JPH07152863A - Character recognizing device - Google Patents

Abstract

PURPOSE:To improve a recognition rate even when a hand-written character with various shapes is a recognition object and also to keep the high recognition rate even at the time of the recognition object with various kinds of construction. CONSTITUTION:A character recognizing device is provided with a characteristic pick-up part 10 picking-out the feature of a picture in character pattern data 54c and outputting it to a recognizing part 20, and the recognizing part 20 discriminating a character code 57e based on the output of the feature pick-up part 10 and character sample data 6A which is previously registered in a storage part 52. The feature pick-up part 10 is provided with a contour line pick-up means 11 picking-up the contour line 11A of character pattern data, a contour line dividing means 12 dividing a contour line 11A into line segments 12A and a character pattern classifying means 13 dividing character pattern data 54c into groups by the number 11k of the contour lines and the number 12m of a dividing points, and the recognizing part 20 is provided with the number of pieces of character code judging means 21 equal to the number of the classified groups.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device,
In particular, the present invention relates to a character recognition device that recognizes handwritten characters.

[0002]

2. Description of the Related Art Conventionally, image data generated by shining light on a character printed or handwritten on a medium such as paper and digitizing the density of the reflected light is compared with sample data of a character registered in advance. A character recognition device that outputs a character code of a character printed or handwritten on a medium by doing so is generally known as an OCR.

Characters to be recognized are classified into printed characters printed by a printer or printed using a computer typesetting machine and handwritten characters written with some writing instrument. Printed characters are characterized by little change in shape and stable presence / absence and direction of character lines, while handwritten characters are characterized by relatively stable relative relationship and direction of character lines. Is.

However, since it is not possible to distinguish similar characters only by the direction of the character lines, it is necessary to collate different characteristics depending on the recognition target. Conventionally, for type alphanumeric characters, handwritten alphanumeric characters or type kanji characters, Character recognition methods for each recognition target have been proposed.

As an example of character recognition using handwritten numbers as a recognition target, as shown in FIG. 10, the character recognition device 50 captures an image including the recognition target and stores the captured image data 52a. The image input unit 51 stored in the unit 52, the cutout unit 53 that extracts the character pattern data 54c from the image data 52a, and the feature extraction unit that extracts the features of the image of the character pattern data 54c and outputs them to the recognition unit 70. 60 and the output of the feature extraction unit 60 and the character sample data 5 registered in the storage unit 52 in advance.
6A and the recognition unit 20 that identifies the character code 57e.

The image input unit 51 is composed of, for example, an image scanner, a CCD camera, etc., and photoelectrically converts an image of a portion including characters printed or handwritten on a medium and outputs it as an analog value. The image input unit 51 is provided with, for example, an A / D converter (not shown), and converts an analog value generated by photoelectric conversion into a digital value and saves it as image data 52a in a predetermined area of the storage unit.

The storage unit 52 stores / reproduces a digital value by using a semiconductor memory, magnetism or light. Recording unit 5
2 includes image data 52a read by the image input unit 51.
Area for storing, and character samples 56A, B,
C. ． ． Is allocated to a storage area for storing and a work area for each unit, and is controlled by a central processing unit (not shown).

The clipping unit 53 adds each pixel of the image data in the vertical direction and the horizontal direction to obtain the projection of the black portion and creates a projection histogram. Character pattern data is created by cutting out a portion of the projection histogram that is equal to or greater than a certain threshold as an area in which characters are present. Such a character cutout portion is described in, for example, Japanese Patent Application Laid-Open No. 1-114992.

The cutout section 53 is also provided with a character pattern data editing section 54 (not shown). The character pattern data editing section 54 edits the character pattern data 54a in response to a request from the feature extracting section 60. Therefore, the configuration and operation of the character pattern data editing unit 54 are different depending on the configuration and operation of the feature extraction unit 60. Here, for example, noise is removed from the character pattern data 54c generated by the cutout unit 53 by a simple filtering process, and the feature extraction unit 6
In response to the request of 0, the circumscribing of the character pattern data 54a is normalized to a fixed size, and the sawtooth noise generated by this normalization is smoothed by a 3 × 3 average value filter.

In the character recognition technology, it is not possible to register all characters as sample data.
Since it generally requires a long processing time as the amount of sample data increases, the processing is realized and speeded up by comparing only the typical character sample 56A.

In recent years, a character has been recognized by obtaining some characteristic of the character sample 56A and character pattern data and comparing the presence or absence of this characteristic and its degree. Various types of feature values have been proposed to be extracted from this character pattern data. For example, it is a method in which sample points are placed on the character pattern data on the two-dimensional coordinates at equal intervals on the coordinates, and the density values of the character patterns at the sample points are compared as characteristic values.

The recognizing unit 20 compares the above-mentioned features extracted by the feature extracting unit 10 from the character pattern data 54c with the features of the character sample data, determines whether the features are the same or similar, and determines the character pattern data. The character code 54c is recognized and output. Specifically, a method of determining the character code of a character sample based on the distribution of feature values based on a statistical method and a method of using a neural network have been proposed.

A conventional example of the feature extraction unit 60 in which character recognition is performed using a "weighted direction index histogram" as a feature value will be described with reference to FIG. In this conventional example,
The feature extraction unit 60 extracts the contour line of the character from the character pattern data 54c. Character pattern data editing unit 54
However, since the character pattern data 54c is binary data, the contour line 60A of the character portion is traced in eight connections based on 1-on and 0-off of the pixel. Here, the contour line is tracked counterclockwise and the hole is tracked clockwise.

Further, the feature extraction unit 60 is configured to detect each contour line 60.
The direction of A is quantized into four directions at intervals of 45 degrees. Here, the lower left direction and the upper right direction are treated the same. This is because the contour line is a closed curve and the starting point is arbitrary, which hinders feature extraction in eight directions. Next, the character pattern data is divided into a total of 49 small areas 61 of 7 × 7. Within each small area 61, a histogram is obtained for each direction of the pixel 62 on the contour 60A included in this small area 61. In this way, since the direction histogram of each contour line quantized in four directions is generated in 49 small regions 61 for one character pattern data 54c, 196
A direction index histogram of a dimension (49 small areas x 4 directions) can be obtained.

Next, this direction index histogram is multiplied by a two-dimensional Gaussian filter (5 × 5) with overlap to obtain 6
Dimension compression is performed in four dimensions (4 vertical x 4 horizontal small regions x 4 directions). The dimension-compressed weighted direction index histogram is output to the recognition unit 70 as a feature value.

In the conventional example using the "weighted direction exponential histogram", the recognition unit 70 uses a pseudo Bayes discrimination function for the determination of the character code 57e. The pseudo Bayes discriminant function is as follows. Where M is the sample mean value, λ
i and φi are the eigenvalues and eigenvectors of the sample covariance matrix, and h2 is a constant. Further, k is the number of main eigenvectors.

[0017]

[Equation 1]

The details of the conventional example using this "weighted direction index histogram" are described in "Handwritten digit recognition by statistical method"("Past, present, and future of handwriting character recognition technology" symposium, April 2, 1993). Sun).

A conventional example using a neural network for the recognition unit 20 will be described. The neural network processing realizes information processing by self-organizing according to the sample data for learning. The information processing based on this neural network does not need to define the processing procedure in detail like the information processing based on the Neumann type, and the organized data is arranged in the neural network that is self-organized in advance by the sample data for learning. Information processing is performed by inputting. Since learning is performed from sample data by self-organization, neural network processing is suitable for information processing such as pattern recognition and classification, and has been used for information processing of image data together with speech recognition technology and speech synthesis technology. It was In particular, various techniques have been proposed for using neural network processing for character recognition, and character recognition devices for handwritten characters have also been studied.

[0020]

However, in the conventional example using the "weighted direction index histogram", since the directional components of the outline of the character are four directions, and there is no distinction between the facing directions, the characteristics are difficult to obtain. However, there is an inconvenience that the recognition rate does not improve more than a certain value. Further, since the characteristic value is not based on the structure and stroke of the character, the deformation of the character is not sufficiently absorbed, and the recognition rate is not improved beyond a certain value.

Further, even in the conventional example of the character recognition device using the neural network processing, for example, in the character recognition using the above-mentioned density value as the characteristic value using the three-layer neural network, the number of intermediate layers is increased. However, it is known that even if the resolution of the image data is detailed, it does not further improve when a certain recognition rate is reached. This indicates that the data that could not be learned is included in the character pattern data, and a character recognition device that supports a variety of handwritten character structures cannot be provided simply by using a neural network for the recognition unit. , Also, the feature value extraction method that maximizes the effect of the neural network is not provided,
There was an inconvenience.

[0022]

It is an object of the present invention to improve the inconvenience of the conventional example, and particularly to improve the recognition rate even when handwriting characters having various shapes are recognized, and the recognition objects have various structures. Also, an object thereof is to provide a character recognition device that can maintain a high recognition rate.

[0023]

According to the present invention as set forth in claim 1 or 2, an image input section for picking up an image including a recognition target and storing the picked-up image data in a storage section; A cutout unit that extracts character pattern data from the data, a feature extraction unit that extracts the features of the image of the character pattern data and outputs it to the recognition unit, and an output of this feature extraction unit and a character sample registered in advance in the storage unit. The character recognition device includes a recognition unit that identifies a character code based on data.

In this character recognition device, the feature extraction unit extracts the contour line of the character pattern data from the contour line extraction means, and based on the feature of the extracted contour line, the contour line is formed into a line segment by a predetermined reference. And a character pattern classifying unit for grouping the character pattern data according to the number of contour lines extracted by the contour line extracting unit and the number of division points divided into line segments by the contour line dividing unit. ing.

Moreover, the recognition section is provided with the same number of character code judging means as the number of groups classified by the character pattern classifying means.

Further, as an example of a predetermined criterion for the contour line dividing means to divide the contour line into line segments, the outermost point list in which a line connecting the outer circumference of the contour line at the shortest and a contact point of the contour line is a dividing point It has a configuration such as having functions. This aims to achieve the above-mentioned object.

According to the third aspect of the present invention, an image including a recognition target is picked up, an image input section for storing the picked-up image data in a storage section, and character pattern data is extracted from the image data. A cutout unit, a feature extraction unit that extracts the features of the image of the character pattern data and outputs the feature to the recognition unit, and a character code based on the output of the feature extraction unit and the character sample data registered in the storage unit in advance. The character recognition device includes a recognition unit that performs identification.

In this character recognition device, the feature extraction unit traces the contour line of the character pattern data to obtain the directional component for each pixel, and the contour line extraction means for extracting the contour line as a closed curve, and the contour line. A contour line dividing unit that divides the contour line into line segments with a contact point between the line connecting the outer periphery of the contour and the contour line as a dividing point, the number of contour lines extracted by the contour line extraction unit, and the contour line dividing unit. And a character pattern classifying unit for classifying the character pattern data into groups according to the number of division points obtained.

Moreover, the recognition unit is provided with the same number of neural network processing units as the number of groups classified by the character pattern classifying unit, and the appearance frequency of the direction component for each pixel of the line segment is grouped by group. In addition, it is equipped with a function for creating input data for a neural network that outputs to a neural network. This aims to achieve the above-mentioned object.

[0030]

When the image input unit reads from the image data medium containing a plurality of character patterns and stores it in the storage unit, the cutout unit reads out the image data and cuts out the character pattern data from the image data in units of one character.

When the character pattern data including the character pattern separated for each character is output, the character pattern data editing unit performs the preprocessing required by the feature extracting unit. The pre-processing here is noise removal or the like.

Next, the contour line extracting means extracts the contour line of the character pattern from the preprocessed character pattern data. Further, the contour line dividing means extracts a division point from the contour line and divides the contour line into line segments. Further, the character pattern classification means determines the character pattern group to which the character pattern belongs based on the number of contour lines and the number of division points of the processed character pattern.

According to the determination of this group, the recognition unit
The character code is determined for each character pattern group. Further, the determined character code is output as a recognition result.

[0034]

The present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the functional arrangement of a character data storage device according to the present invention. The same reference numerals are given to the respective parts and the data having the same properties as in the conventional case, and the description is applied correspondingly. The character recognition device 1 captures an image including a recognition target, and stores the captured image data 52 a in the storage unit 52.
And the character pattern data 54 from the image data 52a.
The extraction unit 53 that extracts c, the feature extraction unit 10 that extracts the features of the image of the character pattern data 54c and outputs the features to the recognition unit 20, the output of the feature extraction unit 10, and the registration in the storage unit 52 in advance. The recognition unit 20 is provided for identifying the character code 57e based on the character sample data 6A.

In the character recognition device 1, the feature extraction unit 10 uses the contour lines 11A, 1 of the character pattern data 54c.
1B, ..., and the contour lines 11A, 11B, ... Are divided into line segments 12A, 12B ,. Contour dividing means 12 for
The contour lines 11A and 11B extracted by the contour line extraction means 11
The number 11k of ... And the contour line dividing means 11 is the line segment 12
The character pattern data 54 is divided into groups 13A, 13A by the number 12m of the division points 12j divided into A, 12B ,.
And a character pattern classification means 13 for dividing into B ...

Moreover, the recognition unit 20 causes the character pattern classification means 13 to classify the groups 13A, 13B,
The number of character code determination means 21A, 21 equal to the number of ...
B, ... are provided.

This will be described in detail. Feature extraction unit 10
By operating each means, the contour line 11A of the character pattern data is extracted and the contour line 11A is divided into line segments. There are various methods of division, for example, the “outermost point list method” described later, or 4 from the upper left.
When moving the line of 5 degrees to the lower right, the contour line 11
Contour line 11A with the contact point with A and the contact point increased or decreased when the number of contact points increases or decreases as a dividing point
May be divided into line segments 12A, 12B, 12C. This division is adopted in the present embodiment as long as the number of division points differs depending on the character structure.

The "outermost point list method" is one of the criteria for dividing the contour line. This will be described with reference to FIGS. FIG. 2 is a conceptual explanatory view showing an example of a predetermined standard for dividing the contour line into line segments. Just contour line 1
When a rubber band is applied to 1A, a portion where the contour line 11A and the rubber band contact each other is convex, and a portion where the contour line 11A and the rubber rubber are apart is concave, and a boundary between the irregularities is an outermost point (division point 12j). Divide the contour line into 12 j
Divided by line segments 12A, 12B, 12C, 12
Let be D.

Specifically, as shown in FIG. 3, the contour line is divided into line segments by first dividing the outermost point (dividing point 12
A point on the contour line 11A, which is a candidate of j), is obtained. This candidate point is a point at which the contour line 11A and each side of the regular hexagon first come into contact when the regular hexagon having the character pattern is reduced. A regular hexagon is an alternative to rubber bands. A straight line passing through the two candidate points is drawn in order from the start point of the contour line 11A.

Since the contour line 11A is calculated counterclockwise, the uneven side of the contour line 11A is uniquely determined with the straight line as the boundary.
Next, the distance from each point on the contour line 11A between the two candidate points to the straight line is obtained. If the point on the concave side where the distance is 1 or more is 1/4 or more of the entire contour line 11A between the two candidate points, the two candidate points are the outermost points and the contour line 11A between the two points is a line. Let it be segment 12A. If not, it is a convex line segment. If the convex line segment 12A is continuous, it is set as one line segment 12A.

The character pattern classification unit 13 converts the character pattern data 54c into the number of contour lines 11k and the number of division points 12
m is classified into groups 13A, 13B, .... The number of contour lines 11k corresponds to the presence / absence of holes in the character pattern, and the number of divisions 12m corresponds to the characteristics of the character structure. As described above, the method of dividing the contour line into line segments may be any method as long as it is a method of extracting the structural feature of the character. However, since the recognition rate of character recognition increases or decreases depending on the method of determining the division points, in the present embodiment,
The "outermost point list method" that can obtain the most desirable recognition rate is adopted.

Further, the character pattern classification means 13 determines the groups 13A and 11k by the number of contour lines 11k and the number of division points 12m.
In order to classify 13B, ..., As shown in FIG. 4, even the same character code 57e may belong to different groups. Contour line 11A that is a handwritten character of "2"
Indicates a character pattern group 13A having one contour line and four division points. Similarly, the contour line 11B belongs to a group 13b having one contour line and six division points. The recognition unit 20 determines the character code 57e for each of the groups 13A, 13B, .... Therefore, handwritten characters that are the same but have significantly different configurations are collated with the character sample data 6A and 6B according to the structure.

The recognition unit 20 uses the character pattern group 13
The number of character code determination means 2 equal to the number of A, 13B, ...
It is equipped with 1A, 21B, .... The character sample data 6A, 6B, ... To be collated is created in advance via the feature extraction unit 10 and stored in a predetermined area of the storage unit 52. Character sample data 6A, 6B, ...
Are grouped in correspondence with the character pattern groups 13A, 13B, .... Character code determination means 21
There are various seed methods for A, 21B, ... And the "pseudo Bayes discriminant function" may be used as in the conventional example. In this case, feature extraction is performed based on the direction index of the line segment, and this feature value is collated. Further, as described later, a neural network may be used to determine the character code.

Next, the processing flow of the character recognition device 1 will be described with reference to FIG. 5, and the properties of various data will be described with reference to FIG.
Will be described with reference to.

When the image input unit 51 reads the image data 52a containing a plurality of character patterns 13n from the medium and stores it in the storage unit 52 (step S1), the cutout unit 53 is formed.
However, the image data 52a is read and the character pattern data 53c is cut out from the image data 52 in units of one character (step S2).

Character pattern 1 separated for each character
When the character pattern data 53c including 3n is output,
The character pattern data editing unit 54 performs the preprocessing requested by the feature extraction unit 10 (step S3). The pre-processing here is noise removal or the like.

Next, the contour line extracting means 11 extracts the character pattern 13 from the preprocessed character pattern data 53c.
The n contour lines 11A are extracted (step S4). Further, the contour line dividing means 12 extracts the dividing points 12j from the contour line 11A to convert the contour line 11A into line segments 12a.
It is divided into A, 12B, 12c, .... Further, the character pattern classification unit 13 determines the character pattern group 13A to which the character pattern belongs based on the number 11k of contour lines and the number 12m of division points of the processed character pattern 13n (step S6).

According to the determination of this group, the recognition unit 20
Is a character pattern group 13A, 13B, 13c, ...
The character code is determined separately (step S7). Further, the determined character code is output as a recognition result (step S8).

The above-described first embodiment corresponding to the first and second aspects of the present invention is not limited by the concrete examples of the contour dividing means 12 and the character code judging means 21A, 21B. What is important is that the contour line is divided into line segments according to a predetermined standard, and the character pattern data is grouped according to the number of contour lines and the number of line segments. This is to determine the character code by using different sample data for each group. By performing feature extraction in units of this line segment, grouping according to the structure of the character, and then performing collation in each group, the shape of the character can be maintained while adopting the conventional character recognition method. Further, the recognition rate of handwritten character recognition having various structures can be improved, and a high recognition rate can be maintained constant.

Further, since the feature extraction unit extracts the feature value in units of line segments, it is possible to perform character recognition based on directional components in eight directions instead of the four directions as in the conventional example. Therefore, the stroke and curvature of the character can be used as the feature value with high accuracy, and the recognition rate is significantly improved.

A second embodiment will be described with reference to FIGS. 7 to 9.

In the second embodiment, a neural network is used to determine the character code in the recognition section 20.
The feature extraction unit 10 is also configured to maximize the effect of this neural network. The configurations and operations of the other parts are almost the same as those of the conventional example and the first embodiment.

In the second embodiment, the feature extraction unit 10 traces the contour line 11A of the character pattern data 54c to obtain the directional component for each pixel, and the contour line extraction means 11 for extracting the contour line 11A as a closed curve. , The contour line 1
A line connecting the outer circumferences of 1A, 11B, ...
The contour line dividing unit 12 that divides the contour line 11A into line segments 12A, 12B, ... Using the contact points with 1A, 11B, ... As the dividing points 12j, the number of contour lines 11k extracted by the contour line extracting unit 11, and the contour line The character dividing means 12 is provided with a character pattern classifying means 13 that divides the character pattern data 54c into groups according to the number of dividing points 12m obtained by the line dividing means 12.

Moreover, the recognition unit 20 is provided with the same number of neural network processing units 22A as the number of groups classified by the character pattern classifying unit 13, and the frequency of appearance of the direction component for each pixel of the line segment is classified by group. A neural network input data creation function 23 for outputting to the neural network processing means 22A for each line segment is provided.

This will be described in detail. In the second embodiment,
The appearance frequency (histogram) of the direction component of the line segment 12A is adopted as a feature value used for character code determination. For example, the contour line extraction unit 12 scans horizontally from the upper left pixel of the character pattern data 54c to generate an ON / OFF array of pixels, and lists the coordinates of the contour line from this array. From the coordinates of the contour line, the direction component of the contour line is obtained as shown in FIG.
Here, since we traced the outline of the outer boundary counterclockwise,
Direction components are generated in eight directions counterclockwise from the division point, as shown in FIG. As shown in FIG. 8A, if the contour line is connected to the lower left, the direction component is "2".
, And the direction component is “3” if connected downward. The closed curve is the uppermost pixel of the curve, and when there are a plurality of pixels of the same height, the direction component is taken with the leftmost pixel as the starting point.

The contour line dividing means 12 divides the contour line into line segments by the same method as in the first embodiment. After the line segment 12A is divided, the contour line dividing unit 12 determines from the data of the direction component attached to the contour line 11A that the line segment 12A shown in FIG.
The appearance frequency (histogram) of the direction component of this line segment 12A is calculated. This is shown in FIG. 8 (C). In this example, a pixel whose contour line advances downward leftward (direction component "2") appears three times, and a pixel whose contour line advances downward (direction component "3") appears twice.

Similar to the first embodiment, the character pattern data dividing means 13 divides the group into the number of contour lines 11k and the number of division points 12m (the number of line segments). By this processing, even the same character can be classified according to patterns having different characteristics. Since the character code is determined based on the character patterns having similar structures, it is possible to perform character recognition after narrowing down the character types that are candidates for correct answers. As a straightforward example, a group of two contour lines and one division point has one character type 0. (Division points are different from "8", which has one hole collapsed)

As shown in FIG. 8, the recognizing unit 20 identifies a character according to the separation of character pattern groups by using a neural network for each group. The direction component table (direction component histogram) for each line segment
The value normalized from 0.0 to 1.0 is the input data to the neural network processing means. Since the direction component has 8 dimensions, the number of line segments (the number of division points) * 8 is the input dimension.

Further, when the neural network processing is used for the cutout section, since the neuron is parallel processing, the neural network having the character cutout function as a unit is fed back from this embodiment having the character recognition function as a unit. You may comprise. For example, Hiragana 1
If there is a character that cannot be written with a stroke (for example, "i"), the cutout may fail. But,
The point to the right of "i" is likely to fail in character recognition. In this case, the neural network according to the present embodiment may be configured to provide feedback to the neural network of the clipping unit as to whether the clipping has failed.

As described above, according to the second embodiment, the parallel processing has a high processing speed, and since the learning is performed from the character sample data by self-organization, the detailed procedure concerning the character recognition will be described. There is no need to write a program), and the information processing by pattern recognition, which boasts high accuracy in information processing related to pattern recognition, brings out the maximum effect of information processing by a neural network and enjoys it. The recognition rate can be kept constant.

Further, this character recognition device may be adopted for character recognition performed by the pen input computer.

[0063]

Since the present invention is constructed and functions as described above, according to this, since the contour line dividing means divides the contour line of the character pattern, the recognition unit divides the line segment into one.
The character code can be collated with the characteristic value as a unit, and therefore, even if the character pattern having various shapes is the recognition target, the collation can be performed based on the contour direction and curvature of the character. Therefore, the performance of collating character codes is significantly improved, and the recognition rate can be dramatically improved. Further, the character pattern classifying unit divides the character patterns into groups according to the number of contour lines of the character pattern and the number of division points, so that the recognition unit can determine the character code for each group of the character patterns. Therefore, even if there are various types of character structures to be recognized, it is possible to dramatically improve the recognition rate by classifying them according to the shape of the character and then collating them in individual groups. It is possible to provide an extremely excellent character recognition device that can be maintained at the above level.

[Brief description of drawings]

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

FIG. 2 is an explanatory diagram showing an operation concept of an example of dividing a contour line in a feature extraction unit.

FIG. 3 is an explanatory diagram showing a specific processing example of the operation shown in FIG. 2 of the contour line dividing means.

FIG. 4 is an explanatory diagram showing an example of character pattern data having the same character code but belonging to different groups.

FIG. 5 is a flowchart showing the operation of the first embodiment.

FIG. 6 is an explanatory diagram showing the nature of data.

FIG. 7 is a block diagram showing a configuration of a second embodiment.

FIG. 8 is an explanatory diagram showing direction components of a line segment,
(A) shows the encoding of the direction component, (B) shows the direction component of the contour line, and (C) is an explanatory diagram showing the appearance frequency of the direction component.

FIG. 9 is an explanatory diagram showing a part of the configuration of a recognition unit in the second embodiment.

FIG. 10 is a functional block diagram showing a configuration of a conventional example.

FIG. 11 is an explanatory diagram showing an operation of a feature extraction unit in a conventional example.

[Explanation of symbols]

 10 Feature Extraction Unit 11 Contour Line Extraction Means 11A Contour Lines 11k Number of Contour Lines 12 Contour Line Division Means 12A Line Segments 12j Division Points 12m Number of Division Points 13 Character Pattern Classification Means 13C Group of Character Pattern Data 20 Recognition Units 21A Judgment Means 22A Neural network processing means 23 Neural network input data creation function 51 Image input section 52 Image storage section 52a Image data 53 Cutout section 54c Character pattern data 56d Character sample 57e Character code

Claims (3)

[Claims] 1. An image input unit that captures an image including a recognition target and stores the captured image data in a storage unit, a cutout unit that extracts character pattern data from the image data, and the character pattern data. A feature extraction unit that extracts the features of the image and outputs it to the recognition unit, and a recognition unit that identifies the character code based on the output of the feature extraction unit and the character sample data registered in the storage unit in advance. In the character recognition device provided, the feature extraction unit divides the contour line into line segments according to a predetermined criterion based on the feature of the extracted contour line and a contour line extraction unit that extracts the contour line of the character pattern data. A character pattern according to the contour dividing means, the number of contours extracted by the contour extracting means, and the number of division points divided into line segments by the contour dividing means. Character recognition means for grouping the data into groups, wherein the recognition section comprises a number of character code determination means equal to the number of groups classified by the character pattern classification means. . 2. The outermost point list function according to claim 1, wherein the contour line dividing means has a function of listing the outermost points of a line connecting the outer circumference of the contour line in the shortest and a contact point of the contour line as a division point. Character recognizer. 3. An image input unit that captures an image including a recognition target and stores the captured image data in a storage unit, a cutout unit that extracts character pattern data from the image data, and the character pattern data. A feature extraction unit that extracts the features of the image and outputs it to the recognition unit, and a recognition unit that identifies the character code based on the output of the feature extraction unit and the character sample data registered in the storage unit in advance. In the character recognition device, the feature extraction unit traces the contour line of the character pattern data, obtains the directional component for each pixel, and extracts the contour line as a closed curve; A contour line dividing unit that divides the contour line into line segments with a contact point between the line connecting the outer circumferences at the shortest and the contour line as a dividing point, and a contour line extracted by the contour line extraction unit. And a character pattern classifying unit that classifies the character pattern data into groups according to the number of division points obtained by the contour dividing unit, and the recognition unit has a number equal to the number of groups classified by the character pattern classifying unit. And the neural network input data creation function for outputting the appearance frequency of the direction component for each pixel of the line segment for each group to the neural network. Character recognizer.