JPH0283688A - Character recognition system - Google Patents

Abstract

PURPOSE:To improve the performance of character recognition by calculating difference between a feature pattern and a reference pattern by performing non-linear matching between each system of scalar quantity in a feature pattern and each system of scalar quantity in a reference pattern corresponding to each system of scalar quantity, and outputting a classification result by performing the sequencing of the difference. CONSTITUTION:A feature extracting part 2 reads in a character pattern signal from a character pattern storage part 1, and outputs the feature of each partial area decided by dividing a character pattern setting the feature pattern described as feature systems in plural directions as a signal. A feature pattern storage part 3 stores a feature pattern signal outputted from the feature extracting part 2, and a classification processing part 5 reads in the feature pattern signal from the feature pattern storage part 3, and reads in a reference pattern signal at every targeted kind of character from a reference pattern storage part 4. Then, the difference at every direction is found in each partial area, and the classification result is decided by performing the sequencing of the difference, and a classification result signal is outputted. In such a way, it is possible to find the difference in which the correspondence of an input character pattern with the reference pattern of each kind of character can be obtained partially, which improves the performance of the character recognition.

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to character recognition technology.

(Prior Art) With the diversification of information processing systems, various manual data processing methods are required. Among these, character recognition technology is being put into practical use as a powerful data input method. However, current character recognition technology is far inferior to humans in terms of character reading performance, and a character reading device with higher character recognition performance is desired.

In order to improve character recognition performance, configure character recognition,
Efforts are being made to improve each of pre-processing, feature extraction processing, classification/identification processing, and post-processing. Among these, regarding the feature extraction process, the effectiveness of directional features is discussed, for example, in the literature (Transactions of the Institute of Electronics, Information and Communication Engineers (D), Vol. J65-D
, No, 5. pp, 550-557. Saito, Yamada, Sansui: [Analysis of handwritten kanji using directional pattern matching method J), etc. The general method for extracting directional features is to first find the distribution of directional components on the character pattern plane, then divide the character pattern into multiple subregions, and integrate the directional components within each subregion. As a result, it is possible to extract a directional feature pattern in which the distribution of directional components is compressed.

In the classification/identification process of character recognition using such directional feature patterns, for example, the literature (IEICE Pattern Recognition and Understanding Study Group Materials, PRU87-104.pp.

41-48.1988.2. As shown in "Handwritten Kanji Recognition Based on Hierarchical Positional Displacement Correction Processing"), simple overlay processing, shift matching processing, etc. are applied. On the other hand, to improve the performance of directional features,
This application is filed by the same applicant and the same inventor as the present application, and was filed on the same day as the present application.
Attempts have been made to improve a feature extraction method that uses two-dimensionally expanding features within a partial region as a set of feature sequences in multiple directions.

(Problem to be solved by the invention) Matching is performed in the classification and identification process of character recognition, but in order to deal with various deformations of characters, it is necessary to overlap patterns partially rather than all at once. This allows for more flexible matching. Therefore, the previous output document (IEICE Pattern Recognition and Understanding Study Group material, PRU
87-104. pp, 44-48゜1988, 2. Yuyuki, Kuninaka: ``Handwritten Kanji Recognition Based on Hierarchical Positional Displacement Correction Processing'') It is conceivable that this method can be applied to feature sequences, and it can perform matching that is more accurate than simple matching, but there is a problem that it cannot deal with local deformation of character patterns.

(Means for Solving the Problem) According to the present invention, there is provided a character pattern storage unit that stores character patterns given as a two-dimensional grid array, and a character pattern is read from the character pattern storage unit and a directional component is determined. , a directional pattern is created, and the characteristics of each partial area are determined by calculating a partial projection in the scanning direction determined for each direction component in the partial area on the directional pattern corresponding to the partial area divided into multiple parts on the character pattern. a feature extraction unit that creates a feature pattern by describing information as a series of scalar quantities for each direction component; a feature pattern storage unit that stores the feature pattern obtained by the feature extraction unit; and a feature pattern storage unit that stores the feature pattern obtained by the feature extraction unit. A reference pattern storage section stores reference patterns for each character type in the same description format as the feature patterns, and a feature pattern is read from the feature pattern storage section, and a reference pattern for each character type is read from the reference pattern storage section. By reading the pattern and performing nonlinear matching between each scalar quantity series of the feature pattern and the scalar quantity series of each reference pattern corresponding to each scalar quantity series, each subregion of the feature pattern and each reference pattern is Calculate the degree of dissimilarity of 1 to 1, calculate the degree of dissimilarity between the feature pattern and each reference pattern from the degree of dissimilarity in each partial region, rank the degree of dissimilarity, and output the classification result. Can be provided.

(Operation) Hereinafter, the principle of the present invention will be explained using the drawings. Second
As shown in the figure, assume that the features of one partial region are projected in four directions: horizontally, upwardly to the right, vertically, and to the left, and are described in each direction as a series of eight elements. . The characteristics at this time are f(i, k) (i=1.2.
..., 8. = 1.2, 3.4), and f(i, 1
) is the element of the horizontal series, 2) is the element of the upward-right series, f(i, 3) is the element of the vertical series, f
Let (i, 4) be an element of the sequence in the upward left direction. The corresponding partial areas of two character patterns F and G (for example, an input character pattern and a reference character pattern) are defined as turtle + k; m + n).
+ g(It k: ”+ n) (it J = 1
v 2+ ”'8; m and n are indices indicating partial regions as shown in Fig. 3, m = 1.-, M; n = 1
．． -, N), the degree of dissimilarity between two series by an effective nonlinear matching method is given by the literature (Journal of the Acoustical Society of Japan, 27.9
．． It has been shown that the DP matching method introduced in ``Continuous word recognition based on temporal normalization of speech using dynamic programming'' by Sakoe and Chiba, pp. 483-500 (1970) is effective. Are known. The DP matching method can be realized by the calculation method using recurrence formulas (1), (2), and (3) shown below.

d(1,1,to;m,n)=lf(1,to;rn,n)
-g(1, to; m, n)l (1)D(i, j, k
; m, n) = d(i, j, k; m, n) d
(i, j-1,; m, n) (2) D(k; m, n
)=D(8,8,to;m,n)(3
) Among the above, D(k; m, n) is the partial area (m, n
) is the degree of dissimilarity of the pattern of the series in the direction of the subregion (m.

The degree of dissimilarity D(m, n) in n) is determined by the calculation method shown in the following equation (4) as the sum of the degrees of dissimilarity of the patterns of the series in each direction.

D(m, n) = ΣD(k; m, n)
(4) k=1 The degree of dissimilarity of the entire pattern is determined by calculating the sum of the degrees of dissimilarity of each partial area as shown in the following equation (5).

The classification process can be realized by performing the above calculations to determine the degree of difference between the human character pattern and the reference pattern for each character type to be read, and ranking them in descending order of degree of difference.

Note that in this description, the plurality of directions is four directions, but the present invention is not limited to this.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of the configuration of the present invention. The character pattern storage section 1 stores human character patterns, and may be a normal storage means. The feature extraction unit 2 reads the character pattern signal from the character pattern storage unit 1, and outputs, as a signal, a feature pattern in which the characteristics of each partial region determined by dividing the character pattern are described as feature sequences in a plurality of directions. It is something. The details will be explained with reference to FIG. The directional distribution extraction section 22 reads the character pattern signal from the character pattern storage section 1, extracts the distribution pattern of the directional component for each direction, and outputs the directional distribution pattern for each direction. A direction extraction algorithm is shown in , and can be easily realized by conventional techniques using ordinary logical operation elements, arithmetic operation elements, memory elements, etc. The directional distribution pattern storage section 23 stores the directional distribution patterns in each direction output from the directional distribution extraction section 22, and may be a normal storage means. The projection area storage unit 24 stores indexes of partial areas, indexes of directions, and position coordinates of projection areas for which partial projections are obtained for each direction of each partial area, and may be a normal storage means. Partial projection extractor 2
5, sequentially reading the index of the partial area, the index of the direction, and the position coordinate of the projection area from the projection area storage unit 4, and reading the corresponding directional distribution pattern from the directional distribution pattern storage unit 3 to obtain the partial projection. This outputs a feature pattern described as projection information in each direction in each partial region, and uses ordinary logical operation elements,
This can be easily realized using conventional techniques using arithmetic operation elements, memory elements, and the like. The feature pattern storage section 3 stores the feature pattern signal output from the feature extraction section 2, and may be a normal storage means. The reference pattern storage section 4 stores, as reference patterns, characteristic patterns for each character type to be read, which are written in the same format as the feature pattern storage section 3, and may be a normal storage means. The classification processing unit 5 reads the feature pattern signal from the feature pattern storage unit 3, reads the reference pattern signal for each character type to be read from the reference pattern storage unit 4, and as explained in the (effect) section, each The degree of dissimilarity of the feature series for each direction in the partial region is determined by DP matching, the degree of dissimilarity for each partial region is determined by the sum of the degree of dissimilarity for each direction, and the degree of dissimilarity for each partial region is determined by the sum of the degree of dissimilarity for each partial region. This method determines the degree of dissimilarity between the pattern and each reference pattern, determines the classification result by ordering the degree of dissimilarity, and outputs a classification result signal, and can be easily realized using technology applied to conventional pattern recognition devices.

As described above, according to the present invention, by applying nonlinear matching to the feature series for each direction in each partial region on the character pattern, it is possible to locally correspond between the human character pattern and the reference pattern of each character type. The degree of dissimilarity can be determined, which is extremely useful for improving character recognition performance.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the configuration of the present invention, and Fig. 2 explains that the characteristics of a partial area are described as a series in the horizontal direction, upward right direction, vertical direction, and upper left direction. FIG. 3 is a diagram showing an example of dividing a character pattern area into a plurality of partial areas, and FIG. 4 is a diagram showing an example of the configuration of the feature extraction section 2. In the figure, 1... Character pattern storage section, 2... Feature extraction section, 3... Feature pattern storage section, 4... Reference pattern storage section, 5... Classification processing section.

Claims (1)

[Claims] A character pattern storage section that stores a character pattern given as a two-dimensional grid array, reads the character pattern from the character pattern storage section, determines a directional component, creates a directional pattern, and divides the character pattern into multiple parts. In the partial area on the directional pattern corresponding to the partial area that has been created, the feature information of each partial area is described as a series of scalar quantities for each directional component by obtaining a partial projection in the scanning direction determined for each directional component. a feature extraction unit that creates a feature pattern by the feature extraction unit; a feature pattern storage unit that stores the feature pattern obtained by the feature extraction unit; A feature pattern is read from a reference pattern storage section storing reference patterns and the feature pattern storage section, a reference pattern for each character type is read from the reference pattern storage section, and each scalar quantity series of the feature pattern and each The degree of dissimilarity of each subregion between the feature pattern and each reference pattern is calculated by performing nonlinear matching between the series of scalar quantities of each reference pattern corresponding to the series of scalar quantities, and the degree of dissimilarity in each subregion is calculated. A character recognition method that includes a classification processing unit that calculates the degree of difference between a feature pattern and each reference pattern, ranks the degree of difference, and outputs a classification result.