JPH02148277A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To improve the performance of pattern recognition by recognizing each character while calculating valuation points indicating the degree of resemblance between an input character pattern and standard patterns in plural slice levels and preferentially gathering higher evaluation points to generate a keyword for word retrieval. CONSTITUTION:A character recognizing part 12 recognizes each character while calculating evaluation points indicating the degree of resemblance between the input character pattern and standard patterns in plural slice levels preliminarily set in relation to the variable density or the like of the character and sends first to fourth character code data as recognition outputs and word division information to a word retrieving part 14. This part 14 preferentially extracts the character candidate having the highest evaluation points from the character candidate group with respect to each character, and words are segmented based on word division information to generate a keyword for word retrieval. Thus, the input character pattern is accurately recognized independently of the variable density of expressed characters.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a pattern recognition device used in, for example, an optical character reading device or a voice recognition device.

(Prior art) In recent years, in the field of pattern recognition technology for characters, speech, etc., the final recognition results of input patterns can be determined by searching databases such as word dictionaries based on the data obtained from pattern recognition results. The method of obtaining it is widely used.

FIG. 10 is a block diagram showing an example of a postal code reading device employing such a pattern recognition method.

As shown in the figure, this device includes a reading section 1 that optically scans a postal letter P on which an address, postal code, etc. are recorded and reads character pattern images on two sides of the letter; A character recognition unit 2 obtains character candidates (character codes) from a character pattern image read by using a recognition method such as a composite similarity method, a word dictionary 3 in which words are registered in advance, and a a word search unit 4 that creates a keyword for word search from the character candidate group and word delimiter information and searches the word dictionary 3 based on this keyword; and an address database 5 that stores address codes corresponding to addresses.
and a database search unit 6 that searches the address database 5 based on the words searched by the word search unit 4.

Next, the operation of this device will be explained.

First, the reading section 1 optically reads the character pattern image on the surface of the postal letter P, and transfers the read data to the character recognition section 2 via A/D conversion.

Thereafter, the character recognition unit 2 calculates an evaluation score indicating the degree of similarity between the human character pattern and the standard pattern using a composite similarity method or the like, performs recognition for each character, and the first
to the fourth character candidate group and word delimiter information are each sent to the word search unit 4.

Next, the word search unit 4 creates a keyword for word search based on the first candidate of each character and the word delimiter information, and searches this keyword and the words in the word dictionary 3, for example, every three character strings. After collation, each word obtained as a search result is converted into a word ID code and output to the address database search section 6.

The database search unit 6 searches the address database 5 using the word ID code as a keyword, thereby obtaining the address code corresponding to the address written on the postal letter P, that is, the address code corresponding to the postal code, as the final recognition result. .

However, such conventional devices have the following problems.

In other words, if the characters recorded on the postal letter P are thin overall, the character recognition unit 2 cannot perform good character recognition under the correct conditions, and character candidates having the characteristics of the input character pattern cannot be recognized. It was difficult to draw out the information faithfully. Moreover, misreading at the character recognition stage has a significant negative effect on the overall processing, resulting in a significant drop in reliability.゛In addition, the word search unit 4 extracts from the word dictionary 3 all the words that match the keyword for every three character strings, so when there are a large number of extracted word candidates, the address search is performed using these as keywords. There is a problem in that searching the database 5 takes an extremely long time.

Furthermore, with conventional devices, the cumulative total for each word is obtained from the evaluation score for each character written as a result of character recognition, and this is divided by the number of characters in the word to obtain the word evaluation value. I list these in order from the word with the highest evaluation value.
After converting to a D code, this was used to search for data in the address database 5.

However, when word evaluation values are determined using this method, the following disadvantages occur.

For example, if the word rJOHNsJ is obtained as a result of a word search for the keyword rJOHNJ, the total character evaluation score obtained for this word is 9°8X.
If it is 4-260 +1EX, the single 5fi rating fdl value will be 7911EX, which is obtained by dividing this by the number of words in the word, 5.

However, in the opposite case, if the word obtained for the keyword rJOHNsJ is rJOHNJ,
Although it is originally preferable that the word evaluation value be equal to the above value, in this case, dividing the cumulative total by 4 results in a higher value (9g + 1EX). Also,
This word's evaluation value may match the value obtained in the case of the keyword rJOHNJ, or even worse, the evaluation value calculated for the keyword "word equal to JOHNSJ" JOHNsJ may be There was also a risk that the score would be lower than the calculated evaluation value because there were more characters.

As described above, conventional devices have a problem in that highly reliable word evaluation values cannot be stably obtained.

Furthermore, conventionally, in the word dictionary 3, character code strings constituting words are registered in ascending order in ASCII code along with word ID codes for database searches, so word searches are always performed using binary search. 17 for all words in word dictionary 3.

Therefore, as the number of registered words in the word dictionary 3 increases, the word search time also becomes longer.

(Problems to be Solved by the Invention) As described above, conventional pattern recognition devices have many problems with accuracy in pattern recognition processing and word searches, and problems with word search efficiency, making it difficult to achieve high reliability. It was difficult to perform pattern recognition stably under these conditions.

The present invention is intended to solve the above-mentioned problems, and is to provide a pattern recognition device that can effectively improve pattern recognition accuracy and word search accuracy, as well as improve word search efficiency and significantly increase reliability. intended to provide.

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above-mentioned object, the first invention provides a pattern reading means for reading an input pattern, and an input pattern and a standard pattern read by the pattern reading means. A pattern recognition means that calculates evaluation points indicating the degree of similarity for each level and recognizes the input pattern for each level, and an optimal pattern based on the evaluation points for each level calculated by this pattern recognition means. The present invention is equipped with an optimum pattern selection means for selecting the optimum pattern.

The second invention also provides a pattern reading means for reading an input pattern, and a method for recognizing the input pattern by calculating evaluation points indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern for each of a plurality of levels. a pattern recognition means for each character, a word storage means in which words are stored in advance, and a keyword for word search by preferentially collecting characters with high evaluation scores for each character from the recognition results of the pattern recognition means. and word search means for searching for words in the word storage means based on the keywords created.

Furthermore, the third invention includes a pattern reading means for reading an input pattern, a pattern recognition means for recognizing the input pattern read by the pattern reading means, a word storage means in which words are stored in advance, and a pattern recognition means. A keyword for word search is created from the recognition results, and this keyword is compared with the words stored in the word storage means for each predetermined number of character strings, and the word with the highest number of matches is used for the final word search. It is equipped with a word search means for outputting the results with priority.

Furthermore, a fourth invention provides pattern reading means for reading an input pattern, pattern recognition means for recognizing the input pattern read by the pattern reading means, word storage means in which words are stored in advance, and recognition of the pattern recognition means. A first word search means that creates a keyword for word search from the result and searches the word storage means based on this keyword; and a first word search means that searches the word storage means based on the keyword; and a second word search means that compares each word and outputs the word with the largest number of matching characters as the final word search result.

Furthermore, a fifth invention includes a pattern reading means for reading an input pattern, and a pattern recognition for recognizing the input pattern by calculating an evaluation score indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern. means, a word storage means in which words are stored in advance, and a first word search means that creates a keyword for word search from the recognition result of the pattern recognition means and searches the word storage means based on the keyword; The cumulative evaluation score for each character in each word searched by this first word search means = 1 is divided by the larger number of search word characters and keyword characters, and the higher value is used as the final score. and a second word search means that outputs the word search results preferentially.

The sixth invention is a combination of the third invention and the fourth invention described above.

The seventh invention is a combination of the third invention and the fifth invention described above.

The eighth invention is a combination of the third to fifth inventions described above.

The ninth invention is a combination of the second to fifth inventions described above.

The tenth invention provides an image input means that inputs an image consisting of a plurality of characters and outputs an image signal, and inputs the image signal and performs appropriate image processing corresponding to each character image. An image processing means that outputs a plurality of character image patterns for each character, and an evaluation score indicating the similarity between each of the plurality of character image patterns and a standard pattern prepared in advance, and based on the evaluation score. a pattern recognition means for recognizing each character; a word storage means in which words consisting of character strings are stored in advance; character candidates with high evaluation scores are selected for each character from the recognition results of the pattern recognition means; The apparatus further comprises word search means for generating a character string by combining character candidates and performing a word search on the word search means using this character string as a keyword.

An 11th invention is based on the above-mentioned 10th invention, by inputting the image signal from the image input means and binarizing it at a plurality of preset slice levels for each character image. The apparatus is equipped with image processing means that outputs a plurality of character image patterns for each character image pattern.

The twelfth invention provides a button reading means for reading an input pattern, a pattern recognition means for recognizing the input pattern read by the pattern reading means, and a word storage in which words are grouped and stored in advance according to predetermined types. A keyword for word search is created from the recognition result of the means and the pattern recognition means, and the type of character string that constitutes this keyword is identified, and based on the identification result, the corresponding group of the word storage means is targeted. and a word search means for performing a word search.

The thirteenth invention provides a pattern reading means for reading an input pattern, a pattern recognition means for recognizing the input pattern read by the pattern reading means, and a word storage in which words are stored in advance in groups according to predetermined types. means and
A keyword for word search is created from the recognition result of the pattern recognition means, the type of character string that constitutes this keyword is identified, and based on the identification result, a word search is performed for the corresponding group of the word storage means. The present invention is equipped with a word search means for performing the following in accordance with predefined priorities for each type of word.

The fourteenth invention provides a pattern reading means for reading an input pattern, a pattern recognition means for recognizing the input pattern read by the pattern reading means, a word storage means in which words are stored in advance, and a recognition result of the pattern recognition means. A keyword creation means for creating a keyword for a word search from and a word search means for outputting words with priority as final word search results.

A fifteenth invention provides a button reading means for reading an input pattern, and a pattern recognition means for recognizing the input pattern by calculating an evaluation score indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern. , a word storage means in which words are stored in advance, a keyword creation means for creating a keyword for a single word search from the recognition result of the pattern recognition means, and a keyword created by the keyword creation means and stored in the word storage means. Divide the cumulative total of evaluation points for each character in each word by the larger number of search word characters and keyword characters, respectively.
The device is equipped with a word search means that outputs the higher value as the final word search result with priority.

(Function) In the pattern recognition device of the first invention, the pattern recognition means calculates evaluation points indicating the degree of similarity between the input pattern input from the pattern reading means and the standard pattern for each of a plurality of levels, and evaluates the input pattern. recognition of each,
Subsequently, the optimum pattern selection means selects the optimum pattern based on the evaluation score for each level calculated by the pattern recognition means, so that accurate recognition results can be obtained regardless of the shading of the input pattern. , thereby improving pattern recognition performance.

In the second invention, the pattern recognition means calculates evaluation points indicating the degree of similarity between the input pattern inputted by the pattern reading means and the signpost pattern for each of a plurality of levels, and recognizes the input pattern for each of the plurality of levels. Next, the word search means creates a keyword for word search by preferentially collecting characters with high evaluation scores for each character from the recognition results of the pattern recognition means, and based on this keyword, the word storage means Since a word search is performed for the word, the optimum word can be extracted from the word storage means with higher reliability.

Further, in the third invention, the word search means creates a keyword for word search from the recognition result of the pattern recognition means,
This keyword is compared with the words stored in the word storage means for each predetermined number of character strings, and the words with a high number of matching matches are prioritized and output as the final word search results. It becomes possible to further narrow down and extract the most suitable words. Furthermore, when performing another data search based on the searched word, it is also possible to shorten the extra data search time.

Furthermore, in the fourth invention, in the first word search means,
After creating a keyword for word search from the recognition result of the pattern recognition means and searching for a word from the word storage means based on this keyword, the second word search means searches the word from the first word search means. Each word and keyword
Since the word with a large number of matching characters is output preferentially as the final word search result, the same effect as the third invention described above can be achieved.

In the fifth invention, the first word search means first creates a keyword for word search from the recognition result of the pattern recognition means, and searches the word storage means based on this keyword. After this, in the second word search means,
The cumulative total of evaluation points for each character in each word searched by the first word search means is divided by the number of characters in that word or the number of keyword characters, whichever is larger, and the higher value is used for the final word search. Since the results are prioritized and output, it becomes possible to stably obtain highly reliable word evaluation values, and it becomes possible to perform word searches with an excellent accuracy rate.

In the sixth invention, the first word search means first creates a keyword for word search from the recognition result of the pattern recognition means, and then combines the keyword and the word stored in the word storage means into a predetermined number of characters. Each column is compared, and words with a high number of matching matches are prioritized and output. After this, the second
In the word search means, each word searched by the first word search means and the keyword are compared character by character, and words with a large number of matched characters are outputted with priority as the final word search result. , words with a higher degree of matching can be extracted from the word storage means, and pattern recognition performance can be improved.

In the seventh invention, the first word search means first creates a keyword for word search from the recognition result of the pattern recognition means, and then combines the keyword and the word stored in the word storage means into a predetermined number of characters. Each column is compared, and words with a high number of matching matches are prioritized and output. After this, the second
In the word search means, the cumulative total of evaluation points for each character in each word searched by the first word search means is divided by the greater number of the number of search word characters and the number of keyword characters, and the higher the value. As the final word search results are prioritized and output, it is possible to stably obtain highly reliable word evaluation values, and to narrow down the words with a higher degree of matching and output them from the word storage means. It is possible to improve pattern recognition performance.

In the eighth invention, the first word search means first creates a keyword for word search from the recognition result of the pattern recognition means, and then combines the keyword and the word stored in the word storage means into a predetermined number of characters. Each column is compared, and words with a high number of matching matches are prioritized and output as search results. After this, in the second word search means, the first tIi
Each word searched by the word search means and the keyword are compared character by character, and the word with the largest number of matching characters is given priority and output as a search result. Next, the third word search means divides the cumulative total of evaluation points for each character in each word searched by the second word search means by the larger number of the search word characters and the keyword characters, respectively. Since the one with the higher value is output with priority as the final word search result, it is possible to obtain effects similar to or even better than those of the sixth and seventh inventions described above.

In the ninth invention, the pattern recognition means generates keywords by calculating evaluation points indicating the degree of similarity between the input pattern and the standard pattern for each of a plurality of levels and recognizing each input pattern. After this, the first
In the word search means, after creating a keyword for word search from the recognition result of the pattern recognition means, this keyword is compared with i11 words stored in the word storage means for each predetermined number of character strings, and the comparison is performed. Prioritize and output words with a high number of matches as search results. Thereafter, the second word search means compares each word searched by the first word search means with the keyword, character by character, and gives priority to words with a large number of matching characters as word search results. Output. Further, the third word search means divides the cumulative evaluation score for each character in each word searched by the second word search means by the larger of the number of search word characters and the number of keyword characters, respectively. Since the one with the higher value is given priority and output as the final word search result, it becomes possible to perform a more reliable word search than in the case of the above-mentioned invention, and significantly improves pattern recognition performance. I can figure it out.

In the tenth invention, the image one-image processing means first performs appropriate image processing corresponding to each character image from the image signal input from the image input means, thereby producing a plurality of characters for each character. Generate image patterns respectively.

Thereafter, the pattern recognition means calculates an evaluation score indicating the degree of similarity with a standard pattern prepared in advance for each of the plurality of character image patterns, and recognizes each character based on the evaluation score. The word search means selects a character candidate with a high evaluation score for each character from the recognition result of the pattern recognition means, combines these character candidates to generate a character string, and performs the processing of this character string.
Effects similar to those of the first and second inventions can be obtained.

Furthermore, in the eleventh invention, the image processing means in the tenth invention described above binarizes the image signal input from the image processing means at a plurality of slice levels preset for each character image. By doing so, a plurality of character image patterns are generated for each character.

In the twelfth invention, the word storage means stores ILT words in advance in groups according to predetermined types, and the word search means creates keywords for word search from the recognition results of the pattern recognition means. , the type of string that constitutes this keyword is identified, and based on the identification result, a word search is performed for the corresponding group in the word storage means, so the number of words in the word storage means is very large. Even if there are many keywords, the word search can be performed only on groups according to the type of character strings that make up the keyword, making it possible to significantly improve the word search.

Furthermore, in the pattern recognition device of the thirteenth invention, the word storage means stores words in advance in groups according to predetermined types, and the word search means selects keywords for word search from the recognition results of the pattern recognition means. After creating
The type of character string that constitutes this keyword is identified, and based on the identification result, a word search is performed for the corresponding group of word storage means according to a predefined priority order for each word type. Therefore, by giving priority to searching for words that appear frequently, processing can proceed more effectively within a limited time.

In the fourteenth invention, first, the pattern recognition means recognizes the input pattern read by the pattern reading means, and then, from the recognition result of the pattern recognition means, the keyword creation means creates a keyword for ll1 word search. . Then, the word search means uses the keyword created by the keyword creation means and each keyword stored in the word storage means.
Each character is compared with one word, and the word with the largest number of matching characters is prioritized and output as the final word search result.

Therefore, according to the present invention, the same effects as the third and fourth inventions described above can be achieved.

In the fifteenth invention, the input pattern read by the pattern recognition means or the pattern reading means is first recognized, and then the keyword creation means creates a keyword for word search from the recognition result of the pattern recognition means. Then, the word search means calculates the cumulative total of evaluation points for each character in the keyword created by the keyword creation means and each word stored in the word storage means, based on the number of characters in the search word and the number of keyword characters, whichever is larger. Divide by a number and output the higher value as the final word search result.

Therefore, according to the present invention, effects equivalent to those of the fifth invention described above can be obtained.

(Example) Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a postal code reading device that employs a pattern recognition device according to an embodiment of the present invention.

As shown in the figure, this postal code reading device includes a reading unit 1 that optically scans a postal letter P on which an address, a postal code, etc. are recorded, and reads a character pattern recognition mechanism on the surface of the letter P.
1, and character recognition that performs character recognition while calculating evaluation points indicating the degree of similarity between the input character pattern from the reading amplification unit 11 and the standard pattern at a plurality of preset slice levels related to the shading of characters, etc. 12, a word dictionary 13 in which words consisting of character strings are stored in advance, and character candidates with high evaluation points for each character are selected from the character candidate group and word delimiter information obtained by the character recognition unit 12, and these a word search unit 14 that combines character candidates to create a keyword consisting of a character string, and performs a word search in the word dictionary 13 based on this keyword;
Address database 1 where code data for each address is saved
5, and a database search unit 16 that searches the address database 15 based on the words obtained by the word search unit 14, forming a pattern recognition mechanism.

Furthermore, this embodiment apparatus includes a stacker section 18 for storing the postal letters P read by the reading section 11 in a journal for each address code, and a letter conveying section for conveying the read postal letters P to the stacker section 18. 19, and a letter conveyance unit 19 to store the postal letter P being conveyed in the corresponding stacker based on the address code output from the database search unit 16.
The pattern recognition mechanism described above is additionally equipped with a conveyance control section 17 for controlling the pattern recognition mechanism.

Next, the operation of this embodiment apparatus will be explained using the flowchart of FIG.

First, the reading unit 11 optically reads the character pattern image on the side of the postal letter P, and converts the pattern image data into A.
/D conversion to the character recognition unit 12 (step 201).

Thereafter, the character recognition unit 12 calculates an evaluation score of 1'', which indicates the degree of similarity between the input character pattern and the standard pattern, at a plurality of preset slice levels related to the shading of the characters, etc. While doing so, each character is recognized, and the recognition results, the first to fourth character code data and word delimiter information, are sent to the word search unit 14 (step 2).
02).

FIG. 3 shows the first to fourth character candidates and the evaluation points assigned to each candidate, which are the character recognition results when the word character string rROO, MJ is written on the postal letter P.

In this figure, the first group of patterns is the character candidates obtained at the normal evaluation level and their evaluation scores, and the second group of patterns is the character candidates and their evaluation scores obtained at the evaluation level assuming the case where the characters are thin. It shows.

After character recognition, the word search unit 14 preferentially extracts character candidates with high evaluation points for each character from the character candidate group obtained in this way, and extracts words based on the word delimiter information. and create keywords for word search (step 203).

At this time, if the same character candidate exists as the first candidate, only the higher evaluation score is left and used for subsequent processing.

FIG. 4 shows first to fourth character string candidates created from the recognition results shown in FIG. 3.

As shown in this figure, as a result of creating character string candidates from the character candidate group shown in Figure 3, here rRO is the first candidate.
The character string ONJ is obtained.

Thereafter, the word search unit 14 uses this character string as the ox-word for the tli word search, and uses this keyword and the word dictionary @1.
Step 2: Match the words in 3 every 3 character strings.
04) The matching words are sequentially extracted from the word dictionary 13 and stored in a predetermined buffer 7 memory (step 205).
.

Note that this word search is generally called a trigram search.

That is, for the character string rROONJ, rR
A search is performed for words containing four types of three-character strings: OOJ roONJ rONRJ rNROJ.

When matching with all the words stored in the word dictionary 13 is completed (step 206), the word search unit 14 then obtains evaluation values for all the buffered words (step 207).

This word evaluation value is obtained by dividing the cumulative total of evaluation points for each character by the number of characters in the word.

For example, if the word rROOMJ is buffered as a word candidate, the word evaluation value for this word is calculated as (90+98+98+88) / 4-92111E
It becomes X.

In this way, the word search unit 14 calculates word evaluation values for all the words extracted from the word dictionary 13, and then outputs the top eight words with the highest word evaluation values as search results ( Step 208).

The word candidates selected in this way are sent to the database search unit 16, and the database search unit 16 searches the address database 15 using each word ■D code corresponding to each word candidate as a keyword (step 2).
09).

Thereafter, the address code output as a search result of the address database 15 is output to the conveyance control section 17, and subsequently, the conveyance control section 17 outputs a stacker control signal corresponding to the address code to the stacker section 18. .

The stacker section 18 stores the postal letter P conveyed by the letter conveyance section 19 in the corresponding stacker based on this stacker control signal (step 210).

Thus, according to the embodiment device according to the first invention, evaluation points indicating the similarity between the input character pattern and the Fj4Q pattern are calculated at a plurality of preset slice levels related to the shading of characters, etc. Since each character is recognized and keywords for word searches are created by collecting those with high evaluation points from the character candidate group for each character, the characters written on the postal letter P are Human character patterns can be accurately recognized regardless of their shading.

Next, a second embodiment will be explained.

The postal code reading device of this embodiment is the same as the device of the above-described embodiment except that it includes a word search section 14 having a new word search function, and other components are the same as those in FIG.

That is, the word search unit 14 in this embodiment device creates a keyword for word search from the character candidate group obtained from the character recognition unit 12, and searches this keyword and words in the word dictionary 13 for each predetermined number of character strings. It is configured so that the words with the highest number of matching matches are prioritized and output as the final search results.

The operation of this embodiment will be explained using the flowchart shown in FIG.

First, the reading section 11 optically reads character pattern images on two sides of a postal letter, and transfers the data to the character recognition section 12 via A/D conversion (step 501).

The character recognition unit 12 recognizes each character while calculating evaluation points indicating the degree of similarity between the manually generated character pattern and the landmark pattern using, for example, the above-mentioned composite similarity method using a plurality of evaluation levels. The character code data and word delimiter information for the first to fourth candidates, which are the recognition results, are sent to the word search unit 14 (step 502).

Thereafter, the word search unit 14 creates keywords from the first to fourth candidates for word search from the character candidate group obtained by the character recognition unit 12 and the word delimiter information (non-chip 503).

Next, the word search unit 14 matches the first candidate keyword and the words in the word dictionary 13 for every three character strings (Step 504), sequentially extracts all the matching words from the word dictionary 13, and searches them in a predetermined manner. The data is stored in the data buffer (step 505), and the number of matches for each word is counted using a count buffer (step 506).

Note that the words in the word dictionary 13 to be searched here are limited to words whose number of characters is within a range of ±1 to the number of characters of the keyword.

Here, for example, if the character string rANGLEsJ is obtained as a keyword for word search, if every three character strings from this keyword are matched with words in the word dictionary 13, 1. as shown in FIG. For discussion, the keywords rANGJ rNGLJrGLEJ rLE
For example, each of the three character strings sJ rEsAJ rSANJ matches the words rANGELESJ and rBNGELESJ four times, and the word exactly equal to the keyword matches six times.

That is, the number of matches counted in this way indicates the degree of matching between the keyword and each word buffered based on the keyword.

In this way, the word search unit 14 performs matching with all words in the word dictionary 13 and counts the number of matches for each word (step 507), and then evaluates each word candidate based on the number of matches (step 507). Step 508), extract the top predetermined number of words from among these word candidates in descending order of the number of matches and convert them into word ID codes;
These are output to the database search unit 16 (step 509).

After this, in the database search unit 16, each word ID
The address database 15 is searched using the code as a keyword (step 510).

The address code output as a search result of the address database 15 is output to the conveyance control section 17, and subsequently, the conveyance control section 17 outputs a stacker control signal corresponding to the address code to the stacker section 18.

Based on this stacker control signal, the stacker section 18 stores the postal letter P conveyed by one letter conveyance section in the corresponding stacker (step 511).

Thus, according to the second embodiment, the keyword for word search and the words in the word dictionary 13 are matched every three character strings, and the word candidates with the highest number of matches are used as the final search word. Since priority is given to output, word dictionary 1
Word candidates with a higher degree of matching can be further narrowed down and extracted from among the word candidates with a higher degree of matching, thereby greatly improving the data search efficiency for the address database 15.

Next, a third embodiment will be described.

This embodiment device is the same as the second embodiment device described above, and further includes a word search unit 14 having a new word search function, and the other components are as shown in FIG. .

Sunaicchiko's il1 word search unit 14 matches keywords and words in the word dictionary 13 for each predetermined number of character strings, and further matches each word candidate with the highest number of matches for each keyword and each character. Then, the word with the highest number of matching characters is output with priority as the final search word.

Next, the flow of word search in this embodiment device will be explained using the flowchart of FIG.

Note that the processing up to step 709 in this flowchart is the same as the processing up to step 509 in FIG. 5, so a description thereof will be omitted.

In step 709, word candidates with a high number of matches are extracted from the word dictionary 13 based on the results of matching every three character strings between keywords and words in the word dictionary 13. In order to select a word candidate with a higher degree of matching from among the word candidates, the keyword and each word candidate are matched character by character (step 7).
10).

After this, the word search unit 14 uses the results of this matching to
- Word candidates up to a predetermined number with the highest number of characters are left as keywords for database search, and other word candidates are cut off (step 711).
).

For example, when the character string rsUITEJ is written on the postal letter P, if an incorrect character such as rSULTEJ is obtained as a keyword, the word dictionary 1 for every 3 character strings
As a result of matching with word 3, the number of matches of 2 becomes the maximum, and a large number of words such as rsUITEJ, rULTERAJ, rRALTESJ, etc. are extracted as word candidates.

When these word candidates and keywords are further matched character by character, rULTERAJ rRALTE
The result is that not a single character matches words such as SJ, but four characters match the word rsUITEJ.

Therefore, in this case, a predetermined number of uppermost words including the word rsUITEJ are finally outputted from the word index section 14 as keywords for database search.

Thus, according to this third embodiment, word candidates with a higher degree of matching can be narrowed down as the final search words from among the word candidates obtained from the trigram search, thereby saving extra search time in subsequent database searches. Can be shortened.

In this example, we have described a device that embeds word candidates with a higher degree of matching from word candidates obtained through a trigram search. The final optimal word may be selected by directly comparing each character with the words in the word dictionary 13.

Next, a fourth embodiment will be described.

This embodiment device is the same as the device of the second embodiment described above, except that it is equipped with a word search section 14 having a new word evaluation value calculation function, and other components are the same as those in FIG. .

That is, the word search unit 14 in this embodiment divides the cumulative evaluation score for each character in the word candidates by the number of word characters or the number of keyword characters, whichever is larger, and selects the word candidate with the higher value as the final word candidate. It is configured to be output as a search word with priority to the database search unit 16.

Next, the flow of word evaluation value calculation in this embodiment device will be explained.

First, a plurality of word candidates are extracted from the word dictionary 13 by trigram search for every three character strings between the keyword and the word in the word dictionary 13.

Thereafter, each word evaluation value is determined from the cumulative total of evaluation points for each character in each word candidate.

The method for obtaining this word evaluation value is as follows.

For example, if the word candidate rROOMJ is extracted from the word dictionary 13 for the keyword [R00NJ, the number of characters in both words is the same, so divide the cumulative evaluation score of each character of this word by the number of characters, and divide this by the number of characters of the word candidate. Use word evaluation value.

That is, in this case, (90+98+98+88) / 4-9211EX
becomes.

Also, if “ROOMsJ” is extracted as a word candidate,
Since the number of characters is different between the two, the word evaluation value is determined as follows.

First, the keyword rROONJ and the word candidate rROOMSJ
Find the evaluation score between each character by matching the first characters of each.

in this case, R: R,, -9011EX O:　O,,-98111EX O: O-, -9811EX N: M - = 88 HEX becomes.

Next, the evaluation points of the following characters are sequentially added starting from the first evaluation point. In this case, add the shorter character length.

As a result, R: R-9011EX/90112XO: 0-98
111EX-128) IEXO: 0 ・9811EX
-1cOIIEXN: M-8811EX-248H
EX Next keyword rROONJ and word candidate rROOM
Find the evaluation score for each letter by combining the last letters of SJ.

in this case, R: 0-5011EX O:　O,,-9811EX O:　M　,,-501[EX N: S -= 50111EX becomes.

Next, starting from the last evaluation point, the evaluation points of the preceding characters are sequentially added.

As a result, R: 0-50112X-18811EXO: 0-9
811EX-138HEXO: M-5011EX-A
OIIEXN: 5-50 HEX-5011EX.

Next, among the values obtained by these two methods, the values obtained by combining the first values are shifted backward by one column, and
Each value of the side obtained by combining these with the last letter and 1
Find the sum of each in a one-to-one correspondence.

That is, 188 HEX -188HEX 90112X+ 138 t(EX -1cII HE
X128 HEX + AOIIEX-1c811EX
IcQ IIEX + 50! IEX-21011
EX241i 11EX-2411112
It becomes X.

Then, the maximum value (248+1EX) is taken as the cumulative total of evaluation points.

To calculate the word evaluation value for this cumulative total, first rROOM
Divide the cumulative total by the number of characters in sJ, and add A to this value as a correction point.
This is done by adding IIEX.

Therefore, the word evaluation value in this case is (2411EX
+50 HEX) / 5+A HEX -7E I
Becomes IEX.

Thereafter, this word evaluation value is compared with the evaluation value (92) IEX) calculated for rROOMJ described above, and the word candidate with the higher value is determined as the final search word.

The method for calculating the word evaluation value described above can be summarized as follows.

■Candidate character length - In the case of keyword length, evaluation value - (sum of evaluation points for each character) / (character length) * 1st
If there is no character candidate among the ~4 candidates, a default value (5011EX) is given.

■ Candidate character length - keyword length ± 1 case * 1 Evaluation value - 3um (candidate character length, keyword length) / W
ax (candidate character length, keyword length) + correction point (^1
11 shi
Maximum value when adding n (candidate character length, keyword length) and adding the cumulative total from the front by shifting it one character backwards + Cumulative evaluation points from the rear ＊＊＊＊＊＊＊＊＊ *** Sum (candidate character length, keyword length) - Max P
S (1) (1≦1≦M+1) Psi (1) - Σ 5 (n) - Cumulative evaluation points from the front n+1 PS2 (1) - Σ 5 (n) - Cumulative A of evaluation points from the rear
10 + 1 However, H - old n (Lngl supplement), Log (keyword)) Log (rate) 2 * character length of string 5 (n) - rating point of nth character PS (1) -PSI (1-1) + PS2 (1)
Between 2≦1≦-PS2(1) I-1
-Psi(1) l-M+1 Thus, according to this embodiment device, the cumulative total of evaluation points for each character of a word candidate is divided by the number of keyword characters or the number of word characters, whichever is larger, and this is used as the word evaluation value. This makes it possible to stably obtain accurate word evaluation values for each word candidate, and to obtain accurate word search results.

Note that the word search in this embodiment is not limited to a device that performs a trigram search, but can also be used in combination with other devices, such as the first embodiment.

Next, a fifth embodiment will be described.

In the postal code reading device of this embodiment, the word dictionary 13 has the configuration shown in FIG. The word search unit 14 identifies the number of characters and character types of the character strings constituting the keyword, and determines the corresponding type of word table in the word dictionary 13 based on these identification results. The system is configured to perform word searches based on keywords in the word table.

In addition, in the trigram search word table mentioned above, one group is made up of words that contain the same three-character string, and the words are further divided into groups according to the position of the word in which the three-character string part is included. has been done.

Next, word search processing in this embodiment will be explained using the flowchart of FIG.

First, when the character recognition unit 12 manually inputs character candidate groups and word delimiter information, which are the recognition results for each character, the word search unit 14
extracts the character code with the highest evaluation score for each character from these character candidate groups, and further creates a keyword for one word based on the word delimiter information (step 801).
.

FIG. 9 shows an example of the character recognition results.

As shown in this figure, the character codes output from the character recognition unit 12 are of four types: numbers, alphabetic characters, lowercase English characters, and special characters, and the candidates for each character are the first
It consists of the fourth candidate. Each of these character candidates has an evaluation score that indicates the degree of similarity to the standard pattern.

Each keyword created from these character candidates is a trigram search word consisting of an alphabetic string of 4 or more characters, a direct word consisting of an alphabetic string of 3 or less characters, and a numeric word consisting of a numeric string or ordinal number of 6 digits or less. , and other special words.

After creating the keyword, the tli word search unit 14 identifies the number and type of characters from this keyword, and determines whether this keyword belongs to the above-mentioned trigram search word, direct word, number word, or special word. do(
Step 802).

If it is determined that the keyword is a trigram search word (step 803), a trigram search for the word is performed in the trigram search word table in the word dictionary 13 (step 8o4).

In addition, in the trigram search here, first, all word candidates where the 3-character string part that matches the keyword and the word exists at the same position in each character string 9 are extracted from the trigram search word table, and then the 3-character strings that match are compared. Shift the position of the column one by one in the keyword and match each word candidate, count the number of matches for each word,
This is done by obtaining the word with the largest number of matches as the first candidate.

If it is determined in step 805 that the keyword is a direct word, a word search is performed in the direct word table in the word dictionary 13 (step 806).

In addition, since this direct word is an alphabetic string of 3 characters or less,
The number of types of character strings that can be constructed is a total of 27 cubed. From this, the word ID code to be obtained as a search result can be directly calculated from the 27 types of alphabetic codes. Therefore, in the direct word search, it is checked whether this word ID code exists in the direct word table, and if it exists, the word ID code is output as a word candidate.

If it is determined in step 807 that the keyword is a numeric word (step 8o6), the numeric word table in the word dictionary 13 is searched for this numeric word (step 8o7).

Note that the word ID code to be obtained as a search result for this numerical word can also be directly calculated from the numerical codes from rOJ to "9".

Trigram search words, direct words, and numeric words are searched in the above manner.

On the other hand, since the frequency of appearance of special words is low, searches for special words are performed at once after all of the above-mentioned search processes for each word have been completed, as in steps 809-812.

This special fit word has an ASCII character code.
They are registered in the special word table along with tit word ID codes in ascending order by code, and searches are performed using binary search.

Thus, according to this embodiment, the word table to be searched in the word dictionary 13 is determined according to the number of characters, character type, etc. of the character string constituting the keyword, and the word according to each word table is determined. Since the search is performed, the word search speed can be effectively increased significantly.

In addition, special words that appear with low frequency are searched at once after completing searches for other types of words with high frequency of appearance, so when processing is performed within a limited time, it is possible to search for important words. It is possible to reduce the amount of unprocessed items.

In addition, in this embodiment, if there is a bias in the types of search words such as trigram search words for special words, words and numeric words for direct i1, etc., other types of words are also prioritized as necessary. It is also possible to perform a word search for each type according to its priority order.

Furthermore, although the above embodiments have been described regarding the recognition of character patterns, the present invention is not limited thereto, and can be similarly used, for example, in an apparatus that recognizes patterns manually inputted by voice or the like.

Although several embodiments have been described above, each of these embodiments can be used in various combinations.

[Effects of the Invention] As explained above, according to the pattern recognition device of the present invention, it is possible to effectively improve pattern recognition accuracy, word search accuracy, and further improve word search efficiency, thereby significantly increasing reliability. Can be done.

[Brief explanation of the drawing]

Fig. 1 is a block diagram for explaining the configuration of a postal code reading device that employs the pattern recognition device of the first embodiment, and Fig. 2 shows the flow of pattern recognition processing in the postal code reading device of Fig. 1. Flowchart, FIG. 3 is a diagram showing examples of character complement groups and evaluation points obtained by the character recognition section in FIG. 1, FIG. 4 is a diagram showing keywords created from the character recognition results in FIG. 3, Figure 5 is a flowchart showing the flow of word search processing in the second embodiment, Figure 6 is a diagram showing the results of trigram search in that embodiment, and Figure 7 is the flow of word search processing in the third embodiment. 8 is a flowchart showing the flow of word search processing in the fourth embodiment, FIG. 9 is a diagram showing the character recognition results in that embodiment, and FIG. 10 is a conventional postal code reading method. 1 is a block diagram for explaining the configuration of the device. 11...reading unit, 12...character recognition unit, 13...
Word dictionary, 14...word sharp teeth section. Applicant: Toshiba Corporation

Claims (15)

[Claims] (1) A pattern reading means for reading an input pattern, and evaluation points indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern are calculated for each of a plurality of levels, and recognition of the input pattern is performed for each of the plurality of levels. What is claimed is: 1. A pattern recognition device comprising: pattern recognition means for performing the pattern recognition; and optimal pattern selection means for selecting the optimal pattern based on evaluation points for each level calculated by the pattern recognition means. (2) A pattern reading means for reading an input pattern, and evaluation points indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern are calculated for each of a plurality of levels, and recognition of the input pattern is performed for each of the plurality of levels. a word storage means in which words are stored in advance, and a keyword for word search by preferentially collecting characters with high evaluation scores for each character from the recognition results of the pattern recognition means. , word search means for performing a word search on the word storage means based on the keyword, a pattern recognition device characterized in that it comprises: (3) pattern reading means for reading an input pattern, pattern recognition means for recognizing the input pattern read by the pattern reading means, word storage means in which words are stored in advance, and words from the recognition results of the pattern recognition means. Creating a keyword for search, comparing this keyword with the words stored in the word storage means for each predetermined number of character strings,
A pattern recognition device comprising: word search means for preferentially outputting a word with a high number of matching matches as a final word search result; (4) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern read by the pattern reading means; a word storage means in which words are stored in advance; and a word from the recognition result of the pattern recognition means. a first word search means that creates a keyword for search and searches the word storage means based on the keyword; and a first word search means that searches the word storage means based on the keyword; 2. A pattern recognition device comprising: a second word search means for comparing each of the words with a large number of matching characters, and outputting a word having a large number of matching characters with priority as a final word search result. (5) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern by calculating an evaluation point indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern; a word storage means storing words; a first word search means for creating a keyword for word search from the recognition result of the pattern recognition means and searching the word storage means based on the keyword; The cumulative total of evaluation points for each character in each word searched by the word search means 1 is divided by the larger number of search word characters and keyword characters, and the higher value is used as the final word search. A pattern recognition device comprising: a second word search means for outputting results with priority; (6) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern by calculating an evaluation point indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern; A word storage means in which words are stored, and a keyword for word search is created from the recognition result of the pattern recognition means, and this keyword is matched with the words stored in the word storage means for each predetermined number of character strings. do,
a first word search means that prioritizes and outputs words with a high number of matching matches; and a first word search means that compares each word output from the first word search means with the keyword for each character, and the number of matched characters. 2. A pattern recognition device comprising: a second word search means for preferentially outputting words with a large number of words as final word search results. (7) pattern reading means for reading an input pattern, pattern recognition means for recognizing the input pattern read by the pattern reading means, word storage means in which words are stored in advance, and words from the recognition results of the pattern recognition means. Creating a keyword for search, comparing this keyword with the words stored in the word storage means for each predetermined number of character strings,
A first word search means that prioritizes and outputs words with a high number of matching matches; and a first word search means that outputs words with a high number of matches, and calculates the total number of evaluation points for each character in each word searched by the first word search means, and calculates the number of characters in the search word. and a second word search means that divides the keyword by the larger number of characters and outputs the higher value as the final word search result with priority. (8) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern by calculating an evaluation point indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern; A word storage means in which words are stored, and a keyword for word search is created from the recognition result of the pattern recognition means, and this keyword is matched with the words stored in the word storage means for each predetermined number of character strings. do,
a first word search means that prioritizes and outputs words with a high number of matching matches; and a first word search means that compares each word output from the first word search means with the keyword for each character, and the number of matched characters. The second option that prioritizes and outputs words with a large number of
The cumulative total of the evaluation points for each character in each word searched by the second word search means and the second word search means is divided by the larger number of search word characters and keyword characters, and the value is calculated as follows: A pattern recognition device comprising: a third word search means for outputting higher word search results with priority given to higher word search results. (9) A pattern reading means for reading an input pattern, and evaluation points indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern are calculated for each of a plurality of levels, and recognition of the input pattern is performed for each of the plurality of levels. a word storage means in which words are stored in advance, a keyword for word search is created from the recognition result of the pattern recognition means, and this keyword and the word stored in the word storage means are predetermined. Match each string of numbers,
a first word search means that prioritizes and outputs words with a high number of matching matches; and a first word search means that compares each word output from the first word search means with the keyword for each character, and the number of matched characters. The second option that prioritizes and outputs words with a large number of
The cumulative total of the evaluation points for each character in each word searched by the second word search means and the second word search means is divided by the larger number of search word characters and keyword characters, and the value is calculated as follows: A pattern recognition device comprising: a third word search means for outputting higher word search results with priority given to higher word search results. (10) an image input means for inputting an image consisting of a plurality of characters and outputting an image signal; and inputting the image signal and performing appropriate image processing corresponding to each of the character images; An image processing means that outputs a plurality of character image patterns for each character, and an evaluation point indicating the degree of similarity between each of the plurality of character image patterns and a standard pattern prepared in advance. a pattern recognition means for recognizing each character based on the recognition results; a word storage means in which words made up of character strings are stored in advance; and a character candidate with a high evaluation score for each character based on the recognition result of the pattern recognition means. A pattern recognition device comprising: word search means for selecting a character string, combining the character candidates to generate a character string, and performing a word search on the word search means using the character string as a keyword. (11) Image input means for inputting an image consisting of a plurality of characters and outputting an image signal; and inputting the image signal and binarizing it at a plurality of preset slice levels for each of the character images. By doing so, the image processing means outputs a plurality of character image patterns for each of the characters, and an evaluation score indicating the degree of similarity between each of the plurality of character image patterns and a standard pattern prepared in advance is calculated. , a pattern recognition means that recognizes each character based on the evaluation score; a word storage means in which words consisting of character strings are stored in advance; and a pattern recognition means that recognizes the evaluation score for each character based on the recognition result of the pattern recognition means. a word search means that selects character candidates with a high value, generates a character string by combining these character candidates, and performs a word search on the word search means using this character string as a keyword; recognition device. (12) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern read by the pattern reading means; a word storage means in which words are stored in advance in groups according to predetermined types; A keyword for word search is created from the recognition result of the pattern recognition means, the type of character string that constitutes this keyword is identified, and based on the identification result, a word is created for the corresponding group of the word storage means. A pattern recognition device comprising: word search means for performing a search. (13) a pattern reading means for reading an input pattern; a pattern recognition means for recognizing the input pattern read by the pattern reading means; a word storage means for storing words grouped in advance by predetermined types; Create a keyword for word search from the recognition result of the pattern recognition means, identify the type of character string that constitutes this keyword, and target the corresponding group of the word storage means based on the identification result. A pattern recognition device comprising: word search means for performing a word search according to a predefined priority order for each word type. (14) pattern reading means for reading an input pattern; pattern recognition means for recognizing the input pattern read by the pattern reading means; word storage means in which words are stored in advance; and words from the recognition results of the pattern recognition means. a keyword creation means for creating a keyword for a search, and a keyword created by the keyword creation means and each word stored in the word storage means, each character-by-character compared, and a word with a large number of matched characters. A pattern recognition device comprising: a word search means for preferentially outputting a word search result as a final word search result; (15) pattern reading means for reading an input pattern; pattern recognition means for recognizing the input pattern by calculating an evaluation point indicating the degree of similarity between the input pattern read by the pattern reading means and a standard pattern; a word storage means storing words; a keyword creation means for creating a keyword for word search from the recognition result of the pattern recognition means; and a keyword created by the keyword creation means and the keyword stored in the word storage means. The cumulative total of evaluation points for each character in each word is divided by the larger number of search word characters and keyword characters, and the word with the higher value is prioritized and output as the final word search result. A pattern recognition device comprising a search means.