a flow chart showing the operation of the present invention, Fig. 3 is a diagram showing one embodiment of Che7 coding, and Fig. 4 is a transition diagram of standard sequential logic. , Figures 5-1 and 5-2 are diagrams showing the definition of directions, Figure 6 is a flowchart of the sequential logic modification process, Figure 7 is a diagram showing the state transition after modification, and Figure 8 is the auxiliary FIG. 9 is a diagram showing a state transition for explaining a rule 1 having a state. FIG. 9 is a diagram showing a state transition for explaining a rule 2 having an auxiliary state. FIG. 10 is a diagram showing state transitions for explaining rule 3 with an auxiliary state, FIG. 11 is a diagram showing state transitions for explaining rule 4 with an auxiliary state, FIG. is a diagram showing identification of category space. Figure 13 is a state transition diagram of pattern "7", Figure 14 is a flowchart related to data collection, Figure 15 is a flowchart for creating a discriminant function, Figures 16IN and 17 are explanatory diagrams of memory maps of sequential logic, Figure 18 is a diagram showing a flowchart of the recognition process, l is the CPU, 7 is the memory for storing sequential logic, 8 is the memory for storing the discriminant function, Figure 3 + Z Kunimata Te..., Fu'? To 6 75th layer identification leap number group life income A

Claims (2)

[Claims] (1) A preprocessing unit that encodes image information converted into a line figure based on the direction of the line figure, and a storage unit that stores sequential logic representing the state of transition in the direction of the line figure encoded by the preprocessing unit. , has a control unit that modifies the standard sequential logic in order to obtain the above recognition result for various image information that should be recognized as the same image information, and the modification operation of the control unit is performed according to the various image information By comparing the corresponding sequential logic and the above standard sequential logic, in the above standard sequential logic, if it is possible to transition from the standard state or auxiliary state to the standard state or auxiliary state with a predetermined number of jumps, that transition is added. A pattern recognition device characterized in that the sequential logic of the various image information is modified to match the standard sequential logic. (2) In claim 1, it is stated that the corrective action adds a transition from the auxiliary state to the standard state, from the auxiliary state or the standard state to the auxiliary state, or from the auxiliary state to the auxiliary state itself. Features a pattern recognition device.