JP2867382B2 - Font determination method in character recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement in a font determining method in a type character recognition apparatus having a recognition dictionary for each font.

(Prior Art) As a method of recognizing a plurality of font characters, a method of recognizing a plurality of fonts by determining a final result using word matching for each of the read results obtained by a plurality of single font reading mechanisms. (IEEE TRANSACTIONS ON COMPUTE
RS.VOL.C-27, NO.8, PP.721 “A Multifont Word Recog
nition System for Postal Address Reading ", hereinafter referred to as Document 1), a method for recognizing using multiple fonts using common features (IEEE TRANSACTIONS ON PATTERN ANALYSI)
S AND MACHINE INTELLIGENCE, VOL.PAMI-9, NO.2, PP.274
“On the Recognition of Printed Characters of An
y Font and Size ", hereinafter referred to as Document 2).

IEICE, Pattern Recognition and Understanding Workshop PRU86
-76 "Hierarchy determination method and its character recognition system PC-OC
R ”(hereinafter referred to as Document 3), character strings and characters are cut out by using positions on a distribution detected as lumps from the frequency distribution of pixels obtained by projecting an image, and discrimination is used for feature extraction. Analysis is used, and judgment is made based on the distance value.

(Problem to be Solved by the Invention) Since Document 1 is based on single font reading, it is easy to create a dictionary, but as a character recognition device for a document, the number of words to be collated increases, and a large-scale device is required. There is a problem that it becomes. Although the method of Document 2 can reduce the size of the character recognition dictionary, it has a problem in that it is difficult to determine a feature that is constant for different fonts and different between character types. If there are no more effective features, there is a problem that it is difficult to improve recognition performance.

(Means for Solving the Problems) According to the invention of the present application, for each recognition target range consisting of an image area having an equal character string interval value or an image area of one character string, each of the first characters of the character string in each of the recognition areas is recognized. The matching value of the judgment processing result by the font-specific dictionary is obtained, and the candidate font is narrowed down by deleting one font having a small matching value by using the obtained matching value or the average matching value from the accumulated matching value, thereby narrowing down the candidate font. In the same manner, the next candidate font is narrowed down by using the matching value of the judgment processing result by the dictionary of each font of the narrowed candidate font or the average matching value from the accumulated matching value, and the above-described processing is performed until the number of candidate fonts becomes one. By repeating the processing, it is possible to obtain a font determination method in a multiple-character type character recognition apparatus in which the last candidate font is a character font in the recognition target range.

(Operation) If a document is to be recognized and a section or one line consisting of a plurality of lines is set as a unit of reading, a single font character is often used in each unit. Single font reading makes it easy to create a dictionary and easily improve recognition performance.
Improve recognition performance by adopting a method in which a character font in a reading unit is first determined in units of clauses or one line, and then character recognition (single font character recognition) is performed using a recognition dictionary for the selected font. It is possible to use a multi-font character recognition method that is easy to measure. As a method of determining the font, a method of using the result of matching of each font in the recognition dictionary can be considered. Reference 3
Has the property that when a character having a different shape from the learning pattern used to create the dictionary is determined, the determination result becomes a large minimum distance value. FIG. 4 is a diagram showing an example of a minimum distance value obtained by collating character data of one font with recognition dictionaries of several fonts. According to FIG. 4, the minimum distance value in the same font is distributed to a relatively small value, but is distributed to a large value in another font. Therefore, a font deciding method that determines a font by comparing the minimum distance value determined by a dictionary created for each font is conceivable. By the way, the input character pattern does not completely match the character pattern used for learning due to the printing state of the character, the state of the image input unit, or the normalization of the image. Therefore, there is a high possibility that erroneous font determination will occur in font selection based on only one minimum distance value. Therefore, a method of judging several characters in all fonts and selecting an optimum font from all the judgment results can be considered. However, in this method, the processing required for determining the font is linearly related to the number of fonts, and thus cannot be used when the number of fonts is large.

However, if the following method is adopted, the font can be determined while suppressing an increase in the processing amount required for determining the font.

The candidate fonts are narrowed down by comparing the minimum distance values obtained by the dictionary determination of all fonts for the first character, and excluding one font having the largest minimum distance value from all fonts. Similarly, the next candidate font is narrowed down by comparing the minimum distance value obtained by determining only the candidate font for the next character. This is a method in which the number of candidate fonts is reduced by repeating the determination for each character, and the font is determined when it becomes one. In this method, the number of candidate fonts decreases each time the determination is repeated, so that an increase in processing can be suppressed.

Also, by comparing the average minimum distance value obtained from the accumulated distance value obtained by accumulating the minimum distance value as the determination result for each character for each font, it is possible to narrow down the candidate fonts. Furthermore, when the minimum distance value and the average minimum distance value are combined and the difference between the minimum distance values for each font is small, the candidate fonts can be narrowed down by comparing the average minimum distance values.

In the above description, the matching is performed using the distance value as the evaluation scale, but the same applies when the correlation value is used as the evaluation scale.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a flowchart showing one embodiment of a font determining method in the character recognition device of the present invention. In the following description, the numbers in parentheses correspond to the numbers in the flowchart.

One recognition target range is an area of a character string image having the same attribute or an image area of one character string. Assuming that character string intervals are attributes, the character string interval values within one recognition target range are equal. When a known technique of determining a character string position by projection as described in Document 3 is used in a character recognition device, the above-described font determination processing can be performed by the processing shown in the flowchart shown in FIG.

First, from the distribution of the number of pixels constituting the character obtained by projecting the image in the horizontal direction (or the vertical direction), the width of the position where the pixels are continuous on the distribution is regarded as a chunk, Using the coordinates that determine the area of the character string, find a plurality of character strings with the same interval value from the interval value between the character strings obtained from the difference between the coordinate values of the adjacent character strings, and calculate the maximum value of the coordinate values of the multiple character strings. Determine the recognition range from the minimum value (10
1). When the interval values are not equal, the coordinate value of one character string is set as the recognition target range.

The font of the recognition dictionary prepared in the character recognition device is set as a candidate font (102). Here, the candidate font number parameter is C. As described above, there is one or more character strings in one recognition target range, and the character string coordinate value indicating the area of each character string is obtained at the time of determining the recognition target range. Read the value (103). From the image in the area indicated by the coordinate value of the read character string, the coordinate value of the character cutout position is obtained by projection in the same manner as in the determination of the character string (104). There are a clipping position detecting method in the literature 3 and many known techniques, and the present invention is not limited by the clipping method, so that further description is omitted. The number of characters detected in one character string and the parameters of character control are M and N, respectively.

When the character control parameter N becomes larger than the number M of characters in the character string, the process proceeds to the next character string in the recognition target range (105).

N in the character string using the recognition dictionary for each of the candidate fonts
The character is determined, and the maximum matching value for each font is output as the determination result (106). The maximum matching value of each font is compared, and one font having a small maximum matching value is excluded from the candidate fonts (the number of candidate fonts is reduced by 1, so the candidate font number parameter is C = C−
1) to determine the next candidate font (107).

Next, the character control parameter N is advanced by one (108), and it is checked whether there are two or more candidate fonts (109). If there are two or more, the above processing is repeated and it is recognized that the number of candidate fonts becomes one. The font of the target area is determined.

FIG. 2 is a flowchart showing another embodiment of the font determining method in the character recognition device of the present invention, and shows a method of combining a matching value and an average matching value for each character.

First, from the distribution of the number of pixels constituting the character obtained by projecting the image in the horizontal direction (or the vertical direction), the width of the continuous position of the pixels on the distribution is regarded as a chunk, and the start and end positions of each chunk are determined. Using the coordinates that determine the area of the character string, find a plurality of character strings with the same interval value from the interval value between the character strings obtained from the difference between the coordinate values of the adjacent character strings, and calculate the maximum value of the coordinate values of the multiple character strings. Determine the recognition target range from the minimum value (20
1). When the interval values are not equal, the coordinate value of one character string is set as the recognition target range.

The fonts of the recognition dictionary prepared in the character recognition device are candidate fonts, and the cumulative matching value of each candidate font is set to zero (202). The number of fonts in the dictionary is set in the candidate font number parameter C. As described above, there is one or more character strings in one recognition target range, and the character string coordinate value indicating the area of each character string is obtained at the time of determining the recognition target range. Read the value (203). From the image in the area indicated by the read coordinate value of the character string, the coordinate value of the character cutout position is obtained by projection in the same manner as the character string determination (204). The number of characters detected in one character string and the parameters of character control are M and N, respectively.

When the character control parameter N is larger than the number M of characters in the character string, the process proceeds to the next character string in the recognition target range (205).

N in the character string using the recognition dictionary for each of the candidate fonts
The character number is determined, and a matching value for each font is output as the determination result (206). The minimum matching value (V1) and the next minimum matching value (V1)
2) and (207). Next, the difference between V1 and V2 is compared with a threshold value (208). When the difference is equal to or greater than the threshold value, a candidate font is selected by excluding the font on the V1 side (210). If the difference is less than the threshold, V1 and V
The average matching value (S1 and S2) from the accumulated matching values of both fonts 2 is compared (209), and the next smaller font is determined by excluding the smaller font from the candidate fonts. Update the cumulative matching value (210, 21
1). For example, if there are three or more candidate fonts,
Fonts (F1) that take the smallest matching value (V1), fonts (F2) that take the next smallest matching value (V2),
It is divided into the remaining font groups (one or more fonts remain). If the difference between V1 and V2 is large, the font (F1) of V1 is deleted (210). If not large, the average matching values S1 and S2 corresponding to the fonts of V1 and V2 are compared (209).
If 1 is equal to or less than the average matching value S2, the font (F1) of V1 is deleted (210), otherwise the font (F2) of V2 is deleted (211). By this processing, the number of candidate fonts is 1
It will decrease one by one. Next, the cumulative matching value is updated by adding the matching values of the candidate fonts, and at the same time, the character control parameter N is increased by 1 (212), and it is checked whether the number of candidate fonts is 2 or more (213). In some cases, the above processing is repeated until the number of candidate fonts is 1
When the number of fonts is reached, the font in the recognition target range is determined.

In the character recognition device, when a font in a certain recognition target range is determined, the recognition target range is determined by the recognition dictionary of the font, and the category name having the maximum matching value is determined as a determination result.

FIG. 3 is a block diagram showing an example of a character recognition apparatus to which the present invention is applied. The control processing of the font determination shown in the flowcharts of FIGS. 1 and 2 is performed by a control means.

The image input means 1 converts a video signal obtained by optically scanning a form into a digital signal and outputs the digital signal. The control means 2 transfers the digital signal output from the image input means 1 to an image storage means 3 and stores the digital signal as an image. When the image is stored, the determination of the recognition target range and the font of the range described with reference to FIG. 1 are performed.

The control means 2 reads the image of the image storage means 3, obtains the coordinate values of the character strings in each recognition target range from the above-described projection processing on the read image, and stores the obtained coordinate values in the coordinate storage means 4. I do. Next, the control means 2 reads out the coordinate values of one character string in one recognition target range in the coordinate storage means 4. Further, the image of the area indicated by the obtained coordinate values is read out from the image storage means 3, a character cutout position is obtained by projection processing, and the obtained cutout coordinate values are stored in the coordinate storage means 4. The control means 2 stores the font of the recognition dictionary stored in the font dictionary storage means 5 as a candidate font.

The control means 2 compares the character number M with the character control parameter N whose initial value is 1, and if the character control parameter N is larger than the character number M, the next character string coordinate value in the recognition target range is calculated. Read and repeat the process up to character segmentation. If the character control parameter N is equal to or less than the number of characters M, the control means 2 controls the character recognition means 6 to perform the Nth character determination process based on the candidate font dictionary stored in the font dictionary storage means 5.

The character recognizing means 6 reads the character cut-out coordinate value of the Nth character stored in the coordinate value storing means 4 based on the character control parameter N, and stores the character image in the area indicated by the obtained coordinate value in the image storing means 3. Then, the maximum matching value, which is the determination result, is output for each candidate font dictionary. The control means 2 receives the maximum matching value for each candidate font output from the character recognizing means 6, compares a plurality of maximum matching values, and removes one or more fonts having smaller values from the candidate fonts storing the fonts. This makes the next candidate font.

The control means 2 updates the value of the control parameter N and repeats the above processing until the number of candidate fonts becomes one.

When the number of candidate fonts becomes one, the control unit 2 controls the character recognizing unit 6 to determine each character image in the recognition target range by using the recognition dictionary of the candidate font, and controls the maximum matching output from the character recognizing unit 6. The category name corresponding to the value is stored in the recognition result storage means 7 as the determination result of each character image.

By controlling the control means 2 to perform the above-described processing for each of the recognition target ranges, the result stored in the recognition result storage means 7 finally becomes the determination result of the entire input form.

Here, it goes without saying that the processing described in FIG. 3 can be performed by software of a personal computer system having a scanner.

(Effects of the Invention) As described above, according to the present invention, a font is determined based on the result of comparison using a single font dictionary, so that special features for font determination are unnecessary. There is little increase in the processing amount required for determination, and reading is performed by single font recognition after font selection, so that dictionary creation is simple and high performance is obtained.

[Brief description of the drawings]

1 and 2 are flowcharts respectively showing an example of a processing procedure in determining a font in the character recognition device of the present invention, FIG. 3 is a block diagram showing an example of the character recognition device,
FIG. 4 is a diagram showing an example of distribution of the minimum distance value for each category corresponding to the maximum matching value which is the result of collation of data of one font with a dictionary of several fonts. In the figure, 1 is an image input unit, 2 is a control unit, 3 is an image storage unit, 4 is a coordinate position storage unit, 5 is a font dictionary storage unit, 6 is a character recognition unit, and 7 is a recognition result storage unit.

Claims (1)

(57) [Claims] 1. A font determining method in a character recognition device for selecting one font from a dictionary prepared for each font and making a determination, wherein the font-based dictionary for the first character in the recognition target range is used for each recognition target. A matching value of the determination processing result is obtained, and a candidate font is selected by using the matching value and excluding at least one font. The second character in the recognition target range is determined according to the dictionary of each font of the selected candidate font. The next candidate font is selected by removing at least one font using the matching value of the determination processing result or the average matching value from the accumulated matching values, and the next candidate font is selected until the number of candidate fonts becomes one. It is characterized in that the character determination processing and the comparison of the matching value or the average matching value are repeated, and the last candidate font is set as the character font in the recognition target range. Font determination method in the character recognition device to.