JPH0322094A - Character recognizing system - Google Patents

Abstract

PURPOSE:To obtain a stable character recognition rate by performing the feature quantity adapting processing where the size of each divided area of a recognition object character is changed for each of prescribed comparison object characters to minimize the difference among feature quantities and using the difference among feature quantities for the discrimination processing. CONSTITUTION:By the control of a feature quantity adapting means 17, the size of each divided area of the recognition object character is changed for prescribed comparison object characters to minimize differences among feature quantities. A dictionary retrieving means 15 compares minimized differences among feature quantities of respective prescribed comparison object characters to output the comparison object character, which corresponds to the minimum difference between feature quantities, as the recognition result. The size of each divided area of the recognition object character is changed to approximate this character to the comparison object character stored in a dictionary 13 in this manner. Thus, deviation and deformation of the recognition object character are easily absorbed when being minute for comparison object characters, and the stable character recognition rate is obtained.

Description

[Detailed Description of the Invention] [Summary] Regarding a character recognition method that recognizes a character by comparing the feature amount of the recognition target character with the feature amount of a comparison target character stored in a dictionary, the position of the recognition target character is In order to absorb minute deviations and deformations and obtain a stable character recognition rate, the image information of the character to be recognized is divided into multiple regions, and the feature values extracted for each divided region are A feature extraction means that obtains the features of the character to be recognized from a dictionary that stores the features of the character to be compared; and a dictionary that stores the features of the character to be compared; In a character recognition method, the size of each divided area of the recognition target character is changed for each predetermined comparison target character, and the size of each divided area of the recognition target character is changed for each predetermined comparison target character.
Perform feature amount adaptation processing to minimize the difference between ll amounts,
Each special ttl. [Industrial Application Field] The present invention provides a feature adaptation means for subjecting the differences between characters m to identification processing. [Industrial Application Field] Compare the features of the characters stored in the dictionary with the features of the comparison target characters stored in the dictionary. ! This paper relates to a character recognition method for character recognition.

[Conventional technology]

In the conventional character recognition method for characters input as image information, image information (character images) corresponding to individual characters is first obtained from the image information of the entire document through a predetermined cutting process (Fig. 5 (l)). .

Next, the special f! is expressed as a numerical string from the shading and outline of each character image. However, in the conventional method, in order to increase the character recognition rate, the rectangular area including the outline is first divided into 8 x 8 small areas, and the characteristic f of each divided area is extracted.
tkl is extracted (Figure 5 (2)).

This feature amount is, for example, a direction code (feature element!) in four directions depending on the connection state of each contour point on the contour line of a character image.
(Figure 5 (3)), and the number (frequency) of each direction code is integrated, and the feature amount for each divided area is (A o, A + , A ! + A 3)
is extracted as Note that the feature amount in the small area shown in FIG. 5(2) is (1, 4, O, O).

Similarly, the feature amounts are extracted from each of the 64 divided regions, and the feature amounts (4×8×8-256 dimensional vector) of the character image are obtained.

The feature quantities of the character images extracted in this way are compared with the feature quantities of the comparison target characters stored in the dictionary in the dictionary search section, and the one with the smallest difference (distance) between the feature quantities is is the recognized character, outputs its character code, and terminates the character recognition process. [Problems to be Solved by the Invention] Such a conventional method extracts each feature from a fixed divided region of a character image, and compares each feature with the corresponding feature in a dictionary to determine the difference. This is a configuration that calculates the difference between feature quantities by integrating the values. Therefore, even if the font of the recognition target character differs slightly from the font of the comparison target character in each divided area, the difference may be accumulated to a large value.

In other words, the font of the character to be recognized (Figure 5 (4).
5)) is different from the font of the comparison target characters stored in the dictionary (Fig. 5 (1). (2)), or there is a slight shift in the character position (shift in the divided area), or the reading When characters are deformed due to changes in conditions (one degree, blurred, etc.), it is not possible to absorb slight fluctuations in the feature amount due to the difference in font, causing a decrease in recognition rate. .

In view of the fact that the character recognition rate is highly dependent on font fluctuations, the present invention can absorb minute positional deviations and deformations of characters to be recognized and obtain a stable character recognition rate. The purpose is to provide a character recognition method.

[Means to solve the problem]

FIG. 1 is a block diagram of the principle of the present invention.

In the figure, a feature amount extracting means 11 divides image information of a character to be recognized into a plurality of regions, and obtains a feature amount of the character to be recognized from the feature amount extracted for each divided region.

The dictionary l3 stores feature amounts of characters to be compared.

The dictionary search means 15 compares the feature amount of the recognition target character with the feature amount of the comparison target character, and outputs the comparison target character with the smallest difference between the feature values as a recognition result.

The feature amount adaptation means 17 performs the following for each predetermined comparison target character:
A feature amount adaptation process is performed to minimize the difference between the feature amounts by changing the size of each divided area of the character to be recognized, and control is performed to subject the difference between the feature amounts to the identification process.

[For production]

According to the present invention, the size of each divided region of a character to be recognized is changed by controlling the feature value adaptation stage 17 so that the difference between the respective feature amounts is minimized for a predetermined character to be compared. The dictionary search means 15 compares the minimum difference between the feature amounts for each predetermined comparison target character, and outputs the comparison target character corresponding to the minimum difference between the feature amounts as a recognition result.

That is, in the method of the present invention, even if the difference between the feature amounts with a certain comparison target character in the initial divided area is minimized, the special @. In this method, if the difference between the feature amounts for other comparison target characters can be minimized through the amount adaptation process, that comparison target character is used as the recognition result.

In this way, the size of each divided area of the recognition target character can be changed to approximate the comparison target character stored in the dictionary l3, so that the displacement or deformation of the recognition target character with respect to the comparison target character can be avoided. If it is minute, it is easily absorbed, and it becomes possible to obtain a more stable character recognition rate than when the divided regions are fixed.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail based on the drawings.

FIG. 2 is a block diagram showing an embodiment of a character recognition device implementing the method of the present invention.

Note that the process of extracting the characteristic tUt of the character to be recognized, the process of character recognition based on the difference between feature amounts based on the dictionary search results, and the process of outputting the corresponding character code are the same as in the conventional method 2.

That is, the character cutout section 21 performs a predetermined cutout process on image information of characters to be recognized manually inputted from an optical reading device, and converts the image information into individual character images. The feature amount extraction unit 23 divides each character image into small regions δ, extracts the feature amount for each divided region, and extracts, for example, a 256-dimensional vector feature amount from one character image. The dictionary 25 stores feature amounts of characters to be compared and their character codes. The dictionary search unit 27 searches the feature quantities of the characters to be compared stored in the dictionary 25, compares each feature quantity with the characteristic of the character image, and determines the one with the smallest difference between the feature quantities. Output the corresponding character code as the recognition result. The output unit 29 converts the character code output from the dictionary search unit 27 into a code corresponding to a predetermined output medium and outputs the code.

The present invention makes the size of each divided area variable when comparing the feature values of the character to be recognized with the feature values of the character to be compared. This is intended to absorb variations in feature amounts and improve character recognition rates.

FIG. 3 is a flowchart showing the processing procedure of the embodiment of the present invention in the feature quantity extraction section and the dictionary search section.

A feature amount extraction unit 23, into which a character image of a recognition target character is manually input, extracts a feature amount of each divided region. The dictionary search unit 25 performs major classification processing using this feature amount. In this major classification process, when searching a dictionary in which the special at of the character to be compared is stored, a somewhat rough special 1fil is used to reduce the amount of calculation, and about 10 to 20 candidate characters are selected. The purpose of this major classification process is to reduce the number of character candidates in the next detailed identification process and reduce the amount of processing required for character recognition.

The feature extraction unit 23 and the dictionary search unit 25 perform detailed identification processing using horizontal feature adaptation processing, vertical feature adaptation processing, and re-identification processing. That is, by moving the dividing line that separates the divided areas horizontally and vertically, the special @. The quantity is compared with the feature quantity of the candidate character (comparison target character), the position of the dividing line that can be most approximated is determined, and the minimum characteristic f! is determined for each candidate character. Find the difference (distance) between I[1.

Here, a specific method of the horizontal direction feature amount adaptation processing will be explained using FIG. 4.

Assuming that the length of the character image in the horizontal direction (X direction) is M (dots), the length in the vertical direction (y direction) is N (dots), and the number of divisions in each direction is 8, the feature quantity XO (κ .y)
, feature amount X(t+j+k), dictionary feature amount X D (
i, j, k) are respectively XO(x, y) (x=0.
1,-,M-1,y=0.1,...,N-1)X (
i, j, k) (i=o, l,..., 7, j=0.
1,... 7, k=0.1,2.3 } X D (i, Lk) (i=o, l,...7, j
=0.1,...,7, k=0.1,2.3). Here, k (=0.1, 2.3) is a possible value of the feature element IXo(x,y) corresponding to the direction code.

Special 1iS! The quantity X (i, j, k) is
．． j), the feature quantity XO(×, y
), and can be expressed as . Here, Z (X O (x, y
), k) is a function that is set to 0 when the feature quantity XO (x.y) is k and is not ISk.

Further, sepx (4) is the value of X corresponding to the i-th dividing line in the X direction, and Sepy (j) is the value of y corresponding to the j-th dividing line in the y direction. For example, M=72,
In the case of N=56, each divided area divided into 8×8 is 9×7 (dots), and sepx(0) =0+s
epx(1)=9, sepx(2)=18. ...
, sepx(8)=72, sepy(0)=O.
sepy(1)=7, sepy(2)=14, ・−
, sepy(8)-56. Therefore sep
x(0) and sepx(8)-1 are the left end 0 and right end 71 of the character image, and sepn(0) and sepx(8)1 are the lower end O and upper end 55 of the character image.

The difference (evaluation function) between the features used for character recognition is the dictionary feature ffiXD(i,Lk) and the extracted feature X
It is defined by the Euclidean distance or urban area distance of (i, j, k). For example, the evaluation function D using city distance is D=ΣΣ:E. IXD(t,j.k) 1X(i
,3. k)1▲”6 J-6 k-6 Now, in Fig. 4 (1), the m-th (m-1
．． 2. ..., 7) with the dividing line d. (=0. 1.2
) to the right, the feature values X(m-1,j+k) and X(+*
, j, k) is as follows, and each characteristic 1vI. The amount is each X (m-1.Lk) - X (m-1.Lk) + ΔX
(+e, j.k)X(m,j,k) ←X(m,j,
k) - ΔX (+m. J l k).

Also, the m-th dividing line in the X direction is d. The amount of change ΔX' (m+J l k) when moving to the left by
and each feature amount is X (s+-1,j.k)←X(a+-Lj,k)-Δ
X'(m.j+k)X (m.j.k) 4-X (
a+, j, k) is converted into Δx' (m+ j+ k).

In addition, d. The feature values other than the divided areas sandwiching the dividing line that moves by the same amount remain unchanged, and as the dividing line moves, the characteristic f! The evaluation function Dd in the divided region portion where l+1 changes can be expressed as follows. Therefore, for the m-th dividing line in the X direction, the movement Id. Determine.

In the horizontal direction feature adaptation processing, the above processing is performed for all dividing lines in the X direction, and each evaluation function Dd is
The adapted feature amount is calculated based on the divided region where the minimum value is obtained.

In addition, in the vertical feature adaptation process, similar processing is performed for all dividing lines in the y direction to calculate adapted feature quantities. Note that since the horizontal and vertical feature adaptation processes are related to each other, comprehensive feature adaptation processing is desirable (Fig. 4 (2)). In the re-identification process, an evaluation function (special ltI [difference between quantities)] is calculated based on the feature quantity adapted for each candidate character, according to the divided area that has been changed to the minimum.

After performing the above processing for all candidate characters, in the recognition character determination process, the evaluation functions (Special 1! Differences between Ir amounts) corresponding to each candidate character are sorted, and the minimum evaluation function among them is sorted. The candidate character corresponding to L'2 1k
Determine as a character.

In this way, the features of the character to be recognized are automatically approximated to the features of the corresponding character in the dictionary, so it is possible to absorb minute deviations and deformations of the target character for vQ recognition with respect to the character to be compared in the dictionary. It's easy.

〔Effect of the invention〕

As described above, according to the present invention, erroneous recognition due to minute shifts or deformations due to variations in character reading conditions or font changes is reduced, and a stable character recognition rate can be obtained.

Note that printer output characters are used as recognition target characters,
Using a dictionary created from Iwata Hoso Mincho typeface characters, approximately 14
The result of character recognition processing performed on the 00 character was 80-85% in the conventional method, but 9% in the method of the present invention.
This can be improved to about 5%, which is extremely useful in practice.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of the principle of the present invention, Fig. 2 is a block diagram showing the structure of an embodiment of a character recognition device that implements the method of the present invention, Fig. 3 is a flowchart showing the processing procedure of the embodiment, and Fig. 4 is Special 1! FIG. 5 is a diagram illustrating a feature amount extraction process of a character image. In the figure, 1 is a feature extraction means, 3 is a dictionary, 5 is a dictionary search means, 7 is a feature adaptation means, l is a character extraction section, 3 is a feature extraction section, 5 is a dictionary, 27 is a dictionary search section , 29 is an output section. Block diagram of the principle of the present invention Figure 1 Character recognition device IIe. Bronok diagram showing an example Fig. 2 m-1 m=1 (1) (2) Fig. 4 explaining a specific method of feature adaptation processing Flowchart showing an example processing procedure Fig. 4 Fig. 3 Fig. (1) ( 2) 2 (3) (4) (5) Figure 5 is a diagram illustrating feature extraction processing of character images.

Claims (1)

[Claims] (1) A feature amount extraction means (11) that divides image information of a character to be recognized into a plurality of regions and obtains a feature amount of the character to be recognized from the feature amount extracted for each divided region, and a feature of a character to be compared. a dictionary (13) in which quantities are stored; and a dictionary that compares the feature quantities of the recognition target character with the feature quantities of the comparison target character and outputs the comparison target character with the smallest difference between the feature quantities as a recognition result. A character recognition method comprising a search means (15), characterized in that, for each predetermined comparison target character, the size of each divided region of the recognition target character is changed to minimize the difference between the feature values. 1. A character recognition method characterized by comprising a feature amount adaptation means (17) that performs a conversion process and subjects differences between respective feature amounts at that time to an identification process.