JPH0319590B2 - - Google Patents

Description

[Detailed description of the invention]

TECHNICAL FIELD OF THE INVENTION The present invention relates to a method for correcting misread characters in a character recognition device equipped with a word dictionary. Conventional technology and problems In character recognition devices that optically read handwritten or printed character strings, the recognition results are displayed on a display, and if there are characters that are incorrectly recognized (misread characters), they can be corrected by instructions from the operator. The function to do this is added. A typical conventional method for correcting misread characters is to prepare the recognition results for each character from 1st to Nth, have the operator indicate which character is misread, and then move to the next rank according to these instructions. If it is a misread character, the recognition result of the next rank is displayed, and in the same way, lower rank character candidates are displayed in order and the operator selects the correct character. . The disadvantage of this method is that it does not utilize the relationship between the misread character and the characters before and after it.
For example, if a single character of a word is misread, meaningless character candidates that do not form an existing word even when combined with other characters are output one after another to the operator just because they are listed as candidates. This is because the operator is forced to make selections, which increases the burden on the operator and lengthens the time required for correction. Therefore, the present inventors previously proposed a method for correcting misread characters using a word dictionary. In this method, when the operator specifies a range of words that include misread characters, the recognition device selects only the words that actually exist in the word dictionary from among the word candidates that are formed by combining candidates for each character within the range. This is what is output. Therefore, character candidates that do not constitute a word are not output at all, so the correction work by the operator is simplified and speeded up accordingly. However, in order to put this method into practical use, it is necessary to pay attention to the following points. That is, words are composed of two or three letters, and there are very many two-letter words.
Therefore, there is a risk that a significant word will be created among the first to Nth candidates for each character of the word, resulting in the operator being instructed that it is not the word. For example, no matter how many words you look up in a word dictionary and output as a word in the dictionary, if that word is combined with a misread character, there is no correct answer. Purpose of the Invention The present invention further improves the word-by-word misread character correction method described above, and corrects the misread characters by using the correct character as is and outputting only words that combine it with other candidate characters for the misread character. This is an attempt to speed up corrections. Structure of the Invention The present invention provides an identification unit that identifies an optically read input character string character by character, a recognition dictionary in which various character candidates referenced by the identification unit are stored in advance, and a and a candidate category buffer that stores recognition results ranked in multiple positions per character, a display section that displays a character string of the first-ranked character in the buffer, and a character string displayed on the display section. In the misread character correction method of a character recognition device, which is equipped with an operation section that allows you to specify the correction part of a character, and a word dictionary that stores various words that are referred to when correcting characters on a word-by-word basis, the operation section allows you to specify the misread characters and the By specifying a range of words containing characters on the display, the correct character is left as is, and other candidate characters for misread characters are combined to create a word with the same word length as the range, and then the word Search the dictionary to retrieve the words that existed in the dictionary, calculate the distance for each of the words by referring to the rank in the buffer regarding the candidate character, and select the word with the smallest distance from the word range that includes the misread character. This will be described in detail below with reference to the illustrated embodiment. Embodiments of the Invention Table 1 shows the recognition results for the input character string "Advances in this field..." recognized by a general character recognition device, that is, the candidate characters ranked 1st to 5th for each character. .

[Table] Ranking of candidate characters (candidate categories) 1, 2,...
are displayed first in order of similarity, with the first place of each character given priority. The operator determines whether the input character string has been correctly read from the display. In the example of Table 1, it can be seen that the fourth character "No" and the ninth character "Shin" in the input character string are misread as "shi" and "su", respectively. Therefore, a correction is necessary. Table 2 shows conventional method 1 that indicates errors character by character.
This figure shows a comparison between conventional method 2, which specifies errors in word units, and the method of the present invention, by instruction method and by selection order.

[Table] Conventional method 1 corrects misread characters on a character-by-character basis. Rank 2 character “Tame” from the candidate category
is displayed instead, and when it is indicated that this is also an error, the character ``end'' of rank 3 is displayed, and the same goes for the characters of lower rank. Since the target "field" is in the 4th place, it will be corrected if the operator performs the selection operation three times. The candidate category for the ninth character "Shin" is also in the third place, so it can be corrected in the second selection. The same applies below, and the lower the rank of the target character, the more selection operations are required. Conventional method 2 is a method using a word dictionary. However, since it only specifies the word range (words that include misread characters) and does not indicate which of the multiple characters in the word is the misread character, it is possible to specify all of the multiple candidate characters for each character. The combination is looked up in a word dictionary and checked to see if the combination is in the dictionary. For example, if you specify a word consisting of the third and fourth letters (corresponding to "field"),
All combinations of character candidates and 4th character candidates (5 x 5
= 25 ways) to find out whether there is a word that has meaning as a word in a word dictionary. In this case, by chance, the only thing that has a meaning in this combination, and therefore the only thing in the dictionary, is "field", so there is no problem (the correct answer can be obtained directly), but the 9th and 10th letters In the case of a word (corresponding to "progress"), the word "Oite", which has a meaning that includes the misread character "su", is ranked first in the selection order.
The problem is that the correct word ``progress'' comes in second place. In this case, "Pursuer" is first displayed, and the operator must indicate that it is not this and have the next menu appear. In the present invention, in addition to the instruction of the word range as in the conventional method 2, an instruction is also provided as to which character is being misread. Although this point is not shown in Table 2, if you specify such misread characters, the search range of the word dictionary will be narrowed and the time will be shortened accordingly. is selected in the first order (described later), making the operator's operation even easier. FIG. 1 is a block diagram of an essential part of an embodiment of the present invention that performs the above-described processing, and FIG. 2 is a block diagram of the entire character recognition device. First, the overall general operation will be explained with reference to FIG. The observation unit 1 has an optical character reading function, and reads (photoelectrically converts) a handwritten or printed input character string. The feature extraction unit 2 detects the features of each character in the read input character string to facilitate subsequent recognition operations. The identification unit 3 compares and collates the output of the feature extraction unit 2 with recognition patterns or parameters in the recognition dictionary 4, calculates distance, and selects candidates from 1st to 5th with the highest degree of similarity for each character. The categories (see Table 1) are stored in buffer 5. The steps up to this point are common to all the methods in Table 2, and the recognition result (character string) in which the first-ranked character is arranged for each character is displayed on the display section 9 through the display control section 8. The operator looks at this display and, if any character is misread, indicates the misread character and the word range containing it. These correction instructions are given, for example, from the keyboard (operation unit) 10. The character recognition device that received the correction instruction is the word matching section 7.
is activated, and characters are corrected for each word by referring to the word dictionary 6. This operation will be explained with reference to FIG. The figure shows an example of the word "progress" in the 9th and 10th letters of Table 1. Since the word range and the misread character are specified from the keyboard 10, the word dictionary 6 that has received the correction instruction information A contains the word B whose word length is "2" and the last character is "Ayumu" (the first character is the misread character "Ayu"). All of the second to fifth candidates for ``su'' (words in which the second character is ``ayu'') are output to the word matching section 7. When there are multiple words B such as "progress", "regression", etc., the word matching unit 7 ranks the words B by referring to the ranks of misread characters in the buffer 5. In other words, if you fix the second character as "Ayumu" and combine the first character with any of the 9th characters in Table 1, five word candidates C will be generated (the first word that is known to be an error can be ignored) ). Therefore, the word matching unit 7 selects the word B from the dictionary 6.
The distance between the word candidate C and the above word candidate C is calculated. At this time 2
Since there is no need to think about the first character "Ayumu", the distance calculation is actually performed only for the first character. First, when we calculate the distance between word B "Shin" and word candidate C, we find that the first character "Shin" is in the third candidate ranking, so we give the value of distance (K-1) to the ranking K. Then, the distance D of "progress" is "2". From a similar idea, the word B “regression”
A distance "4" is given for . After distance calculations are performed for all of the words B in this manner, the word with the smallest distance (in this case, "progress") is displayed on the display section 9. Therefore, as in conventional method 2 shown in Table 2, the word ``suite'' that includes a misread character before ``progress''
is never output. Effects of the Invention As described above, according to the present invention, not only the word range including the misread character but also the character itself is specified, so the search range of the word dictionary is narrowed and the processing time is shortened. , since the character candidate ranking at the time of recognition regarding the misread character is also referred to, there is an advantage that the correct answer rate of the output word for correction is increased.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of main parts showing one embodiment of the present invention,
FIG. 2 is a block diagram of the entire character recognition device to which the present invention is applied. In the figure, 1 is an observation unit, 2 is a feature extraction unit, 3 is an identification unit, 4 is a recognition dictionary, 5 is a candidate category buffer,
6 is a word dictionary, 7 is a word matching section, 8 is a display control section, 9 is a display section, and 10 is an operation section.

Claims (1)

[Scope of Claims] 1. An identification unit that identifies an optically read input character string character by character, a recognition dictionary that is referenced by the identification unit, and a recognition dictionary that is recognized by the identification unit and ranks each character in multiple positions. A candidate category buffer that stores the assigned recognition results, a display section that displays a character string of the first-ranked character in the buffer, and an operation section that allows you to specify where to modify the character string displayed on the display section. In a misread character correction method of a character recognition device, which includes a word dictionary storing various words to be referred to when correcting characters on a word-by-word basis, the misread character and the range of words including the misread character are displayed by the misoperation unit. By specifying the word in the dictionary, the correct character is left as is, and other candidate characters for the misread character are combined to create a word with the same word length as the range, and the word is searched in the word dictionary and found in the dictionary. The method is characterized by extracting the words that have been read, calculating the distance for each of the words by referring to the rank in the buffer with respect to the candidate character, and displaying the word with the smallest distance as another candidate in the word range that includes the misread character. A method for correcting misread characters.