JPH01282691A - Character recognizing system - Google Patents

Abstract

PURPOSE:To enhance reading accuracy without the increase of the number of recognizing logics by obtaining a discrepancy degree by means of a distance calculation according to the degree of the deformation of an input character pattern. CONSTITUTION:A primary classifying part 3 inputs a feature pattern from a feature extracting part 1, inputs a reference pattern from a reference pattern memory part 2, the discrepancy degree for a primary classification is calculated, and the discrepancy degree between the character code of a candidate character type and the first candidate character type is outputted. A secondary classifying part 4 inputs the feature pattern from the feature extracting part 1, inputs the discrepancy degree between the character code of the candidate character type and the first candidate character type, inputs the reference pattern of the candidate character type from the reference pattern memory part 2, sets the parameter value of dislocation to correct misregistration by the discrepancy degree of the first candidate character type, the discrepancy degree of the secondary classification is calculated, and outputs the character code of the character type, at which the discrepancy degree of the secondary classification becomes minimum, as a recognition result. Thus, the reading accuracy can be enhanced without increasing the number of the recognizing logics.

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to character recognition, and more particularly to handwritten character recognition methods.

(Prior art and its issues) With the diversification of information processing systems, various data input methods are required, and character recognition technology is being put into practical use as a powerful data input method. However, current character recognition The reality is that technology is far inferior to human reading ability in terms of character reading performance, character types to be read, and environmental constraints in which text is read.

In order to improve character reading performance, various improvements have been made to feature extraction methods and matching methods. As a result, carefully written characters, even kanji that were thought to be difficult to read, are now becoming quite readable. However, sufficient reading performance has not yet been achieved for characters with large deformations, and changing the recognition logic to read characters with large deformations may result in misreading of carefully written characters. There are problems such as: The reason for this is that the same recognition logic is used to read carefully written characters with different shapes and characters with large deformations. Therefore, it is conceivable to solve this problem by providing a plurality of recognition logics depending on the degree of deformation for carefully written characters and deformed characters.

The second solution is effective when there are only a few types of characters to be read, such as numbers and alphabetic characters, but when there are many types of characters, such as a single kanji, the problem arises that the number of recognition logics becomes enormous.

As described above, the conventional technology has had the problem of making it possible to read characters with high accuracy even when there is wide dispersion in the degree of deformation without any significant increase in the number of recognition logics.

(Means for Solving the Problems) According to the present invention, candidate character types are determined in the primary classification, and the degree of dissimilarity is calculated for the candidate character types in the secondary classification while correcting local positional deviations of character patterns. In a character recognition method, a feature extraction unit generates a feature pattern from a quantized character pattern, a reference pattern storage unit stores a reference pattern of a readable digit type, and a feature pattern is input from the feature extraction unit. Then, the degree of difference between the reference pattern of each character type in the reference pattern storage unit is detected by the first calculation method, and the candidate character types are determined by ranking them in descending order of the degree of difference. a primary classification unit that outputs a character code and a degree of difference of the first candidate character type; a feature pattern output from the feature extraction unit; a character code of the candidate character type output from the primary classification unit; The degree of dissimilarity of the first candidate character type is input, the reference pattern of each candidate character type is read from the reference pattern storage section, and the shift is performed for local positional deviation correction based on the degree of difference of the first candidate character type. The value of the range is determined, the degree of dissimilarity is determined by a second calculation method that corrects local positional deviation using the value of the shift range as a parameter, and the character of the candidate glyph type that has the minimum degree of difference according to the second calculation method is determined. It is possible to realize a character recognition method characterized by comprising a secondary classification unit that outputs a code as a recognition result, and the character recognition method is characterized in that the character recognition method is characterized in that the character recognition method is characterized by comprising a secondary classification unit that outputs a code as a recognition result, The degree of deformation can be determined, and the secondary classification section can perform recognition processing using a dissimilarity detection calculation means corresponding to the recognition logic corresponding to the degree of deformation.

(effect) The principle of the present invention will be explained below.

The feature pattern obtained from the input character pattern is expressed as f(1, jl
, the reference pattern of character type C is g I Cl (1, j) tosult, f (i, 318g lc'(i, jl,!
: tnear) The degree of dissimilarity D(f, , IcI) is defined using, for example, the city block distance of the following formula.

D(f, g”') = Σ:E If(1,j)-1
”' (i, jllll feature pattern (i, jl) After calculating the degree of dissimilarity between the reference pattern of each character type C, it is ranked according to the degree of dissimilarity, and the character types CI from the first place to the first place are ....

The primary classification unit outputs the character code of C1, the degree of difference D (f, g""l) between the first candidate character type C1 and the input character pattern. If the shape has few external shapes, the degree of dissimilarity D(t, g I Cl l ) between the input character pattern and the character type C1 takes a small value, and if the input character pattern has a shape with a large deformation, the degree of dissimilarity D(t, g I Cl l ) takes a small value. D (f, g+cl)) takes a large value.

Apart from such a formula for calculating the degree of dissimilarity, a formula for calculating the degree of dissimilarity that corrects the local positional shift of the pattern and has the range of shift as a parameter has been proposed.
Annual International Conference on Pattern Recognition Gross Seeding Volume 2 1)l
], 770-773 Shun Tsukumo and Kouasai “Nonlinear Matching Method Four
Han F1 Rinchiodo Character Recognizance”
(Jam Tsukumo and Ko
Asai“Non-Linear Matchl
ng 1 lathod For Handprin
tcdCharacter RecoHiton'7
th Inter++ationalConfere
nce on Patter++ Recognize
ition ProcediBs.

Vol, 2 pp, 770-773. '84) etc.

For example, the invention "Similarity Detection Device" (Japanese Patent Application No. 125811/1982) filed by the same applicant as the present application is one example of realization. Let the shift parameters be x, y, and the degree of dissimilarity including the parameters is D(f, g'c': x, y)
In this case, X and y are parameters that absorb the deformation of the character pattern. For recognition of character patterns with small deformation, the displacement parameters X and y may have small values.
Increasing the values of the shift parameters X and y is more effective for recognizing character patterns with large deformations, but this increases the amount of calculations and may lead to misreading if a character pattern with small deformations is input. get up. Therefore, when a character pattern with a small deformation is input, the difference calculation is performed by reducing the values of the shift parameters If the dissimilarity calculation can be performed by increasing the value of , recognition accuracy can be improved. The degree of dissimilarity D (f1g+cll) of the candidate character type obtained at the -component neck is a quantity that represents the scale of deformation of the input character pattern, so the values of the smoothing parameters D(f
, g! By setting it as a function of 6X=X (D
(f, g”■)) y=Y (D (f, g”■)) As above,
According to the present invention, the degree of difference can be determined by distance calculation according to the degree of deformation of the input character pattern, and reading accuracy can be improved by adopting the method of the present invention.

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

The feature extraction unit 1 inputs a character pattern as a signal 11, generates a feature pattern used in classification processing, and outputs it as a signal 12, which can be easily realized using technology applied to conventional character recognition devices.

The reference pattern storage section 2 stores reference patterns for each character type, and may be a normal storage means.
A reference pattern is read as a signal 14 from the reference pattern storage unit 2, a degree of dissimilarity for the -component classification is calculated, and the character code of the candidate character type and the degree of dissimilarity of the candidate character type in the -th position are outputted as a signal 13. , can be easily realized using technology applied to conventional character recognition devices. The secondary classification unit 4 reads the feature pattern as a signal 12 from the feature extraction unit 1, reads the character code of the candidate character type and the degree of dissimilarity of the primary candidate character type from the -component neck part 3 as a signal 13, and stores the reference pattern. The reference pattern 'It of the candidate character type is read as a signal 15 from the unit 2, and the value of the shift parameter for positional deviation correction is determined based on the degree of dissimilarity of the candidate character type in the second position, and the degree of difference of the secondary classification is calculated. The character code of the character type with the minimum degree of difference in the secondary classification is output as a signal 16 indicating the recognition result, the details of which will be explained next.

FIG. 2 is a block diagram showing the configuration of a specific example of the secondary classification section 4.

The shift parameter determining unit 5 reads the character code of the candidate character type and the degree of difference of the second candidate character type as a signal 13,
A functional operation is performed to determine the value of the shift parameter according to the value of the degree of difference, and the character codes of the two r candidates are output as a signal 18, and the value of the shift parameter is output as a signal 17. The reference pattern selection unit 6 reads the character code of the candidate character type as a signal 18, selectively inputs the reference pattern of the candidate character type indicated by the character code as a signal 15 from the reference pattern storage unit 2, and selects the input character type. The reference pattern is sent to signal 19.
The similarity calculation unit 7 reads the smoothing parameter of the input character pattern as t=17, determines the shift parameter for similarity calculation, reads the characteristic pattern of the input character pattern as signal 12, and reads it as signal 19. This system calculates the degree of dissimilarity between the reference pattern and the reference pattern for secondary classification, and outputs the character code of the candidate character type and the degree of dissimilarity value as a signal 20. Detection device” (Special Rp. 1986-1258)
This can be realized using the technology used in No. 11) etc. The minimum dissimilarity detection unit 8 outputs the character code of the candidate character type and the dissimilarity value to a signal 2.
0 is input, the candidate character type with the minimum dissimilarity value is detected, and its character code is output as signal 16.

(Effects of the Invention) As described above, according to the present invention, it is possible to realize character recognition that calculates the degree of dissimilarity according to the degree of deformation of input character patterns, without increasing recognition logic compared to conventional methods. This greatly helps improve character reading accuracy.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the structure of an embodiment of the present invention, and FIG. 2 is a block diagram showing the structure of a specific example of the secondary classification section in the embodiment of FIG. In the figure, 1 is a feature extraction unit, 2 is a reference pattern storage unit, 3 is a primary classification unit, 4 is a secondary classification unit, 5 is a shift parameter determination unit, 6 is a reference pattern selection unit, 7 is a similarity calculation unit, 8 is a minimum difference detection unit.

Claims (1)

[Claims] In a character recognition method that determines candidate character types in the primary classification and calculates the degree of dissimilarity for the candidate character types in the two-character classification while correcting local positional deviations of character patterns, quantization is a feature extraction section that generates a feature pattern from the character pattern that has been read, a reference pattern storage section that stores reference patterns of the character types to be read; The degree of dissimilarity with the reference pattern of each character type is detected by the first calculation method, and candidate character types are determined by ranking them in descending order of the degree of difference, and the character code of the candidate character type and the first candidate are determined. a primary classification unit that outputs the degree of difference between the character types; the feature pattern output from the feature extraction unit; the character code of the candidate character type output from the primary classification unit; and the degree of difference between the first candidate character type. is input, reads the reference pattern of each candidate character type from the reference pattern storage unit, determines the value of the shift range for the local position shift candidate based on the degree of dissimilarity of the first candidate character type, and calculates the shift range. The degree of dissimilarity is calculated using a second calculation method that corrects local positional deviation using the range value as a parameter, and the character code of the candidate character type that minimizes the degree of difference according to the second calculation method is output as a recognition result. A character recognition method characterized by comprising a next classification section.