JPH02264388A - Character recognition post-processing device - Google Patents

Abstract

PURPOSE:To accurately and rapidly estimate a correct character by associating a word from a recognition candidate character, comparing the respective characters of the word with the recognition candidate character, and estimating the correct character from the recognition candidate character. CONSTITUTION:A retrieve part 2 associates the word having a certain character in any position, and retrieves the word as a retrieve key in the sequence form the candidate character at high recognition priority. Next by comparing each character included in the word with the recognition candidate character by a collating part 4, the likelihood of each word is calculated, and based on the likelihood of each word, the correct character is estimated by a collating part 5. Thus even when the correct character is not included in the recognition candidate character in the local character position, the corresponding word is associated and retrieved, and by not comparing the all combinations of the recognition candidate characters but only comparing the character included in the retrieved word with the recognition candidate character in the corresponding position, the correct character can be accurately and rapidly estimated.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a character recognition post-processing device that estimates a correct character by referring to a word dictionary from recognition candidate characters obtained as output from a character recognition device. be.

[Prior Art] Conventional devices of this type have adopted the following method. In other words, in the method of matching words and recognition candidate characters,
All combinations of recognition candidate characters were covered, and a combination of recognition candidate characters with the highest degree of matching was found among the combinations. Therefore, when estimating the correct character from a recognition result with a length of N characters and M recognition candidate characters at each character position, it is necessary to consider 811 combinations.
If N=M=10, it was necessary to match the words by assuming a huge number of combinations, 1010, so there was a problem in that the matching took a long time. Furthermore, in the method of searching for a word from recognition candidate characters, a method has been used in which a word is searched starting from the matching start character position. For this reason, if the correct character does not exist among the recognition candidate characters at the matching start character position, the matching between the searched word and the recognition candidate characters will always fail, and the ability to estimate the correct character will be significantly reduced. There was a problem.

[Problems to be Solved by the Invention] An object of the present invention is to provide post-character recognition processing that improves correct character estimation ability and shortens processing time in a method of searching for words from recognition candidate characters and matching the words with recognition candidate characters. It is about providing the means.

[Means for Solving the Problems] The present invention includes a search unit that associates a word with a certain character in any character position and searches for a word as a search key in order of candidate characters having a high recognition rank; The most important feature is that it has a matching unit that calculates the likelihood of each word by comparing each character in the list with recognition candidate characters, and a processing unit that estimates the correct character from the likelihood of each word.

[Operation] The present invention searches for words by associating words even if the correct character is not necessarily included in the recognition candidate characters at some character positions, and searches for the correct character without examining all combinations of recognition candidate characters. It acts to estimate .

[Example] FIG. 1 is a detailed diagram of the present invention, in which 1 is a memory that stores recognition candidate characters, 2 is a search unit that searches for a corresponding word using the recognition candidate characters as a search key, and 3 is a word. 4 is a collation unit that matches and collates the word obtained by the search unit 2 with the recognition candidate characters stored in the memo + 71; 5 is a collation unit that receives the results of the collation unit 4 and estimates the most likely word. A processing unit 6 is a memory that stores the verification results.

To operate this, candidate characters are first input into memory l from an external character recognition device. Memory 1 has a string length of N
If the candidate character corresponding to each character is M character, (
It can be configured with a memory capacity of NXMX2) bytes.

Next, the search unit 2 operates. FIG. 2 is a diagram explaining the operation of the search unit 2, in which 20 is an example of recognition candidate characters for the character string "Yokosuka City, Kanagawa Prefecture", 21 and 22 are matching start positions, 23 and 2
4 is the operating range. In FIG. 2, the search unit 2 sets 5=11 in the character string direction, that is, the character position, to the matching start position 21, and sets the test range n in the character string direction and the test range m in the recognition ranking direction ( In the example in Figure 2, n=
3, m=5). Next, among the candidate characters included in the inspection range 23, the character “Ki” with recognition rank R=1,
Words are searched from the dictionary section 3 using "na" and "su" as search keys and sent to the matching section 4.

FIG. 3 is a diagram showing the construction method of the dictionary, where 30 is a heading section and 31 is a word notation section. From Figure 3, word notation section 31
Since the structure is such that words containing the characters in the heading part 30 are accommodated, the heading W
It is easy to understand that words can be associated with the letters &30. Although this dictionary structure increases the capacity, there is a rule of thumb that the average character string length of a word is about 2 to 3 characters, so the degree of increase in capacity is at most 2 to 3 times.

When the collation unit 4 receives a word from the search unit 2, it collates the word with candidate characters stored in the memory 1, and calculates the likelihood of each word. FIG. 4 is a diagram showing a search word and an example of calculating the likelihood. 40 is an example of a word searched for "Kore", 41 is an example of a word searched for "NA", and in this example, It is assumed that the word searched for "su" is empty. The detailed operation of FIG. 4 will be explained below.

The matching unit 4 calculates Wt, W for ranks 1 to 5 of recognition candidates.
3. W3. w4. Set Ws. In this embodiment, a weight of W1=2'-' is assigned to the rank i. That is, w,=16. W, = 3°W3:4. W,=2.
It is assumed that wl=l.

First, attention is paid to the word 40 that was searched for "tori", and the beginning of the word is aligned with the matching start position 21 in FIG. Regarding the words "catcher" and "whaling", the second letter "hand"
Since there is no matching character among the recognition candidate characters for the corresponding character position, the likelihoods of these characters are K=w, -w, =Q
becomes. Similarly, for the words "whaler J" and "arrest", the respective likelihoods are K=-16 and K=-32.

Next, by focusing on the word 41 searched for "Nara" and performing the same operation as above, "Nara", "Kanagawa",
The likelihood of "Kanagawa system" is = -32, K = 8, respectively.
, K=24. Note that the weight of each rank is w, ≧W2
...≧W@>0 (m is the number of candidate rankings) Any numerical value may be used as long as it satisfies the relationship, and the value can be changed depending on the characteristics of the character recognition device.

The matching unit 4 notifies the processing unit 5 of each word and the calculation results of their respective likelihoods, and the processing unit 5 operates. The processing unit 5 detects a word having a positive value and the maximum likelihood among the words received from the matching unit 4. At this time, if a word with a positive maximum likelihood exists, that word is written into the memory 6 as a definite word. In the example of FIG. 4, the word "Kanagawa-kei" is detected.

Since the word "Kanagawa-kei・" has a length of 4 characters, the position in FIG. 2 is updated by 4 characters from the matching start position 21 (that is, as S←S+4) and a new matching start position 22 is searched. The search unit 2 is notified and the search unit repeats the same operation as described above. On the other hand, if the matching unit 4 is unable to detect a word with a positive likelihood, it updates the recognition ranking by 'l' (that is, as R-R+1) and sends a new Request a word search. With the above repeated operations,
If R=5 and a word with a positive maximum likelihood cannot be detected, the first candidate character at the matching start position is written to the memory 6 as an unknown word, and the matching start position is
The CPU 1 updates the recognition order by l' (ie, S←S+1) and sets the recognition rank to the initial value (ie, R-1), and requests the search unit 2 to search for a new word. The above operation is
The process ends when the processing unit 5 writes a character string with the same length as the character string input into the memory 1 into the memory 6.

As is clear from the above operations, compared to the conventional technology,
Even if the correct character is not included in the recognition candidate characters at some character positions, you can search by association with the corresponding word, and it corresponds to the characters included in the searched word rather than all combinations of recognition candidate characters. An improvement is that the correct character can be estimated simply by comparing the recognition candidate characters at the position.

[Effect of the invention] As explained above, words can be associated with recognition candidate characters,
The correct character is estimated from the recognition candidate characters by comparing each character of the word with the recognition candidate characters, so even if the correct character is missing at some character positions (especially at the beginning of the word), the correct answer is accurate and fast. It has the advantage of being able to estimate characters.

[Brief explanation of drawings]

FIG. 1 is a detailed explanatory diagram of the present invention, FIG. 2 is an explanatory diagram of the operation of the search section 2, FIG. 3 is an explanatory diagram of the configuration of the dictionary section 3, and FIG. 4 is an explanatory diagram of the operation of the collation section 4. . 0... Recognition candidate character example 1122... Verification start position, 3.24... Search range 0... Heading part, 1... Notation part, 0.41... Word and likelihood calculation example .

Claims (1)

[Claims] In a character recognition post-processing device that estimates a correct character from a character string that is N characters long and has M recognition candidate characters for each character position, n(n
≦N) A search unit that searches for a word containing the candidate character for each candidate character from the first rank to the mth rank (m≦M) among the candidate characters up to the character position; For each rank, w_1≧w_2≧
...Set a positive value that satisfies the relationship ≧w_m>0,
For each word searched by the search unit, if there is a candidate character that matches the character in the i-th rank of the same character position for each character included in the word, a positive value w_i is added to find the matching candidate. If the character does not exist -w_i
1. A character recognition post-processing device comprising: a matching unit that calculates the likelihood of the word by adding a negative value; and a processing unit that estimates a correct character from the likelihood obtained by the matching unit.