JPH0789364B2 - Pattern recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern recognition device as a Japanese input device, and particularly to the input of a pattern in which many very similar character groups such as handwritten kanji exist. The present invention relates to a suitable pattern recognition device.

[Conventional technology]

Generally, in a character reading device, there is a method of identifying an input unknown pattern by matching the input unknown pattern with a standard pattern, converting the input unknown pattern into a characteristic code string, and passing through a sequential logic circuit. However, in any of the methods, it may be difficult to determine which category an input unknown pattern belongs to because it is accepted as two or more standard patterns. Usually, such a pattern is rejected as a misleading pattern.

There are the following three problems that must be solved in order to put the handwritten kanji recognition device into practical use.

(1) Since handwritten Chinese characters are complicated and have various transformations, a recognition method that can tolerate these characteristics is necessary.

(2) The recognition target character type has more than one digit compared to alphanumeric characters, so a recognition method that can reduce the processing amount and dictionary capacity is necessary.

(3) Since many handwritten Chinese characters have very similar character groups, a recognition method that can identify them is necessary.

Regarding these problems (1) and (2), many methods have been proposed up to now and it is considered that they have reached a certain level, but (3) remains as the last problem. ing.

On the other hand, as described above, the character recognition method is roughly classified into a structural analysis method that handles characteristic points such as end points, bifurcation points, and inflection points of a character pattern, and a pattern matching method that handles a character pattern globally in a two-dimensional manner. Can be separated.

A conventional device for recognizing alphanumeric characters performs recognition processing by the structural analysis method as described in Japanese Patent Laid-Open No. 51-118333, and when a plurality of candidate categories remain, the same characteristics as the recognition stage are obtained. It is known that a similar character is identified by extracting and analyzing the detailed structure of a character pattern. Since this device makes a distinction by paying attention to a fine structure of a character pattern such as an end point, a branch point, and a refraction point, it can exhibit a high discrimination system even for similar characters. However, when applied to handwritten kanji, no consideration was given to the increase in dictionary capacity (memory capacity) due to the variety of the glyphs.

Further, as another conventional technique for recognizing handwritten Chinese characters, as described in JP-A-58-207183, recognition processing is performed by a pattern matching method focusing on the background portion of a character pattern, and a plurality of candidates are obtained. It is known that when a category remains, similar characters are identified by using the same features as in the recognition stage and focusing on different parts between similar categories. This technology distinguishes by focusing on the global characteristics of the character pattern without being restricted by the fine structure of the character pattern, thus preventing an increase in dictionary capacity due to the variety of handwritten Chinese character shapes. I can. However, no consideration has been given to the extraction of the fine structure of the character pattern that reflects the difference between similar characters.

[Problems to be solved by the invention]

Kanji handwriting has a large number of strokes compared with alphanumeric characters, so that it has the property that image distortion that collapses the topology such as crushing and blurring of character line segments is noticeable. For example, it has been known from an experiment conducted on the handwritten character “Ta” that the number of loops, which is one of the structural features of the character pattern, is distributed almost evenly from 1 to 4.

For this reason, it was thought that the structural analysis features that were effective in identifying similar characters for handwritten alphanumeric characters could not be applied to handwritten Kanji characters as they are because the topology at the feature points was used. .

The prior art described in the former publication does not take into consideration the variety of characteristic points of handwritten Chinese characters, and a special standard pattern must be prepared for each similar character pair of each transformation. There was a problem.

On the other hand, the similar character identification method that focuses on the background part of the character pattern scans each point of the background part vertically and horizontally, measures the number of times the character line segment and the defined angle intersect, and this is set to each point. Features are extracted by allocating. Further, the similar characters are identified by paying attention to different parts between the similar characters in this feature space. Therefore, it is known that this feature, by its nature, can extract the global feature of the character pattern relatively stably, but cannot extract the structure of the character stroke in detail.

For this reason, the feature that paid attention to the background part, which was effective in handwritten kanji recognition using the entire character pattern, cannot express the minute structure of the character pattern. Has been done.

The prior art described in the latter publication does not consider the extraction of a fine structure of a character pattern that reflects the dissimilarity between similar characters, and when identifying similar characters, high identification accuracy cannot be expected. There was a point.

The present invention, when identifying similar characters in handwritten Chinese characters, suppresses an increase in the dictionary capacity due to the variety of handwritten Chinese characters' glyphs, and then a fine structure of a character pattern reflecting the differences between similar characters. In order to extract the same, it is an object of the present invention to provide a pattern recognition device for identifying a similar character by using the difference position information of the similar character and the pattern information for each direction obtained from the direction component of the character line segment.

Here, the differential position information is defined as position information of different shapes of similar candidate pairs. The pattern information for each direction will be described in detail in the section [Means for Solving Problems] below.

[Means for solving problems]

The above object is achieved by utilizing the property that the pattern information for each direction obtained by the recognition unit expresses a fine structure of a character line segment. That is, in order to extract the contour from the input character pattern and remove the noise component, the direction and strength are determined using each contour point and the preceding and following point sequences, and this is assigned to each direction plane. Next, in order to absorb the positional deviation of the character line segment, blurring processing is applied to each direction surface, and this is used as the characteristic amount of the entire character pattern. Further, using the difference pattern information determined for each candidate character type pair, different portions are cut out from the characteristic amount of the entire character pattern, and are set as the characteristic amount expressing the difference for each candidate character type pair. Finally, pair determination is sequentially performed using these feature amounts to identify similar characters.

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

FIG. 3 is a conceptual diagram illustrating directional feature extraction representing the principle of the present invention. First, contour points and direction components are extracted from the preprocessed normalized character pattern (a) (b).
Next, in order to remove noise components such as line edge noise, the direction and intensity of each contour point are determined using the contour point of interest and the contour point sequence surrounding it. Also, in order to create characteristic faces by direction,
The strength according to this determination is assigned to the corresponding direction surface (c), and the completed direction surface is used as the direction-specific feature surface. Further, in order to prevent the displacement of the character line segment and to compress the feature information amount, blurring processing is applied to the feature surface for each direction (d). Here, convolution is performed using a 6 × 6 window function and the amount of information is compressed to 1/16, but this is not the case as long as the same effect is obtained.

FIG. 2 is a conceptual diagram illustrating a configuration example of a similar character identification dictionary representing the principle according to the present invention. The example in the figure is (ball) 2
Value pattern 1, Binary pattern 2 of (king), Directional feature 3 of (ball) extracted by the above method, Directional feature 4 of the same (king) 4, Similar character pair (ball) and (king) 5 shows a difference pattern 5 of. In the present invention, the recognition processing is performed using the directional feature, and when a plurality of candidate categories remain, pairwise determination is sequentially performed using the feature amount of the part that represents the difference between each pair of candidate character types, and the last is left. It is configured to output the category as the recognition result. That is, the directional characteristic 3 of (ball) has a feature amount at the position of the difference pattern 5 of (ball) and (king). On the other hand, the directional characteristic 4 of (king) has no feature quantity at the same position.

As is clear from this example, even a pair of similar characters that cannot be distinguished by the characteristic amount of the entire character pattern can be easily identified by using the difference pattern of the directional characteristics that the details of the character pattern represent.

[Action]

After selecting a candidate category using the directional features of the entire input unknown pattern, using the differential position information and the directional feature indicating the parts of different shapes of similar candidate pairs, further narrowing down the candidate categories sequentially, The last remaining category is output as the recognition result.

As a result, it is possible to make a determination using the difference portion of the features that can express the fine structure of the character, and it is possible to accurately identify the similar character.

Further, since the recognition standard pattern and the similar character recognition standard pattern can be shared, without increasing the dictionary memory,
Similar character identification logic can be added.

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

FIG. 1 is a block diagram showing a configuration of an embodiment of a pattern recognition apparatus according to the present invention, in which 11 is a character observing section, 12 is a recognizing section, 13 is a similar character identifying section, 21 is a photoelectric converting section, 22 is A /
D conversion unit, 23 preprocessing unit, 24 directional feature extraction unit, 25 matching unit, 26 feature dictionary, 27 candidate character type storage unit, 28 difference information extraction unit, 29 difference pattern dictionary, 30 Is a character type identification unit, 31 is a determination unit, 32 is a control unit, and 33 is an output terminal.

In FIG. 1, a character pattern written on a paper surface or the like is converted into a video signal by the photoelectric conversion unit 21, and then is sampled and quantized by the analog / digital conversion unit 22 to form a binary mesh pattern.

The photoelectric conversion unit 21 and the analog / digital (A / D) conversion unit 22 are collectively referred to as a character observation unit 11.

The observed character pattern is subjected to a series of preprocessing such as clipping, noise removal, and normalization in the preprocessing unit 23, and becomes a normalized pattern.

To cut out, because the processing unit for recognition etc. is one character at a time,
This refers to the process of extracting one character from a group of character patterns on the paper surface, and usually cuts out an area of about 100 × 100 mesh containing one character.

The noise removal refers to a process of removing a pattern other than a character pattern such as dust adhering to a paper surface. This is difficult to distinguish from the original pattern of characters such as dakuten, and various contrivances have been made in the past.

Finally, the normalization process refers to a process of aligning characters to a certain size in order to facilitate recognition, and a technique of aligning with a circumscribing frame or a technique of aligning with the center of gravity is selected according to the recognition technique.

Next, the preprocessed normalized pattern is added to the directional feature extraction unit 24, and the directional feature extraction unit 24 creates four feature patterns for each direction.

The directional feature extraction unit 24 first performs contour extraction. That is, the raster scan is sequentially performed from the upper left of the character pattern downward, and the contour is traced along the character pattern from the point hitting the character pattern. At the same time, the direction and coordinates of each contour point are described in the contour table prepared in advance. According to this procedure, all the contours for one character are described in the contour table.

Next, when the contour table for one character is completed, four characteristic patterns for each direction are created. That is, the direction and intensity of each point are obtained from the created contour table by using the contour point of interest and the contour point sequence surrounding it. When the direction and strength of the focused contour portion are obtained, the strength is entered in the predetermined coordinates of the corresponding direction-specific feature surface. According to this procedure, a characteristic pattern for one character for each direction is created.

Furthermore, in order to prevent misalignment of character line segments and compress the amount of feature information, blurring processing is added to the feature surface for each direction, and re-sampling is performed to remove unnecessary meshes to reduce the amount of information. Compress. Four directional features are created according to this procedure.

The extracted directional features are added to the matching unit 25, and the matching unit 25
At, the degree of similarity with the standard pattern stored in the feature dictionary 26 is calculated.

The matching unit 25 first stores the four directional features in the input feature buffer. Next, the standard patterns stored in the feature dictionary 26 are sequentially read, the degree of similarity with the features of the input character pattern is obtained, and those satisfying the preset evaluation value are added to the candidate character type storage unit 27 together with the category information. .

The directional feature extraction unit 24, the matching unit 25, and the feature dictionary 26 are collectively referred to as a recognition unit 12.

The candidate character type storage unit 27 sequentially selects pairs of similar characters from the similar character group determined by the recognition unit 12 to be likely. This similar character is the same as in the explanation of the principle of FIG.
Corresponds to (ball)-(king).

The selected similar character pair is added to the difference pattern dictionary 29, and the difference pattern corresponding to the similar character pair is selected in the difference pattern dictionary 29.

On the other hand, the directional feature of the input character pattern extracted by the directional feature extraction unit 24 is also added to the difference information extraction unit 28. In addition, the standard pattern of similar character pairs is also used in the feature dictionary 26.
And is added to the difference information extraction unit 28.

The difference information extraction unit 28 stores the directional feature of the input pattern in the input feature buffer. Also, the standard pattern of similar character pairs is stored in the standard pattern buffer. Next, the difference pattern of the similar character pair stored in the difference pattern dictionary 29 is read, and the difference portion is cut out from the directional feature corresponding to the input character pattern and the standard pattern. This difference is
In the explanation of the principle of FIG. 2, the feature quantity corresponds to the black background portion of the difference pattern 5.

The cut-out directional characteristics of the input character pattern and the standard pattern are added to the character type pair identifying unit 30, and the character type pair identifying unit 30 calculates the degree of similarity between them.

The character type pair identification unit 30 obtains a similarity between the partial directional characteristics of the input character pattern and each standard pattern of the similar character pair, and determines the similarity together with the category information of the similar character pair. Added to.

The determination unit 31 compares the similarity between pairs of similar characters and determines the character types excluded from the candidate character types. The deleted character type is added to the candidate character type storage unit 27 and deleted from the candidate character type storage unit 27.

According to the above procedure, pair determination is sequentially repeated, and when the candidate character type is narrowed down to one character type, the character type is output from the output terminal 33.

Candidate character type storage unit 27, difference information extraction unit 28, difference pattern dictionary
29 and the character type identification unit 30 are collectively referred to as a similar character identification unit 13.

As described above, in the description of FIG. 1, the control unit 32 supplies the control signals to the respective units of the circuit, and the signals for notifying the states of the respective units of the circuit are given to the control unit 32. However, it is not necessary to understand the present invention. These are not necessary and are not described for simplicity.

From the above description, according to the present embodiment, after selecting a candidate category in the recognition stage, the similar character is identified by using the differential position information of the similar character pair and the directional characteristic, and the fine structure of the character can be expressed. In addition, since the similar character identifying standard pattern and the identifying standard pattern can be shared, the similar character can be identified with extremely high accuracy without adding a dictionary memory.

〔The invention's effect〕

As described above, according to the present invention, by adopting the directional feature, it becomes possible to express the fine structure of the character pattern without using the structural analysis feature. Even in the identification of similar characters requiring a difference, it is possible to identify a similar character group with extremely high accuracy without increasing the dictionary capacity.

Further, by adopting the directional feature as the feature for identifying similar characters, the feature can be shared with the standard pattern for recognition, so that a similar character identifying logic can be added without increasing the dictionary memory. It is possible to provide a pattern recognition device which has advantages and is excellent in function except for the drawbacks of the conventional example.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, FIG. 2 is a conceptual diagram showing an example of a similar character identification dictionary showing the principle of the present invention, and FIG. 3 is the same principle of the present invention. It is a conceptual diagram of the feature extraction showing. 11 …… Character observation section, 12 …… Recognition section, 13 …… Similar character identification section, 21 …… Photoelectric conversion section, 22 …… A / D conversion section, 23 …… Preprocessing section, 24 …… Directivity characteristics Extraction section, 25 ... Matching section, 26 ...
Feature dictionary, 27 ... Candidate character type storage unit, 28 ... Difference information extraction unit, 29 ... Difference pattern dictionary, 30 ... Character type pair identification unit, 31
…… Judgment part, 32 …… Control part, 33 …… Output terminal.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Yoshihiro Yokoyama, inventor Yoshihiro Yokoyama, 292 Yoshida-cho, Totsuka-ku, Yokohama, Kanagawa Prefecture, Ltd. Microelectronics Equipment Development Laboratory, Hitachi, Ltd. (72) Koichi Okazawa, Totsuka, Yokohama, Kanagawa 292 Yoshida-cho, Tokyo, Hitachi, Ltd. Microelectronics equipment development laboratory (72) Inventor Shozo Kadota 2880 Kunifuzu, Odawara, Kanagawa Stock company Hitachi Ltd., Odawara factory (56) References JP-A-61- 147385 (JP, A) JP-A-59-154579 (JP, A)

Claims (1)

[Claims] 1. A directional feature extraction unit for observing an unknown input pattern and forming a feature pattern for each direction from a directional component of a character line segment, and each standard pattern are also stored as feature patterns for each direction. The similarity between the feature dictionary, the feature pattern for each direction of the input pattern formed by the directional feature extraction unit, and the feature pattern for each direction of each of the standard patterns stored in the feature dictionary is calculated. With respect to the standard pattern stored in the feature dictionary, the matching unit that determines the standard pattern having the similarity that satisfies a preset evaluation value as a similar character, and the feature of each of the two standard patterns by direction. The difference pattern dictionary in which the difference patterns of the patterns are stored and the arbitrary two similar character pairs of the similar character group determined by the matching unit are fetched and A candidate character type storage unit that reads out a difference pattern corresponding to the similar character pair from the differential pattern dictionary, a feature pattern for each direction of two standard patterns that are similar character pairs in the candidate character type storage unit, and the input pattern. And a difference information extraction unit for extracting the portion of the position corresponding to the difference pattern read from the difference pattern dictionary in each of the feature patterns, and for each direction of the input pattern. A portion cut out from the characteristic pattern by the difference information extracting unit and a portion cut out from the characteristic information for each direction of one of the two standard patterns forming the similar character pair by the difference information extracting unit. The degree of similarity and the portion cut out by the difference information extraction unit from the feature pattern for each direction of the input pattern A character type pair identifying unit that obtains the degree of similarity with the part cut out by the difference information extracting unit from the characteristic pattern for each direction of the other standard pattern of the two standard patterns forming similar character pairs; Comparing the two similarities obtained by the pair discriminating unit, the above-mentioned 2 which is regarded as a similar character pair in the candidate character type storage unit is compared.
Of the two standard patterns having a lower degree of similarity is determined, and the characteristic pattern for each direction of the standard pattern determined to have a lower degree of similarity is deleted from the matching section. As long as a plurality of similar characters are stored in the matching unit, it is determined by the matching unit by repeating a series of processes by the candidate character type storage unit, the difference information extraction unit, the character type pair identification unit, and the determination unit. Narrow down similar characters,
A pattern recognition apparatus characterized in that a standard pattern as the similar character remaining in the matching section at the end is used as a recognition result of the input pattern.