JP2917396B2 - Character recognition method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a character recognition system in optical character recognition.

[Prior Art] With the diversification of information processing systems, various data input methods are required, and practical application of powerful data input methods of character recognition technology is progressing. However, the current character recognition technology is far inferior to humans in character reading ability, and a character reading device having higher recognition ability is desired.

In order to enhance the character recognition ability, improvements are being made in each of a pre-processing method, a normalization method, a feature extraction method, a classification / identification method, and a post-processing method that constitute the character recognition method. Conventionally, in most cases, the normalization method uses linear normalization based on alignment of entire characters and enlargement / reduction of graphics. However, for the recognition of handwritten kanji, etc., the effectiveness of nonlinear normalization, which partially scales up and down a character image and aligns it in units of strokes that compose a character, is, for example, the IEICE Pattern Recognition Research Group. Research report PRU87-104 (1988.2), Tsumo, Tanaka:
"Handwritten Kanji Recognition Based on Hierarchical Positional Correction"
And others. FIG. 2A shows an example of a character pattern obtained by ordinary linear normalization, and FIG. 2B shows an example of a character pattern obtained by nonlinear normalization.

[Problems to be Solved by the Invention] Nonlinear normalization is very effective for reading kanji having a large number of strokes and a complicated shape, but is relatively effective for characters with relatively few shapes such as numbers and few strokes. In other words, there is a problem that the distortion of the character shape deformed by the normalization is too large and adversely affects character recognition. Considering the usage of the character reading device, there are far more requests to read non-kanji characters such as numbers together with handwritten kanji characters than to read only handwritten kanji characters independently.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition method which does not adversely affect the reading of characters having simple shapes such as non-kanji characters while retaining the advantage of nonlinear normalization in handwritten kanji recognition.

[Means for Solving the Problems] A character recognition method according to the present invention comprises: an input pattern storage unit for storing a character image expressed as a two-dimensional lattice-shaped digital image; First normalization means for performing a normalization process by alignment and uniform enlargement / reduction of size and outputting a first normalization pattern; reading a character image from the input pattern storage means; A second normalization unit that performs a non-linear normalization process by appropriate scaling to output a second normalization pattern and a second feature pattern; and a first normalization pattern output from the first normalization unit. And a second normalization pattern output from the second normalization means, and extract a feature for character recognition from each of the normalization patterns to output a first feature pattern. A first character recognition dictionary for storing a reference pattern for a character type defined corresponding to the first normalization processing; and a character defined corresponding to the second normalization means. A second recognition dictionary storing a reference pattern of the species, a calculation for determining a similarity measure between the first feature pattern and each reference pattern of the first recognition dictionary, and a second feature pattern And a classification / identification unit configured to determine a similarity measure for all character types by determining a measure of similarity with each reference pattern of the second recognition dictionary and determine and output a recognition result. Have been.

[Operation] Hereinafter, the principle of the present invention will be described.

If it can be determined in advance whether to perform normal normalization or non-linear normalization on the input character pattern, the above problem can be solved. However, mixed reading is impossible. It is also conceivable to create two normalized character patterns by performing normal normalization and nonlinear normalization, and perform feature extraction, classification and identification for each. However, it is necessary to have two types of reference patterns for recognition. In addition, there arises a problem that the matching calculation between the reference pattern and the input pattern is doubled.

Therefore, when performing non-linear normalization, reference patterns obtained after performing normal normalization are registered in the first recognition dictionary for character types whose character shape is greatly distorted, and nonlinear normalization is performed for other character types. The reference pattern obtained after the execution is registered in the second recognition dictionary to cope with the capacity problem of the recognition dictionary. Also, normal normalization will be referred to as first normalization, nonlinear normalization will be referred to as second normalization, and a feature pattern obtained by performing feature extraction after the first normalization will be referred to as a first feature pattern. If a feature pattern obtained by performing feature extraction after the second normalization is referred to as a second feature, the first feature pattern performs a matching calculation with a reference pattern of the first recognition dictionary,
The second feature pattern performs collation calculation with the reference pattern of the second recognition dictionary, and addresses the problem of collation calculation.

Embodiment FIG. 1 is a block diagram showing an embodiment of the configuration of the present invention.

This character recognition system includes an input pattern storage unit 1, a first normalization unit 2, a second normalization unit 3, a feature extraction unit 4, a first recognition dictionary 5, and a second recognition dictionary. 6 and classification / identification means 7.

The input pattern storage unit 1 reads and stores a character pattern expressed as a two-dimensional lattice digital image as a signal 10, and may be a normal storage unit.

The first normalization means 2 reads a character image as a signal 11 from the character pattern storage means 1, performs a size normalization process by alignment and enlargement / reduction, and outputs a normalization pattern as a signal 12. This is a normal linear normalization means performed by character recognition.

The second normalization means 3 reads a character image as a signal 11 from the character pattern storage means 1, performs non-linear normalization, and outputs a normalized pattern as a signal 13. This can be realized by the conventional technique described in the specification of Japanese Patent No. 185826 “pattern normalization method”.

The feature extraction means 4 receives the first normalized pattern signal 12 and the second normalized pattern signal 13, performs feature extraction on both normalized patterns, and extracts the features of the first normalized pattern into the first normalized pattern. And outputs the feature of the second normalized pattern as a second feature pattern signal 15,
It can be realized by the conventional technology of character recognition. The feature extraction may be, for example, the one described in Japanese Patent Application No. 63-237064 by the same applicant, "feature extraction method", but is not limited thereto, and feature extraction generally used in character recognition can be used. is there.

The first recognition dictionary 5 stores a reference pattern of a simple character type having a shape such as a non-Kanji character, and may be an ordinary storage unit.

The second recognition dictionary 6 stores a reference pattern of a complicated character type having a shape such as a kanji, and may be an ordinary storage unit.

The classifying / identifying means 7 reads the first feature pattern signal 14, the second feature pattern signal 15, the first reference pattern signal 16, and the second reference pattern signal 17, and reads the first feature pattern and the first feature pattern. A similarity measure for all character types is calculated by performing a matching calculation that determines a measure of similarity with each reference pattern, and performing a matching calculation that determines a measure of similarity between the second feature pattern and each second reference pattern. To determine the recognition result, and the signal 18
And can be realized by the conventional technology of character recognition.

 Next, the operation of this embodiment will be described.

The first recognition dictionary 5 stores in advance reference patterns of simple character types such as non-Kanji characters, and the second recognition dictionary 6
, A reference pattern of a complicated character type having a shape such as a kanji is stored in advance.

The input pattern storage means 1 reads and stores a character pattern signal 10 to be recognized. First normalization means 2
Reads the image signal 11 from the character pattern storage unit 1, performs a size normalization process by scaling up and down, and outputs a normalized pattern signal 12. On the other hand, the second normalization means 3 reads the character image signal 11 from the character pattern storage means 1, performs nonlinear normalization, and outputs a normalized pattern signal 13. The feature extraction means 4 receives the first normalized pattern signal 12 and the second normal pattern signal 13, performs feature extraction on both normalized patterns, and outputs the first feature pattern signal 14 and the second A characteristic pattern signal 15 is output.

The classifying / identifying means 7 converts the first feature pattern signal 14 and the second feature pattern signal 15 from the feature extracting means 4, the first reference pattern signal 16 from the first recognition dictionary 5, the second recognition dictionary 6, the second reference pattern signal 17 is read, and a matching calculation is performed to determine a similarity measure between the first feature pattern and the first reference pattern, and the second feature pattern and the second reference pattern are compared with each other. A similarity measure for all character types is obtained by performing a collation calculation to determine a measure of similarity of, and a recognition result is determined, and a signal 18 is output.

[Effects of the Invention] As described above, according to the present invention, it is possible to utilize the effect of nonlinear normalization while avoiding the adverse effects of distortion on character types in which the shape of characters is greatly distorted by nonlinear normalization. It is possible to provide a character recognition method that maintains high recognition performance even in the case of mixed reading of kanji and non-kanji, and is very useful for realizing a character recognition device.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention, and FIG. 2 is a diagram showing a difference between linear normalization and nonlinear normalization. 1 ... input pattern storage means 2 ... first normalization means 3 ... second normalization means 4 ... feature extraction means 5 ... first recognition dictionary 6 ... second recognition dictionary 7 ... Classification and identification means

Claims (1)

(57) [Claims] An input pattern storage means for storing a character image represented as a two-dimensional lattice digital image, and a character image is read from the input pattern storage means.
A first normalization unit that performs a normalization process by alignment and uniform scaling of the size and outputs a first normalization pattern, and reads a character image from the input pattern storage unit,
A second normalization unit that performs a non-linear normalization process by partially enlarging and reducing the character image and outputs a second normalization pattern, and a first normalization pattern output from the first normalization unit. A second normalization pattern output from the second normalization means is input, and a feature for character recognition is extracted from each of the normalization patterns to obtain a first feature pattern and a second feature pattern. A feature extraction means for outputting, a first recognition dictionary for storing a reference pattern for a character type defined corresponding to the first normalization processing, and a feature dictionary defined corresponding to the second normalization means. A second recognition dictionary for storing reference patterns for character types, a calculation for determining a similarity between the first feature pattern and each reference pattern of the first recognition dictionary, and a second feature pattern And a measure of similarity between each reference pattern and the second recognition dictionary A character recognition system comprising classification / identification means for determining a similarity measure for all character types, determining a recognition result, and outputting the result by performing calculation.