JPS60153579A - Forming method of dictionary for character recognition - Google Patents

Abstract

PURPOSE:To form a highly accurate dictionary automatically by encoding the features of a character, and storing only feature codes excluding inhibited features out of feature codes including essential features in a dictionary to use the stored feature codes as study data. CONSTITUTION:When a CPU11 executes a program stored in a RAM12, a counter CO is initially set up. The contents of the counter CO are compared with the essential feature codes of the character, and when the essential feature is specified as ''0'' or ''1'' and the bit of the corresponding counter CO coincides with the essential feature code, an inhibited feature table 15 is referred. When the combination of the counter CO and the corresponding bit in the inhibited feature table coincides with each other, the succeeding character is selected as the inhibited feature. In case of inconsistency, the contents of the counter CO are stored in a RAM13 as study data. Thus, the contents of the counter CO which are stored in the RAM13 become automatically generated (formed) effective study data.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a method for creating a dictionary for character recognition used when a computer reads characters and figures handwritten on a form.

<Prior art and its problems> FIG. 1 schematically shows an example of a conventional character recognition device as a block diagram. Characters and figures drawn by hand on the form 1 are optically read by an image sensor 2 such as a COD, and the optical signal is converted into an electrical signal and input to the F converter 3, where it is converted into an A/
The signal is D-converted and input to the preprocessing circuit 4. This preprocessing circuit 4 binarizes the optically read characters and figures, that is, divides them into white and black backgrounds, and then performs so-called preprocessing such as noise removal and line thinning. A feature extraction circuit 5 following the preprocessing circuit 4 extracts all character features of the preprocessed character graphics, such as character endpoints, character branching points, character intersections, bending points, depressions, and loops. The character features extracted here are expressed in the same format as a memory means called a dictionary 6, which stores character features of standard patterns encoded in advance. Next, the dictionary matching circuit 1 compares the character feature code extracted by the feature extraction circuit 5 with the standard code in the dictionary 6, and when the two match, outputs the ID of the standard pattern having that code, and If you do not, it will be treated as a so-called reject.

In such a handwritten character recognition device, the characters entered on the form differ depending on the scribe, even if they are of the same type.
It is necessary to prepare multiple dictionaries called subcategories for the same character type. Figure 2 VC, character 1
An example of the dictionary for Ako is shown, each time ω)0
The subcategories indicated by V are 1 each with code 10111.
It is expressed as 00100.11101. Here, the bit number 0 becomes logic 1 when the number of connected components of the character/figure is 1, and becomes logic 0 when it is not 1. Similarly, the bit number l becomes logical 1 when there is a loop in the character figure, and becomes logical 0 when there is no loop. Bit number 2 is a left concave υ,
It means whether or not there is a dent in the character figure when viewed from the left side. If there is a left dent, the logic is 1, and if there is no dent, the logic is 0. Bit number 3 becomes logic 1 when the number of character endpoints is three, and becomes logic 0 when there are not three. Bit number 4tX becomes logic 1 when there is a branch in the character figure, and becomes logic 0 when there is no branch. A dictionary is a dictionary in which the necessary number of codes for each character is stored in advance in a ROM or the like. Of course, C's dictionary is the second
Although not shown in the figure, it is composed of a series of information including a so-called ID code and other auxiliary information indicating that the character having such a feature code is the character "A". Then, in the feature extraction circuit 5 described above, the percentage of the character/figure read by the image sensor 2 is determined by the second
Conventionally, such a dictionary, which is extracted in the format shown in the figure, coded, and input to the dictionary matching circuit, is created by a person for each character and figure to be recognized. Was.

The amount of time it takes to create this dictionary is enormous, and when developing a character recognition device, it took about 50 minutes to create this dictionary.
It took more than % of the time and effort to do so. Therefore, in order to solve this problem, we asked an unspecified number of reporters to write in all the characters that are subject to gR, read these characters with a character recognition device, and read all the output from the feature extraction circuit described in Ar1. Dictionary V as standard code? - Dictionaries began to be created using the so-called Gaku11 method.

However, in such conventional methods, the accuracy of the completed dictionary depends on the accuracy of character feature data from collected written characters, and so-called omissions and omissions may occur. Even if the contents of the input characters are the same, they are processed in the same way as data to be learned, so the problem remains that a lot of time is spent on processing.

<Purpose of the invention> Therefore, the purpose of the invention is to create all highly accurate learning data in one period, and to fully automatically create highly accurate and small-capacity words based on this learning data. VC aims to provide an improved dictionary creation method for character recognition.

Structure of the Invention> In the method for creating a dictionary for character recognition according to the present invention, a character feature code equivalent to the input character is automatically generated and created by a computer or the like. At this time, for each valley character, the character is designated as an all-required feature that the character always takes within any allowable deformation range, and the character feature code is automatically generated by a computer or the like within the designated deformation range. As a result, the Y of the dictionary, which was a problem with the conventional method mentioned above.
? The i[ problem is now resolved. In other words, in the conventional method, the quality of the dictionary depended on the number and contents of each name, but with the method according to the present invention,
All character feature codes are generated and created within the allowable deformation range, so there are no problems with omissions or omissions in the dictionary.

In this invention, in addition to specifying such an essential feature code, a feature that cannot be obtained in character structure within any allowable deformation range on a character feature code having an essential feature is specified as a prohibited feature code. For example, the characters in Figure 2 “
Regarding "A", a character whose loop is 1 and the number of connected components is not 1 cannot be expressed by a character graphic image. Such a feature code that cannot be expressed by a character/graphic image may occur if the above-mentioned electronic computer or the like is configured only to automatically generate feature codes and specify essential features. Although this is not a problem in terms of character recognition, there is a problem in that the size of the gado becomes extremely large. Therefore, in addition to specifying the required feature code, it is prohibited! By specifying the characteristic code, you can quickly store the storage space and the memory valley S-1.
You will be able to create learning data for creating . This learning data is used in the same way as the input (learning) data of a conventional learning automatic dictionary creation device, and becomes data for creating an actual dictionary for the character recognition system.

In the detailed description of the present invention below, an example of creating learning data in the case of creating a dictionary composed of five feature codes will be described.

As shown in Figure 2, in each subcategory of the letter [A], the left dent of bit number 2 is always a logical 1.
Deroru. Therefore, the required feature code for the character “A” is
It is a combination with other 4 bits in which bit number 2 is 1, and it is sufficient to generate this code. In a similar way, the required feature code for the character
As shown in the figure, bit number 1 and bit number 2 are always 0 in combination with the other 3 bits, and this code can be generated. As a result, the letter “i”
For, the required l from a to h as shown in Figure 4.
RF fine code is generated. These codes are different from the essential feature codes for the letter "A" shown in FIG. Therefore, it is understood that the characters "A" and "°I" are recognized separately in the dictionary matching circuit T shown in FIG. 1.

In the essential feature codes for the letter "A" shown in FIG. 4, the codes from code b to g do not constitute the letter "I" within any allowable deformation range.

In these codes, bit numbers 3 and 4
The combinations have features of 0.1 and 1°0, respectively. Therefore, codes with these characteristics are learned and removed from the data as codes that do not constitute a character figure [in other words, that is, a code that cannot be expressed as a character figure. That is, this is the prohibited feature code in this invention.

The above-mentioned storage of the essential characteristic code, υ, prohibited sexual characteristic code, etc. in the RAM, etc. is carried out based on the tables shown in FIGS. 5 and 6, respectively. In the m5 diagram, the x mark is 1'! ! : or O, so-called any f, where l indicates that the required feature is logic 1, and 0 indicates that the essential feature is logic 0. In the case of l, it means that the corresponding character % enemy is always 1'''C7.
0 means always 0. Regarding the table shown in FIG. 6, which specifies the closure characteristic code, each code has the same meaning as in FIG. 5. 〇 FIG. 11 is a CPU for executing a program stored in a RAM 12;
Reference numeral 13 is a RAM''c receptacle for storing the contents of the counter CO, that is, learning data, reference numeral 14 is a table memory for specifying essential feature codes, and reference numeral 15 is a table memory for specifying prohibited feature codes.

The eighth FA is a flowchart showing an example of a program for implementing the present invention stored in the RAM 12, which will be described below. First, in step 21, a counter CO is initially set.

This initial set is to set the 5-bit counter to all O's, or ooooo. It is then compared in step 22 with the essential feature code of the corresponding character in FIG. If the required feature is specified as 0 or 1 here, step 2 if the corresponding counter CO bits match.
The process moves to step 3, and if they do not match, the process moves to step 25. When proceeding to step 23, the prohibited feature table shown in figure $6 is referred to, and a comparison is made to see if the corresponding bit combination of the counter CO matches the corresponding bit combination of the prohibited Q4 characteristic table. When a match is found, proceed to step 25 as a prohibited 4M song. Doesn't match Z')
If so, the process goes to step 24, and the entire contents of the color/color CO are stored in the RAM 13 of FIG. 7 as learning data. As shown in FIG. 6, there are cases where multiple types of combinations of the same character/figure are designated as prohibited features, so in the comparison operation U of step 23, these multiple prohibited! It is assumed that the comparison and matching matching operation is performed in response to the error.

The contents of the counter CO thus stored in the RAMI 3r become effective learning data that is automatically generated (created) in this invention.Next, in step 25, the contents of the counter CO stored in the RAMI
+IL to O', update the feature code.

This operation is completed by generating 1, that is, a maximum of 2'=32 types of %fine codes, which continue until the contents of the counter CO become 00000 again in step 26. The generation and creation of learning data is completed when the above operations are performed only for the literary society targeted for dictionary creation. This determination is made in step 2γ. Note that step 28
If you select the next character with , of course,
The essential features corresponding to the characters to be processed in Figure 5 and the 6th
The corresponding prohibition percentage of 1 is referenced during the comparison operations of steps 22 and 23, respectively.

<Effects of the Invention> As described above, in the method for creating a dictionary for character recognition according to the present invention, when creating a dictionary by learning by encoding the percentage characteristics of optically read characters and figures, each character recognition target is For a character figure, features that can always be obtained within any allowable deformation range and features that cannot be obtained within any allowable deformation range are designated as essential features and prohibited features, respectively, and the required features are determined. Among the included feature codes, only the feature codes that do not include prohibited features are stored in the dictionary and automatically
'll: Since this is used as learning data when performing g, it is possible to automatically generate and create a highly accurate and small-capacity dictionary in a short period of time.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing an outline of the character recognition device. Figure 2 is a diagram showing an example of the feature extraction code for the character "A". Figure 3 is the feature extraction code for the character "A". Figure 4 shows a list of essential feature codes for the character "i", and Figure 5 shows the memory contents of a table specifying essential feature codes in an embodiment of the present invention. FIG. 6 is a diagram showing the memory contents of a table for specifying prohibited 4+fine codes in one embodiment of the present invention, and FIG.
FIG. 8 is a block diagram showing the outline of the hardware configuration in the example, and FIG. 8 is a flow chart showing the operation in one embodiment of the present invention. Patent Applicant Tateishi Electric Co., Ltd. Agent Patent Attorney Tetsuji Iwakura (1 other person) Figure 5 Figure 6 Figure 7 Figure 8

Claims (1)

[Claims] Characteristic code generation means for reading characters and figures on a form by fully optical scanning, extracting and encoding character features of the read characters and figures, and standard codes stored in a dictionary of characteristic codes generated by said means; In a character recognition device equipped with collation means for comparing and collating and outputting the results, means for specifying character features that can always be obtained within any allowable deformation range of each character figure to be recognized as essential features. and a means for specifying character features that cannot be acquired within any allowable deformation range as prohibited features;
The characteristic code from the micro code generation means is I! (ii) means for determining whether the essential feature is included; means for determining whether the feature code determined by the essential feature determining means to include the prohibited feature; means for storing a feature code determined not to include a prohibited feature by the prohibited feature determining means; and f:[I? ft.
E., a character recognition dictionary creation method in which all feature codes stored in the storage means are used as learning data when the dictionary is fully automatically created.