JP3100786B2 - Character recognition post-processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition post-processing system capable of improving the accuracy of character recognition of handwritten input characters and the like in a character recognition device.

[0002]

2. Description of the Related Art A conventional post-character recognition technique will be described below. For the sake of explanation, it is assumed that a set of character categories to be recognized are kanji and katakana. It is assumed that the input pattern sequence is written on a form without a frame. Also,
In this conventional example, a case where a word dictionary is used is taken as an example, and a word dictionary composed of nine words shown in FIG. 2 is used as an example for explanation.

[0005] First, image data of the word “Yamamura Shoten” shown in FIG. 3 is input. Subsequently, a process of cutting out individual character patterns is performed. In this example, since the second character “village” is separated from the side, it is not possible to determine whether this is one character or two characters only by the information of the image data, and FIG. 4A and FIG. It is assumed that two cutout results are obtained as follows.

Next, each of the cut-out character patterns is subjected to character recognition, and character positions estimated from the cut-out results of the individual character patterns as shown in FIGS. ) And a candidate character are combined (this is called a candidate character matrix).

Next, the candidate character matrix created in this way is compared with a word dictionary, and a word having a large number of matching characters is selected as a correct answer (FIG. 6). In this example, FIG.
Match character number of "Yamamura Shoten" is 3
And "Yamamura Shoten" is output correctly.
This is because the estimated character position in FIG. 5B matches the character position of the word in the word dictionary (this is called the correct character position).

[0006]

However, in the character recognition post-processing method according to the above-mentioned prior art, in comparing a word in a word dictionary with a candidate character matrix, a character at a corresponding character position is compared. If the correct character positions do not match, the word matching fails.

An example is shown below. FIG. 7 shows the result of extracting characters from one example of image data of “Micro Corporation”. These are "stock" and "formula", "kai" and "sha"
This is an example in which the character cutout fails at these two locations because of contact, and dirt is mixed between “ku” and “b”. As shown in FIG. 8, the candidate character matrix obtained from the cut-out result and the words in the word dictionary have a misaligned estimated character position and correct character position, so that there is no matching character and word matching fails.

The present invention has been made to solve the above problems, and a first object of the present invention is to provide a character recognition post-processing method capable of accurately performing word matching and improving recognition accuracy. The second object of the present invention is to provide a character recognition post-processing method capable of executing the above-described character recognition post-processing at high speed.

[0009]

In order to achieve the first object, the present invention provides a character recognition device for recognizing an input character pattern sequence according to the present invention. Character extracting means for extracting an individual character pattern from a character, character recognizing means for character recognizing the individual character pattern output from the character extracting means, a key character added to a word as an index, and a character position of the key character. A key character dictionary, key character matching means for comparing a character recognition result output from the character recognition means with the key character dictionary, and obtaining a candidate word and a character position of the key character from a corresponding key character; A candidate word that is likely to be a correct answer among candidate words obtained by the character matching means is evaluated and selected at the character position of the corresponding key character, and is selected and output. The word evaluation means has a configuration to be used.

In order to achieve the second object,
According to a second aspect of the present invention, in the character recognition device for recognizing an input character pattern string, a character cutout unit that cuts out an individual character pattern from the input character pattern string, and a character output unit that outputs the character pattern. Character recognition means for recognizing individual character patterns, key character dictionaries summarizing key characters having significance assigned to each word as an index and character positions of the key characters, and characters output from the character recognition means Key character matching means for comparing a recognition result with the key character dictionary to obtain a candidate word and a character position of the key character from a corresponding key character; and a key character dictionary registered in the key character dictionary among the key characters constituting each word. A non-key character dictionary that summarizes the non-key characters and the character positions of the non-key characters for each word, and the key character matching unit. The information of the non-key character not registered in the key character dictionary of the candidate word is obtained from the non-key character dictionary, and the information of the non-key character is compared with the character recognition result output from the character recognition means. A non-key character matching means for obtaining a character position of a character; and a key character matching means and / or
Alternatively, a candidate word evaluation unit that evaluates, selects, and outputs a character position obtained by the non-key character matching unit is used.

[0011]

According to the character recognition post-processing method of the present invention, a character pattern is individually cut out from an input character pattern string, character recognition is performed, and word matching is performed based on the character recognition result. A character is searched to select a candidate word, and the candidate word is evaluated based on the character position information of the corresponding key character to select a word having a high possibility of a correct answer. In this way, by performing word matching without using character position information at the time of character extraction, even if the character position estimated from the character extraction result is deviated from the correct character position, word matching is possible, and recognition accuracy is improved. Let it.

Further, by selecting a key character having a high importance as a key character to be added to the key character dictionary as an index of the candidate word, the number of candidate words obtained as a result of the key character collation can be reduced, and the word collation can be reduced. The processing speed of selecting a candidate word having a high possibility of a correct answer is improved.

[0013]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a first embodiment of the character recognition post-processing system of the present invention. In FIG. 1, reference numeral 1 denotes a character extracting unit, 2 denotes a character recognizing unit, 3 denotes a key character dictionary, 4 denotes a key character collating unit, and 5 denotes a candidate word evaluating unit. This embodiment shows an embodiment of the first aspect of the present invention.

First, a character pattern string is inputted to the character extracting means 1, and an individual character pattern is extracted. Then, the individual character patterns are recognized by the character recognition means 2. As an example of input data, it is assumed that the image data of “Micro Corporation” in FIG. 7 is used, and the processing results of the character extracting unit 1 and the character recognizing unit 2 are the candidate matrices shown in FIG.

Next, the candidate character matrix is input to the key character collating means 4. The key character matching means 4 creates a search key table composed of candidate characters and estimated character positions as shown in FIG. 9 from the candidate character matrix. Then, this search key table is collated with the key character dictionary 3. Next, a configuration example of the key character dictionary 3 will be described.

As an example of the description, a case will be described in which all characters constituting a word are assigned to the word as key characters. First, a word code is assigned to each word in the word dictionary. Next, each word is decomposed into characters, and a data structure (referred to as individual character information) in which each character is combined with a word code and a correct character position is created. FIG. 10 shows an example in which individual character information of each character of each word is created from the word dictionary of FIG. Next, all the individual character information is collected for each character, and this is used as a key character dictionary. An example is shown in FIG.

When the search key table shown in FIG. 9 is compared with the key character dictionary shown in FIG. 11, a direct comparison result shown in FIG. 12 is obtained. The key character collating means 4 arranges these for each candidate word and creates a candidate word table as shown in FIG. As an example of candidate word information in the candidate word table, character position information including an estimated character position and a correct character position is used.

Next, the candidate word evaluation means 5 calculates an evaluation score for each candidate word in the candidate word table. There are various methods for calculating the evaluation score, and the following method is used as an example for explanation.

First, a value obtained by subtracting the correct character position from the estimated character position (referred to as a character position difference) is calculated from each character position information. Next, for each numerical value of the character position difference,
Calculate the difference score defined by the following method.

Assuming that the numerical value of the character position difference is X and the difference score of X is F (X), F (X) is given by the following equation: F (X) = the number of X + the number of (X−1) × 0.5+ (X
+1) × 0.5. Next, the largest numerical value in the obtained set of F (X) is set as the evaluation score of the candidate word.

FIG. 14 shows the process of calculating the evaluation score from the example of FIG. For example, for the candidate word “Micro Corporation”, the difference score of the character position difference−2 is 2.
5. The difference score of the character position difference -1 is 2, and the evaluation score is the maximum value 2.5 among these.

The candidate word evaluation means 5 outputs the candidate word having the highest evaluation score as a correct answer. In this case, "Micro Corporation" is correctly selected. As described above, according to the present embodiment, even if the positions of the words in the word dictionary and the character positions estimated from the character segmentation result are shifted, word matching can be performed, so that the recognition accuracy can be improved.

Next, a second embodiment of the present invention will be described.
This embodiment shows an embodiment of the second aspect of the present invention.

In the first embodiment, the method of restricting key characters to be added to a word is not described. In some cases, the number of candidate words obtained as a result of key character collation increases, and the processing amount increases. There is a danger of becoming.

As an example, in the description of the first embodiment, it is assumed that data of "Yamamura Shoten" in FIG. 3 is input as an input character pattern string. Character segmentation means 1
5A and 5B are obtained as a result of the character recognition unit 3. The character recognizing means 2 creates the search key table of FIG. 15 from the candidate character matrix of FIG. 5B.

The direct result of the comparison between the key character dictionary 3 and the retrieval key table of FIG.
6, and the candidate word table to be created is shown in FIG. As can be seen from FIG. 17, since the “quotient” and “store” of the character recognition result are used in seven words in the word dictionary of FIG. 2, the number of candidate words is less than that of the description in the first embodiment. It is much larger than if you were.

The present embodiment has been made in view of the above, and by limiting key characters to be assigned to words according to the importance as a search key, post-recognition processing with high recognition accuracy can be performed at high speed. This is an example of the operation.

FIG. 18 is a block diagram showing a configuration of a character recognition post-processing system according to the second embodiment of the present invention. In FIG. 18, 1 is a character segmenting means, 2 is a character recognizing means, 3 is a key character dictionary, 4 is a key character collating means, 60 is a non-key character dictionary, 70 is a non-key character collating means, and 5 is a candidate word evaluating means. It is.

First, a character pattern string is input to the character extracting means 1, and individual character pattern extracting processing is performed. Then, each character pattern is read by the character recognition means 2.
Recognized by The processing up to this point is the same as that of the first embodiment. When the image data of “Yamamura Shoten” in FIG. 3 is used as an example of the input data, the processing results of the character cutout unit 1 and the character recognition unit 2 are shown in FIG. The candidate character matrix shown in FIG. 5A and FIG. 5B is obtained.

Next, the candidate character matrix is input to the key character matching means 4. Here, for the sake of explanation, the candidate character matrix shown in FIG. The key character matching means 4 creates the search key table of FIG. Next, a configuration example of the key character dictionary 3 in the present embodiment will be described.

As described in "Problems to be Solved by the Invention", there are characters used in many words in the characters constituting the word dictionary, and these are used as key characters in all the words. If assigned, there is a risk of leading to an increase in the number of candidate words. Therefore, in this embodiment, it is considered that the importance of a character constituting each word as a search key is evaluated, and only a character having high importance is selected as a key character. There can be various measures of importance as a search key, but for explanation, the reciprocal of the number of words in which the character is used is referred to as importance.

For each character in the word dictionary of FIG.
Figure 1 summarizes the number and importance of words used
9 Here, when "importance is 0.4 or more" is used as an example of a standard adopted as a key character, characters other than "island", "quote", and "store" in the characters in FIG. FIG. 20 shows a key character dictionary adopted as characters. FIG. 21 shows a direct result of the collation result between the search key table of FIG. 15 and the key character dictionary of FIG. 20, and FIG. 22 shows a candidate word table created from this.

The non-key character collating unit 70 collates the candidate word table output from the key character dictionary 3 with the non-key character dictionary 60. The non-key character dictionary 60 stores information of characters (character information and non-key character information, which are composed of characters and correct character positions) among characters constituting each word which are not registered in the key character dictionary. FIG. 23
Shows an example. The non-key character matching unit 70 first searches the non-key character dictionary 60 for non-key character information for each candidate word included in the candidate word table, and matches the non-key character information with the search key table.
FIG. 24 shows the result of comparing the candidate word table of FIG. 22 with the search key table of FIG. 21 using the non-key character dictionary of FIG. Comparing this with FIG. 17, it can be seen that the number of candidate words has been significantly reduced from seven to two. Therefore, the time required for the subsequent processing performed on each candidate word is also greatly reduced. That is, the processing speed of the character recognition post-processing can be improved.

Next, the candidate word evaluation means 5 calculates the evaluation score of each candidate word in the same manner as in the first embodiment, for the result of the non-key character matching means 70, and determines the candidate word having the highest evaluation score. Output as a correct answer (an example is shown in FIG. 25). In this example, "Yamamura Shoten" is correctly selected.

[0036]

As is clear from the above description, according to the character recognition post-processing method of the present invention, even if the character position estimated from the word in the word dictionary and the character cutout result is shifted from the correct character position. Since word matching can be performed, recognition accuracy can be improved.

According to the second aspect of the present invention, in particular, the number of candidate words as a result of key character collation can be reduced by selecting a key character to be assigned to a word as a search key. The processing speed of the processing can be increased.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a character recognition post-processing method according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of a word dictionary used for explanation in an embodiment of the present invention and a conventional example.

FIG. 3 is a diagram showing image data of the word “Yamamura Shoten” as an example of input data;

FIGS. 4A and 4B are diagrams showing an example of a cutout result obtained by a character cutout process from the image data of FIG. 3;

FIGS. 5A and 5B are diagrams showing candidate character matrices obtained by recognizing the character segmentation results of FIGS. 4A and 4B;

FIG. 6 is a diagram showing a comparison between the candidate character matrix of FIG. 5B and the word “Yamamura Shoten”;

FIG. 7 is a diagram illustrating an example of a character cutout result using image data of “Micro Corporation” as an example of input data;

FIG. 8 is a diagram showing matching between a candidate character matrix obtained from FIG. 7 and the word “Micro Corporation”;

FIG. 9 is a diagram illustrating an example of a search key table including candidate characters and estimated character positions.

FIG. 10 is a diagram showing an example in which individual character information of each character of each word is created from the word dictionary of FIG. 2;

FIG. 11 is a diagram showing an example of a key character dictionary in the first embodiment.

12 is a diagram showing an example of a direct collation result obtained by collating the search key table of FIG. 9 with the key character dictionary of FIG. 11;

FIG. 13 is a diagram showing an example of a candidate word table according to the first embodiment;

14 is a diagram showing an example of calculating an evaluation score of each candidate word in the candidate word table of FIG.

FIG. 15 is a view showing a search key table obtained from the candidate character matrix of FIG. 4 (b).

FIG. 16 is a diagram showing a direct result of collation between the key character collating means 3 and the retrieval key table of FIG. 15;

FIG. 17 is a diagram showing a candidate word table created from FIG. 16;

FIG. 18 is a block diagram illustrating a configuration of a character recognition post-processing method according to a second embodiment of the present invention.

19 is a diagram showing a table summarizing the number of words used and the degree of importance for each character in the word dictionary in FIG. 2;

FIG. 20 is a diagram showing an example of a key character dictionary in the second embodiment.

21 is a diagram showing a direct result of a collation result between the search key table of FIG. 15 and the key character dictionary of FIG. 20;

FIG. 22 is a diagram showing a candidate word table created from FIG. 21;

FIG. 23 is a diagram showing an example of a non-key character dictionary in the second embodiment.

FIG. 24 is a diagram showing a result of comparing the candidate word table of FIG. 22 with the search key table of FIG. 21;

FIG. 25 is a diagram showing a result of calculating an evaluation score of each candidate word in FIG. 24;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Character cut-out means 2 ... Character recognition means 3 ... Key character dictionary 4 ... Key character collation means 5 ... Candidate word evaluation means 60 ... Non-key character dictionary 70 ... Non-key character collation means

Continuation of the front page (56) References JP-A-4-96889 (JP, A) JP-A-2-264388 (JP, A) JP-A-2-121078 (JP, A) JP-A-61-136182 (JP) , A) JP-A-59-188783 (JP, A) "Transactions of the National Convention of the Information Processing Society of Japan," No. 34 (Early 1987) 3 p. 1845-1846 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/72 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A character recognition apparatus for recognizing an input character pattern sequence, comprising: a character extracting unit for extracting an individual character pattern from the input character pattern sequence; and character recognizing an individual character pattern output from the character extracting unit. A character recognition unit, a key character dictionary summarizing key characters assigned to the word as an index and character positions of the key characters, and matching the key character dictionary with a character recognition result output from the character recognition unit. Key character matching means for obtaining a candidate word and the character position of the key character from the key character; and a candidate word having a high possibility of being a correct answer among candidate words obtained by the key character matching means. And a candidate word evaluation means for evaluating and selecting at a position and outputting the selected word. 2. A character recognition device for recognizing an input character pattern sequence, comprising: a character extracting unit for extracting an individual character pattern from the input character pattern sequence; and character recognizing the individual character pattern output from the character extracting unit. A character recognition unit, a key character dictionary that summarizes key characters having importance assigned to each word as an index and character positions of the key characters, a character recognition result output from the character recognition unit, and the key character dictionary. Key character matching means for obtaining a candidate word and a character position of the key character from the corresponding key character by collating, non-key characters not registered in the key character dictionary among the key characters constituting each word, and the non-key A non-key character dictionary in which character positions of characters are grouped for each word; and the key sentences of candidate words obtained by the key character matching means. Non-key character information not registered in the character dictionary is obtained from the non-key character dictionary, and the non-key character information is compared with the character recognition result output from the character recognition means, and the character position of the corresponding non-key character is determined. Non-key character matching means to be obtained, and a word having a high possibility of being a correct answer among the candidate words is evaluated and selected and output at a character position obtained by the key character matching means and / or the non-key character matching means. A character recognition post-processing method characterized by using candidate word evaluation means.