JP2564276B2 - Pattern recognition method and device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition device, and more particularly to an improvement of a character recognition device that employs a large classification logic suitable for handwritten Chinese character recognition.

[Conventional technology]

A conventional character recognition device is configured to classify or identify an input character by matching by comparing each element as a stroke density and a vector of a function, as described in JP-A-55-134484. This stroke density function is a function that expresses the complexity of the character graphic with respect to the coordinate axes, views the character from the horizontal or vertical coordinate axis direction, and counts the number of character lines that form the character along the coordinate axis.
It is extracted as a vector on the coordinate axis of the character graphic.

 Here, the difference between classification and identification will be described.

Classification refers to the operation of determining multiple candidates including the input category from all target category sets, and is often used to increase the processing efficiency of the entire recognition system in a kanji recognition device with a large number of target character types. It is a method. That is, the number of target candidates is limited by using a standard pattern having a small amount of information for classification before comparison between a standard pattern having a relatively large amount of information for identification and an input pattern, and then identification is performed. Is. In some cases, the classification is also called a large classification.

Identification refers to the operation of uniquely determining the input category from all the target category sets, and it is necessary to capture the detailed features of the input pattern, so a standard pattern with a relatively large amount of information is prepared. In some cases, identification is also called recognition.

Since the above-mentioned conventional technique can extract a stroke density function having an arbitrary amount of information, it can be applied to both classification and identification. That is, if the quantization accuracy of the stroke density function is increased, a fine portion of the character structure can be expressed, which is a feature suitable for identification. Also, if the quantization accuracy is small, it is possible to represent the overall shape of the character structure, which is a feature suitable for classification.

Further, in the above-mentioned conventional technique, the character line segment of the Chinese character has a very simple structure to a complicated one, and uses this various character line density distribution. Due to this property, categories having different character line density distributions can be easily separated. That is, it is possible to classify only the categories in which the number and the position of the character line segments projected on the coordinate axes are different. However, when extracting the stroke density function, the direction of the character line segment, which is one of the basic constituent factors of the character, has not been considered.

[Problems to be Solved by the Invention] In the above-described conventional technology, the directionality of a line segment, which is one of the basic constituent factors of a character line segment, is not reflected in the feature amount, and it is the same from figures of different shapes. There was a problem of so-called degeneracy, in which the feature of was extracted. That is, it means that characters having different shapes occupy the same position in the feature space and it becomes impossible to distinguish between these characters. Therefore, such degeneracy is considered to be a serious defect in recognizing characters. The problem of this prior art will be described with reference to FIG.

FIG. 2 shows an example in which the same stroke density function is extracted from different figures. Figure 2 (a) is an L-shaped figure
22 and the stroke density function 21 extracted on the 3 axis are shown,
FIG. 2 (b) shows the stroke density function 23 extracted on the three axes as with the shaded figure 24. The stroke density function is
In order to count the number of intersections of line segments on the cross section orthogonal to the coordinate axes, a stroke density function corresponding to a height of 1 is extracted for both figures on all three axes.

As described above, in the stroke density function, the same feature may be extracted from figures having different shapes, and this property has a possibility of being a serious defect in constructing a recognition device.

An object of the present invention is to solve the above-mentioned problems when extracting the characteristics of handwritten Chinese characters, so that not only the same characteristics are extracted from characters of different shapes, but also the characteristics to be extracted are not only complicated and their positions. It is an object of the present invention to provide a pattern recognition device capable of extracting features reflecting the directionality of character line segments and recognizing handwritten Chinese characters at high speed and with high accuracy.

[Means for solving problems]

The first aspect of the present invention, as means for solving the above-mentioned problems, is to obtain pattern information to be a recognition target by observing an input pattern, and use the recognition target pattern information as a standard for each predetermined category. In the pattern recognition method for comparing the pattern information with the pattern information to classify the input pattern into the category to which the input pattern belongs, the input pattern is scanned for each raster, and the contour tracing is performed along the input pattern from the intersection with the input pattern. Obtain the direction, and from the direction of the contour obtained above, the X and Y coordinate axes and two 45
For each of the four oblique axes, the direction component of the contour parallel to each axis is extracted, the number of the extracted direction components is counted, and the direction component of the contour of the input pattern counted for each of the four axes is calculated. It is characterized in that the input patterns are classified into categories to which they belong.

Further, a second invention of the present application comprises: a means for observing an input pattern; a means for extracting features from the input pattern obtained by the observation to extract pattern information to be recognized; and the recognition pattern information,
In a pattern recognition device provided with means for classifying the input pattern into categories to which the input pattern belongs, by comparing with standard information for each predetermined category, the input pattern is scanned for each raster, and the input pattern is detected from the intersection with the input pattern. A means for determining the direction of the contour by tracing the contour along with the X and Y coordinate axes and two 45
Means for extracting the direction component of the contour parallel to each of the four oblique axes and counting the number of the extracted direction components, and the contour of the input pattern counted for each of the four axes. Means for comparing the input pattern with a feature dictionary according to the number of directional components and classifying the input pattern into a category to which the input pattern should belong.

In each invention of the present application, the classification result may be collated with the identification result of the recognition target pattern in which another type of feature is extracted, which is obtained by another aspect, to make the final determination.

[Action]

In order to achieve the above-mentioned object, the pattern recognition method and apparatus of the present invention utilize the property that the contour reflects the direction of the character line segment. That is, the contour and its direction are obtained from the input character pattern, and only the direction component of the contour parallel to two or more coordinate axes is counted for each coordinate axis, and this is used as the feature amount. Here, the two or more coordinate axes indicate a vertical axis, a horizontal axis, an oblique axis, and the like.

With this improvement, it is possible to extract the features that reflect not only the complexity of the input character and its position but also the directionality of the line segment, and it is possible to extract the features that can easily identify the characters of different shapes. In this specification, this is called a modified linear density feature expressing the complexity and directionality of a character graphic.

 Hereinafter, the principle of the present invention will be described with reference to the drawings.

FIG. 2 shows an example in which the same stroke density function is extracted from different figures as described above. FIG. 1 shows an example in which different modified linear density features are extracted from the same figure as in FIG. Figure 1 (a) shows an L-shaped figure 12.
And the deformed linear density feature 11 extracted on the three axes, and FIG. 1 (b) shows the deformed linear density feature extracted on the three axes similarly to the shaded figure 14.

Since the deformed linear density feature counts only the direction component of the contour parallel to two or more coordinate axes for each coordinate axis, different features are extracted from both figures unlike the stroke density function. That is, in the L-shaped figure, the deformed linear density feature corresponding to the height 2 is extracted on the vertical axis and the horizontal axis, and is not extracted on the diagonal axis. On the other hand, the shaded figure has a height of 2 on the diagonal axis.
The deformed linear density feature corresponding to is extracted and is not extracted on the vertical axis and the horizontal axis.

As is clear from these examples, even a character figure that cannot be distinguished by a stroke density function that does not consider the directionality of character line segments can be easily distinguished by using a modified linear density function that reflects the directionality. .

Thus, according to the present invention, the contour and its direction are obtained from the input character pattern, and only the direction component of the contour parallel to the two or more coordinate axes is counted for each coordinate axis. Features that reflect not only the position information but also the directionality of line segments can be extracted, and in particular, the classification ability or identification ability for handwritten Chinese characters can be dramatically improved. Thus, with a few improvements,
The practical advantage of improving the ability to classify or identify features is tremendous.

〔Example〕

Embodiments of the present invention will be described with reference to the drawings. Note that the following embodiments are examples applied to recognition of handwritten Chinese characters, but of course they can be applied to other pattern recognition.

<Structure of Embodiment> FIG. 3 shows the structure of an apparatus for carrying out the pattern recognition method of the present invention, which is an embodiment of the pattern recognition apparatus.

In the figure, the apparatus of this embodiment includes a character observing section 51 that functions as an observing means for an input unknown pattern, a pre-processing section 33 that pre-processes an observed character pattern into a normalized pattern, and the above normalization. A large classification unit 52 that classifies and selects target candidates by extracting features from patterns to form recognition target pattern information and comparing the pattern information with a standard pattern having a small amount of information for classification; A candidate buffer 37 that stores and holds candidates, and a recognition unit that extracts a feature from the normalized pattern and compares the feature with a standard pattern having a relatively large amount of information while referring to the target candidate to obtain a similarity. 53, a determination unit 41 that determines whether the category is a category to which an input pattern belongs from the similarity and the category information obtained by the recognition unit 53, and control that controls the operation of each unit. Configured with a 42.

The character observing section 51 includes a photoelectric conversion section 31 and an analog / digital (A / D) conversion section 32.

The large classification unit 52 includes a modified linear density extraction unit 34 and a matching unit 35.
And a feature dictionary 36. The characteristic dictionary 36 stores standard patterns for classification that have a relatively small amount of information.

The modified linear density extraction unit 34 is configured, for example, as shown in FIG. That is, the character pattern buffer 61 that temporarily holds the normalized pattern, and the contour extraction circuit 62 that extracts the contour of the normalized pattern in each direction.
And a direction plane buffer 63 holding the direction-specific extracted data
And a rotation circuit 64 for rotating the data for each direction surface of the direction surface buffer to make the coordinate axes common, and an adder 65 for counting the number of contours for each direction surface, that is, the modified linear density feature.
And a modified linear density buffer 66 that stores and holds the counted modified linear density features.

<Operation of Embodiment> Next, the operation of the present embodiment will be described with reference to the drawings.

In FIG. 3, the character pattern written on the paper or the like is converted into a video signal by the photoelectric conversion unit 31,
The analog-to-digital conversion unit 32 performs sampling and quantization to form a binary mesh pattern.

The observed character pattern is subjected to a series of preprocessing such as clipping, noise removal, and normalization in the preprocessing unit 33, and becomes a normalized pattern.

The cutout is a process of taking out one character from the character pattern group on the paper because the processing unit of recognition or the like is for each character, and usually cuts out an area of about 100 × 100 mesh including one character.

Further, noise removal refers to a process of removing a pattern other than a character pattern such as dust adhering to a paper surface, etc. This is difficult to distinguish from a character's original pattern such as a dakuten, and usually various measures are taken. There is.

Finally, the normalization process refers to a process of aligning characters to a certain size for facilitating recognition, and a technique of aligning with a circumscribing frame or a technique of aligning centroids is selected according to the recognition technique.

Next, the preprocessed normalized pattern is added to the modified linear density extraction unit 34, and the modified linear density extraction unit 34 extracts modified linear density features on the coordinate axes corresponding to the vertical, horizontal, and diagonal directions. It

The extracted deformed linear density feature is added to the matching unit 35,
The matching unit 35 calculates the similarity with the standard pattern for large classification stored in the feature dictionary 36.

The matching unit 35 first reads the modified linear density features of the input character pattern counted on the four coordinate axes from the feature buffer. Next, the standard patterns for large classification stored in the feature dictionary 36 are sequentially read out, the degree of similarity with the features of the input character pattern is obtained, and those satisfying the preset criteria are stored in the candidate buffer 37 together with the category information. Add.

On the other hand, the preprocessed normalized pattern is also added to the recognition feature extraction unit 38, and the recognition feature is extracted in the feature extraction unit 38.

Any feature can be used as the extracted feature as long as one character type can be recognized from the candidate character group narrowed down by the large classification. For example, in order to reflect the detailed structure of the input character pattern, it is conceivable that the modified linear density feature has a finer quantization accuracy.

The extracted recognition feature is added to the matching unit 39, and the matching unit 39 has a similarity between the recognition standard pattern stored in the feature dictionary 40 and the standard pattern stored in the candidate buffer 37. It is calculated.

The matching unit 39 first reads the recognition feature extracted from the input pattern from the feature extraction unit 38. Next, referring to the candidate buffer 37, the corresponding standard patterns stored in the feature dictionary 40 are sequentially read out, the similarity between the features of the input character pattern is obtained, and the similarity and the feature dictionary are stored. It is added to the determination unit 41 together with the category information that exists.

The determination unit 41 first selects the category with the highest degree of similarity using the degree of similarity and the category information. Next, the absolute value of the difference between the similarity having the next largest value and the maximum similarity is calculated. If the absolute value of this difference is larger than a preset threshold value, the category having the maximum similarity is accepted. That is, it is determined that the category having the maximum similarity is the input character category, and the category is output. On the contrary, if the absolute value is smaller than the preset threshold value, the category having the maximum similarity is rejected. That is, it is determined that the category having the maximum similarity is not the input character category, and the rejection code is output.

Next, details of the operation in the modified linear density extraction unit 34 will be described.

FIG. 4 is a detailed block diagram showing the configuration of an example of the modified linear density extraction unit.

In FIG. 4, the preprocessed normalized pattern is sent to the contour extraction circuit 62 via the character pattern buffer 61. Here, four character patterns are created for each direction and added to the four direction surface buffers 63.

The contour extraction circuit 62 first performs contour extraction. That is, the scanner is sequentially scanned downward from the upper left of the character pattern for each raster, and from the point of hitting the character pattern, contour tracing is started along the character pattern. At the same time, the direction and coordinates of each contour point are detected, and the character pattern is described on the corresponding direction surface. When the creation of the direction surface for one character is completed through the above procedure, the character patterns separated in each direction are added to the corresponding direction surface buffer 63.

Next, the character pattern applied to the direction cap 63 is applied to the adder 65 via the rotation circuit 64. Adder
At 65, the deformed linear density features for each coordinate axis are counted and finally added to the deformed linear density buffer 66.

In the rotation circuit 64, the modified linear density feature of each coordinate axis is added by an adder.
In 65, rearrange so that the same procedure can be performed.
For example, no rotation is applied to the directional surface for extracting the modified linear density feature on the ordinate, and 90 ° is applied to the directional surface for extracting the abscissa feature. According to the same procedure, the calculation directions of the remaining two planes are adjusted, the number of contours in each direction is counted by the adder 65, and stored in the modified linear density buffer 66.

In the above description, the deformed linear density feature is extracted via the direction plane, but it may be extracted directly from the normalized character pattern.

Further, in the description of FIGS. 3 and 4, the control unit 42 supplies a control signal to each unit of the circuit, and a signal notifying the state of each unit of the circuit is given to the control unit 42. Since these are not always necessary, they are not shown and described for simplification.

As is clear from the above description, according to this embodiment,
The feature is extracted by finding the direction component of the character pattern and counting only the direction component of the contour parallel to at least two coordinate axes for each coordinate axis, and the directionality of the character line segment is reflected in the feature for major classification. As a result, it has the effect of classifying handwritten kanji with extremely high efficiency.

<Other Embodiments> In the above embodiment, both the large classification unit 52 and the recognition unit 53 are provided. However, in the large classification unit 52, when the candidates are narrowed down to one, the recognition unit 53 is omitted. May be

Further, in the above embodiment, the case of recognizing handwritten Chinese characters has been described, but the present invention is not limited to this, and can be applied to other pattern recognition.

〔The invention's effect〕

According to the present invention, it is possible to reflect the directionality of a character line segment that effectively expresses the characteristics of a handwritten Chinese character in the characteristics for large classification or identification, so that a large amount of processing is required for the handwritten Chinese character. In recognition, it is possible to recognize with extremely high speed and high accuracy.

[Brief description of drawings]

FIG. 1 is a principle diagram of a feature extraction method in a pattern recognition method according to the present invention, FIG. 2 is a principle diagram of a conventional feature extraction method, and FIG. 3 is an embodiment of a pattern recognition method and apparatus according to the present invention. FIG. 4 is a block diagram shown in FIG. 4, and FIG. 4 is a block diagram showing an example of a feature extraction unit which is a main constituent part of the above embodiment. 11 …… Deformed linear density feature 12 …… L-shaped figure, 13 …… Diagonal line figure 21 …… Stroke density function 22 …… L-shaped figure, 23 …… Diagonal line figure 31 …… Photoelectric conversion part, 32 …… A / D Conversion unit 33 …… Preprocessing unit, 34 …… Modified linear density extraction unit 35 …… Matching unit, 36 …… Feature dictionary 37 …… Candidate buffer, 38 …… Feature extraction unit 39 …… Matching unit, 40 …… Features Dictionary 41 …… Decision unit, 42 …… Control unit 51 …… Character observation unit, 52 …… Large classification unit 53 …… Recognition unit 61 …… Character pattern buffer 62 …… Contour extraction circuit, 63 …… Direction plane buffer 64 ...... Rotating circuit, 65 …… Adder 66 …… Modified linear density buffer

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Yoshihiro Yokoyama, inventor Yoshihiro Yokoyama, 292 Yoshida-cho, Totsuka-ku, Yokohama Inside Microelectronics Device Development Laboratory, Hitachi, Ltd. (72) Inventor Koichi Okazawa 292 Yoshida-cho, Totsuka-ku, Yokohama Shares Company Hitachi, Ltd. Microelectronics Equipment Development Laboratory (72) Inventor Shozo Kadota 2880 Kozu, Odawara-shi Hitachi Ltd. Odawara Koba

Claims (2)

(57) [Claims] 1. A pattern information to be recognized is obtained by observing an input pattern, the pattern information to be recognized is compared with standard pattern information for each predetermined category, and the category to which the input pattern belongs should be compared. In the pattern recognition method to classify, the input pattern is scanned for each raster, contour tracing is performed along the input pattern from the intersection with the input pattern to obtain the contour direction, and from the obtained contour direction, X, Y With respect to each of the coordinate axes and the four axes of the two 45-degree oblique axes, the direction components of the contour parallel to the respective axes are extracted, the number of the extracted direction components is counted, and the input pattern counted for each of the four axes is extracted. A pattern recognition method characterized by classifying an input pattern into a category to which the input pattern should belong, according to the number of direction components of the contour of the. 2. A means for observing an input pattern, a means for extracting pattern information to be a recognition target by extracting features from the input pattern obtained by the observation, and the recognition pattern information for each predetermined category. In a pattern recognition device having means for classifying the input pattern into a category to which the input pattern belongs, the input pattern is scanned for each raster, and contour tracing is performed along the input pattern from the intersection with the input pattern. Means for performing the contour direction and extracting the direction component of the contour parallel to each of the four axes of the X and Y coordinate axes and the two 45-degree oblique axes from the obtained contour direction, The input pattern is characterized by means for counting the number of the extracted directional components, and the number of directional components of the contour of the input pattern counted for each of the four axes. Pattern recognition apparatus characterized by comprising a means for classifying the compared category should have the input pattern belongs as written.