JPH0259504B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character reading device that reads characters written on a form etc. and outputs the character code of the character. classification) method.

When recognizing multi-character characters such as kanji, a hierarchical identification method is generally used in which the characters to be identified are first narrowed down to a decimal number, and then further detailed characteristics are used for identification. This first process is called major classification, and it is desirable that the process itself be simple and that the number of candidate characters selected in the major classification be as small as possible.

FIG. 1 is a block diagram showing an embodiment of the present invention, but the one in FIG. 1 excluding the threshold value table 7 shows a substantially conventional device.
3 is a scanning means, 3 is a feature extraction means, 4 is a similarity calculation means, 5 is a recognition dictionary, 6 is a classification means, and 8 is an identification means.

FIG. 2 is a diagram showing characters written in the frame of form 1, and 9 shows an example in which the kanji character "田" is written. Since the conventional device is well known, a detailed explanation thereof will be omitted, but the character 9 written on the form 1 is scanned by a photoelectric conversion device such as a television camera, and as a result of this scanning, the input character is It becomes a pattern and is memorized. The feature extraction means 3 extracts features from the input character pattern according to a predetermined rule.

On the other hand, for all the characters to be read by this device, features are extracted for the character pattern of the standard sentence shape of each character, and the extracted features are stored in the recognition dictionary 5 in correspondence with the character code of each character. There is.

The features extracted by the feature extraction means 3 are compared with the features stored in the recognition dictionary 5 by the similarity calculation means 4 to calculate the similarity.

Figure 3 shows an example of the calculated similarity.
0 temporarily indicates the character represented by the character code in the recognition dictionary as a kanji, and 11 indicates each character 10.
12 indicates the similarity to the kanji character ``田'' among the characters 10, and 13 indicates the similarity to the kanji character ``田''.

One of the conventional methods of selecting candidate characters from the similarities 11 outputted from the similarity calculation means 4, for example as shown in FIG. It was a matter of selecting. In the example in Figure 3, if N = 5, the five characters ``field'', ``country'', ``figure'', ``ma'', and ``onna'' are candidate characters, but the disadvantage of this method is that each time a new similarity is calculated, , the character code order must be sorted according to the similarity order, and when there are many characters to be recognized, it takes a lot of time to perform the sorting, which slows down the recognition process. It was hot.

The second method among the conventional methods is to find the maximum similarity from the similarity for each character, and use the judgment value obtained by subtracting a predetermined fixed threshold from this maximum similarity. This is a method of selecting characters with a degree of similarity as candidate characters. In this method, it is not necessary to rearrange the order of the characters in order of similarity, but it is sufficient to obtain a judgment value from the maximum similarity and compare this judgment value with the degree of similarity for each character.

However, in this second method, the same threshold value is used for all characters, so this threshold value is too large for some character groups, resulting in too many candidate characters. However, this threshold value was too small, resulting in an increased rate of misclassification.

This invention was made in order to eliminate the above-mentioned drawbacks of conventional devices, and it determines a threshold value for a group of characters classified by the similarity of glyph shapes from the dispersion of similarity for each character, and The purpose of this invention is to shorten the time required for classification processing without reducing classification ability by storing threshold values and performing major classification using the stored threshold values.

Embodiments of the invention will be described below with reference to the drawings. As mentioned above, FIG. 1 is a block diagram showing one embodiment of the present invention, and the operation up to the output point of the similarity calculation means 4 is as described above.

The contents of the threshold table 7 can be determined as follows. In other words, for one character, a plurality of deformed characters are prepared, which are deformed from the standard character form, and which are deformed but within a permissible range that can be easily and correctly read by humans. However, by calculating the degree of similarity of these modified characters, it is possible to know the classification of degree of similarity for the character. Next, a threshold value corresponding to the character group classified based on the similarity of letterforms is determined using the above classification. Then, the threshold value determined for each character group is stored in association with the characters constituting the character group. For example, the threshold value TH _M for the character group M is
It can be the maximum value of the variance value of the characters belonging to . Then, the value of the threshold THM is set as the threshold _THcn corresponding to the character _Cm belonging to the character group M. It is a well-known fact in this field that characters are generally easily misread as specific characters with similar shapes. Therefore, a set of characters that are likely to be mutually misread is defined as a character group. As a result, the threshold value can be reduced for a character group in which the dispersion of similarity is small with respect to character shape deformation. On the other hand, for a group of characters with a large variance in similarity, the threshold value can be set large so that even if the character group is deformed and written, the character group will not be omitted from the selection of candidate characters.

Thresholds are determined for characters of all character codes existing in the recognition dictionary 5, and are stored in a threshold table 7 in correspondence with the character codes.

FIG. 4 shows the thresholds read out from the threshold table 7 by the classification means 6 in correspondence with the characters in FIG. 3, where the same symbols as in FIG.
5 is the threshold value for the kanji character "田".

The classification means 6 knows that the maximum similarity calculated by the similarity calculation means 4 is 0.80 (FIG. 3), and sets the threshold 15 for the character "田" giving this similarity to 0.04 from the threshold table 7. and read out, 0.80
Using −0.04=0.76 as the judgment value, candidate characters 16 are the characters “Ta”, “Kuni”, and “Zu” with a similarity of 0.76 or more (Figure 5)
Determine as.

If the kanji ``田'' is written with a significant deformation, or if there is noise in the character pattern, the similarity of the kanji ``田'' will be 0.80-0.04 = 0.76, and the other characters will be in the 3rd place.
If it is as shown in the figure, the maximum similarity is
The result is 0.78 (``Kuni''), and the judgment value is 0.78-0.04 = 0.74, so ``Onna'', ``Figure'', ``田'', ``Kuni'', and ``Ma'' are candidate characters, but ``田'' is excluded from the candidate characters. There are very few opportunities.

The identification means 8 inputs candidate characters and determines whether to select one of the characters or refuse to read it, but this operation is the same as that of conventional devices, so a description thereof will be omitted. do.

As described above, in this invention, a threshold value determined in advance for each character group based on the similarity of character shapes is set as a threshold value for each character, and a threshold value corresponding to the character that gives the maximum degree of similarity is used for classification determination. Therefore, it is possible to effectively narrow down candidate characters without increasing classification errors, and there is no need to rearrange the order of characters based on high similarity, so classification can be performed at high speed. can.

Although the examples of FIGS. 2 to 5 show only Chinese characters, the characters that can be read by this invention are not limited to Chinese characters.

As described above, according to the present invention, the processing time for major classification can be significantly shortened without reducing recognition accuracy.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of input characters, and FIG. 3 is a diagram showing an example of similarity calculated by the similarity calculation means of FIG. 1. , FIG. 4 is a diagram showing threshold values corresponding to the characters shown in FIG. 3, and FIG. 5 is a diagram showing selected candidate characters. 1... Form, 2... Scanning means, 3... Feature extraction means, 4... Similarity calculation means, 5... Recognition dictionary,
6... Classification means, 7... Threshold table, 8... Recognition means.

Claims (1)

[Scope of Claims] 1. A scanning means for scanning and photoelectrically converting characters written on a form, etc., and storing the processing result as an input character pattern; a feature extracting means for extracting the features of the standard glyphs of all the characters to be read; and a feature extracting means for extracting the features of the character patterns extracted according to the above-mentioned predetermined rules from the input character pattern regarding the standard glyph shapes of all the characters to be read, and storing the features of the character patterns in correspondence with the character code of each character. a recognition dictionary that calculates the similarity between the features extracted by the feature extraction means and the features stored in the recognition dictionary; By calculating the similarity between each feature extracted for multiple types of deformed glyphs that have been deformed within the allowable range from the standard glyph shape of the character and the features of the character in the recognition dictionary. a threshold table that stores a threshold value determined for each character group predetermined based on the similarity of character shapes, in correspondence with the character code of each character, based on the obtained dispersion value of the similarity degree regarding the character; and The threshold value for the character that gives the maximum value among the degrees of similarity calculated by the calculation means is read from the threshold value table, and the value obtained by subtracting the read threshold value from the maximum value is used as a judgment value, and the value obtained by subtracting the read threshold value from the maximum value is used as the judgment value. a classification means for selecting as a candidate character a character having a degree of similarity equal to or higher than the above-mentioned judgment value among the degrees of similarity thus calculated; and one character is determined according to a predetermined rule from among the candidate characters selected by the classification means. A character reading device comprising: a recognition means for processing the characters as either undecidable or undecidable;