JPS5851309B2 - Character selection processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character selection processing method, and more particularly to a character selection processing method by a post-processing unit in a character recognition device that recognizes a combination of a plurality of character type groups.

In conventional character recognition devices that read multiple character groups, a common method was to use format control to switch between separate recognition logics prepared for each character group, and to separate fields according to the character type.

That is, as shown in (:) Japanese Patent Application Laid-open No. 52-100936, format information such as the number of characters and character types of each field on a form is stored in advance in the device, and when reading, the field position to which each character belongs is stored. One that switches to recognition logic while confirming, and (1i) Japanese Patent Application Laid-open No. 53-1834
As shown in Publication No. 5, when writing on a form, a specific symbol determined for each character type group is added to the beginning of each field, and when reading, the specific symbol is first read and recognized according to that symbol. There are some that switch the logic and consider the field to be the same until the next symbol is read.

However, these are all predetermined types of characters that must be written in each entry field, and are used as is for address fields where kana characters and numbers are mixed, product name fields where kana characters, alphabets, numbers, etc. are mixed. could not be used.

For this reason, there was a drawback that it increased the burden on the scribe, such as having to code part of the address and product name and write it in numbers.

Apart from this, there is also a method (iii) of reading a plurality of character type groups into one group and treating the same as a single character type group.

This method reads mixed fields (for example, confectionery, numbers, and kana characters in the same field) without using information other than the shape of the individual characters, but it reads characters that are very similar to each other (for example, "7" and "ri"). , '5' and 'S', etc.), the psychological burden on scribes, such as having to train them to write emphasizing the differences, increases, and the scale of their cognitive logic becomes significantly larger. There were drawbacks.

In addition, the context of the read characters as seen in (IVI Patent Application 1986-86025) is checked, and if there is a grammatically unacceptable combination, it is assumed that one of the characters has been misread, and one or both characters are There is also a method of performing post-processing such as rejecting the ``Dakuten'', but this is only effective for very limited categories such as ``Dakuten'' and 9 Handakuten'', and is not general.

The present invention eliminates these shortcomings and provides a character selection processing method for performing lethal context processing, thereby eliminating the need to increase the psychological burden on the scribe. The aim is to realize a character recognition device that can read characters that include kana characters.

Another object of the present invention is to achieve a high recognition rate by using a plurality of small-scale recognition logics for dedicated fields in a character recognition device that reads a plurality of character type groups in a mixed manner.

The outline is to prioritize each subcategory included in the recognition logic, and based on this, select candidate categories obtained from the recognition unit, and to apply knowledge specific to grammar rules and reading forms. It is characterized by the fact that it utilizes this method to re-select promising candidate subcategories based on the character type and context of the category name.

FIG. 1 shows an embodiment of the present invention, in which 1 is a character recognition unit, 2 is a character recognition unit;
3 is a candidate category selection means based on priority, 3 is a candidate memory that stores the candidate category name and priority for each character, 4 is a character pointer, and 5 is a character type control means that selects a candidate category based on the contents of the candidate memory 3. What determines candidate categories;
12 is a result memory that stores three types of information for each character: candidate category name, priority, and character type code; 6 is a forward subcategory that determines a candidate category for a semi-selected character based on the character type code of the immediately preceding character; A processing unit 13 is a knowledge table storing permissible/prohibited information on the context of two adjacent characters; 1 is a forward knowledge processing unit which checks the context between the character and the immediately preceding character using the knowledge table 13; 8 is a backward subcategory processing unit that re-determines the candidate category of the immediately preceding character based on the character type code of the character; 9 is a backward knowledge processing unit that corrects the candidate category of the character concerned and the immediately preceding character; 10 is a field end detection unit that determines whether the character in question is the end of the field; 11 is a result output unit that corrects the latter candidate category; 11 is a result output unit; This indicates that the contents of the file are output as the reading results.

Next, the operation of the above embodiment will be explained in order.

In this embodiment, "0" is assigned to each subcategory in recognition logic according to the definition shown in FIG. 5, based on the degree to which it can be confused with other categories.

Three levels of priority, "1" and "2", are assigned in advance.

From now on, 5O2 will be used as a generic name for the subcategories for each priority level.
It will be called SIt82.

Further, it is assumed that the higher the subcategory is in the upper row of FIG. 5, the higher the priority.

First, the recognition unit 1 sequentially transfers category names and priorities corresponding to subcategories that match the recognition logic for each character on the form.

The category selection means 2 which receives these as input selects a candidate category with a high priority for each character and writes it into the candidate memory 3.

An example of the candidate memory 3 is shown in FIG.
PCi, 2): Priority, i=1. ....

5) is a memory capable of storing candidates for n characters in a context field (an entry area on a form that is semantically related).

FIG. 3 shows the processing mode of the category selection means 2, in which 31 is an example of a character on a form, 32 is an example of a candidate output from the recognition unit 1, and 33 is a candidate selected by the category selection means 2. This is written in the part of the candidate memory 30 corresponding to the character.

For example, as a result of processing character example 31 in the recognition unit, the character “wa”
If it is the character "ri", it will be recognized as having a priority of "1" corresponding to Figure 5, and if it is the character "ri", it will be recognized as having a priority of "2", and the character "7" will be recognized. Suppose that it is recognized as having a priority of "1".

In this case, the characters “wa” and “7” are added to the candidate memory 3.
written to.

Candidate category information for the entire context field is stored in candidate memory 3.
When stored in the character type control unit 5 based on the contents of the candidate memory 3 and the knowledge table 13 in order from the first character of the field.
, forward subcategory processing unit 6, forward knowledge processing unit 7,
Processing is performed by the backward subcategory processing section 8 and the backward knowledge processing section 9, and the respective results are written into the result memory.

The character pointer 4 specifies the character to be processed in the five processes described above, and its role is to specify the first character at first, and to specify the next character in order when the field end detection unit 10 does not detect an end. have.

The role of the character type control section 5 is to provisionally determine a candidate category for each character based on the contents of the candidate memory 3 (however, if the category cannot be determined, it is rejected), and to write it into the result memory 12.

When the character pointer 4 specifies the th character in the current field, the corresponding content of the candidate memory 3 PtC in FIG. 2 P t (lt J ), where i = 1, 2
．． ...5゜j''''1, 2), the candidate category information Dt of the character is determined according to FIG. 6, and is written into the result memory 12.

Here, one embodiment of the structure of the result memory 12 stores the following three pieces of information for each character as shown in FIG. R,, D
t(2) is the priority (0, ~, 3), Dt(3) is the character type code (in this example, O: don't care, 1: alphabet, 2:
symbol, 3: kana, and 4: symbol) are respectively assigned.

In addition, in Figure 6, "S" in D7(1) is Pt
means the candidate category name that exists within.

Also, in case of rejection, there is also a category! , l l[J, and four levels of priority are defined as shown in FIG.

Note that the glue jets in the upper row of FIG. 7 have the highest priority, and from now on, Ro is used as a generic name for the glue jets for each priority level.

They are called R,, R2, and R3.

For example, when the identification result of the th character of the money plan is that there is only one subcategory with priority "1", there will be one Sl in Pt, and according to Figure 6, the result will be as shown in the fifth column from the bottom. The memory Dt stores the category name S to which the subcategory belongs, and
The priority level "1" and the character type code of the category are respectively stored.

Next, the forward subcategory processing unit 6 processes the immediately preceding character (t-
The ambiguity of the Dt of the character is corrected based on the character type code (Dt-, (3)) of the first character.

The Dt(1,2) of the character is 82, R,. In either case R2 (when the contents of the resulting memory of the character concerned are considered ambiguous) D), the immediately preceding character is determined from the contents Pi of the candidate memory 3 according to Fig. 8 based on -+(3). D t (1) in the form of drawing candidate categories of the same character type.
, D t(2) , and D t<3> are updated.

Naori 7 (1, 2) is result memory D 7 (1) and D
The result Dt(2) is shown together with I(2) to give a candidate category name list with priority, and in FIG.
i in the term means the priority of the matched candidate category.

For example, if the character type of the immediately preceding character is a number, one subcategory of numbers and one subcategory of katakana are extracted for the character, and the character satisfies multiple candidate subcategories of numbers and katakana at the same time. For example, when R,), according to Figure 8, as shown in the 7th column from the bottom, D, /7+(3)\0 and Dl-+(3)
Since the number of categories for the same character type is one for numbers, numbers are selected as the candidate category for the character.

Next, the forward knowledge processing unit 7 uses the category name (Dt
-t(1)) and the knowledge table 13 to correct the ambiguity of Dt of the character.

The content of Dt(1,2) of the character is 8. ．． 82. In R2, the immediately preceding character Dl-, (1, 2) is S.

, Sl. At any time in S2, D,! + Based on (1) and according to the condition of the content Pt of the candidate memory 3 according to FIG. 9, D t(1) , D t(2) , D t
Update (3).

In Figure 9, D,! In the item (3), j means the character type code corresponding to the category of allowable characters in Pt.

Note that the knowledge table 13 instructs whether to prohibit or permit the context of various combinations of all categories to be read, depending on the purpose and use of the form.

For example, in the case of a form in which an amount is to be entered, the knowledge table 13 includes conditions such as prohibiting kana characters after ``¥'' (yen sign) and allowing numbers.

At this time, if the immediately preceding character is determined to be n Yn, and the character satisfies multiple candidate subcategories of numbers and katakana at the same time, or is jet R2, then according to Figure 9, the allowable number of characters in Pt is a number candidate. There is one category, and numbers are selected as candidate categories for the character.

Next, the backward subcategory processing unit 8 corrects the ambiguity of the immediately preceding character DF based on the character type code (Dt(3)) of the character.

The forward subcategory processing unit 6 attempts to select a candidate category that is considered to be the correct answer for the character based on the character type code of the immediately preceding character, but here, processing is performed in the opposite direction.

The immediately preceding character paste, /, -+ (1t 2 ) is R2, and the character D t (1,2) is S.

When either j s or ts2, the content of candidate memory 3 P , /= , according to FIG. 10 based on D7(3)
Based on the conditions of Dz-1(1), Dz-t(2
), Dz −+ (3) is updated.

Next, the backward knowledge processing unit 9 processes the category name (I
M(1)) and the knowledge table 13, the ambiguity of the immediately preceding character Dt is corrected.

Although the forward knowledge processing unit 1 attempted to select a candidate category based on the category concatenation relationship based on the content of the category of the immediately preceding character, here, the process is performed in the opposite direction.

The immediately preceding character, Dt, (1, 2) is R2, and the character paste, 4 (1, 2), is So, Sl.

At any time in S2, the result is determined again from the content PL of the candidate memory 3 according to FIG. 11 based on D41), and Dt-1('') t D, /, -1(2) t
Update Dl-□(3).

Next, the field end detection unit 10 determines whether or not the character is the last character of the context field, and regarding the next character,
The above-mentioned processes are repeated, or the contents of the result memory 12 are outputted as a read result through the result output section 11.

FIG. 12 is an explanatory diagram illustrating a character selection process using a specific example in which the writing is based on the writing in margins in this embodiment.

51 is the entered character on the form, 52 is the result obtained from the recognition unit,
53 is the content of the candidate memory, 54 is the result memory after character type control, 55 is the control condition regarding forward and backward subcategories and execution of knowledge processing, and 56 is the result memory after all processing is completed.

The following is a step-by-step explanation based on these. Entry character 51
is ``Kanagawaken'' (seven characters of Kanagawa Itoretsu kana).

It is assumed that the recognition result 521 for each character output from the recognition unit is the illustrated candidate category and its priority.

Character selection processing is started based on these results.

That is, first, only the higher-rank categories are written in the candidate memory 53 by the category selection means 2.

Here, 521 “ya”, 523 r7J, 525 “
)” were removed.

Next, control the character type for the first character.

Based on the contents of 531 and the conditions shown in FIG. 6, the candidate category name, priority, and character type code (3: kana) are written in 541.

Next, the control conditions for executing each process in the forward direction and backward direction regarding the character are as shown in 551, and the ◯ marks are the execution points.

Here, 55 is a condition related to forward subcategory processing, forward knowledge processing, backward subcategory processing, and backward knowledge processing from the top.

Since 541 is the first character, forward knowledge processing is performed to check only whether the category of the character can appear first, and the result is 561.

Next, processing regarding the character 512 is performed. When character type control is performed based on 532, 542 is obtained from the conditions shown in FIG.

Since the priority of 542 is "2", each process in the forward direction as shown in 552 is executed, but since the immediately preceding character is not rejected (
From result 561), each process in the reverse direction is not executed.

In this specific example, since the results of the characters subsequent to 542 can be determined by forward processing in the same way, reverse processing is not performed.

Based on the conditions shown in Figures 7 and 8, 56
Get 2.

Since the process of obtaining 563 from 533 is obvious, the explanation will be omitted.

Regarding the character 514, if character type control is performed based on 534, 544 is obtained from FIG.

The priority of 544 is the highest (D(2)=O), so it is 5.
44, 564 is obtained without executing each process.

Note that if the priority is "0", that category will be adopted with the highest priority even if the separation principle is not followed.

Next, regarding the process related to 515, in 535 “
The different character types of ``wa'' and ``7'' appear simultaneously, and as a result of character type control, 5450 ways are rejected, priority is 1, and character type code O (don't care).

Based on 545, only two forward processes are performed as in 555, and as a result, a category name similar to the character type code of the immediately preceding character is selected again, and 'Wa' is written in 565 as a result.

The following processing based on 536°537 is also executed in the same procedure to obtain 566.567.

If 51 is one context field, after these processes, 56 category names are output as the reading results.

As explained above, according to the present invention, when the result obtained by the recognition unit is rejected due to the appearance of multiple candidates or an ambiguous low-priority subcategory appears, the result immediately before or after the character in question, that is, the category name, By inquiring about character type codes, etc., it is possible to select and extract a single candidate and eliminate misreading.

Furthermore, the above character selection processing method allows highly accurate mixed reading to be achieved without changing the content or expanding the scale of the recognition logic for ground character types.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a block diagram of a candidate memory, FIG. 3 is an explanatory diagram explaining the processing of the candidate category selection means, and FIG. 4 is a block diagram of a result memory. FIG. 5 is an explanatory diagram for explaining the definition regarding subcategory priority, FIG. 6 is an explanatory diagram for explaining the content of character type control processing, FIG. 7 is an explanatory diagram for explaining the definition for project priority, and FIG.
The figure is an explanatory diagram explaining the contents of forward subcategory processing,
Fig. 9 is an explanatory diagram explaining the contents of forward knowledge processing;
FIG. 0 is an explanatory diagram explaining the contents of backward subcategory processing, FIG. 11 is an explanatory diagram explaining the contents of backward knowledge processing,
FIG. 12 shows an explanatory diagram for explaining the processing according to the present invention using a specific example. 1... Recognition unit, 2... Candidate category selection means, 3.
... Candidate memory, 4... Character pointer, 5... Character type control unit, 6... Forward subcategory processing unit, 7...
Forward knowledge processing unit, 8... Reverse subcategory processing unit, 9... Backward knowledge processing unit, 10... Field end detection unit, 11... Result output unit, 12... Candidate memory, 13...Knowledge table, 31...Example of written characters on form, 32...Judgment result of recognition unit, 33...Candidate memory, 51...Example of written characters on form, 52...
Judgment result of the recognition unit, Kaku: Candidate memory, 54: Result memory after character type control processing, 55: Control conditions regarding process execution, 56: Result memory after character selection processing.

Claims (1)

[Claims] 1. In a character recognition device that reads target characters that are composed of multiple character type groups,
The recognition unit of the character recognition device is configured to output a plurality of candidate categories including a plurality of candidate subcategories, and the post-processing unit compares the priorities of the candidate subcategories and selects a candidate category with a high priority. A character type control process to select, a subcategory process to select a candidate category that is considered to be the correct answer for the character based on the connection relationship between the candidate category of the character immediately before or after the character and the candidate category of the character, and or knowledge processing that selects a candidate category that is considered to be the correct answer for the character based on the category itself's connection relationship between the candidate category of the character immediately after, the candidate category of the character, and the candidate category of the character; The character reading device is configured such that the priority given to each subcategory to select candidate subcategories obtained from the recognition unit can be set in advance. A character selection processing method characterized by being classified into a subcategory group that agrees with and can be clearly distinguished from all other categories, a subcategory group that can be clearly distinguished from all categories of an allopathic character group, and others.