JP2002183667A - Character-recognizing device and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognizing device for quickly executing recognizing processing at a high recognizing rate by setting an optional character set as a recognizing object character set, and a recording medium for recording data and a program for executing the character-recognizing processing. SOLUTION: This character-recognizing device investigates a document of a character-recognizing object, designates a character set proper as a recognizing result as the recognizing object character set, and inputs a pattern intended for reading from a scanner. A characteristic of an image of the recognizing object character set is extracted, and the extracting result is dimensionally compressed. A made characteristic vector is checked with a standard pattern stored in a pattern dictionary, the standard pattern, having a distance near an input pattern in a characteristic quantity space is selected, only a character corresponding to the designated recognizing object character set among the selected standard pattern is selected, and is used as a recognizing result. Thus, an easily mistaking character set, such as a similar character is designated as the recognizing object character set, and the character can be highly accurately recognized at a high speed.

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 and a recording medium, and more particularly, to an arbitrary character set as a character set to be recognized, and to quickly perform a recognition process at a high recognition rate based on language information. 1. Field of the Invention The present invention relates to a character recognition device that performs the above-described processing, and a recording medium that stores a program and data that perform the character recognition processing.

[0002]

2. Description of the Related Art Generally, in a character recognition apparatus, an unknown character pattern is compared with a standard pattern stored in a dictionary, and the character code of the standard pattern having the highest similarity is determined. Character code,
As a method of calculating the similarity, a pattern feature is converted into a multidimensional feature vector, and in a feature space,
A method for obtaining the minimum distance between patterns is widely used.

In a conventional character recognition device,
In order to reduce the calculation cost for finding the minimum distance between patterns, roughly identify features with a small number of dimensions (large classification)
And only those that have been determined to have a high degree of similarity as a result of the
Further, a method of performing identification (detailed classification) using a multidimensional feature amount is generally performed.

[0004] This method is to reduce the total collation cost by reducing the number of characters to be subjected to the detailed classification, and it is necessary to perform the large classification only the number of times of all the characters.

Further, as a method of reducing the collation cost per one time, an algorithm (dimensional compression) for reducing the number of dimensions of a feature vector by discriminant analysis or KL expansion without reducing the discriminating ability has been proposed. (Seiichi Nakagawa, Speech Recognition by Probability Model, pp.15)
-16, IEICE). With this method, if the number of dimensions of the feature vector is reduced, the amount of calculation for finding the minimum distance between the feature vectors can be reduced.

[0006] General character recognition is performed as shown in FIG. That is, an image reading device such as a scanner inputs a pattern to be read (character recognition) (step S1), and extracts image features such as direction codes and contour features (step S2). Next, since the feature extraction result is represented as a multi-dimensional vector, the feature extraction result is subjected to dimensional compression to reduce the number of dimensions so that the discrimination ability does not decrease, and a feature vector is created (step S).
3) Compare the created feature vector with a standard pattern (feature vector) stored in the dictionary (step S)
4). Then, from the collation result, a standard pattern having a distance close to the input pattern is selected in the feature amount space (step S5), and a character code corresponding to the selected standard pattern is stored and managed as a recognition result (step S5).
6).

However, in a character recognition apparatus for reading a general Japanese printed document, as a set of characters to be recognized, there are many types exceeding tens of thousands, such as JIS first-level character codes. ing. For this reason, the number of times of matching between the unknown pattern and the standard pattern is thousands of times, and the cost required for the matching process is enormous.

Conventionally, a pattern recognition device has been proposed which focuses on the difference between two similar characters and emphasizes the features of the difference (see Japanese Patent Publication No. 7-89364).

In addition, the present applicant first adds the character pattern to the pattern dictionary every time the character is erroneously recognized,
Alternatively, there has been proposed a character recognition device for determining whether to add to a confusion table (Japanese Unexamined Patent Application Publication No. Hei.
No. 62020).

[0010] Conventionally, there has been proposed a character type selection method, a character recognition device, and a recording medium for reducing a processing amount by first performing a rough matching process and then performing a more detailed matching process (particularly). See JP-A-10-222611).

Conventionally, for each pattern obtained from an image and for each combination obtained when a combination of two patterns is temporarily integrated, an index indicating the magnitude of the possibility of misidentification is calculated, and the combination of the two patterns is calculated. On the other hand, comparing the indices of misrecognizability before and after integrating them and judging the magnitude of misrecognition when combining them, sorting them in order from the most similar pattern combination, and reducing the possibility of misrecognition Create a similar category identification dictionary that integrates the determined combinations, starts calculating the index again with the combined pattern, and repeats until there are no more pattern combinations that can integrate the collimating lens, thereby creating a similar category dictionary with good identification accuracy. An apparatus and a method have been proposed (see JP-A-11-203413).

[0012]

However, in such a conventional character recognition device, the conventional character recognition is performed at high speed and with high accuracy, and the improvement of the recognition performance for similar characters is improved. Needed.

That is, in the pattern recognition device described in Japanese Patent Publication No. Hei 7-89364, the feature of the difference portion is emphasized by focusing on the difference portion between two similar characters. However, there is a problem in that the inspection by all pairs of combinations is required, and as the number of similar characters increases, the number of pairs to be combined increases exponentially, and the processing time required for the inspection becomes longer.

In the character recognition device described in Japanese Patent Laid-Open No. 5-62020, every time a character is erroneously recognized, the character pattern is added to a pattern dictionary or added to a confusion table. Therefore, the performance can be gradually improved, but the recognition performance of a large number of similar characters cannot be collectively improved, and thus it is necessary to improve the recognition performance.

Further, in the character type selection method, character recognition device and recording medium described in Japanese Patent Application Laid-Open No. 10-222611, a collation process is first performed roughly, and then a more detailed collation process is performed. The main purpose is to reduce the amount of processing, and there is a problem that recognition accuracy cannot be improved for similar characters that are erroneously recognized even in detailed processing.

Further, the similar category dictionary creating apparatus and method described in Japanese Patent Application Laid-Open No. 11-203413 is a technique for general countermeasures for similar characters, and is a character of a similar character set. For similar characters that can select the correct recognition result,
Unnecessary increase in processing time for detailed identification,
There is a problem that the processing speed becomes slow.

That is, in a normal document, a word that causes confusion due to similar characters tends to be avoided. For example, the scholarly character is a similar character, so the word “school” exists, but the word “character school” hardly exists. That is, if the word recognition result is “school” or “character school” and there is only the word “school”, there is no need to identify the similar character set (study-character) in detail. If it is known that the character is any one of the similar character sets, the correct recognition result “school” can be selected. Therefore, even in such a case, the conventional technology for detailed identification has a problem that unnecessary processing time is increased and the processing speed is reduced.

As described above, in a character recognition apparatus intended to read a general Japanese type print document, a set of characters to be recognized includes more than several thousand types such as JIS first-level character codes. Handles many types. For this reason, the number of times of matching between the unknown pattern and the standard pattern is thousands of times, and the cost required for the matching process is enormous.

However, in a specific document, the characters to be recognized may be limited. For example, in a certain product slip, only arithmetic numerals will appear in the price column, and only combinations of alphabets, arithmetic numerals and specific symbols will appear in the product code column.

As described above, only a specific character type (for example, kanji, hiragana, katakana, a symbol, a kanji numeral, an arithmetic numeral, etc.) may appear in a document to be recognized.

In addition, there is a case in which a rule that one character is not used for a character having a similar shape that is confusing to human eyes as shown below is used in order to avoid confusion of reading.

I (uppercase letter of alphabet i) l (lowercase letter of alphabet L) 1 (number) 0 (number) O (uppercase letter of alphabet o) o (lowercase letter of alphabet O) 2 (number) z (uppercase letter of alphabet Z) Z (lowercase letter of the alphabet z) Examples of character strings that are easy for humans to misread "12345""12345""1z345" Or it is not fixed when the character recognition device is created,
It may change each time the character recognition device is used.

That is, as compared with the case where the pattern dictionary is created and the character recognition device is created on the assumption that the character to be recognized is fixed as in the related art, a method that supports the variable character set to be recognized is used. However, since the information of the character set to be recognized can be used, high recognition accuracy can be expected.

Therefore, the present invention is to collectively obtain characters which are easily misrecognized such as similar characters in advance and perform detailed recognition specialized for similar characters in addition to ordinary recognition. It is intended to improve recognition performance for similar characters.

The field in which the character recognition device is used is as follows:
In advance, if it is limited to some extent, similarity characters to be identified and similar characters that do not need to be identified are obtained by using linguistic knowledge in the relevant field, and similarity characters are reduced while reducing the processing time required for similar character recognition. The purpose is to improve the character recognition performance.

Specifically, the invention according to claim 1 performs predetermined image processing on a character image to extract image features, compares the extracted features with an image feature storage dictionary, and stores the image feature in the image feature storage dictionary. When performing character correction by performing error correction based on the stored linguistic information, a character set to be recognized among character images can be arbitrarily set as a recognition target character set, and the set recognition target character set By performing character recognition processing using only the image feature storage dictionary for similar characters, similar characters that are easily misrecognized, etc., are investigated in advance and found, and image feature storage specialized for discrimination between similar characters is performed. A dictionary is created, and a character set that is likely to be misidentified, such as a similar character, among character images to be recognized is designated as a character set to be recognized, and an image feature storage dictionary created only for the character set to be recognized is used. It is intended to perform character recognition, to provide a character recognition device capable of to improve the recognition accuracy of such erroneous recognition easily similar character performs character recognition with high accuracy and high speed.

According to the second aspect of the present invention, the character set to be recognized can be appropriately changed, and character recognition processing is sequentially performed on the changed character set to be recognized, so that the recognition accuracy of similar characters and the like which are easily erroneously recognized can be improved. It is an object of the present invention to provide a character recognition device capable of performing character recognition with high accuracy and high speed.

According to the third aspect of the present invention, a set of characters to be recognized can be set for each character to be recognized, and a character recognition process is performed on the set of characters to be recognized. It is an object of the present invention to provide a character recognition device capable of recognizing recognition accuracy of similar characters and the like more quickly and performing high-accuracy and higher-speed character recognition.

According to the fourth aspect of the present invention, the character set to be recognized can be set based on the linguistic information stored in the image feature storage dictionary. When limited, it is possible to determine the necessity of detailed identification of similar characters using linguistic knowledge specific to the field, and to reduce the amount of processing required for detailed identification without reducing the recognition accuracy of similar characters. It is an object of the present invention to provide a character recognition device that can reduce the number of characters and perform character recognition with higher accuracy and higher speed.

According to a fifth aspect of the present invention, a plurality of characters can be collectively set as one group as a recognition target character in the same manner as character data, based on language information stored in an image feature storage dictionary. If words appearing in a document subject to character recognition are fixed or the frequency of occurrence is clear, the presence / absence and frequency of such words are determined using linguistic knowledge to identify similar characters in detail. A character recognition device capable of determining whether or not the character recognition is necessary, reducing the amount of processing for detailed identification without lowering the recognition accuracy of similar characters, and performing more accurate and faster character recognition. It is intended to be.

According to a sixth aspect of the present invention, there is provided a character recognition processing program for executing the character recognition processing according to any one of the first to fifth aspects, and the character recognition processing such as the image feature storage dictionary. By recording the data necessary to execute the program, the character recognition processing program is executed using the existing hardware resources, and a character set that is likely to be misidentified such as the similar character is recognized from among the character images to be recognized. Designate as a character set to be recognized, perform character recognition using the image feature storage dictionary created only for the character set to be recognized, and improve the recognition accuracy of similar characters that are easy to misrecognize to achieve high accuracy and high speed. It is an object of the present invention to provide a recording medium capable of performing character recognition.

[0032]

According to a first aspect of the present invention, there is provided a character recognition apparatus which performs predetermined image processing on a character image to extract image features, and compares the extracted image features with an image feature storage dictionary.
In a character recognition device that performs character correction by performing error correction based on linguistic information stored in the image feature storage dictionary, a character set to be recognized among the character images can be arbitrarily set as a character set to be recognized. The above object is achieved by performing a character recognition process using only the set of recognition target character sets using the image feature storage dictionary.

According to the above arrangement, a predetermined image processing is performed on the character image to extract the image feature, collate the image feature with the image feature storage dictionary, and extract the language information stored in the image feature storage dictionary. When performing character recognition by performing error correction based on a character set to be recognized among character images, the character set to be recognized can be arbitrarily set as a character set to be recognized, and an image feature storage dictionary is provided only for the set character set to be recognized. Is used to perform character recognition processing. Therefore, a similar character or the like that is easily misrecognized is searched and found in advance, and an image feature storage dictionary specialized for discriminating between similar characters is created and recognition is performed. A character set that is easily misidentified such as similar characters in character images is specified as a character set to be recognized, and character recognition is performed using an image feature storage dictionary created only for the character set to be recognized. It can be, to improve the recognition accuracy of such erroneous recognition easily similar character can perform character recognition with high accuracy and high speed.

In this case, for example, as described in claim 2, the character recognition device enables the character set to be recognized to be appropriately changed, and sequentially performs the character recognition process on the changed character set to be recognized. It may be performed.

According to the above configuration, the character set to be recognized can be changed as appropriate, and the character recognition process is sequentially performed on the changed character set to be recognized. The character recognition can be performed with high accuracy and high speed.

In addition, for example, the character recognition device can set the character set to be recognized for each character to be recognized, and can set the character set to be recognized. The character recognition processing may be performed.

According to the above arrangement, for each character to be recognized,
The recognition target character set can be set, and character recognition processing is performed on the set recognition target character set. Therefore, the recognition accuracy of similar characters and the like that are likely to be erroneously recognized can be more quickly recognized, and high accuracy can be achieved. In addition, character recognition can be performed even faster.

Further, for example, as set forth in claim 4, the character recognition device enables the character set to be recognized to be set based on the language information stored in the image feature storage dictionary. The character recognition processing may be performed on the set of recognition target characters.

According to the above configuration, the character set to be recognized can be set based on the linguistic information stored in the image feature storage dictionary, so that the field of the document to be subjected to character recognition is limited. Linguistic knowledge specific to the field, determine whether or not detailed identification of similar characters is necessary, and reduce the amount of processing required for detailed identification without reducing the recognition accuracy of similar characters. Thus, character recognition can be performed with higher precision and at higher speed.

Further, for example, as described in claim 5, the character recognition device collects a plurality of characters into one group based on the language information stored in the image feature storage dictionary. The data may be set as the recognition target character similarly to the data, and the character recognition processing may be performed on the set recognition target character set.

According to the above configuration, based on the language information stored in the image feature storage dictionary, a plurality of characters can be collectively set as one group as a recognition target character similarly to character data. If the words appearing in the document subject to character recognition are fixed or the frequency of occurrence is clear, it is necessary to determine the occurrence and frequency of such words using linguistic knowledge to identify similar characters in detail. It is possible to determine the presence or absence of the character and reduce the processing amount for the detailed identification without lowering the recognition accuracy of the similar character, and perform the character recognition with higher accuracy and higher speed.

According to a sixth aspect of the present invention, there is provided a recording medium including a character recognition processing program necessary for executing the character recognition processing according to any one of the first to fifth aspects and the image feature storage dictionary. The object is achieved by recording data necessary for executing the character recognition processing program.

According to the above configuration, the present invention is characterized in that:
Since the data necessary for executing the character recognition processing program such as the character recognition processing program required to execute the character recognition processing described in any one of the above and the image feature storage dictionary or the like is recorded, the existing data is recorded. Executes a character recognition processing program using hardware resources, designates a character set that is likely to be misidentified such as similar characters among character images to be recognized as a character set to be recognized, and creates only the character set to be recognized. Character recognition can be performed using the image feature storage dictionary thus obtained, and the accuracy of recognition of similar characters or the like which are likely to be erroneously recognized can be improved to perform high-accuracy and high-speed character recognition.

[0044]

Preferred embodiments of the present invention will be described below in detail with reference to the accompanying drawings. It should be noted that the embodiments described below are preferred embodiments of the present invention, and therefore, various technically preferable limitations are added. However, the scope of the present invention is not limited to the following description. The embodiments are not limited to these embodiments unless otherwise specified. In particular, the present invention can be widely applied to a character recognition device in which a character set to be recognized can be limited by checking a document to be recognized in advance.

FIGS. 1 to 4 show a first embodiment of a character recognition apparatus according to the present invention. This embodiment corresponds to claim 1 of the present invention. FIG. 1 is a conceptual diagram of a character recognition device 1 to which a first embodiment of the character recognition device of the present invention is applied.

In FIG. 1, a character recognition device 1 includes a pattern dictionary 2, a pattern input unit 3, a feature extraction unit 4, a dimension compression unit 5, a pattern comparison unit 6, a minimum distance pattern selection unit 7, a minimum pattern holding / managing unit. 8, a character selection unit 9, a recognition target character designation unit 10, a recognition target character set holding unit 11, and the like.

The pattern dictionary (image feature storage dictionary) 2
A dimensionally compressed result of the feature extraction of the learning sample is stored as a feature amount vector (standard pattern).

The pattern input means 3 is a means for inputting an image intended to be read by a scanner or the like, and the feature extracting means 4 extracts image features from the image input by the pattern input means 3. This feature extraction means 4
Is expressed as a multidimensional vector.

The dimension compressing means 5 reduces the number of dimensions of the extraction result of the feature extracting means 4 so that the discriminating ability does not deteriorate, and the pattern matching means 6 compares the feature vector created by the dimension compressing means 5 with the pattern dictionary. Then, the distance between the two in the feature amount space is obtained by collating with the standard pattern held in 2.

The minimum distance pattern selecting means 7 selects a standard pattern close to the input pattern based on the result of the comparison by the pattern comparing means 6, and in this case, a plurality of standard patterns may be selected.

The minimum distance pattern holding / managing means 8 holds / manages the standard pattern selected by the minimum distance pattern selecting means 7, and the character selecting means 9 is held / managed by the minimum distance pattern holding / managing means 8. Of the candidate recognition results, the result corresponding to the specified character set is selected.

The recognition target character specification means 10 specifies in advance a set of characters required as a recognition result. For example, the recognition target character designating means 10 may use a character type such as {number}, {alphabet}, {hiragana}, {katakana}, or {A,
B, C}, {1, 2, 3}, etc., are enumerated and specified. Such designation can be easily realized by using a regular expression.

The recognition target character set holding means 11 holds the character set specified by the recognition target character specification means 10.

The character recognition device 1 is specifically realized by software using the hardware of a computer 20 as shown in FIG. 2, for example. That is, the computer 20 has a CPU (Central Processing Unit).
) 21, memory 22, display device 23, hard disk 24, keyboard 25, CD (Compact Disc) -ROM
(Read Only Memory) drive 26, floppy (registered trademark) disk (FD) drive 27, and communication device 28
The main parts are connected by a bus 29.

A CD-ROM or floppy disk as a recording medium storing the document recognition software program of this embodiment and necessary data such as dictionary data of the pattern dictionary 2 is inserted into the CD-ROM drive 26 or the FD drive 27. The character recognition software program and necessary data are read into, for example, the hard disk 24, and the CPU 21
The character recognition software program in the memory 22
The character recognition device 1 shown in FIG. 1 described above can be realized by executing while using as a work memory.

The computer 20 has a CPU 21
Under the control of, various commands are input from the keyboard 25 or a pointing device such as a mouse (not shown), and necessary information is displayed and output on the display device 23.

The communication device 28 includes a telephone line and a LAN (Lo
The computer 20 is connected to various information processing devices such as other computers and storage media via a communication line connected to the communication device 28 via the communication device 28. Transfer information between the two.

As shown in FIG. 3, the character recognition apparatus 1 includes a plurality of computers PC1 to PCn (PC1 to PCn in FIG. 3) connected via a communication line such as the Internet.
(Only PC3 is described). That is,
Each means constituting the character recognition device 1, particularly, the pattern dictionary 2 is stored in one or a plurality of computers PC1 to PCn connected via a communication line, and the respective computers PC1 to PCn connected via the communication line When executing character recognition processing with a character recognition software program mounted on the hard disk or the like, each unit of the character recognition device 1 is referred to as a plurality of units such as referring to the pattern dictionary 2 of the other computers PC1 to PCn. The character recognition device 1 can be realized in a form of being distributed to the computers PC1 to PCn or stored in each other.

Next, the operation of the present embodiment will be described. As shown in FIG. 4, the character recognition apparatus 1 according to the present embodiment first examines a document to be subjected to character recognition by the recognition target character designating unit 10 and determines a character set appropriate as a recognition result. As a recognition target character specifying means 1
The recognition target character set designated as 0 is held in the recognition target character set holding unit 11 (step S101), and a pattern intended to be read is input from the pattern input unit 3 which is an image reading device such as a scanner (step S10).
2).

Next, the character recognition device 1 includes a feature extracting means 4
To extract image features (step S103). The feature extracting means 4 performs the feature extraction of the image using various types such as a direction code and a contour feature, and the feature extraction result is expressed as a multidimensional vector. The number of dimensions is reduced by dimensionally compressing the extraction result of the feature extracting means 4 so that the discriminating ability is not reduced by the dimension compressing means 5 (step S10).
4).

Next, the pattern matching means 6 compares the created feature vector with a standard pattern (feature vector) stored in the pattern dictionary 2 (step S1).
05), the minimum distance pattern selecting means 7 selects a standard pattern having a distance close to the input pattern in the feature amount space, and the minimum distance pattern holding / managing means 8 holds / manages the selected standard pattern. (Step S1
06).

Then, the character selection means 9 is specified by the recognition target character specification means 10 and stored in the recognition target character set storage means 11 among the standard patterns stored and managed by the minimum distance pattern storage and management means 8. Only the characters corresponding to the set of characters to be recognized are selected, and are set as a recognition result 12 (see FIG. 1) (step S107).

As described above, the character recognition device 1 of the present embodiment performs predetermined image processing on a character image to extract the features of the image, compares it with the pattern dictionary 2, When performing character correction by performing error correction based on the stored linguistic information, a character set to be recognized among character images can be arbitrarily set as a recognition target character set, and the set recognition target character set The character recognition process is performed by using the pattern dictionary 2 only for this.

Therefore, a similar character or the like which is easily misrecognized is obtained by examining in advance, and a pattern dictionary 2 specialized for discriminating between similar characters is created. A character set that is easily misidentified is designated as a character set to be recognized, and character recognition can be performed using the pattern dictionary 2 created only for the character set to be recognized. Character recognition can be performed with high accuracy and high speed by improving the accuracy.

FIGS. 5 and 6 show a character recognition apparatus according to a second embodiment of the present invention. This embodiment corresponds to claim 2 of the present invention.

This embodiment is applied to a character recognition device similar to the character recognition device 1 of the first embodiment. In the description of the present embodiment, the first embodiment will be described.
The same components as those of the embodiment are denoted by the same reference numerals, and detailed description thereof will be omitted.

FIG. 5 is a conceptual diagram of a character recognition device 30 to which the second embodiment of the character recognition device of the present invention is applied.
Like the character recognition device 1 of the first embodiment, the character recognition device 30 of the present embodiment also includes a computer 20 shown in FIG. 2 or a plurality of computers PC1 connected by a communication line shown in FIG. ~ PCn and the like.

In FIG. 5, a character recognition device 30 includes a pattern dictionary 2, a pattern input device 3, a feature extraction device 4, a dimension compression device 5, a pattern collation device similar to the character recognition device 1 of the first embodiment. 6, a minimum distance pattern selecting unit 7, a minimum pattern holding / managing unit 8, a character selecting unit 9, a recognition target character designating unit 10, a recognition target character set holding unit 11, and the like.
And learning sample feature vector holding means 32 and the like.

The pattern dictionary 2, pattern input means 3, feature extraction means 4, dimension compression means 5, pattern collation means 6, minimum distance pattern selection means 7, minimum pattern holding / management means 8, character selection means 9, recognition target The character designating unit 10 and the recognition target character set holding unit 11 are the same as those in the first embodiment, and a description thereof will be omitted.

The dimension compressing means 31 is based on the character set specified by the recognition target character designating means 10 and held in the recognition target character set holding means 11, ie, the feature vector of the learning sample, that is, the feature vector before the dimension compression. Dimensionally compress a vector.

The learning work feature vector holding means 32 holds the result of feature extraction of the image of the learning sample as a multidimensional vector.

Next, the operation of the present embodiment will be described. The character recognition device 30 according to the present embodiment converts feature vectors of all characters before dimension compression into learning sample feature vectors 32.
Each time the recognition target character set is changed, a dimension compression process is performed, and then an actual recognition process is performed, whereby the dimension compression is performed with high accuracy, and the recognition rate is further improved.

That is, in the character recognition device 1 according to the first aspect, only the characters specified by the recognition target character specifying means 10 are selected from the recognition results, but the pattern dictionary 2 is subjected to the dimensional compression. In such a case, the precision of the dimensional compression may deteriorate. For example, dimensional compression is to minimize the average error before and after dimensional compression for the feature vectors of all characters, or
Since the conversion axis that satisfies the criteria such as the maximum separation between categories is obtained, if the character set to be recognized changes, the conversion axis cannot be accurately compressed unless the conversion axis is obtained again. .

Therefore, in this embodiment, all characters
The feature vector before the dimensional compression is held in the learning sample feature vector holding means 32, and every time the character set to be recognized changes, the dimensional compression processing is performed, and then the actual recognition processing is performed, so that the dimensional compression can be performed accurately. Then, the recognition rate is improved.

That is, the character recognition device 3 of this embodiment
0, as shown in FIG. 6, first, it is checked whether or not the character set to be recognized is changed (step S201). When the character set to be recognized is changed, character recognition is performed in order to newly specify the character set to be recognized. The target document is examined, a character set appropriate as a recognition result is specified as a character set to be recognized, and the character set to be recognized specified by the character to be recognized specifying means 10 is held by the character set to be held holding means 11 ( Step S202). The character recognition device 30 uses the dimension compression means 31 based on the specified character set to be recognized.
The feature vector held by the learning sample vector holding unit 32 is dimensionally compressed, and the pattern dictionary 2 is created again (step S203).

After re-creating the pattern dictionary 2, the character recognition device 30 inputs a pattern intended to be read from the pattern input means 3 which is an image reading device such as a scanner (step S204).

On the other hand, if the character set to be recognized is not changed in step S201, the character recognition device 30 does not need to newly re-dimensionally compress the pattern dictionary 2, so the process proceeds to step S4 and the pattern input means 3 , A pattern intended to be read is input (step S204).

When the character recognition device 30 performs pattern input, the feature extracting means 4 extracts features of the image (step S205). The feature extracting means 4 performs the feature extraction of the image using various types such as a direction code and a contour feature, and the feature extraction result is expressed as a multidimensional vector. After performing the feature extraction, the character recognizing device 30 dimensionally compresses the extraction result of the feature extracting means 4 by the dimension compressing means 5 based on the character set designated in step S202 so that the discriminating ability does not decrease. To reduce the number of dimensions (step S206).

Next, the pattern matching means 6 matches the created feature vector with a standard pattern (feature vector) stored in the pattern dictionary 2 (step S2).
07), the minimum distance pattern selecting means 7 selects a standard pattern having a distance close to the input pattern in the feature space, and the minimum distance pattern holding / managing means 8 holds / manages the selected standard pattern. (Step S
208).

Then, the character selection means 9 is designated by the recognition target character designation means 10 and held in the recognition target character set holding means 11 among the standard patterns held and managed by the minimum distance pattern holding / management means 8. Only the characters corresponding to the set of characters to be recognized are selected and set as a recognition result 12 (step S209).

As described above, the character recognition apparatus 1 of the present embodiment enables the character set to be recognized to be changed as appropriate, and sequentially performs character recognition processing on the changed recognized character set.

Therefore, it is possible to further improve the recognition accuracy of a similar character or the like which is likely to be erroneously recognized, and it is possible to perform the character recognition with high accuracy and high speed.

FIG. 7 is a diagram showing a third embodiment of the character recognition device of the present invention. This embodiment corresponds to claim 3.

The present embodiment is applied to a character recognition device similar to the character recognition device 1 of the first embodiment, and in the description of this embodiment, if necessary, Description will be made using the reference numerals used in the first embodiment as they are.

The character recognition apparatus 1 of this embodiment holds the feature vectors of all the characters before the dimension compression in the learning sample feature vector 32, and performs the dimension compression processing every time the character set to be recognized changes. Thereafter, by performing actual recognition processing, dimension compression is performed with high accuracy, and the recognition rate is further improved.

That is, in the character recognition device 1 of the first embodiment, as described above, only the character specified by the recognition target character specifying means 10 is selected from the recognition results. When the pattern dictionary 2 is dimensionally compressed, the precision of the dimension compression may be reduced.
For example, dimensional compression is based on a criterion such as minimizing the average error before and after dimensional compression, or maximizing the separation between categories for feature vectors of all characters. Since the conversion axis that satisfies is determined, if the character set to be recognized changes, the dimension cannot be accurately compressed unless the conversion axis is determined again.

Therefore, the character recognition device 1 of the present embodiment
Is a pattern dictionary 2 obtained by dimensionally compressing a frequently used character set for each character set in advance.
Are prepared, and at the time of character recognition, a prepared character set is selected to select and use the pattern dictionary 2 which has been subjected to optimal dimensional compression for the character set.

That is, as shown in FIG. 7, when the character recognition processing device 1 is specified by the recognition target character specification means 10 to narrow down the set of characters to be recognized in the order of A → B (step S301). Then, a pattern intended to be read is inputted from the pattern input means 3 which is an image reading device such as a scanner (step S302).

When a pattern is input, the character recognition device 1
Extracts the features of the image by the feature extraction unit 4 (step S303), and sets the extraction result of the feature extraction unit 4 as a recognition target character set A (step S304).

Next, the character recognition apparatus 1 sets the dictionary for the character set A to be recognized as the current pattern dictionary 2 (step S305), and the dimension compression means 5 sets the dictionary to the specified character set A for recognition. Based on this, the extraction result of the feature extraction means 4 is dimensionally compressed so that the discrimination ability does not decrease (step S306).

Next, the pattern matching means 6 compares the created feature vector with the set current standard pattern (feature vector) among the standard patterns stored in the pattern dictionary 2 (step S307). The minimum distance pattern selecting means 7 selects a plurality of standard patterns having a distance close to the input pattern in the feature space, and the minimum distance pattern holding / managing means 8 holds the selected plurality of standard patterns. Then, it is created and managed as a recognition result candidate (step S308).

Next, the character recognition device 1 sets the recognition target character set to B by the recognition target character designation means 10 (step S309), and sets the dictionary for the recognition target character set B as the current pattern dictionary 2. It is set (step S310).

When the dictionary for the character set B to be recognized is set as the current pattern dictionary 2, the character recognizing device 1 outputs the result extracted in step S303 to the dimension compressing means 5,
Based on the recognition target character set B, the dimension is compressed so that the discrimination ability does not decrease (step S311).

Next, the pattern matching means 6 determines in step S31 only the results belonging to the character set B to be recognized among the recognition candidates selected in step S308.
1 is compared with the current standard pattern (feature vector) set above among the standard patterns stored in the pattern dictionary 2 (step S31).
2) The minimum distance pattern selecting means 7 selects a plurality of standard patterns having a distance close to the input pattern in the feature amount space, and the minimum distance pattern holding / managing means 8
The plurality of selected standard patterns are held, created and managed as recognition result candidates (step S313).

As described above, the character recognition apparatus 1 according to the present embodiment enables a character set to be recognized to be set for each character to be recognized, and performs a character recognition process on the set character set to be recognized. ing.

Therefore, the storage capacity can be reduced without retaining the feature vectors before the dimension compression for all the characters and without performing the dimension compression processing for recreating the pattern dictionary every time the character set changes. In addition, it is possible to further improve the recognition rate by using a pattern dictionary that has been subjected to optimal dimensional compression for a character set while reducing processing time.

In this embodiment, the character set B, which is a subset of the character set A, is designated as the final recognized character set, and the selection is made in two stages. Instead, the selection may be made in more stages.

FIG. 8 is a diagram showing a fourth embodiment of the character recognition apparatus according to the present invention. This embodiment corresponds to claim 4 of the present invention.

Note that the present embodiment is applied to a character recognition device similar to the character recognition device 1 of the first embodiment, and in the description of this embodiment, if necessary, Description will be made using the reference numerals used in the first embodiment as they are.

The character recognition apparatus 1 according to the present embodiment focuses on the fact that it is not necessary to change the character set and recognize it in multiple stages for characters that can be stably recognized. By specifying the character set, the selection process can be set more finely without performing the recognition process unnecessarily in multiple stages.
The recognition accuracy has been improved.

That is, in the character recognition device 1 of the third embodiment, a character set which is generally used is assumed and prepared in advance, and a multi-stage recognition process is performed to improve the recognition rate. However, characters that can be stably recognized can be recognized at a high recognition rate without performing multi-stage recognition by changing the character set.

Therefore, the character recognition device 1 of the present embodiment
By specifying a character set to be recognized for each character, the selection process is set more finely without performing the recognition process unnecessarily in multiple stages, thereby improving the recognition accuracy.

That is, the character recognition device 1 of the present embodiment
First, as shown in FIG.
In step S401, it is specified that the character set A to be recognized is narrowed down to each character (A → each character) (step S401), and a pattern intended to be read is input from the pattern input unit 3 which is an image reading device such as a scanner (step S401). Step S402).

When a pattern is input, the character recognition device 1
Extracts the features of the image by the feature extraction unit 4 (step S403), and sets the extraction result of the feature extraction unit 4 as a recognition target character set A (step S404).

Next, the character recognition device 1 sets the dictionary for the character set A to be recognized as the current pattern dictionary 2 (step S405), and the dimension compression means 5 sets the dictionary to the specified character set A to be recognized. Based on this, the extraction result of the feature extraction means 4 is dimensionally compressed so that the discrimination ability does not decrease (step S406).

Next, the pattern matching means 6 compares the created feature vector with the current standard pattern (feature vector) set above among the standard patterns stored in the pattern dictionary 2 (step S407). The minimum distance pattern selecting means 7 selects a plurality of standard patterns having a distance close to the input pattern in the feature space, and the minimum distance pattern holding / managing means 8 holds the selected plurality of standard patterns. Then, it is created and managed as a recognition result candidate (step S408).

Next, the character recognition device 1 sets the recognition target character set X for each recognition target character based on the recognition result of step S408 by the recognition target character designation means 10 (step S409). The dictionary for the set X is set as the current pattern dictionary 2 (step S41).
0).

When the dictionary for the character set X to be recognized is set as the current pattern dictionary 2, the character recognizing device 1 outputs the result extracted in step S403 to the dimension compressing means 5,
Based on the recognition target character set X, the dimension is compressed so that the discrimination ability does not decrease (step S411).

Next, the pattern matching means 6 determines only the result belonging to the character set X to be recognized among the recognition candidates selected in step S408, in step S411.
Is compared with the set current standard pattern (feature vector) among the standard patterns stored in the pattern dictionary 2 (step S41).
2) The minimum distance pattern selecting means 7 selects a plurality of standard patterns having a distance close to the input pattern in the feature amount space, and the minimum distance pattern holding / managing means 8
The selected plurality of standard patterns are held, created and managed as recognition result candidates (step S413).

As described above, the character recognition device 1 according to the present embodiment collects a plurality of characters into one group based on the language information stored in the pattern dictionary 2 in the same manner as the character data. It is possible to set as.

Therefore, if a word that appears in a document to be subjected to character recognition is fixed or the frequency of occurrence is clear, the presence / absence and frequency of the word are determined using linguistic knowledge to determine the similar character. It is possible to determine the necessity of the detailed identification, reduce the processing amount for the detailed identification without lowering the recognition accuracy of the similar character, and perform the character recognition with higher accuracy and higher speed.

In the case of the present embodiment, when a word that appears in a manuscript used by the character recognition device is known to some extent, for example, when the word is selected as a headword in a Japanese language dictionary By investigating words that are known to appear, an important similar character set that can improve recognition accuracy can be obtained.

As a method for obtaining the similar character set,
For example, a method of collecting feature amount vectors by clustering or a method of totaling errors in recognition results to create a confusion matrix can be used. As the clustering, an LBG algorithm (see "Algorithm of Pattern Recognition and Learning" written by Uesaka and Ozeki, Chapter 7, Bunichi Sogo Publishing, etc.) and the like can be used.

For example, if there is a similar character set {right, door, stone}, and words that are expected to appear in the manuscript include “Migita”, “Toda”, and “Ishida”, even if the word Even if there is an opportunity for correction, {right, door, stone}
It can be understood that the correct character cannot be obtained unless the character is correctly identified. That is, the similar character set {right, door, stone}
It is a similar character set important for identification.

Such a set is obtained by replacing a character included in a word with a similar character set and examining whether there is a character that is the same as another word.

For example, in the above example, the three become identical as follows.

"Mita" → {right, door, stone} + field "Toda" → {right, door, stone} + field "Ishida" → {right, door, stone} + field From this result, recognition processing is performed. Previously, a similar character set important for identification can be determined in advance.

Therefore, by creating a dictionary for identifying these important similar character sets and referencing them from a normal pattern dictionary, it is possible to perform detailed identification following normal recognition. , The recognition accuracy can be improved.

Such a similar character set dictionary is, for example,
FIG. 9 shows an example of dictionary records of “koi”, “bonito”, and “salmon”, which can be dealt with by creating a dictionary of a similar character set that does not require detailed identification.

In the case of the present embodiment, on the contrary to the above case, if the words appearing in the original used by the character recognition device 1 are fixed, similar characters are identified in detail. May not be necessary.

For example, now, a similar character set {bonito, salmon,
It is assumed that there is a carp 単 語 and it is known that only “first bonito”, “salt salmon”, and “koiguchi” do not appear in the manuscript as words including characters in the similar character set.

In this case, the input pattern is bonito, salmon,
Even if any of the similar character sets of Koi is not identified in detail, the ambiguity is resolved by the adjacent character and word information as follows.

First + bonito, salmon, carp ○ First bonito × first salmon × first carp Salt + bonito, salmon, carp × salted bonito ○ Salted salmon × salted carp bonito, salmon, carp + mouth × bonito × Therefore, similar character sets that do not need to be distinguished, such as {bonito, salmon, carp}, are collectively defined as one standard pattern (feature vector) in the pattern dictionary 2.
The number of standard patterns can be reduced, and the number of times of matching with the input pattern can be reduced. As a result, the processing time can be reduced without lowering the recognition accuracy.

As a standard pattern of a similar character set that does not require detailed identification, for example, a pattern shown in FIG. 10 showing the dictionary record can be used.

The invention made by the present inventor has been specifically described based on the preferred embodiments. However, the present invention is not limited to the above, and various modifications can be made without departing from the gist of the invention. It goes without saying that it is possible.

[0126]

According to the character recognition apparatus of the first aspect of the present invention, a predetermined image processing is performed on the character image to extract the image feature, and the extracted image feature is compared with the image feature storage dictionary. When performing character correction by performing error correction based on linguistic information stored in a storage dictionary, a character set to be recognized among character images can be arbitrarily set as a character set to be recognized, and the set recognition target is set. Since the character recognition process is performed only for the character set by using the image feature storage dictionary, similar characters that are easily misrecognized are determined by examining in advance, and images specialized for distinguishing between similar characters An image feature storage dictionary that creates a feature storage dictionary and designates a character set that is likely to be misidentified such as similar characters among character images to be recognized as a character set to be recognized, and creates only the character set to be recognized. It is possible to perform character recognition can be performed character recognition, to improve the recognition accuracy of such erroneous recognition easily similar character with high accuracy and fast in use.

According to the character recognition apparatus of the second aspect of the invention, the character set to be recognized can be changed as appropriate, and the changed character set is sequentially subjected to character recognition processing. The recognition accuracy of characters and the like can be further improved, and character recognition can be performed with high accuracy and at high speed.

According to the character recognition device of the third aspect of the present invention, a set of characters to be recognized can be set for each character to be recognized, and the character recognition process is performed on the set of characters to be recognized. In addition, the recognition accuracy of similar characters or the like which are easily misrecognized can be more quickly recognized, and character recognition can be performed with higher accuracy and higher speed.

According to the character recognition apparatus of the present invention, the character set to be recognized can be set based on the language information stored in the image feature storage dictionary. When the field of the original document is limited, linguistic knowledge unique to the field is used to determine whether detailed identification of similar characters is necessary, and to reduce the accuracy of similar character recognition without lowering the accuracy. The amount of processing required for identification can be reduced, and character recognition can be performed with higher accuracy and at higher speed.

According to the character recognition device of the present invention, based on the linguistic information stored in the image feature storage dictionary, a plurality of characters are grouped into one group in the same manner as the character data. When words appearing in the document to be subjected to character recognition are fixed or when the frequency of occurrence is clear, the presence or absence and frequency of the occurrence are determined by using linguistic knowledge. , It is possible to determine the necessity of detailed identification of similar characters, and to reduce the processing amount required for detailed identification without lowering the recognition accuracy of similar characters, and perform character recognition with higher accuracy and at higher speed. be able to.

According to the recording medium of the invention described in claim 6,
A character recognition processing program required to execute the character recognition processing according to any one of claims 1 to 5, and data necessary to execute the character recognition processing program such as the image feature storage dictionary are recorded. Since the character recognition processing program is executed by using the existing hardware resources, a character set that is easily misidentified such as the similar character in the character image to be recognized is designated as a character set to be recognized.
Character recognition can be performed using the image feature storage dictionary created only for the character set to be recognized. it can.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram of a character recognition device to which a first embodiment of the character recognition device of the present invention is applied.

FIG. 2 is a block diagram of a computer that realizes the character recognition device of FIG. 1;

FIG. 3 is a system configuration diagram when the character recognition device of FIG. 1 is implemented by a plurality of computers connected to a network.

FIG. 4 is a flowchart illustrating a character recognition process performed by the character recognition device of FIG. 1;

FIG. 5 is a conceptual diagram of a character recognition device to which a second embodiment of the character recognition device of the present invention is applied.

FIG. 6 is a flowchart showing a character recognition process performed by the character recognition device of FIG. 5;

FIG. 7 is a flowchart illustrating a character recognition process performed by the character recognition device according to the third embodiment of the present invention.

FIG. 8 is a flowchart illustrating a character recognition process performed by a character recognition device according to a fourth embodiment of the present invention.

FIG. 9 is a diagram showing an example of a dictionary record of a similar character set dictionary.

FIG. 10 is a diagram showing an example of a dictionary record of a standard pattern dictionary of a similar character set.

FIG. 11 is a flowchart showing a conventional general character recognition process.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 2 Pattern dictionary 3 Pattern input means 4 Feature extraction means 5 dimensional compression means 6 Pattern collation means 7 Minimum distance pattern selection means 8 Minimum pattern holding / management means 9 Character selection means 10 Character recognition target designation means 11 Characters to be recognized Set holding means 20 Computer 21 CPU 22 Memory 23 Display device 24 Hard disk 25 Keyboard 26 CD-ROM drive 27 Floppy disk drive 28 Communication device 29 Bus PC1-PCn Computer 30 Character recognition device 31 Dimensional compression means 32 Learning sample feature vector holding means

Claims (6)

[Claims] An image feature is extracted by performing predetermined image processing on a character image, collated with an image feature storage dictionary, and an error correction is performed based on language information stored in the image feature storage dictionary. A character set to be recognized among the character images can be arbitrarily set as a character set to be recognized, and the image feature storage is performed only for the set character set to be recognized. A character recognition device for performing character recognition processing using a dictionary. 2. The character according to claim 1, wherein the character recognition device enables the character set to be recognized to be appropriately changed, and performs the character recognition processing on the changed character set to be recognized sequentially. Recognition device. 3. The character recognition device according to claim 1, wherein the character set to be recognized can be set for each character to be recognized, and the character recognition processing is performed on the set of character set to be recognized. The character recognition device according to claim 1 or 2. 4. The character recognition device according to claim 1, wherein the character set to be recognized is set based on the language information stored in the image feature storage dictionary. 4. The character recognition device according to claim 1, wherein a recognition process is performed. 5. The character recognition apparatus according to claim 1, wherein a plurality of characters are collectively set as one group as the character to be recognized in the same manner as character data, based on the language information stored in the image feature storage dictionary. The character recognition device according to claim 4, wherein the character recognition processing is performed on the set of the recognition target character set. 6. A program for executing the character recognition processing program required for executing the character recognition processing according to claim 1 and the character recognition processing program such as the image feature storage dictionary. A recording medium for recording necessary data.