JPH01255080A - Optical character reading system - Google Patents

Abstract

PURPOSE:To reduce reading omission without increasing misreading by dynamically changing a feature value for the dispersion of a pattern and making larger the similarity degree. CONSTITUTION:In a matching retrial part 18 and an increase and decrease propriety table 19, the feature values are successively taken off from the features of an unknown pattern and a dictionary, and the total matching distance is obtained. Next, the feature value for each feature number stored in a table is taken off. Then, the value obtained by adding or subtracting the fine value to or from the feature value is substituted for a function, to matching distances are obtained, and the smaller matching distance is obtained. The feature value stored into the table and the matching distance are subtracted from the whole matching distance, and further, the obtained smaller matching distance is added. These processings are successively executed, the corrected matching distance is obtained, and sent to a decision part 15, and when the distance is smaller than the value set beforehand, it is decided to be a candidate character, and outputted from an answer output part 16.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an optical character reading system, and more particularly to an optical character reading system suitable for character recognition of deformed patterns in which the font is significantly deformed.

[Conventional technology]

In conventional optical character reading systems, when characters are recognized in a deformed pattern in which the typeface is significantly deformed, the deformed pattern is registered in a dictionary. That is, in order to absorb the variation in each characteristic value '· with respect to the deformed pattern, the reading range of each characteristic value was widened. As related to such prior art, for example, there is an invention disclosed in Rima No. 169289/1989.
ILI [Problem to be Solved by the Invention] In the above-mentioned conventional technology, since the variations in each feature value are diverse, the dictionary must be huge in order to absorb all the variations in each feature value of the deformation pattern. There was a problem with this. Also, 1.Using such a dictionary will reduce the number of non-readers, but it is a mistake.

There was a problem that the number of readers increased.

The present invention was made in view of the problems of the prior art described above, and it is an object of the present invention to provide an optical character reading method that does not require a huge dictionary and can reduce misreading without increasing misreading. It is in.

[Means to solve the problem]

The optical character reading method of the present invention extracts a plurality of feature values from an input character pattern, and extracts each of the plurality of features/feature values for each character stored in a dictionary prepared in advance and each of the above extracted feature values. This method is applied to the optical character reading method, which calculates the similarity for each feature, sums up each calculated similarity to obtain the overall similarity, and determines candidate characters. For at least one of the features with particularly low similarity, the extracted feature value is increased or decreased.

The similarity was calculated again using this recalculation.

It is characterized by determining the overall similarity using the similarity and determining candidate characters.

[Effect] 1.
Generally, a dictionary for one character is made up of many features, and if the degree of similarity between some features happens to be extremely small, the dictionary becomes illegible. However, in the case of the same category, the degree of similarity increases by slightly increasing or decreasing the feature value. On the other hand, in the case of different categories and 1 gori (characters are different), the degree of similarity does not increase. Therefore, by slightly increasing or decreasing the feature values, the inherent discrimination ability of the dictionary is degraded.

That is very unlikely. Therefore, increasing or decreasing the feature value by a small amount for features with low similarity that cannot be matched will absorb the irregularity of the pattern features, which may increase misreading. Unreading can be reduced without any problems.

〔Example〕

The present invention will be described below in more detail in 1.1 according to the embodiments shown in the accompanying drawings. In FIG.

Light is irradiated onto the reading area of the form, and the reflected light from the form is received and converted into an electrical signal. The light reflected from a form reflects the shading information of the characters, figures, and patterns on the form, and the electrical signal is the pattern.

This is a signal that corresponds to the mode of operation.

The preprocessing unit 12 performs a process of binarizing the electrical signal and storing the result in an internal pattern memory.

Furthermore, in this preprocessing section 12, the pattern memo 11
．． .

, 6.

A pattern for one character is cut out from the blank and sent to the feature extraction section 13.

The feature extraction unit 13 extracts features such as the number of inner rings and outer rings, and the positional relationship between line segments for one character's human cover pattern,
Perform rough pattern classification. Next), the matching unit 14 extracts features that are effective for reading the classified patterns, and matches the extracted feature values with the dictionary feature values in the dictionary to obtain a matching distance. Ma.

The etching distance is a property that is opposite to the above-mentioned similarity. .

The smaller the value, the more similar it is.

It means that. Pairs with characters registered in the dictionary.

All the matching distances are smaller than a predetermined value, and all the matching distances are selected as candidates.
1° determination unit 15
If multiple candidates are output, P.

The output unit 16 performs processing to narrow down the candidates based on a determination, further determines whether the final candidate is valid, and outputs the answer.
give it to

In the above, the matching unit 14 and the determining unit 15 select 51 candidates.
, 1.4.

If the output is not output, the matching retry section 1B performs moral degree matching. Due to this now.

Characters that were not output as candidates until then are output as candidates, and as a result, there is an effect that reading is improved.

Next, the operation of the matching retry unit 1 day and the increase/decrease table 19 will be explained using FIGS. 2 and 5. As shown in Figure 2, the characteristics of unknown pattern 21
The feature values x i (i=1.2, 5.-)− are sequentially extracted from the dictionary features 22, and the feature values y i (i=
1.2,5. ) are taken out one after another, and the results are displayed.

Substituted into the number f(xi, yi), the matching distance.

The 0 matching distances Di for which Di (i=1.2, 3, . . . ) are determined are sequentially added, and the total matching distance is determined. At this time, as shown in Figure 2.

feature numbers in descending order of matching distance Di.

i and matching distance Di are sorted in table 25. In this embodiment, a maximum of five items are sorted, but this number may be arbitrary.

Next, as shown in FIG. 3, the table 25 stores 2. .

For each feature number l, extract the feature value xi of the unknown pattern feature 21, and find the same and feature value xi.

For number i, take the feature value yi from the dictionary features 22.

put out. Next, the function f □・(xi±
Δd, yi) and the two matching distances d1.
Find d2. Next, the two matching distances d
1. Find the smaller pine cutting distance of d2.

Then, what is stored in the table 25 and the mark y1,1 matching distance Di are subtracted from the overall matching distance/distance, and the above processing is further performed.

The smaller ma-soching distance (, d 1 or d
.

2) is added. The above process is executed sequentially for the feature number 1 stored in the table 25 to obtain the corrected overall feature distance D'. The 1° corrected matching distance D' thus obtained is sent to the determination section 15 and corrected.

It is determined whether the matching distance [)' is smaller than a predetermined value, and if it is smaller, the character is selected as a candidate character.

Here, the increase/decrease possibility table 19 has the feature value xiv2L△
Whether or not it is possible to increase or decrease by d is stored for each character feature value, and the matching retry unit 18 performs processing by referring to the increase/decrease table 19. This depends on the type of character. is the feature value Δd.

This is because an increase or decrease may actually lead to erroneous determination5. Such characters include, for example, the numbers ``1-1''
and similar characters like ``ノ'' in katakana. This increase/decrease permission table 19 has an increase/decrease flag and a ``1'' to indicate that the increase is allowed or a decrease is allowed, respectively, and a 0 to indicate that the value is not allowed. The ILI increase/decrease possibility table 19 is for one feature number.

It is composed of 2 bits.

Through this processing, characters that were not output as candidates in the conventional technology are now output as candidates, and as a result, the reading efficiency is improved.

Ru.

〔Effect of the invention〕

According to the present invention, it is possible to dynamically change the feature value in response to pattern variations, thereby reducing the number of misreadings without increasing the number of misreadings. 11.7.

There is.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIGS. 2 and 3 are explanatory diagrams showing the operation of the embodiment shown in FIG. 1. 11...Scanning section/photoelectric conversion section, 12...Preprocessing section/
Feature/extraction unit, 14... Matching unit, 15... Judgment unit, 16... Answer/e output unit, 17... Dictionary unit, 1
8...Matching try section-19...Increase/decrease possibility table. , 8

Claims (1)

[Claims] 1. Extract multiple feature values for the input character pattern,
Based on the feature values for each character stored in a dictionary prepared in advance and the feature values extracted above, the similarity is calculated for each feature, and the overall similarity is calculated by summing each similarity, and the candidate is selected. In the optical character reading method that determines characters, the similarity is calculated again using the extracted feature value increased or decreased for at least one of the features with low similarity, and this recalculated similarity is An optical character reading method characterized by determining candidate characters by determining overall similarity.