JPS62130481A - Character recognition system - Google Patents

Abstract

PURPOSE:To reduce the number of checks, and to reduce the lowering of a recognition speed due to an increase in the number of character fonts by converting a feature of one kind with plural mapping conversions, and using an obtained conversion feature for the check. CONSTITUTION:The output of a photoelectric transducing part 10 is given to a feature extraction part 20, and a feature value is outputted. A discrimination part 30 consists of a discrimination mapping storage part 31, a feature conversion part 32, a dictionary storage part 33, and a distance calculation part 34, and outputs the category name of a mapping value to make the minimum distance between the mapping value of an input pattern and that of a recognition objected category at the dictionary storage part 33. A discrimination control part 40 selects discrimination parts 51-5n based on the category name from the discrimination part 30, and an effective category is outputted from selected discrimination parts 51-5n. A discrimination result editing part 60 performs the edit processing of the category name that is the output from a discrimination part 50.

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention is a character recognition method suitable for cases where the recognition target is characters of multiple fonts and recognition speed is an issue (prior art and its problems).Statistical pattern Linear discriminant analysis is a field of recognition theory, and is calculated by discriminantly analyzing the feature values extracted from the standard pattern of the recognition target category. It can be used for recognition by using the mapping average value of the number of axes of the discriminant mapping as an element of the recognition dictionary for each category.

Linear discriminant analysis is based on Makoto Nagao's ``Pattern Recognition and Shape Processing'' (March 1980, Iwanami Shoten), pages 34 to 36.
As shown on the page, it is an extension of Fisher's discriminant function for two classes to the case of multiple classes.
It can be thought of as a linear mapping that reduces the dimensionality from the set of discriminant functions and at the same time emphasizes the separation between classes.

M to extract from the standard pattern with the number of target categories
In the case of a covariance matrix, the linear discriminant mapping of M-dimensional feature values corresponding to individual features is to solve the following eigenvalue problem.

Σ8A-Σ, AΔ where Σa: inter-category co-public variance matrix (XN dimension) ΣW: within-category co-public variance matrix
(NXN dimensions) A: Eigenvector matrix (N
XN dimension) 85m1n(K-1, M) The row vector of the eigenvector matrix to be found becomes the axis of the discriminant mapping.

Identification by discriminant analysis uses the mapping average value obtained by mapping the standard pattern of each category as the recognition dictionary as described above, and uses the dictionary category name obtained by linear mapping the feature values extracted from the unknown pattern as the matching result. shall be.

The method of using discriminant analysis for character recognition is described in Nobuyuki Otsu, “Mathematical Research on Feature Extraction in Pattern Recognition,” Electronics Research Institute Report No. 818, July 1988, pages 188 to 191. The results of one discriminant analysis for characters of one font have been reported.

When there are multiple fonts to be recognized, the mapped values of the feature values from the standard pattern are dispersed, and identification using a method that uses the average value as a dictionary results in a decline in identification performance for patterns that are close to different categories.

When characters from multiple fonts are to be recognized, it is easy to prepare recognition dictionaries for each category of each font in order to recognize them without reducing accuracy. However, this method has a problem in that the recognition speed decreases as the number of fonts to be recognized increases. In order to increase the recognition speed, it is sufficient to reduce the number of dictionaries used for matching and the number of times the distance or similarity of each dictionary is calculated, so the following method is used. In the first stage, we prepare a dictionary made of decimal features that are less affected by the font, and in the next stage, we prepare a dictionary that distinguishes between similar categories in the first stage, find candidate categories by matching with the dictionary in the first stage, and then A well-known method is to check only the dictionary of the candidate category.

Also, the characteristics of the first stage and the characteristics of the second stage are generally used with different characteristics. When there are multiple feature extraction mechanisms, the problem is how to combine features.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide a character recognition method that solves the above-mentioned problems by performing step-by-step verification using only one feature.

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention provides a means for solving the above-mentioned problems. This is a character recognition method that determines the input pattern by comparing it with the converted feature values of the target standard pattern. Mapping conversion is performed using multiple vectors that are the axes of the discriminant mapping determined by discriminant analysis of the feature values. A category name is determined by matching using the conversion feature value obtained from the conversion, a second-stage mapping transformation is selected based on the obtained category name, and the feature value from the input pattern is matched using the conversion feature value found by the selected mapping transformation. The method is characterized in that the category name is obtained by using , and the obtained category name is used as the category name of the input pattern.

(Operation) In the present invention, since the transformed features obtained by transforming one type of feature through multiple mapping transformations are used for matching, only one feature extraction mechanism is required.Also, as described later, the mapping transformation axis is determined by discriminant analysis. By using it as the axis of the discriminant mapping, there is no need for a dictionary for each character font, so the number of times of matching can be reduced, and there is little reduction in recognition speed due to an increase in the number of character fonts.

(Example) Examples of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention, and the recognition target pattern is a character pattern.

The photoelectric conversion unit 10 scans the form with characters printed on it, converts the black dots of the form into binary values of "do" and the white dots of "0", and converts the obtained binary image into a binary image. Each character is cut out and output sequentially.The feature extraction unit 20 aligns the binary pattern output from the photoelectric conversion unit 10, and obtains a value that is the sum of products of the binary pattern and the two-dimensional weight mask at a predetermined position. The value of the sum of products at a plurality of positions is output as a feature value using the feature element as a feature element.Therefore, the output feature corresponds to a pattern obtained by blurring a binary pattern (blur pattern).

For example, the size of the mask M is 5X5, and the binary pattern is F.
'(i, j), the blurring pattern B(k, 1) is expressed by the following equation.

When the blur patterns B(k, 1) are arranged in a certain order, they become feature values, which can be expressed in the form of a vector.

The discrimination unit 30 includes a discriminant mapping storage unit 31, a feature conversion unit 32,
It consists of a dictionary storage section 33 and a distance calculation section 34.

The discriminant mapping storage unit 31 stores mapping vectors corresponding to the axes of mapping transformation, and the feature conversion unit 32 converts the feature values that are output from the feature extraction unit 20 with the mapping vectors of the discriminant mapping storage unit 31, and converts them into inputs. Output the mapped value of the pattern.

Assuming that the feature value is F and each axis of the mapping is H9, the shape is a vector with the number of dimensions corresponding to the number of axes of the mapping.

F-(F,,F,,...,F.)-1Hl-'
-(tb□, Hl!+...+HIa)-1 to 1'
-(K,, ni□,...,ni,)K, -[H,,
F co-H, + XF, +1(, *XF, +----+H&,
×F.

The dictionary storage unit 33 stores the mapping value obtained from the standard pattern to be recognized and the category name as a pair, and stores the mapping value and category name as a pair.
4 calculates the distance between the mapped value of the input pattern output from the feature converter 32 and the mapped value of the recognition target category in the dictionary storage unit 33, and outputs the category name of the mapped value that provides the minimum distance.

The distance S12 between the mapped values 1 and 1 is determined by the following equation.

SI! - [K'-K", K"+K" but [A, B
] represents the inner product of vectors A and B.

51 to 50 are second-stage determination units, and the determination control unit 40
controls to select which discriminating section to be driven from among discriminating sections 51 to 50 based on the category name output from the discriminating section 30. Here, the discriminating units 51 to 5n have the same configuration as the above-described discriminating unit 30, and include a discriminant mapping storage unit, a feature conversion unit, a dictionary storage unit, and a distance calculation unit, and perform similar processing. Therefore, valid category names can be output only from the discriminating section selected by the discriminating section 40 among the discriminating sections 51 to 50. The determination result editing section 60 edits the category name output from the determination section 50. That is, the determination result editing unit 60 outputs, as the recognition result of the input pattern, if there is only one type of category name from the determining unit 50, that category name is output, and if there is more than one, a rejected category name indicating that it could not be determined.

FIG. 2 is a flowchart showing one method for creating dictionaries to be stored in the discriminating units 30 and 51 to 5n. Features are extracted from the standard binary pattern of the recognition target category, and the feature value for each pattern is determined. Next, from the feature values found, F is the ratio of the between-category variance value and the within-category variance value for each feature element.
Find the ratio, and select a certain number of features or F in descending order of the F ratio value.
Features whose ratio value is greater than or equal to a certain value are selected. Discriminant analysis is performed on the feature values of only the selected feature elements to determine each axis of the mapping. Furthermore, the feature values for each category are converted to each axis of the obtained mapping, and a mapping average value is obtained by averaging the converted values. The category names and mapping average values determined as described above are used as a dictionary.

Each of the discriminating sections 51 to 5n of the discriminating section 50 has a function of making a discrimination within a set of categories in which there is a possibility that the first discriminating section 30 alone may make an error.

FIG. 3 is a flowchart illustrating an example of the procedure for classifying standard patterns of recognition target categories necessary for dictionary creation in the discriminating units 51 to 5n. Based on the discrimination results obtained, a misreading matrix is created and classification is performed by pairing mutually incorrect category names.

In the above description, the feature value is a blur pattern, but the present invention is not limited to a method using this feature.

(Effects of the Invention) As explained above, in the present invention, one type of feature is converted by multiple mapping transformations and used to collate the converted feature values obtained, so only one feature extraction mechanism is required, and Since it does not require a dictionary, recognition speed is improved compared to the conventional method of preparing a dictionary for each character font and checking it.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing an embodiment of the character recognition method of the present invention, Fig. 2 is a flow chart showing one method for creating a first-stage dictionary, and Fig. 3 is a standard pattern necessary for creating a second-stage dictionary. 2 is a flowchart illustrating an example of a classification method. In the figure, 10 is a photoelectric conversion unit, 20 is a feature extraction unit, and 3
0 is a first-stage discriminator, 31 is a discriminant map storage, 32 is a feature conversion unit, 33 is a dictionary storage unit, 34 is a distance calculation unit, 51
to 50 are the respective discrimination units in the second stage.

Claims (1)

[Claims] In a character recognition method that determines an input pattern by comparing the converted feature values obtained by mapping the feature values from the input pattern with the converted feature values of a standard pattern to be recognized that has been converted and stored in advance, mapping conversion converts the feature values. Perform discriminant analysis using multiple vectors that are the axes of the discriminant mapping, find the category name by matching the transformation feature values obtained in the first mapping transformation, and select the second mapping transformation based on the obtained category name. Then, find the category name by matching the feature values from the input pattern with the transformed feature values found by the selected mapping transformation,
A character recognition method characterized in that the obtained category name is used as the category name of an input pattern.