JPS63178379A - Optical character reader - Google Patents

Abstract

PURPOSE:To improve the accuracy of recognition and to shorten a recognizing time by reading out a separation feature based on two candidate characters extracted from a decision table from a separation feature table and comparing the separation feature with a character pattern to decide a character. CONSTITUTION:A character deciding part 15 extracts two proposed characters from the decision table 14 at first, and then reads out a separation feature separated as the result of superposing both the characters from the separation feature table 12. A character pattern corresponding to a character to be read out from a document is compared with the separation feature to decide the character. When the character can be recognized, the character to be read out is converted into a character code corresponding to the character and the code is outputted. If the character is not the candidate character extracted from the table 14 is decided, a proposed character other than the character is extracted from the table 14 and similar deciding operation is executed. Consequently, the recognition accuracy can be improved and the recognizing time can be shortened.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an optical character reading device (hereinafter abbreviated as OCR), and particularly to a method for optically reading and recognizing printed characters.

(Prior Art) FIG. 8 is a block diagram showing a general configuration of OCR. In the figure, a reading target 82 printed on a form 81
The printed characters are converted into electrical signals by a photoelectric conversion section 83 and stored in a line memory (not shown), and are cut out character by character by a character cutting section 84 and a character pattern for each character is stored in a pattern memory 85. Then, the feature extraction unit 86 reads out the character pattern for one character from the pattern memory 85 and extracts the necessary features from the character pattern. Next, the extracted features are compared with a recognition dictionary 88 in which feature patterns for each character are registered in advance by a character determination section 87, and the recognition result is output as a character code.

Next, a conventional recognition method performed in the feature extraction section 86 and character determination section 87 shown in FIG. 8 will be explained. Here, as an example, a case will be described in which five characters A to H are to be recognized.

The conventional recognition method is performed as follows. First, as shown in Fig. 9, when the characters to be recognized are overlapped with each other at the center of the characters, there is a black area where they do not overlap (the character pattern is quantized into a black/white binary pattern).
"1" for the black text part and "0" for the white background part
shall be. ), regions that are sufficiently stable against various variations such as thickening or thinning of the pattern are extracted as features.

F3 shown in FIG. 9 is a feature that allows characters A and E to be stably identified. FIG. 10 shows a mask pattern called feature FINF4 created by the same operation for the combination of remaining target characters. It is assumed that an unknown character pattern A is stored in the pattern memory, and the center of the character matches the center of the pattern memory. This pattern memory has the 10th
The mask patterns shown in the figure are overlapped, and the number of points constituting the character where the feature region of Fl-F4 and the character part of the unknown character pattern A overlap is counted for each of F1 to F4. F
The counted value for each INF4 is compared with a threshold value determined in advance by an appropriate method, and if it is larger than the threshold value, it is determined that the characteristic exists and is set to "1", and if it is smaller than the threshold value, it is determined that the characteristic does not exist and is set to "0". ” is quantized.

In the example shown in FIG. 9, F1=0, F2=1, F3=0, F
4=indeterminate.

The characteristic results obtained in this way are shown in Figure 8, character determination unit 8.
7, and the recognition result is output as a character code based on the following recognition logic and feature results created in advance by an appropriate method such as experimentation.

A=Shield-F2・Shield, B=F1-F2・F3・F4
゜C=F1-■・F4. D=F1・Takumi・F4.

E=Fl・F2・F3・■ In the example shown in FIG. 9, A=0・1・O=1. B=0・1・
0・(undefined)=O, C=O・1・(undefined)=O, D=0
・1・(undefined)=0. E=0.1.0.(undefined)=0, and the character code "A" corresponding to the character A is output. In this way, unknown pattern A is recognized as the character "A".

(Problem to be Solved by the Invention) However, in the conventional method described above, all features are counted, so when the number of characters to be recognized increases and the number of features necessary for identification increases due to the combination formula, the feature commission is The processing time for this method also increases, which poses a problem in that the recognition time increases significantly. Therefore, in order to avoid increasing the recognition time, it is possible to reduce the number of features by substituting a single feature for multiple features, but this method results in a decrease in feature stability, that is, in recognition accuracy. As described above, in the conventional method, recognition time and recognition accuracy have a contradictory relationship, so it is difficult to satisfy both recognition time and accuracy.

The present invention is intended to solve these problems, and is an excellent OCR that can improve recognition accuracy and shorten recognition time.
The purpose is to provide

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention is an OCR system that optically reads and recognizes characters to be read on a form. A pattern memory that stores character patterns created by converting them into values, and an arbitrary 2
A separation feature table in which separation features are created and registered in advance that can separate two read target characters by overlapping them at the center of the character, a judgment table for storing candidate characters, and two tables extracted from this judgment table. A feature of the present invention is that it includes a character determination unit that reads a separation feature based on a candidate character from a separation feature table, and compares and matches the separation feature with a character pattern to determine a character.

(Function) According to the present invention having the above configuration, first, the character determination section extracts two candidate characters from the determination table, and
The separation feature that separates the characters from each other by overlapping them is read from the separation feature table.

Then, a character pattern corresponding to the character to be read on the form is compared and determined based on the separation characteristics. Therefore, if it can be recognized, the resulting character is converted into a corresponding character code and output. However, if it is determined that the candidate character is not the candidate character taken out from the judgment table, candidate literature other than that character is taken out from the judgment table and the same judgment operation as described above is performed.

Therefore, the present invention can solve the above problems and provide an excellent OCR that can improve recognition accuracy and shorten recognition time.

(Example) Hereinafter, one embodiment of the present invention will be described based on the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the figure, 11 is a pattern memory, 12 is a separation feature table, 13 is a separation feature map, 14 is a determination table,
15 is a character determination section. The pattern memory 11 is composed of mxn meshes as shown in FIG. 2, and an unknown pattern, for example, a character pattern that may be the letter A, is converted into a binary value of "1" for the character part and "0" for the background part. Quantized and stored. Each point on this pattern memory 11 is
Specified by address and Y address.

FIGS. 3(a) and 3(b) are diagrams showing the appearance when two characters forming a category bear are superimposed. Figure 3(a)
3(b) shows the state when B and C are superimposed, and the areas that distinguish each other in both figures (hereinafter referred to as
(referred to as separation features) are indicated by diagonal lines. Let fAE be a separation feature that distinguishes A and E, and fBC be a separation feature that distinguishes B and C. This separation feature may be composed of a single point as shown in FIGS. 3(a) and 3(b), or may be composed of a region composed of a plurality of points.

FIG. 4 is a diagram showing a conceptual tree of the character determination method.

In the tree shown in the same figure, the characters to be recognized are 5 from A to E.
Assume that the character and unknown pattern is "D". In addition, the loco shows characters that are considered candidates for each school in the tree,
The part in C indicates the separation feature.

Initially, all characters are candidates. First, assume that the separation feature fAB that separates the category bears of characters A and B is examined, and it is determined that the characters are not, for example, B. In that case, arrow ■
The new candidates are A, C, D, excluding B,
It will be 4 characters H.

Next, the separation feature fcD of category bear CD is examined.

Since the unknown pattern is D, unlike the case of fAB, it always branches in the direction of D (in the direction of the arrow ■), and the new candidates are A, D, and E. Similarly, fAE+fDE branches in the direction of the arrow ■→■, and the unknown pattern is determined to be a character string.

In order to realize a character determination method using such a tree, the separation feature table 12 and the separation feature map 1 shown in FIG.
3 and the determination table 14 are configured as shown in FIG.

The determination table 14 includes a fixed area in which character codes A to E to be recognized are stored, and a work area in which predetermined character codes are stored in accordance with the progress of classification based on separation features.

For example, one category bear CD is taken out from the determination table 14 by the character determination unit 15 in FIG. 1, and the separated feature map 13 is drawn using the category bear as an address. The separation feature map 13 stores the start address of the separation feature fCD of the category bear in the separation feature table 12, and the start address is loaded into an address counter (not shown). The character determination unit 15 increments the contents of the address counter by 1, reads out the contents of the separation feature table shown in FIG. 6, and checks the value of the point on the pattern memory 11 of FIG. 1 indicated by the X and Y addresses. If the point value is 1, candidate character code C is stored in the work area of the determination table, and if it is 0, candidate character code D is stored in the work area of the determination table. The situation is shown in FIG. As can be seen from the figure, the character determination unit 15 in FIG. A and D are stored in the work area of the determination table 14 in order. Also, E in the fixed area and A at the top of the work area (this is based on the separation feature result of the category bear AB in the fixed area mentioned above) become the category bear AE, and the separation feature is examined, and the result is is stored in the work area. As the operation progresses in this way, the number of candidate character codes left in the determination table decreases, and the last remaining character code becomes the determination result for the unknown pattern, and the operation is similar to the operation using the tree shown in Figure 4. There will be a roar that this will be done.

(Effects of the Invention) As explained above, according to the present invention, the feature used for determination may be a feature that separates only two characters to be recognized (category bear), and unlike conventional methods, one feature can be used to separate multiple recognition target characters. Compared to separating characters, the reliability of the features is higher and higher recognition accuracy is expected. In addition, when there are N characters to be recognized, the number of branches to pass through to obtain a determination result, that is, the number of features that are actually processed, is N-1, and if the number of characters to be recognized increases by M, The processing time increases by at most 100%, and compared to the conventional method where the increase in the number of features, that is, the processing time increases by M times, the processing time is shorter and the processing speed is significantly faster. Be expected. Furthermore, when the number of characters to be recognized increases, it is only necessary to increase the category bear separation feature between the increased characters and the characters before the increase, and that feature does not have any effect on the features before the increase. Therefore, the number of target characters can be easily increased.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing the configuration of the pattern memory in FIG. 1, and FIG. 3(a)
, (b) is a diagram showing the state when two characters forming a category bear are superimposed, FIG. 4 is a diagram showing a conceptual tree of the determination method according to this embodiment, and FIG. FIG. 6 is a diagram showing the contents of the separation feature table in FIG. 1, and FIG. 7 is a diagram showing the determination operation in the determination table in FIG. 1. , FIG. 8 is a block diagram showing the general configuration of OCR, FIG. 9 is a diagram showing the state of conventional feature extraction,
FIG. 1O is a diagram showing a mask pattern in a conventional method. 11--Pattern memory, 12--Separation feature table, 13--Separation feature map, 14-Determination table, 1
5--Character determination section. v11 Confirmation M Tomo Yo Ritsuyori 1 Missing Qi 7-Jl/ no Uchigaku No. 6
Figure $ to Figure Half 1 Constant Chi 7 Tsushin no Mushi 1 Constant Weight Kai Figure 7

Claims (1)

[Scope of Claims] In an optical character reading device that optically reads and recognizes characters to be read on a form, a pattern that stores a character pattern created by binarizing data obtained by optically reading characters to be read. A memory, a separation feature table in which separation features are created and registered in advance to separate any two characters to be read by overlapping them at the center of the character, and a determination table for storing candidate characters. and 2 taken out from the judgment table.
An optical character reading device comprising: a character determination section that reads a separation feature based on the two candidate characters from the separation feature table and compares and matches the separation feature with the character pattern to determine a character.