JP3115459B2 - Method of constructing and retrieving character recognition dictionary - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for constructing and retrieving a character recognition dictionary for performing post-processing for automatically correcting the result of character recognition.

[0002]

2. Description of the Related Art Conventionally, in character recognition for recognizing handwritten characters or the like by an optical method or the like, post-processing for collating a character recognition result with a word dictionary is performed in order to automatically correct a recognition error. ing. In order to perform such post-processing quickly, it is necessary to quickly search the word dictionary. For this reason, simply searching the word dictionary using the first character as a search key, for example, in the case of an English dictionary, there are many words starting with “a” and few words starting with “x”. As a result, the number of words varies, and an efficient search cannot be performed. Therefore, there is a method of performing a character search by constructing a dictionary capable of searching using a character at a specific position in a word, for example, the character of the second character as a search key (Japanese Patent Publication No. 4-38026).

On the other hand, there is a method in which a word dictionary is divided using a hash method, a hash value is assigned to each divided word group, and a search is performed for a divided word group equal to a hash value obtained as a result of character recognition.

[0004]

However, the above-mentioned prior art has the following problems. That is, even when a character at a specific position of a word is used as a search key, the number of words registered in each search key still varies. Therefore, when words registered for each search key are divided by the same division number for each search key by the hash method, a large difference occurs in the number of words registered for each hash value of each search key. In some cases.

[0005] As a result, when the dictionary is unnecessarily subdivided, the data portion such as the hash value assigned to each divided word group becomes unnecessarily large, and the use efficiency of the memory deteriorates. On the other hand, if the division of the dictionary is too large, there is a problem that a high-speed search cannot be performed even by a search using the hash method.

For example, in the example shown in FIG. 4, when the second character of each word is used as a search key, the number of words registered in the search key "a" is 2,469 words, and the search key "f""
Are 62 words. If it is attempted to divide such a dictionary into 64 words by hash value in order to search for the words registered in the search key "a" sufficiently fast by using the hash value, the words registered in the search key "f" are also divided into 64 words. You have to do it. However, only 62 words are registered for the search key "f", and it is meaningless to divide it into 64 words. For the above reasons, a simple combination of the technique of Japanese Patent Publication No. 4-38026 and the hash method could not search the word dictionary at high speed.

The present invention has been made in view of the above points . A code of a search key and a hash value are added to each word.
Registering at high speed while efficiently using memory.
It is an object of the present invention to provide a method of constructing a character recognition dictionary capable of searching words and a method of searching for the words .

[0008]

According to the method of constructing a character recognition dictionary of the present invention , each character at a predetermined common specific position in each word is used as a search key, and the specific position of each word is determined.
Search for all words whose characters match each search key
And the number of words belonging to each search key
And the number of words is registered for each search key.
-The search key with the larger number of words registered for each
Code corresponding to each search key so that it becomes a short code
And the more words registered for each search key
Of each word belonging to each search key so as to have a long bit length
A hash value is calculated, and a check is performed for each word.
It is characterized in that a search key code and a hash value are added and registered.

[0009] The method of the search character recognition dictionary of the present invention, character
The result of character recognition of the character pattern to be recognized
Of the candidate characters to be
As a search key, all units that match the code of the search key
A sentence to be searched for a word from the character recognition dictionary and subjected to character recognition
By the candidate character obtained as a result of character recognition of the character pattern
Calculate the hash value of the candidate character and use it as the search key code
Of all words searched based on
The hash value of the word and the hash value of the candidate character are equal
For all words, there is a word that matches the candidate character
Whether or not there is a word that matches the candidate character
The candidate character as a recognition result.
Things.

[0010]

In the method for constructing a character recognition dictionary according to the present invention,
For example, as the number of words to search becomes as uniform as possible for the English word dictionary, as well as the search key and the second character of the character of each word, a single belonging to each search key
Count the number of words in a word. And a search key with many words
Has a short bit length and a search key with a small number of words has a small bit length.
Is assigned to each search key. Also,
The bit length of the hash value is increased for search keys having a large number of registered words, and is shortened for search keys having a small number of registered words. This allows each word to be
The bit lengths of the code and the hash value are constant.

In the character recognition dictionary search method of the present invention, when a character pattern recognized by an optical character reader or the like is used to correct the recognition result using the word dictionary configured as described above, the following is performed. Search the dictionary. The word dictionary is searched using the candidate character of the second character pattern as a search key, and the ASCI of the candidate character of each character pattern is searched.
A hash value is obtained from the total value of the I code values in the same manner as described above, and a word group including a word suitable for correcting the character recognition result is searched. If an appropriate word is not searched from the word group, another word group belonging to the same search key is searched.

[0012]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block diagram of an embodiment of an apparatus to which a method for configuring a character recognition dictionary according to the present invention is applied. The illustrated device includes a search key extracting unit 120, a word counting unit 500,
It comprises a division number determination unit 600, a hash value calculation unit 320, a dictionary creation unit 700, and the like. The search key extraction unit 120 extracts a character at a specific character position as a search key for each word to be registered in the word dictionary 420. This specific character position is, for example, the second character in English or the like. The device of FIG. 2 described later recognizes the character pattern at this position,
The word dictionary 420 is searched using the candidate character as a search key.

The word number counting section 500 includes a search key extracting section 1
At 20, the number of words registered in each search key is counted using the search keys extracted for each word. In general, since the appearance rate of characters is biased, the number of words registered by a search key differs considerably. Therefore, in order to divide these words into a word group having a uniform number of words, the word number counting unit 500 counts all words registered for all search keys.

The number-of-division determining unit 600 includes a word number counting unit 50.
According to the number of words registered in each search key obtained by 0,
The number of words to be registered in each search key is determined so that the number of words registered in each hash value of each search key is substantially uniform. Thereby, even if the number of words registered in each search key is significantly different, it is possible to effectively combine a search using a hash value with a search using such a search key. The number of divisions for each search key determined by the number-of-divisions determination unit 600 is used in the division number table 220 for use in searching the word dictionary in the apparatus of FIG.
Is stored in The hash value calculation unit 320 determines whether the word dictionary 4
For each word to be registered in the search key extraction unit 120,
The hash value of the word is determined using a hash value calculation method corresponding to the number of divisions corresponding to the search key extracted in step (1). A specific example of this calculation method will be described later.

The dictionary creation unit 700 uses the search key obtained by the search key extraction unit 120 and the hash value obtained by the hash value calculation unit 320 as a search code to convert each word registered in the word dictionary 420 into a word dictionary. 420. FIG.
The word dictionary configured by the apparatus of FIG. 2 is searched by the apparatus of FIG.

FIG. 2 is a block diagram of an embodiment of an apparatus to which the character recognition dictionary search method according to the present invention is applied. The illustrated device includes a search key extraction unit 110, a division number search unit 210,
It comprises a hash value calculation unit 310, a dictionary search unit 410 and the like. The search key extraction unit 110 extracts, as a search key, a character code (first candidate character) that is a result of character recognition at a specific position in a word from a character recognition result of an optical character reader or the like. When a plurality of characters are listed as candidate characters as a result of character recognition, the first candidate character may be used as a search key first, and the characters subsequent to the second candidate may be used sequentially as search keys.

The number-of-divisions search unit 210 searches the number-of-divisions table 220 for how many words registered in the word dictionary 420 by the search key characters extracted by the search key extraction unit 110 are divided by hash values. Division number table 22
For each word registered in the word dictionary 420, 0 stores the number of divisions indicating how many words registered in the search key are divided by the hash value. The hash value calculation unit 310 calculates a hash value from a character recognition result or the like using a hash value calculation method predetermined so as to realize the number of divisions, based on the number of divisions searched by the number of divisions search unit 210. .
For example, if the division number is “8”, the remainder obtained by adding the character codes (such as ASCII codes) of the first candidate characters of the recognition result of each character of the word and dividing the sum by “8” is used as the hash value. . Such a hash value calculation method may be any other method.

The dictionary search unit 410 searches words registered in the search key obtained by the search key extraction unit 110 from the word dictionary 420 in order from the hash value obtained by the hash value calculation unit 310. As a result, if the character recognition result is correct, the search key and the hash value are both correct, and the word is quickly searched by the word dictionary 420. Therefore, when the character recognition result is correct, the post-processing is performed at high speed.

Next, a specific processing configuration example of the division number determination unit 600 will be described in detail with reference to FIG. As shown in FIG. 3, in the Huffman code assignment processing 610, codes are assigned by the Huffman method using the number of words registered for each search key. The Huffman method is a method of performing data compression by allocating a short code to data having a high appearance rate and allocating a long code to data having a low appearance rate. According to the Huffman method, a code having a short bit length is assigned to a search key having a large number of registered words, and a code having a long bit length is assigned to a search key having a small number of registered words. For example, as shown in FIG. 4, "00" is assigned to the search key "a" having the largest number of registered words.
Is assigned. Also, “110” is assigned to the search key “e” having the second largest number of registered words. And
The search keys “d” and “f” with a small number of registered words include “10”
“111101” and “01111100” are assigned.

In the example of FIG. 4, “00” such as “001” is used.
Are not assigned to any code other than "a".
Therefore, if "00" is detected, "a" can be detected without detecting the subsequent bit. That is,
Even if an arbitrary bit is connected after the code “00” of “a”, only the code portion can be extracted. Although only a part of the search key is shown in FIG. 4, codes (not shown) are similarly assigned to 26 letters of the alphabet, periods, commas, and the like. In the hash value bit length determination processing 620, a value obtained by subtracting the bit length allocated to the code of the search key in the Huffman code allocation processing 610 from the predetermined bit length is set as the bit length of the hash value. Therefore, the bit length of the hash value is
A search key with many registered words becomes longer, and a search key with less words becomes shorter. That is, as the number of words registered in the search key increases, the number of divisions by the hash value increases.

When the code of the search key and the number of divisions by the hash value are determined by the above processing, the sum of the code length of the search key and the bit length of the hash value becomes constant. Therefore, regardless of the number of words registered in the search key, the total bit length of the field occupied by the code of the search key and the hash value in the word dictionary can be made constant.

Next, a specific example of a procedure for configuring a dictionary and a procedure for searching the configured dictionary will be described. In this example, the second character of the word is used as the search key. In calculating the hash value, the remainder obtained by dividing the sum of the ASCII codes of the characters of the word by the number of divisions is used. FIG. 4 shows an example of a process of determining the number of divisions into which words registered in each search key are divided.

First, in the search key extracting unit 120 of FIG. 1, the second character is extracted as a search key for all registered words. Then, in the next word counting unit 500,
For each search key, the number of words from which the search key character is extracted is counted. As shown in FIG. 4, for example, the word whose search key is “a” is 2469 words. The words with the search keys "c" and "d" are 288 words and 99 words, and the other words are as shown in the figure. As can be seen from the illustrated example, the number of words registered by the search key differs, and the difference is quite large.

Next, in a process 610 of FIG. 3, a code is assigned to the character of each search key by the Huffman method using the number of words described above. That is, a short code is assigned to a character having a large number of words, and a long code is assigned to a character having a small number of words. The code determined in this manner has a shorter bit length such as "00" when more words are registered than the other search key of 2469 words such as the search key "a". Also, when a word such as the search key “f” is registered, which is 62 words less than the other search keys, the bit length becomes longer like “01111100”. Then, the bit length of the code assigned to the search key in the process 610 is subtracted from the bit length predetermined in the process 620 to obtain the bit length of the hash value. For example, if the predetermined bit length is “8”, in the case of the search key “a”, the bit length “2” of the code “00” assigned to the search key is subtracted from “8”, and the bit of the hash value is The length is 6 bits, and the number of divisions is “64” as the sixth power of “2”.

Next, in the processing 620 for each search key, divides the words registered in each of the search key by split number determined as described above. Then, as shown in the lowermost column of FIG. 4, the difference in the number of words registered in each hash value of each search key becomes small. For example, for the search keys “a” and “f”, the number of registered words is about 40 times different between “2469” and “62”.
The difference between "9" and "62" is reduced to about 1.6 times, whereby the search by the hash method can be performed effectively, and stable search performance can be realized.

FIG. 5 shows a specific procedure for constructing a word dictionary, taking the word "read" as an example. First, the search key extraction unit 120 extracts “e” as the second character of “read” as a search key. According to FIG. 4, when the search key is "e", the search key code is "110" and the number of divisions by the hash method is "32". Next, the hash value calculation unit 320
ASC of each character of "r", "e", "a", "d"
"412", which is the sum of II codes, is divided by "32"
The remainder “28”, ie, “11100” in binary, is calculated as the hash value. Thus, the word "rea
d ”is a search key code“ 110 ”and a hash value“ 11100
"In the word dictionary 420.

In this example, the word to be searched is divided into 32, and the hash value is represented by 5 bits. The search key code is 3 bits of "110", which is 8 bits in total with the hash value. Thus, the total of the bit lengths of the search key code and the hash value is always 8 bits. Therefore, by synthesizing the search key code and the hash value and using them as one key as a whole, a memory area for the search key code and a memory area for the hash value are secured in consideration of the bit length difference. There is no need. Therefore, the memory efficiency becomes very good.

FIG. 6 shows an example of the dictionary search process. In the dictionary search, the hash value is calculated by performing search key extraction and hash value calculation on character recognition results of an optical character reader or the like, not on registered words. Other than this,
This is the same as the dictionary construction process. Therefore, the dictionary search procedure is the same as the procedure shown in FIG. Hereinafter, this procedure will be described. First, the search key extraction unit 110 searches for the word 2
The first candidate character “e” of the recognition result of the character is extracted as a search key. Next, the division number search unit 210 searches the division number table 220 for a search key code and a division number corresponding to the search key “e”, and respectively retrieves the search key code “110”.
And the division number “32”.

Then, the hash value calculation unit 310 calculates the A of the first candidate characters "r", "e", "a", and "d" of each character.
The binary number “11100” of the remainder “28” obtained by dividing “412” of the sum of the SCII codes by the division number “32” is calculated as a hash value. The search key code "110" thus obtained
First, a search for a word is started from the part of the hash value "11100" of. In this case, as long as the first candidate character obtained as a result of the recognition is correct, “read” of the target word is stored in the portion of the hash value at which the search is first started. As a result, a high-speed search can be performed. On the other hand, when the first candidate characters obtained as the recognition result are incorrect, the word to be searched is not usually stored in the portion of the hash value calculated from the first candidate characters. When there is no word to be searched, words other than the hash value of the search key code are compared with the first candidate character of the recognition result, as in the conventional method.

As described above, when the character recognition result is correct, of the words of the search key obtained by the search key extraction unit 110 of the word dictionary 420, the words of the hash value obtained by the hash value calculation unit 310 Since a word to be searched is found in the word, the word is searched uniformly and at high speed regardless of the number of words registered in the search key code. still,
In the above-described embodiment, the case where the English dictionary is constructed and searched has been described. However, the present invention is not limited to this, and it is needless to say that the present invention can be applied to Japanese, French and other languages.

[0031]

As described above, according to the method of constructing and retrieving a character recognition dictionary of the present invention, a method of performing a word search using a character at a specific position in a word as a search key,
Since the dictionary is configured and searched by combining with the hash method, the words registered in each search key are divided almost uniformly for each hash value, and the number of divisions is too large, and the amount of code data for search is large. Can be prevented from increasing and the use efficiency of the memory can be prevented from deteriorating. Further, it can be prevented that the number of divisions is too small to perform efficient search. Therefore, it is possible to shorten the dictionary search time when the character recognition result is correct. Further, by using a Huffman code or the like as a search key code, it becomes possible to connect a hash value to the search key code and store it in one area,
The use efficiency of the memory can be further improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of an apparatus to which a method for configuring a character recognition dictionary according to the present invention is applied.

FIG. 2 is a block diagram of an embodiment of an apparatus to which the character recognition dictionary search method of the present invention is applied.

FIG. 3 is an explanatory diagram of a division number determination processing procedure.

FIG. 4 is an explanatory diagram of an example of a division number determination process.

FIG. 5 is an explanatory diagram of a dictionary configuration processing example.

FIG. 6 is an explanatory diagram of a dictionary search processing example.

[Explanation of symbols]

 110, 120 search key extraction unit 210 division number search unit 220 division number table 310, 320 hash value calculation unit 410 dictionary search unit 420 word dictionary 600 division number determination unit 700 dictionary creation unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-62-278689 (JP, A) V. Eiho and two others, "Data Processing Series 11 Data Structure and Algorithm" First Edition, Baifukan Co., Ltd., March 10, 1987, p106-p112 and p331-336 (58) Fields surveyed (Int. Cl. ⁷ , DB name ) G06K 9/72 G06F 17/30 412 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A method according to claim 1 , wherein each character at a predetermined specific position in each word is used as a search key, and all words whose characters at the specific position of each word match the search key are used as search keys. , The number of words belonging to each search key is counted, and the number of words is registered for each search key. The search key having a larger number of words registered for each search key has a bit length A code corresponding to each of the search keys is assigned so as to have a shorter code length. A hash of each word belonging to each of the search keys is set such that the longer the number of words registered for each of the search keys is, the longer the bit length becomes. A method for constructing a character recognition dictionary, comprising: calculating a value; adding a code of the search key corresponding to each word and the hash value to each word; 2. A method for retrieving a character recognition dictionary according to claim 1, wherein the character pattern is a character pattern to be recognized. Of these, the predetermined character at the specific position is used as a search key, all words that match the code of the search key are searched from the character recognition dictionary, and the result of character recognition of a character pattern to be subjected to character recognition is obtained. The hash value of the candidate character is calculated based on the candidate character to be obtained, and among all the words searched based on the code of the search key, the hash value of all the searched words is equal to the hash value of the candidate character. Is searched for a word that matches the candidate character, and if there is a word that matches the candidate character, the candidate character is used as the recognition result. A method for searching a character recognition dictionary.