JP3173363B2 - OCR maintenance method and device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an OCR maintenance method for optimally setting a recognition dictionary and parameters of an OCR device.

[0002]

2. Description of the Related Art As important techniques in document reading processing, there are techniques for extracting and recognizing characters. In general, character extraction processing is performed using statistics such as character chunks and histograms of inter-character spaces, and distance values or similarities obtained as a result of character recognition. It is known that the tuning performance is improved by tuning the contribution degree of. For example, in the case of a fixed-pitch printed document, the method of Tsuji et al. ("Character separation", IEICE technical report, PRL83-66)
Is suitable for keeping the distance between the cutout positions as constant as possible. In the case of a printed document or a handwritten document with a proportional pitch, Ariyoshi's method ("Character segmentation from Japanese printed document by dynamic hypothesis generation / verification", IEICE technical report, PRU9
As described in 3-47), a method of placing importance on the result of character recognition and the like rather than the distance between cutout positions is generally used. The setting of the degree of contribution of each statistic can be controlled by expressing the degree of contribution as a parameter. Such a method is described by the method of Ishidera et al. (“Character string recognition method using two-dimensional arrangement information”, electronic information IEICE Technical Report, PRU95-10
8) is also used.

In order to increase the character recognition rate, not only a character recognition algorithm but also a method of creating a recognition dictionary (= standard pattern of each category created from learning data) is important. Generally, a recognition dictionary is created by changing the contents of a dictionary for many character images so that the recognition result is correct. However, recognition of unlearned characters may be erroneous. Therefore, as a method of additionally learning characters, Japanese Patent Laid-Open Nos.
-314303 and the like are disclosed. In these inventions, the recognition result when a document image is input is displayed on both sides, and misread characters in the recognition result specified by the operator are learned to improve the system performance.

[0004]

However, in the conventional character segmentation processing, parameters are determined by trial and error, and it has been very difficult to optimize the parameters for a large amount of data.

Further, Japanese Patent Application Laid-Open No.
Dictionary learning methods such as 24783 and JP-A-5-314303 are not suitable for system maintenance using a large amount of input data because the operator needs to extract misread characters or misread words.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems. The present invention relates a recognition result of a document image to a correct text, and automatically extracts a misread character or a character segmentation error portion. Another object of the present invention is to provide a method for easily realizing optimization of a character extraction parameter and a recognition dictionary.

[0007]

In order to achieve the above object, according to the present invention, a result of cutting out and recognizing characters in a document image and a correct text in which characters in the document image are converted into character codes are moved. The distance between the recognition result and the correct text and the correspondence between the characters are determined by comparing them with the dynamic programming method, and the erroneous recognition part is extracted from the correspondence, and the recognition dictionary of the erroneously recognized character so that the distance is as small as possible. Modify the extraction parameters.

Further, according to the present invention, the character recognition result is represented by A =
_{{A 1, a 2, ...} , a i, ..., a I}, the correct answer text _{B = {b 1, b 2} , ..., b j, ..., b J} and, among character codes by the following formula When the distance d (i, j) is defined, the character string matching means defines the cumulative distance of the dynamic programming by the following recurrence formula g (i, j), and the distance g (I, j) J), and the correspondence between individual characters at that time is output.

[0009]

(Equation 3)

When the character recognition result is compared with the correct answer text, for each character, (1) correct reading character, (2) misread character,
(3) There are three cases of character extraction error.
Of these, the character codes for (1) and (2) correspond one-to-one, and the character codes for (3) usually correspond to one-to-n (or n-to-1). Therefore, these locations can be specified from the difference between the character codes of the characters associated by the dynamic programming and the shape of the optimal correspondence path.
In particular, when calculating the cumulative distance value of dynamic programming, the path extending in the horizontal and vertical directions corresponds to the number of characters changes,
In other words, since a character segmentation error is represented, in addition to the normal distance value, a penalty distance value is added to the path extending in these directions from the grid point where the character code correctly corresponds, and the character segmentation error portion is correctly identified. Be able to extract. If the recognition dictionary of the erroneously recognized character is corrected or the cutout parameters are adjusted so that the cumulative distance value of the dynamic programming becomes as small as possible, the OCR
The character extraction performance and the character recognition performance of the system can be further improved.

[0011]

Embodiments of the present invention will be described below in detail. FIG. 1 is a block diagram showing an example of a recognition dictionary maintenance method employing the maintenance method of the present invention. In the present embodiment, the document image database 1
0, correct text 11 of each document image, character recognition unit 12,
It comprises a character string collating unit 13, a maintenance unit 14, and a character recognition dictionary 15.

The document image database 10 is a set of document images used for system maintenance, and the correct answer text 11 is a file storing character codes of characters included in each image. One correct text corresponds to one image data.

The character recognition unit 12 is a processing unit that performs layout analysis, line segmentation, character segmentation, character recognition, etc., performs a recognition process after extracting a character area, and outputs a character code as a character recognition result.

The character string collating unit 13 collates the character recognition result output from the character recognizing unit 12 and the correct text 11 corresponding to the input document image line by line by dynamic programming. The formula for the dynamic programming is shown below.

[0015]

(Equation 4)

[0016] where the function OR is at least one value of the two arguments returns the value 1 when is 0, otherwise is a function that returns 0, this term is the normal dynamic programming Is different. An example in which a character code string is collated by ordinary dynamic programming will be described with reference to FIG. A typical cumulative distance value is

[0017]

(Equation 5)

When a character recognition result 20 and a correct answer text 21 are given, a correct path 22 and an incorrect path 23 are extracted as optimally corresponding paths. In general character string matching, when an input pattern sequence contains a noise pattern (insertion) or an important pattern is missing (missing), that is, when the pattern in the input pattern sequence is any of the standard patterns In many cases, it is difficult to strictly specify the correspondence between patterns. However, in the case of the present invention, each character in the recognition result can be assumed to correspond to any one of the characters in the correct character string, so that it is possible to more accurately correspond. In the present invention, the function OR is used for a path extending in the horizontal or vertical direction from the grid point where the character code matches.
Is added, and a path in an oblique direction, that is, a path in which the correct-reading character code corresponds to the one-to-one correspondence as much as possible is extracted. When this method is applied to the character code strings 20 and 21, the path indicated by the path 22 is extracted as the path that gives the minimum cumulative distance value g (I, J) = 4 , and a correct correspondence relationship can be obtained.

From this correspondence, it is possible to determine a recognition error portion such as a misread character or a cutout error. If n character codes in the recognition result correspond to one character code of the correct text, a cutout error (over-segmentation) occurs,
If one character code in the recognition result and n character codes of the correct text correspond, a cutout error (over-integration),
If the recognition result and the character code in the correct answer text correspond one-to-one and the character codes are different, it can be determined that the character is misread. In the example shown in FIG. 2, (recognition result, correct answer) error pattern (coast, information) misreading (report, report) (process, process) (physical, logical) (no,, no) (development, departure) (exhibition, exposition) (1), (1), (2), (1), (2), (1), (2) are obtained.

FIG. 3 shows an example in which cutout errors and misread characters occur continuously. In such a case,
Even if any of the routes indicated by the dotted lines are selected, the accumulated distance values become equal, and an accurate corresponding route cannot be obtained. Therefore, when specifying a misrecognized part, n (≧ 2) in the recognition result
If characters are erroneously recognized consecutively and m (≧ 2) characters in the correct text correspond thereto, the m characters are not considered as maintenance information, or the user is notified of the fact. I do.

However, if the basic performance of the OCR is low and erroneous recognition is likely to occur continuously, maintenance information can be increased to some extent by modifying the definition of the distance between character codes. In the above-described embodiment, the distance between character codes is represented by binary values. However, a set of characters having similar shapes (for example, “ba” and “ぱ”) and a set of characters in which OCR is likely to be mistaken are selected in advance. If the inter-code distance D _sim corresponding to these characters is defined as 0 <D _sim <1, then
The correct route can be obtained to some extent. In the case of FIG. 3, the correct answer path 33 is uniquely obtained. Therefore, when the formula for calculating the cumulative distance is set in this way,
Only locations where erroneously recognized characters other than erroneously recognized characters having similar shapes or characters that are easy to read erroneously are not used for maintenance.

The maintenance unit 14 corrects or additionally registers a recognition dictionary of erroneously recognized characters or adjusts a cut-out parameter so that the cumulative distance of the dynamic programming is as small as possible.

Correction of the recognition dictionary is mainly performed for characters that have been misread, but if the character cutout algorithm uses the result of character recognition, learning for characters that have been cut out incorrectly (over-segmented) is also effective. If there are many cutout errors, the parameters are minutely changed. By performing such a correction so that the total value of the accumulated distance values for about several tens of lines becomes smaller, the reading performance of the system can be improved.

[0024]

As described above, according to the present invention, the character recognition performance and the character segmentation performance of OCR can be effectively improved. In addition, in the present invention, O
The correct text and the recognition result are collated using dynamic programming to which an R function term is added. In this method, when there is no insertion or deletion of a pattern between two patterns, that is, when a pattern included in one pattern is always the other This has an effect in matching a pattern sequence corresponding to a certain pattern. Therefore, character recognition results obtained from a plurality of character recognition devices,
Alternatively, the present invention can be applied to character string collation in a case where character recognition results obtained from a plurality of overlapping document images are combined to generate a character recognition result as a whole.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a character recognition dictionary maintenance device.

FIG. 2 is a diagram showing a correspondence between a recognition result and a correct answer text.

FIG. 3 is a diagram illustrating a case where characters are continuously erroneously recognized;

[Explanation of symbols]

Reference Signs List 10 Document image database 11 Correct text 12 Character recognition unit 13 Character string collation unit 14 Maintenance unit 15 Character recognition dictionary 20 Character recognition result (for one line) 21 Correct text (for one line) 22 Correct optimal path 23 Same as correct path Path that gives the cumulative distance value 30 Character recognition result 31 Correct answer text 32 Corresponding path to be extracted 33 Corresponding path to be extracted by the modified cumulative distance calculation formula

Continuation of the front page (56) References JP-A-2-250188 (JP, A) JP-A-5-128299 (JP, A) JP-A-6-176197 (JP, A) JP-A-6-195519 (JP) JP-A-6-251204 (JP, A) JP-A-61-296480 (JP, A) JP-A-4-98586 (JP, A) "Pattern recognition and learning algorithms" Yoshinori Uesaka, Kazuhiko Ozeki Author, pp. 91-108 (especially, see pages 6.5 and 6.5, pages 100-102) Published on May 6, 1992, Bunichi Sogo Publishing Co., Ltd. (2nd printing) (58) Field surveyed (Int.Cl. ⁷ , DB name) G06K 9/68 G06K 9/62 640

Claims (2)

(57) [Claims] 1. A character cutout recognizing means for cutting out and recognizing characters in a document image, a character recognition result output by the character cutout recognizing means,
A character string matching unit that compares the character in the document image prepared in advance with a character-coded correct text by a dynamic programming method, and outputs a distance between the character recognition result and the correct text, and a correspondence relationship between the characters. And a system maintenance means for extracting a misrecognized portion from the correspondence and correcting a recognition parameter of a misrecognized character or a cut-out parameter so that the distance becomes as small as possible. a ₁ , a ₂ , ..., a _i , ...,
a _I }, and the correct answer text is B = {b ₁ , b ₂ ,.
_j ,..., b _J }, and when the inter-character code distance d (i, j) is defined by the following equation, the character string matching means calculates the cumulative distance of the dynamic programming by the following recurrence equation g (I,
j), and outputs g (I, J) as the distance,
An OCR maintenance device characterized by outputting the correspondence of each character at that time. (Equation 1) [OR () returns 0 if both arguments are 1 ;
Otherwise, a function that returns 1 is shown. min is a function that returns the minimum value of the calculation results of the three equations shown on the right side of the parentheses (｛). w indicates a positive constant. ] 2. A dynamic programming method for comparing a result of extracting and recognizing characters in a document image with a correct text prepared by converting characters in the document image into character codes, and comparing the result with the correct answer. The distance from the text and the correspondence between the characters are obtained, the erroneously recognized portion is extracted from the correspondence, and the recognition dictionary or cutout parameter of the erroneously recognized character is corrected so that the distance becomes as small as possible. A = ｛a ₁ , a ₂ , ..., a _i , ...,
a _I }, and the correct answer text is B = {b ₁ , b ₂ ,.
_j ,..., b _J }, and when the character code distance d (i, j) is defined by the following equation, the cumulative distance of the dynamic programming is defined by the following recurrence equation g (i, j). , The distance as g
An OCR maintenance method, wherein (I, J) is obtained, and a correspondence between individual characters at that time is obtained. (Equation 2) [OR () returns 0 if both arguments are 1;
Otherwise, a function that returns 1 is shown. min is a function that returns the minimum value of the calculation results of the three equations shown on the right side of the parentheses (｛). w indicates a positive constant. ]