JP3466903B2 - How to create a multi-template dictionary for character recognition - Google Patents

Description

Detailed Description of the Invention

[0001]

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for creating a multi-template dictionary for a character recognition engine. When dealing with multi-font typed characters and handwritten characters in character recognition, the recognition rate deteriorates. Therefore, as one method to deal with this, there is a method of providing multiple dictionaries for each character type (multi-template method).

It is desired to create such a multi-template dictionary without taking time.

[0003]

2. Description of the Related Art Regarding the technology of character recognition, a proposal for creating a multi-template for a dictionary was published in a paper "Construction of Multi-Template Dictionary for Character Recognition Considering Distribution between Categories" (IEICE Transactions-II Vol.J79- D-II No.9
pp.1525-1533 September 1996). The contents will be outlined below with reference to FIGS. 6 to 8.

In order to investigate the distribution between categories (character types), principal component analysis is performed on the set of centroids of feature amounts between each character type, and the centroid vector and individual sample patterns are plotted to confirm the distribution shape. In this case, the horizontal axis is the first principal component and the vertical axis is the second principal component, and the distribution of the character feature amount when the directional line element feature amount is used is wide.

Basic techniques for creating a multi-template dictionary and
Then, the character type adaptive class considering the distribution pattern of the character pattern
The Stirling method is used. FIG. 6 is an explanatory diagram of the conditions for division.
Yes, in FIG. Is the character type c1 (for example, hiragana
"A") and c2 (for example, hiragana "o")
The following shows a typical example in which misrecognition occurs. Two ellipses
j and k are clusters (ten) of the characters c1 and c2, respectively.
Plate) and the center of gravity of cluster j is m^j,cluster
the center of gravity of k is m ^kIs. Of the sample pattern of the character c2
Focusing on one point x, this point x is originally the character c.
2, but the distance d from the center of gravity of the character c1_jIs the character c2
Distance to the center of gravity of_kShorter pattern of points x
It is mistakenly recognized as c1. In this way, there is a possibility of misrecognition
Increase the number of clusters by dividing the category containing clusters
You

The condition for dividing a category is a cluster k.
The center of gravity m ^k , the eigenvalue of the i-th principal component is λ _i ^k , and the eigenbell (unit vector) is α _i ^k . Since the eigenvalue represents the variance on the principal component axis, use the appropriate constant α
Two points (1) and (2) can represent both ends of the feature region on the i-th principal component axis.

[0007]

[Equation 1]

Here, α = 3.3. This is a value that "99.9% of the patterns belonging to the cluster are included when the distribution on the axis is a normal distribution. Considering up to the Lth principal component axis, some i (1≤i When the following equations (3) and (4) hold for ≤ L), the i-th principal component axis of the cluster k is defined as "the axis that may be misrecognized", and the category including the cluster k is defined as Consider it necessary to split.
E here is the Euclidean distance.

[0009]

[Equation 2]

Where c is the character type of cluster k, and j
(c) is the set of clusters of all character types except c, and N (c)
Is the number of clusters forming the character type c. M is a constant and is set to prevent one category from being divided more than necessary. A. of FIG. Distances d _j and d _k
Is the following equations (5) and (6).

[0011]

[Equation 3]

Referring to FIG. In this case, the point x is erroneously recognized as the character c1, but in this case d _j <d _k and the above equation (3) is satisfied, so the cluster k is divided. B. of FIG. When k is divided into _two clusters (templates) k ₁ and k ₂ as in the above, the distance from the same point x to the center of gravity of the cluster k ₁ becomes shorter, and the character c2 is correctly recognized. Conceivable. Note that k1 and k2 are the same character c2
(Category) is recognized, and for example, it corresponds to Mincho type "O" and Gothic type "O".

The algorithm of the character type adaptive clustering method is defined as follows. First, the feature quantity (vector) is obtained from various sample patterns, and the set of feature quantities for each character type is regarded as one cluster. Examine whether there is an axis that satisfies Eqs. (3) and (4) for each cluster. Then, of (k, j) satisfying the equation (3), the one having the largest eigenvalue λ _i ^k of the i-th principal component is searched (the criterion for selecting the cluster to be divided is the number of samples included in the cluster). , The number of clusters of the character type including that cluster is increased by 1 (the target cluster is not divided into two, but the entire category is divided again).

Here, the value of λ _i ^k represents the ease of misreading, and this is defined as the misreading degree. This operation is expressed by Equation (3) and Equation (4).
Repeat until there are no more satisfying clusters. The centroid of each finally obtained cluster is calculated and used as a dictionary template. The K-means method was used as the dividing method. k
Is a cluster of character type c, the initial cluster center for subdivision is m ^k ± εα _i ^k and the center of gravity (if any) of a cluster of c other than k (ε is a minute length 0.001). Is).

FIG. 7 shows how to take the initial cluster center,
A. of FIG. If it is determined that the cluster k ₁ of k ₁ is divided when the first principal component is considered, the center of the initial cluster is ± ± from the center of gravity of the cluster k ₁ toward the first principal component of k ₁ .
Two points moved by ε are set, and all vectors are clustered using the K-means method. The result is B. become that way. Hereinafter, B. The first principal component of the cluster k ₂ of C.
Assuming that the second principal component of the cluster k ₂ is selected by the above judgment method, finally D. It becomes a cluster configuration like.

FIG. 8 is a processing flow for estimating a conventional division target category. This processing flow is as shown in FIGS.
It is executed in the division as shown in. First, i = 0 (S1 in FIG. 8), i is set as a misreading category (character type to be recognized) (S2), and then j = 0 (S1).
3) and j are misread categories (category j when the character type i is misread as j) (S4). Where i is j
If it is misread, the misreading degree of i is updated (step S6). If not misread and after the processing of step S6,
j is incremented by 1 (S7 in FIG. 8), it is determined whether or not j has reached the number of categories (the number of character types to be recognized) (at step S8), and if not, the process returns to S3 and the same process is performed for the next j. To do. If it has reached i, i is incremented by 1 (S9 in FIG. 7) and the process moves to the next category. Then, it is determined whether i reaches the number of categories to be recognized (S10 in FIG. 8). Returning to S2, the same processing as above is executed. When it reaches, the category having the highest degree of misreading for each misreading category is divided (at step S11). In this way, the category to be divided is selected and divided.

According to FIG. 8 above, the loop is doubled (for i and j) for the misread category check,
Assuming that the number of categories is N (for example, N = 3600 in some cases), the number of loops becomes the order of N squared and the processing time becomes long.

FIG. 9 is an explanatory diagram of the number of times of the conventional misread category check. As shown in FIG. 8, when the horizontal axis is the misread category and the vertical axis is the misread category, i ×
Recalculated all j times.

[0019]

In the conventional method described above, when a dictionary tool is created by simply dividing the template according to the basic method when creating a dictionary, the input data increases when a large amount of data is input. The execution time of the misrecognition check exceeds the practical limit, making it virtually impossible to execute.

The present invention provides a method for creating a multi-template dictionary for character recognition, which is capable of creating a highly accurate multi-template dictionary in a short processing time even if a large amount of data is input, which was impossible with the prior art. The purpose is to

[0021]

FIG. 1 shows the principle configuration of the present invention. In the figure, 1 is a previous division result holding unit, 1a is a previous division category holding means, 1b is a category-corresponding misread category holding means, and 2 is a first misreading for determining a misreading category in which a division target category is a misreading category. Checking unit, 2a is a specific misreading checking unit that is activated when the misreading category is the same as the misreading category held in 1b, and checks misreading, and 3 changes the misreading category and misreads it as a division target category The second erroneous reading check unit for making the determination, 4 is a detecting unit for detecting the category of the maximum erroneous reading degree, and 5 is a holding unit for holding information such as the erroneous reading degree updated by the respective units 2, 2a and 3 above. .

First, a category (template dictionary) division process is performed by the conventional method, and the previous division result holding unit 1
The previous division category holding means 1a holds the division target category (referred to as mc), and further holds the misread category having a high misreading degree for each category in the misread category holding means 1b. After that, the first misreading check unit 2 is activated to change each misreading category (referred to as i), determine whether the misreading category is misread with the division target category, and if misreading, update the misreading degree. Then, it is recorded in the information holding unit 5. And misreading category (i)
When the misread category and the division target category are the same category, the specific misreading check unit 2a is activated, the misread category is changed, and it is checked whether or not the misread category (i) is misread as the misread category. When misreading, the misreading degree is updated and recorded in the information holding unit 5.

When the operation of the first misreading checking unit 2 or the specific misreading checking unit 2a is completed, the second misreading checking unit 3 is activated, and the category (mc) held by the previously divided category holding unit 1a is set as the misreading category. The misread category is changed to determine whether the misread category is misread as mc. If misread, the misreading degree is updated and recorded in the information holding unit 5. Finally, the detection unit 4 detects the maximum misreading degree category from each updated misreading degree recorded in the information holding unit 5. The division is performed for the detected category.

As described above, conventionally, the template (dictionary) is divided by repeating the processing in the unit of the square of the number of categories N, but since the calculation only needs to be performed in the unit of N times,
The amount of calculation can be reduced and a highly accurate multi-template dictionary can be created.

[0025]

2 is a basic processing flow of the present invention. In FIG. 2, first, the distribution state of the character feature amount is investigated (S1 in FIG. 2). In this case, the feature amount (vector or the like) corresponding to each character (category) is extracted. Next, it is determined whether the two character types are erroneously recognized (S2). The process ends when it is found that no misrecognition occurs, but if misrecognition occurs, a new template (dictionary for character recognition) is created for the misrecognized character type (S3 in FIG. 2), and the new template is taken into consideration. The inter-species distribution is recalculated (S4). After this recalculation, the process returns to step S2, and it is determined again whether the two character types are erroneously recognized. The loop process of S2 to S4 and S2 is repeated until there is no misrecognition between the two character types.

FIGS. 3 and 4 are process flows (1) and (2) of the embodiment of the present invention, and FIG. 5 is a diagram showing an example of checking misreading categories according to the present invention. Once started, the first division is performed by the conventional method (S1 in FIG. 3). This conventional method is the method shown in FIG. 7 above, and when the category to be read is 3500, the processing of the number of times of squaring is repeated by the processing of the double loop. Next, the category that is the division target in the first time is stored as a variable mc (see FIG. 3).
S2). Next, for each category (for example, for the character type i to be read), i is the other character type j
If misread and stores each j as the misreading category. Each misreading degree calculated at this time is stored in a memory (not shown) (S3 in FIG. 3). The above S1 to S3 are
Corresponding to the pre-processing for determining the division target category according to the present invention, necessary data as a result of the first processing is stored, and in the second and subsequent processings thereafter, data that does not change from the first processing is processed. Instead, the processing is shortened by processing only the changed parts.

That is, it is roughly divided into a process of an A block and a process of a B block which follows it, and the A block is composed of an A1 block or an A2 block.

At A1 of the A block, the misread category is changed from i to N. That is, first, i = 0 is set (S4 in FIG. 3), and i is set as a misreading category (target character type) (S5). Next, it is determined whether or not this i becomes the misread category = mc (i is misread as mc) (S6 in FIG. 3). Here, it is determined whether the misread category of i and mc stored in the previous process match the same category, and if they match, the process moves to A2, but if they do not match, mc is changed. The misread category is set (S7 in FIG. 3), and it is determined whether or not the category i is misread as the category mc (S8). When misreading, when the misreading degree for mc is higher, the misreading degree of i is updated (see FIG. 3).
S9), and the misread category of i is updated (S1 of the same).
0). After that, i is incremented by 1 (S11 in FIG. 3), and it is determined whether i reaches the number of categories (S12). If i is not reached, the process returns to S5, and if i is reached, group B processing is performed.

At step S6 in the above A1,
If the misread category of i matches mc, the processing of the A2 group is executed. In A2, the misread category is set to j instead of mc, the misread category j is changed to N, and it is determined whether or not i is misread as j. If i is misread as some category and the misreading level for that category is higher, the misreading level of i is updated.

That is, regarding the values such as the misreading degree for the category changed by the previous division, if these are the misread categories, only the changed misreading degree is changed, and other values are used. Is not subject to processing because it is not changed. That is, processing is performed only for each category in which the misread category i that is the target of division is the misread category.

First, j = 0 is set (S13 in FIG. 3), j
Is the misread category (S14), and it is determined whether i is misrecognized as j (S15). Here, in the case of erroneously recognizing i as j, if the misreading degree for j is higher, the misreading degree of j is updated (at step S16) and the misread category of j is updated (at step S17). Then, j is incremented by 1 (the same S1
8), j is determined whether it exceeds the number of categories (S1 in FIG. 3).
9) If not exceeded, the process returns to step S13 to repeat the loop operation. When j reaches the number of categories, i is incremented by 1 (at step S20) and the process returns to step S12 of group A.

When it is determined in S12 that j has reached the number of categories, the group B shown in FIG. 4 is entered.
In group B, the category m that was the target of the previous division
With c as the misread category, the misread category j misread by this category is changed to N, and it is determined whether or not mc is misread as j.

That is, mc is set as the misread category (S21 in FIG. 4), j = 0 is set (S22 in the same), and j is set as the misread category (S23). Next, it is determined whether mc is misread as j (S24 in FIG. 4). When misreading is performed, if the misreading degree for j is higher, the misreading degree of mc is updated (S25 in the same), and the misread category of mc is set. It is updated (at step S26). Subsequent to S24 and S26, j is incremented by 1 (see FIG. 4).
S27), and it is determined whether or not j exceeds the number of categories (S28), and if it does not exceed j, the process returns to S23 to repeat the loop operation. When j reaches the number of categories, the category having the largest misreading degree for each misreading category is divided (S29 in FIG. 4).

Since the processing of A2 is executed only rarely, the actual processing time is the processing time of the loop of block A1 + the processing time of the loop of block B. Since it is only once, the execution time is reduced to N orders.

The processing of FIGS. 3 and 4 will be described with reference to the example shown in FIG. A misreading category i is arranged in each line of FIG. 5, and each line of (1) to (9) is a character type, which corresponds to a character type such as “A”, “I”, and “U”. Misread categories are placed in each column, and (1) to (9)
Each column of has the same character type as (1) to (9) of each row, and expresses the mutual relationship except that the characters of the same number intersect. The line (3) in FIG. 5 (i = 3) is the category to be divided (S2 in FIG. 3). This i is set as the misread category = mc (= 3), and this category is set as the misread category as shown in the third column of FIG.
The processing of group A1 is executed. In the example of FIG. 5, at this time, the misreading category i is misread as the character mc of (3).
Are two categories, line (5) and line (8). That is, the processing of group A2 in FIG. 3 is executed for the shaded areas of the fifth row and the eighth row in the horizontal direction, and the misreading degree is updated. Expressed by the black circle at the intersection with the third column,
Indicates that there is a misread category.

Further, the processing of the group B for obtaining the degree of misreading when mc is misread as a misreading category and this mc is misread as another category is performed in the column direction with respect to the shaded area in the third row in FIG. The crossing point becomes the misread category, and the misreading degree is updated. In this example, the category in 3 rows is misread as the category in column (8) (represented by a black circle).

[0037]

According to the present invention, the misreading degree used for estimating a character type that may be misrecognized is stored, and only the changed portion is recalculated to check the misrecognition. The calculation amount of can be significantly reduced. This makes it possible to speed up the process of dividing the categories and create a highly accurate template dictionary at high speed.

[Brief description of drawings]

FIG. 1 is a diagram showing a principle configuration of the present invention.

FIG. 2 is a basic processing flow of the present invention.

FIG. 3 is a diagram showing a processing flow (No. 1) of the embodiment.

FIG. 4 is a diagram showing a processing flow (No. 2) of the embodiment.

FIG. 5 is a diagram showing an example of checking a misread category according to the present invention.

FIG. 6 is an explanatory diagram of a dividing condition.

FIG. 7 is a diagram showing how to take the center of an initial cluster.

FIG. 8 is a diagram showing a processing flow for estimating a conventional division target category.

FIG. 9 is an explanatory diagram of the number of times of conventional misread category check.

[Explanation of symbols]

1 Previous division result holding unit 1a Previously divided category holding means 1b Misreading category holding means 2 First misreading check section 2a Specific misreading check section 3 Second misreading check section 4 detector 5 holding part

─────────────────────────────────────────────────── ─── Continuation of the front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06K 9/00-9/76

Claims (2)

(57) [Claims] 1. In a character template multi-template dictionary creation method, a category (mc) that was previously divided, a holding unit that holds misreading degree and misread character as misread information at the time of previous division, and a misread category are set. Fixed to the previous division category (mc), the misread category is checked for misreading only for all categories, and the first check unit for updating and holding the misread degree and the misread character according to the check result, Fixed to, the misreading category is checked for misreading only for all categories, and the second check part holds the misreading degree and the misread character according to the result of the check, and the misreading degree for each of the held misreading categories And a detection unit for detecting a category having the largest one, and dividing the category having the largest misreading degree. Multi-template dictionary creation method for character recognition to. 2. The specific misreading check unit according to claim 1, further comprising a specific misreading check unit that is activated when the misread category of the misread category and the division target category are detected in the first check unit. The misreading checking unit changes the misreading category, checks whether the misreading category is misread as the misreading category, and if misreading, updates the misreading degree and records it. How to create a multi-template dictionary.