JPH1153474A - Character string recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To shorten a character recognition time, to make character recognition and word matching well-balanced, and to improve efficiency by dividing the whole character set into two partial character sets and storing them in a matrix dictionary. SOLUTION: In a 1st recognizing process, a character matching and discrimination part 25 performs matching against all characters which belong to the 1st partial character set in the matrix dictionary 26. A work matching and discrimination part 28 tries to determine all characters by a work matching process by using a 1st candidate character group which possibly includes errors as it is. In a 2nd recognizing process, a binary image obtained by cutting the patterns of characters which can not be determined again is inputted from an input terminal 21 and reexecution is performed up to the extraction of a feature matrix; and characters are matched and discriminated over the 1st partial character set and 2nd partial character set of the matrix dictionary 26, and a specific number of candidate characters put in order as a result are found and written so that corresponding parts of the candidate character group stored in a candidate character memory 27 are updated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a character recognition method for reading a character pattern of a document or the like.

[0002]

2. Description of the Related Art A conventional character recognition method of this type is disclosed in the above-mentioned document (A): Japanese Patent Application Laid-Open No. Sho 57-23185, "Character Recognition Method". It is a block diagram of an apparatus. A character pattern of a character to be read is input from an input terminal 1 to a character pattern memory 2 as a binary image. The character frame dividing unit 3 detects a circumscribed frame of the input character pattern, and divides the internal region into a fixed number in each of the horizontal and vertical directions. FIG. 5A shows an example of an input character pattern and character frame division. In this example, the cells are divided into a matrix of 7 × 7 = 49 cells.

The horizontal sub-pattern extracting section 4 extracts horizontal stroke components from an input character pattern to create a horizontal sub-pattern. FIG. 5B shows a horizontal sub-pattern (H
SP). Similarly, for each of the vertical, diagonally left, and diagonally right directions, sub-patterns such as those shown in FIGS. 5C, 5D, and 5E are created by extracting the stroke components in the direction from the input character pattern.

The feature matrix extraction unit 8 divides these four sub-patterns HSP, VSP, LSP, and RSP into a matrix of squares using the division points obtained by the character frame division unit 3, and calculates strokes in each square. , The lengths of (a) horizontal matrix H, (b) vertical matrix V, (c) left diagonal matrix L, and (d) right diagonal matrix R as shown in FIG.

A matrix of 4 × (7 × 7) elements obtained by merging these elements is called a feature matrix because it represents a feature that can identify what character (category) it is. This feature matrix expresses the characteristics peculiar to the category such as the direction, position, and length of the stroke constituting the character pattern, but the character size, line width, variation in the character shape, etc. do not affect the value of the feature matrix. Has been converted to For this purpose, line width calculation, non-linear division of a matrix, blur processing, normalization processing, and the like are included, but descriptions of these processing methods are omitted.

The character collation / identification unit 9 creates a feature matrix obtained from an input character pattern and a standard character pattern for each category (character) of the entire character set to be read in the same procedure as described above. The feature matrix is collated with a feature matrix stored in the matrix dictionary 10 (hereinafter referred to as a standard matrix), and a category having a standard matrix most similar to the feature matrix is selected.

[0007] The distance between the feature matrix and the standard matrix in the dictionary memory is determined by the difference between the corresponding elements by two.
The distance is evaluated based on the sum of squares (or its square root), and a certain number of categories are selected as candidate characters by ranking the objects whose distances are close to 0.
Output from 1.

Next, FIG. 7 shows a character string using a combination of the above-described character recognition method and the character recognition post-processing method disclosed in the above-mentioned document (B): Japanese Patent Laid-Open No. 1-205288 "Character post-processing method". It is a block diagram of a recognition device. In this character string recognition device, a character pattern of an input character string to be read is input from an input terminal 21, and only one character is stored in a character pattern memory 22. The sub-pattern extraction unit 23 combines the blocks 3 to 7 in FIG.
Extract the sub-patterns as shown in Thereafter, the recognition of one character is performed by the feature matrix extraction unit 24, the character collation identification unit 25, and the matrix dictionary 26 in the manner described above. As a result, for example, the character names (character codes and the like) of the candidate characters up to third place And data of similarity (distance, etc.) and character type (discrimination of kanji, kana, symbols, etc.) are obtained and stored in the candidate character memory 27.

FIG. 8 is a flowchart of a conventional character string recognition using this character string recognition device. In FIG. 8, the above-described recognition processing in units of characters (step S101) is repeated until a delimiter (such as a punctuation mark) appears in the input character string in the determination branch processing (step S102). Here, it is assumed that the delimiter is easily recognized and uniquely determined.

Next, the word collating and identifying section 28 stores a candidate character memory.
The input character string is divided into several matching regions from the 27 candidate character groups (step S103 in FIG. 8). The division into the collation regions is performed in order to prevent a large number of meaningless word candidate character strings from being created in the collation with the next word dictionary. FIG.
A conventional method of obtaining a collation area from character type data of a candidate character will be described using an example of one input character string.

[0011] As shown in FIG.
The type of character type is assigned to each bit position of the 8-bit character type flag, and the character type data of each candidate character is represented by a character type flag as shown in FIG. In this example, the first-place candidate character “レ” of the input character “レ” is katakana 20H, the second place “shi” is hiragana 10H, and the third parenthesis mark is the symbol 02H. The character type flag of the unknown input character "shi" is set to 32H which is the logical sum of them. Thus, character type flags (106) are obtained for all unknown input characters as shown in FIG.

Next, from the character position P = 1, the character type flags are accumulated by logical AND in the forward direction. This is called a character type connection flag (107) and indicates which character type is continuous from the first character position. Then, the process proceeds to a position where the character type connection flag becomes 00H. However, there is an exception that if it changes (possibly) from Kanji (40H) to Hiragana (10H) even at 00H, it is reset to 10H. This is because the kanji and the following hiragana are not separated. Next, the character type connection flag is set to 00H.
From the character position (P = 4), the character type flag is accumulated in the reverse direction without exception (reference character search 108), and the position (P = 3) where this flag becomes 00H is set in the first collation area. The end point is set, and the next character position (P = 4) is set as the start point of the second collation area. Hereinafter, three matching regions can be determined by the same processing procedure.

Thereafter, as shown in the processing (step S104) and the processing (step S105) of FIG. 8, candidate character strings are selected by combining candidate characters in each of the obtained matching regions, and a predetermined word dictionary is selected. The character string is compared with 29, and each character of the input character string is determined based on the result of the comparison. The character determination method is
FIG. 11 shows the first collation area “character” in FIG. 9 as an example. Let us consider a case where six words as shown in the figure are extracted from a word dictionary for a word candidate character string composed of combinations of candidate characters. The word lengths and average candidate ranks of these six extracted words are as shown in the table in FIG.
For example, the word length of “character” is 2, and the candidate ranks of “sentence” and “character” are 2 and 1, so the average candidate rank of “character” is
1.5. If the word length is the longest and the word is determined from the word having the lowest average candidate rank, "character" is selected first, and then the remaining "wo" is selected, and all characters are determined.

[0014]

However, in the above character recognition method, the feature matrix extracted from the input character pattern is compared with the standard matrices of all the categories stored in the matrix dictionary. Was spending a long time processing. For example, if 3000 different Chinese characters are registered in the matrix dictionary,
In spite of the fact that it contains many characters that rarely appear in ordinary sentences, the probability of occurrence is not zero, so it is conventionally compared with the entire dictionary, that is, 3000 matrices Was.

In the above-described character string recognition method, a collation area is determined in a subsequent stage by using a candidate character group of an input character string, and collation with a word dictionary is performed in the collation area, thereby efficiently performing the entire character string. Although the characters can be determined, there is a problem of imbalance between the processes that the load of the character recognition process based on the feature matrix matching in the previous stage has not been reduced yet.

SUMMARY OF THE INVENTION It is an object of the present invention to provide an efficient character string recognition method in which character recognition time by matrix matching is reduced, and character recognition and word matching work in a well-balanced manner.

[0017]

In order to achieve this object, according to the first aspect of the present invention, the d direction (d) including at least the horizontal and vertical directions from the character pattern of each character of the input character string.
A sub-pattern extraction unit that extracts d sub-patterns by extracting stroke components of 2 or more constants, and divides a character circumscribed frame of the character pattern into an M × N rectangular area (M and N are constants). a matrix extraction unit for obtaining a feature matrix composed of d M × N matrix elements from the d subpatterns; identifying the character by comparing the feature matrix with a standard matrix stored in a matrix dictionary; A character collating and identifying unit for selecting candidate characters, and a word candidate character string created from a candidate character group of each character of the input character string, and collated with a character string stored in a word dictionary. A character string recognition method comprising: a word collating and identifying unit that determines each character of the character string; wherein the character set of the matrix dictionary is a first partial character set composed of characters having a high frequency of appearance. In the first recognition processing of the input character string, the image data is divided into a second partial character set composed of characters having a low appearance frequency and stored in the matrix dictionary in advance.
The character collation / identification unit compares the feature matrix of each character with a standard matrix of all the characters in the first partial character set to select a first candidate character group. A word candidate character string is created using the candidate character group of the above, and if all the characters of the input character string can be determined, the recognition processing of the input character string is terminated, and the characters that cannot be determined in the input character string Exists, in the second recognition process of the input character string, the character collation / identification unit converts the feature matrix of the undecidable character into a standard matrix of all the characters in the first and second partial character sets. To select a second candidate character group, and the word collation / identification unit creates a word candidate character string using the second candidate character group, and determines all characters of the input character string. It is characterized by the following.

According to a second aspect of the present invention, in the first aspect of the present invention, in the first recognition processing of the input character string, the character collating and identifying section stores the characteristic matrix of each character in the first character string. Performs matching with a standard matrix of all the characters of the partial character set and a threshold test for similarity. If there is a character in the input character string for which a candidate character cannot be selected by the threshold test, the characteristics of the character Comparing the matrix with a standard matrix of all the characters of the first and second partial character sets without performing a similarity threshold check;
It is characterized in that a candidate character group for the second time is selected.

Further, in order to achieve the above object, according to the third aspect of the present invention, the character collating and identifying section comprises K identical collating and identifying circuits (K is a constant) and a character collating control section. Dividing the character set of the matrix dictionary into the first partial character set and the second partial character set and storing them in the matrix dictionary in advance; and performing the sub-pattern extraction in the first recognition processing of the input character string. Unit and the matrix extraction unit obtain a feature matrix for each K character of the input character string and distribute the obtained feature matrices to the K matching identification circuits, wherein the K matching identification circuits respectively store different feature matrices in the first portion. A first candidate character group is selected by collating in parallel with a standard matrix of all characters in the character set.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION

[Explanation of First Embodiment] [Explanation of Configuration] The configuration of the character string recognition device of the first embodiment of the present invention is the same as that of FIG. Also a matrix dictionary
Although the contents of the standard matrix for each category are the same as those of the conventional technology, the present embodiment is characterized in that the entire character set is divided into two partial character sets and stored in the matrix dictionary 26. is there. The first partial character set stores a standard matrix of characters having a high appearance frequency, and the second partial character set stores a standard matrix of other characters separately. Here, an appearance frequency or an appearance probability in character units is considered for a general Japanese sentence.

A typical example in which the appearance frequency differs greatly depending on the character is a kanji. FIG. 12 shows an example in which a kanji set is divided into two parts by appearance frequency. This is JIS Level 1 Kanji 29
Of the 65 characters, 1006 educational kanji characters were divided as (a) the first partial character set, and the other 1959 kanji characters were divided as (b) the second partial character set. , But its appearance frequency in general documents is
The former is much higher than the latter, both in terms of characters and in sets.

Although the non-kanji characters (alphabetic characters, numbers, kana, and symbols) are slightly seen in the bias of the appearance frequency, they are not as remarkable as the kanji characters. In particular, alphabetic characters, numbers, katakana, and hiragana have a small difference in the appearance frequency among the respective character types, and the appearance frequency of each character is higher than that of the kanji. It should belong to the subcharacter set. Accordingly, the second partial character set is a kanji having a low appearance frequency and a symbol or a special character having a low appearance frequency.

In general, it is assumed that the number of characters in the first partial character set is 1 / m of the number of characters in the entire dictionary, and the appearance probability of the characters in the second partial character set in standard text is 1 / p. Then, numerical values such as m = 3 and p = 30 apply, and it can be said that the larger these values, the greater the bias in the appearance frequency.

[Description of Operation] The character string recognition device of FIG. 7 operates according to the flowchart of FIG. 1 in the present embodiment. In processing step S111, the character pattern of the input character string to be read is cut out character by character and input from the input terminal 21 in FIG. Sub-pattern extraction, feature matrix extraction, and character collation identification are executed for each input character, and each time, for example, the candidate characters up to the third place are stored in the candidate character memory 27.
This is performed up to the position delimited by the punctuation of the input character string (or the position where the candidate character memory 27 is saturated) (step S112). The above is referred to as a first recognition process of the input character string. In the first recognition process (step S111), the character collating and identifying unit 25
Of all characters belonging to the first partial character set.

At this stage, if the number of characters in the first partial character set is 1 / m of the whole, the collation identification time is 1 / m of the conventional method. However, all of the candidate characters obtained here are characters belonging to the first partial character set, and if the input character string contains characters of the second partial character set at a rate of one in p characters, Then, the character does not include the correct candidate character.

Next, the word collation / identification unit 28 attempts to determine all the characters by word collation processing using the first candidate character group possibly containing the error. Although the processing procedure is the same as the method described in the related art, a method for efficiently executing this processing will be described below. First, in order to divide an input character string into a plurality of matching regions, character type flags of all candidate characters are required.
It is assumed that the matrix dictionary 26 stores character type data in the form of an 8-bit character type flag in addition to a standard matrix and a character code for each category (character). Thus, the process of determining the character type from the character code and creating the character type flag can be omitted. Each time a plurality of candidate characters are determined for each input character, the logical sum of the character type flags of all the candidate characters is obtained, and a character type connection flag in the character string direction is also obtained.

The collation area is cut out using the character type connection flag. In the example of FIG. 9, the character type connection flag is set to 00H at the time when the recognition processing of “light” at the character position P = 4 is completed. Is determined. Similarly, the character type connection flag is set to 0 at the time of recognition of the "strength" at the character position P = 8.
It becomes 0H, the second collation area “collate” is determined by the backward search, and the third collation area “output” is determined when “.” At the character position P = 11 is recognized as a delimiter. As described above, the collation area determination processing is the first character recognition processing.
Can be executed in parallel with 111.

If the delimiter character is recognized in the processing step S112, it is sufficient to confirm the collation area division result in the processing step S113, and thereafter, for each collation area (step S11).
4) A process of creating a word candidate character string, collating with a word dictionary, and determining a character is performed (step S115). Here, if the input character string does not include a character of the second partial character set, both candidate characters and correct characters are within the range of the first partial character set. All characters can be determined in the same manner as in the conventional method. This is a case where the processing step S116 is always Yes, and the processing for each collation area is completed.

Next, if the input character string includes characters of the second partial character set, all of the candidate characters
Contains an error because it is selected from the sub-character set of. This error can be determined if the word matching process does not find a word candidate character string composed of candidate character groups of the input character string in the word dictionary 29, since the character cannot be determined. Judge that it is detected. This is a case where the processing step S116 is No, and the character cannot be determined because it belongs to the second partial character set or is in the same collation area as the character belonging to the second partial character set.
Are determined to belong to the partial character set, and a new candidate character is selected by performing recognition processing again on those characters.

This is referred to as a second recognition process (step S11).
7). Assuming that the ratio of the character to be subjected to the second recognition process in the standard sentence is 1 / n, n is smaller than the above-described p, and a numerical value such as n = 10 is applicable. In this case, the second of the characters
The second recognition process will be performed.

In the second recognition processing step S117, a binary image obtained by cutting out the pattern of the character which cannot be determined again from the input terminal 21 in FIG. 7 is input, and the process is repeated until the feature matrix is extracted. Character collation identification is performed over the entire first partial character set and second partial character set of the dictionary 26, and a predetermined number of ranked candidate characters are obtained as a result. Write in the form of updating the corresponding part of the character group. The above is repeated for all undecided characters in the collation area (step S118).

Thereafter, returning to the processing step S115, if the correct candidate character is included in the new candidate character, a word candidate character string using this is found in the word dictionary 29,
Unknown character is determined. In this case, there is no need to redo the collation area division (step S113) using the second candidate character group, and the second character recognition and word collation can be performed again in the collation area including the undecided character. good. In the first recognition processing (step S111), even if a kanji belonging to the second partial character set is recognized as a non-kanji, if a kanji exists in another candidate character, the determination of the collation area is performed. This is because mistakes can be avoided. Therefore, it can be considered that the collation area has been correctly obtained by using the first candidate character group.

The second recognition processing step S117
After that, if all the characters cannot be determined as a result of the second word collation processing step S115, the second first candidate character is output as a solution.

As described above, the first recognition processing is 1 / m of the conventional technique, and the second recognition processing is 1 / n of the conventional technique.

[Explanation of Second Embodiment] [Explanation of Configuration] The configuration of a character string recognition apparatus according to a second embodiment of the present invention is the same as that of FIG. Also a matrix dictionary
The contents of 26, that is, the contents of the standard matrix for each category, and the point that the storage method is stored by being divided into two partial character sets are the same as those in the first embodiment. The difference from the first embodiment is that a threshold value regarding similarity (distance) is set for each category of the first partial character set of the matrix dictionary 26. If the distance in the feature space between the feature matrix of the input character and the standard matrix of the matrix dictionary 26 exceeds this threshold, it is determined that the two matrices are clearly different categories (characters). Are set. In that case,
Characters in the matrix dictionary 26 having the standard matrix cannot be candidate characters.

Conversely, if the distance between the above matrices does not exceed the threshold, both matrices are not necessarily of the same category, but the characters in the matrix dictionary 26 having the standard matrices can be candidate characters. As described above, the threshold check can be performed for all the characters in the first partial character set. However, the recognition process can be performed without performing the threshold test.

[Explanation of Operation] The character string recognizing device shown in FIG. 7 operates in the same manner as the first embodiment as a whole in the present embodiment, but the first recognition processing portion is the same as that in FIG. It operates according to the flowchart. That is, in the first recognition processing step S121, a feature matrix is obtained for each character of the input character string and is compared with the standard matrix of all the characters belonging to the first partial character set. Is different from the first embodiment in that when the distance exceeds a threshold, the character is not considered as a candidate character. Therefore, the number of candidate characters for a certain input character may be three or less of the designated number, or in an extreme case, no single character may be obtained. In the latter case, the result is No in the processing step S122, and in the processing step S123, the character recognition processing is redone as a selection target up to the second partial character set, so that a new candidate character is obtained.

Since an input character having no candidate character in the first partial character set is regarded as a character included in the second partial character set, the character collation / identification unit 25 uses the feature matrix to Character matching is performed on the entire first partial character set and second partial character set of the dictionary 26. The reason why the collation including the first partial character set is performed is to select and select the candidate characters from the entire set obtained by combining the first and second partial character sets again. In this re-recognition process, the similarity of the candidate characters is not compared with a threshold to reduce the number of candidate characters (threshold check). This is because the correct answer is included in the candidate character, so that the same recognition processing as in the conventional method may be performed, and the correct answer was found in the first partial character set. This is because if a character is erroneously excluded by the threshold check, it must be picked up as a candidate character. In this case, it is assumed that the collation area determination processing proceeds using the character type flag of the new candidate character.

Conversely, if there is at least one character in the matrix dictionary that can be a candidate character whose distance between the matrices does not exceed the threshold value for a certain input character (when processing step S122 is Yes), Since a fixed number (for example, a maximum of three) is selected as candidate characters from those having the shortest distance, it is unnecessary to execute the re-recognition processing step S123. If there is a correct answer in the second partial character set, but there is a similar character in the first partial character set and an incorrect candidate character is selected, the error is detected at the word matching stage. Therefore, the correct character may be picked up in the second recognition processing.

Therefore, after the first recognition process described above, the process proceeds to the processing step S113 in the flowchart of FIG. 1 to create a word candidate character string for each collation area, collate with a word dictionary, and input character string. If a decision is made but still cannot be made, the second decision is made as described in the first embodiment.
After performing the second recognition processing step S117, the final determination of the input character string is performed.

[Explanation of Third Embodiment] [Explanation of Configuration] FIG. 3 is a diagram showing the configuration of a character string recognition apparatus according to a third embodiment of the present invention, which includes an input terminal 21, a character pattern memory 22, a sub-pattern. Extraction unit 23, feature matrix extraction unit
24, a character collation / identification unit 25, a matrix dictionary 26, a candidate character memory 27, a word collation / identification unit 28, a word dictionary 29, and an output terminal 30. The character collation identification section 25 includes K collation identification circuits 31,
And a character collation control unit 32. K is an arbitrary integer.

The difference from FIG. 7 is that K collation identification circuits 31 are provided, and these operate in parallel. The character collation control unit 32 is connected to the K collation identification circuits 31 via a bus 33 to control the parallel operation, and the address line 34
The dictionary 26 is accessed through.

The contents of the matrix dictionary 26 and the method of storing the same are the same as those in the first embodiment. The matrix dictionary 26 contains standard matrices for all categories of all character sets to be read as described in the first embodiment. As in the first sub-character set
36 and a second partial character set 37.

[Explanation of Operation] The character string recognition apparatus shown in FIG. 3 basically operates according to the flowchart shown in FIG. 1. However, the matrix recognition in the first recognition processing step S111 and the second recognition processing step S117 is performed. , The recognition processing is performed in parallel for a maximum of K characters.

From the input terminal 21, the character pattern of the character string to be read is sequentially input. 1 in character pattern memory 22
When the characters are loaded, the sub-pattern extraction unit 23 obtains a sub-pattern, and the matrix extraction unit 24 obtains a feature matrix. This feature matrix is sent to any one of the K verification and identification circuits 31 via the bus 33. For example, the feature extraction result of the first character is sent to the first collation identification circuit # 1.

Next, the character pattern of the second character is input terminal 21
And loaded into the character pattern memory 22,
The feature extraction process is executed, and the resulting feature matrix is sent to the second verification / identification circuit # 2 via the bus 33. When the feature extraction processing of the K-th character pattern is completed in this way, the K matching identification circuits 31 have different input character feature matrices and execute the matching identification processing in parallel under the control of the character matching control unit 32.

Here, the first recognition processing step S11
In 1, each collation identification circuit 31 performs collation only with the first partial character set 36 of the matrix dictionary 26. With respect to the standard matrices of all the categories included therein, the distances from the input feature matrix are calculated, and they are sorted and ranked from those closest to 0, and a certain number of categories are selected. In the parallel processing, the address line 34 is connected to the first
Is controlled so as to access the partial character set 36, and dictionary data is simultaneously read from the matrix dictionary 26 into the K matching identification circuits 31 through the data lines 35. Each collation identification circuit 31 selects a predetermined number (for example, three) of candidate characters from the first partial character set of the matrix dictionary for the input characters. The results of the K collation identification processes (for example, 3K candidate characters) are transferred from the bus 33 to the candidate character memory 27.

The above-described K-character recognition processing is performed until a delimiter appears in the input character string (step S11).
2). After the first recognition process, the word matching / identifying unit 28 selects a word candidate character string, matches with a word dictionary, and determines an input character string for each matching area. The final decision is made after performing the second recognition process. This second recognition processing step S11
Also in 7, matrix matching is performed in parallel with a maximum of K characters using K matching identification circuits.

In this embodiment, since K characters are processed in parallel, the average processing time per character is 1 / K.
For example, if K = 4, it is reduced to one fourth of that of the first embodiment. However, this is a reduction in processing time for only character collation identification using the matrix dictionary 26, and does not necessarily reduce processing time for feature extraction processing of input character patterns or processing time for word collation.

The present invention is suitable for use in character string recognition of Japanese documents containing kanji and kana with many types of characters and large differences in the appearance frequency between the characters. If a high-speed large-capacity dynamic RAM or the like is used for the matrix dictionary and the word dictionary, and the other parts are configured by a dedicated logic integrated circuit, a small-scale and high-performance character string recognition device can be realized.

[0051]

According to the first aspect of the present invention, a stroke component in the d direction (d is a constant of 2 or more) including at least the horizontal and vertical directions is extracted from the character pattern of each character of the input character string, and d is extracted.
A sub-pattern extracting unit for obtaining the number of sub-patterns, and dividing the character circumscribed frame of the character pattern into M × N rectangular areas (M and N are constants), and comprising d M × N matrix elements from d sub-patterns A matrix extraction unit that obtains a feature matrix; a character matching identification unit that identifies the character by comparing the feature matrix with a standard matrix stored in a matrix dictionary and selects one or more candidate characters for each character;
A word matching identification unit that creates a word candidate character string from the candidate character group of each character of the input character string, compares the character string with a character string stored in a word dictionary, and determines each character of the input character string based on the matching result. In the character string recognition method, the character set of the matrix dictionary is divided into a first partial character set consisting of characters having a high frequency of appearance and a second partial character set consisting of characters having a low frequency of appearance. In the first recognition processing of the input character string stored in advance, the character collation / identification unit compares the feature matrix of each character with the standard matrix of all the characters in the first partial character set, and performs the first candidate character A group is selected, and the word collation / identification unit creates a word candidate character string by using the first candidate character group. If all the characters of the input character string can be determined, the input character string recognition processing is performed. Finished and said If there is a character that cannot be determined in the input character string, in the second recognition processing of the input character string, the character collation / identification unit converts the feature matrix of the character that cannot be determined into first and second partial character sets. , A second candidate character group is selected by comparing with the standard matrix of all the characters, and the word collation / identification unit creates a word candidate character string by using the second candidate character group. This method is characterized by determining all the characters. In the first recognition processing, only the first partial character set having a high appearance frequency in a sentence and a small number of characters in a dictionary is checked, and its recognition is performed. Only when an error is detected at the time of word matching from the result, the entire dictionary including the second partial character set with a large number of characters is compared in the second recognition process. (1 / m) + (1 / n ). For example, if m = 3 and n = 10, it is reduced to 0.43 times. Of course, the processing time other than the character collation identification is not shortened, but the word collation area is determined during the first character collation processing, and the word collation processing is performed by using the word collation area to efficiently generate the character string. Since the decision has been made, the above-mentioned effect of shortening the time is considered to be effective.

According to the second invention, in the first recognition processing of the input character string, the character collation / identification unit sets the characteristic matrix of each character to be similar to the standard matrix of all the characters in the first partial character set. If a character with which a candidate character cannot be selected by the threshold check exists in the input character string, the feature matrix of the character is changed to a total of the first and second partial character sets. This method is characterized in that the character string is compared with a standard matrix of characters without performing a similarity threshold test, and a first candidate character group is selected. Since an error in character selection is detected and a correct candidate character is selected by re-recognition processing, a word candidate character string selection processing using an incomplete first candidate character group as in the first embodiment, Word matching Management, and it can be avoided by performing the second round of recognition. Therefore, the processing time is shorter than in the first embodiment. Further, even if an error cannot be detected in the first recognition process and the re-recognition process, the error is detected in the second recognition process, so that the same result as in the first embodiment can be obtained in a short processing time. Can be

According to the third aspect of the present invention, the character collating / identifying unit is composed of K identical collating / identifying circuits and a character collating control unit, and the character set of the matrix dictionary is combined with the first partial character set. 2 and stored in a matrix dictionary in advance, and in the first recognition processing of the input character string, the sub-pattern extraction unit and the matrix extraction unit obtain a K-characteristic matrix for each K character of the input character string. Distributed to the collation identification circuits, and the K collation identification circuits collate different feature matrices in parallel with the standard matrix of all characters in the first partial character set to select a first candidate character group. In the first character recognition, only the first partial character set needs to be compared, and in the second recognition processing, only a part of the input character string needs to be compared. In the most time-consuming can be shortened collation identification processing time, further the K matching identifying a character per collation identification processing time since the are poised circuit K by the character recognition
It can be shortened by a factor of one. If m and n described above are used, (1 / m + 1 / n) * 1 / K of the conventional method is obtained. Since the K matching identification circuits simultaneously use common data from one matrix dictionary, K matching dictionaries are not required and the circuit configuration is simple.

[Brief description of the drawings]

FIG. 1 is a flowchart of character string recognition according to a first embodiment of the present invention.

FIG. 2 is a flowchart of character string recognition according to a second embodiment of the present invention.

FIG. 3 is a block diagram illustrating a configuration of a character string recognition device according to a third embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a conventional character recognition device.

FIG. 5 is a diagram showing an example of an input character pattern and a sub pattern.

FIG. 6 is a diagram illustrating an example of a feature matrix.

FIG. 7 is a block diagram illustrating a configuration of a character string recognition device.

FIG. 8 is a flowchart of conventional character string recognition.

FIG. 9 is an explanatory diagram of a method of obtaining a matching area.

FIG. 10 is an explanatory diagram of a character type flag.

FIG. 11 is an explanatory diagram of word matching.

FIG. 12 is a diagram showing an example of division of a matrix dictionary.

[Explanation of symbols]

 Reference Signs List 21 input terminal 22 character pattern memory 23 sub-pattern extraction unit 24 feature matrix extraction unit 25 character collation identification unit 26 matrix dictionary 27 candidate character memory 28 word collation identification unit 29 word dictionary 30 output terminal

Claims (3)

[Claims] 1. A d direction including at least horizontal and vertical from a character pattern of each character of an input character string (d is a constant of 2 or more)
And a sub-pattern extraction unit for extracting d sub-patterns by extracting the stroke components of the character pattern, and defining a character circumscribed frame of the character pattern as an M × N rectangular area (M,
N is a constant), a matrix extraction unit that obtains a feature matrix composed of d M × N matrix elements from the d subpatterns, and compares the feature matrix with a standard matrix stored in a matrix dictionary to write characters. And a character collating and identifying unit for identifying one or more candidate characters for each character, a word candidate character string created from a candidate character group of each character of the input character string, and a character string stored in a word dictionary. And a word collating and identifying unit that collates and determines each character of the input character string based on the collation result. In the first recognition process of the input character string, the character string is divided into a second partial character set composed of characters having a low appearance frequency and stored in the matrix dictionary in advance. Serial character collating identification unit first against the entire character of the standard matrix of the first portion character set the feature matrix of each character
A second candidate character group is selected, and the word collation / identification unit creates a word candidate character string using the first candidate character group, and if all characters of the input character string can be determined, the input is performed. When the character string recognition processing is completed, and there is a character that cannot be determined in the input character string, in the second recognition processing of the input character string,
The character collation / identification unit compares the feature matrix of the undecidable character with a standard matrix of all characters in the first and second partial character sets to select a second candidate character group, A word candidate character string is created using the second candidate character group, and all characters of the input character string are determined. 2. In the first recognition processing of the input character string, the character collation / identification unit checks a characteristic matrix of each character with a standard matrix of all characters of the first partial character set and a threshold test for similarity. When there is a character whose candidate character cannot be selected by the threshold check in the input character string, the feature matrix of the character is set to the standard of all the characters in the first and second partial character sets. The character string recognition method according to claim 1, wherein the first candidate character group is selected by performing matching with a matrix without performing a threshold value check of similarity. 3. The character collation / identification unit is composed of K identical collation / identification circuits (K is a constant) and a character collation control unit, and the character set of the matrix dictionary is combined with the first partial character set and the first partial character set. 2 and is stored in the matrix dictionary in advance, and in the first recognition processing of the input character string, the sub-pattern extraction unit and the matrix extraction unit are characterized by a unit of K characters of the input character string. A matrix is obtained and distributed to the K number of matching / identifying circuits, and the K number of matching / identifying circuits collate different feature matrices in parallel with the standard matrix of all characters of the first partial character set in the first time. The character string recognition method according to claim 1, wherein a candidate character group is selected.