JPH0632086B2 - Pattern recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern recognition apparatus that targets characters or voices and also uses word matching using a previously registered word dictionary.

[Background of the Invention]

Conventionally, in a pattern recognition device, for example, a character recognition device, particularly when a character having a large number of character types, such as a Chinese character, is targeted for recognition, the recognition result is used to concatenate information and linguistic information of characters before and after word matching. It is known to use a process for improving performance by utilizing (see, for example, JP-A-58).
-39378). In this case, the combination of words to be recognized generally differs depending on the purpose of use of the user. For this reason, in the conventional character recognition device, the word dictionary used for word matching is composed of words in a limited range such as an address and a name, and therefore the usage of the target is limited. In this way, the character types to be recognized (for example, the types that are distinguished by the character code such as “East” and “Kyo” and are referred to as “character types” below) are limited according to the word dictionary used. However, both the word dictionary and the recognition dictionary for character recognition are fixed, and no particular consideration is given to how they are used by the user.

On the other hand, in the character recognition device for alphanumeric characters and katakana symbols, the recognition performance can be improved by switching the contents of the recognition dictionary according to the character type set to be read, such as the number mode for reading only numbers or the mixed mode for reading all character types. Is known to improve. As one example, the user specifies the contents of the character type set using a flag or the like,
As a result, the recognition performance is improved by limiting the recognition target character type. However, like Kanji,
When there are many recognition target character types (1000 to 3000 character types), it is very troublesome for the user to specify each character type by setting a flag corresponding to each character type. .

[Object of the Invention]

An object of the present invention is to provide a pattern recognition device which can easily limit the recognition target character type according to the purpose of use of the user and improve the recognition performance, excluding the above-mentioned drawbacks of the prior art. .

[Outline of Invention]

In order to achieve this object, the present invention makes it possible to register a word dictionary according to the purpose of use of the user and to limit the recognition target character type of the recognition dictionary according to the character type of the word dictionary. There are features.

Here, the processing procedure of the present invention will be schematically described with reference to FIG.

In the figure, the user first registers the word to be read from a keyboard or the like in advance, and inputs information for creating a word dictionary. Next, only different character types are extracted from the registered word dictionary (analysis of appearance character types), and a recognition target character type table is automatically created.
In the conventional OCR, it is known to use flag information in order to define a character type used by a user from about 150 types of alphanumeric characters and katakana symbols. In this case, it is necessary to set a flag for each character type in the definition of the character type set, and when selecting the necessary character type from about 150 types, the number of character types is small, so there is almost no problem. However, there is a drawback that it takes a lot of time to select the required character type when there are more than 1000 character types such as Kanji.

In the present invention, in order to eliminate this drawback, as described above,
One of the points is to automatically register the recognized character type set from the recognition dictionary according to the words registered for use.

In the present invention, when recognizing a character, for example, J
A character type for recognition, which is intended for a relatively large character type set that is generally used, such as the IS first level is used. Therefore, when registering a used word, it is necessary to check whether there are character types that are not registered in the recognition dictionary, and if there are character types that do not exist in the recognition dictionary, it is necessary to change the registered word. There is. In order to deal with such a case, it is necessary to prepare some recognition dictionaries up to JIS Level 2 and some with special characters added as options.

Here, the reason for using a relatively large character type set as a recognition dictionary for character recognition will be described. In general, it is more efficient to construct the recognition dictionary according to the character type set in terms of processing amount and recognition performance. However,
To reconstruct the recognition dictionary according to the character type set,
It is not enough to simply collect the required standard patterns of the character type, and the structure of the target pattern space changes due to the change of the character type set. Therefore, it is necessary to adjust the parameters related to the interrelation such as the judgment threshold value. Actually, it is difficult to automatically perform this parameter adjustment, and adjustment is usually performed by trial and error, which is an important know-how of each manufacturer. Therefore, it is advantageous in terms of development efficiency to fix the recognition dictionary corresponding to the character set as much as possible.

The result of character recognition as described above is output in the form of a recognition result candidate character table for each input word.
Only character types existing in the recognition target character type table are extracted from the recognition result candidate character table, and a corrected candidate character table is created. By this processing, only the recognition target character type becomes the recognition result without changing the recognition dictionary.

Then, using the already registered word dictionary, word matching is performed on the corrected candidate character table, and the most probable word is output as the recognition result.

FIG. 5 shows an example in which different character type sets are extracted from the registered word dictionary. In this example, four words are registered, and the circled "day", "month" ...
The 11 characters of "Kaku" are different character types, and these are the character types to be recognized.

FIG. 6 shows an example of processing from character recognition to word matching in FIG. FIG. 6 (a) shows the recognition result candidate character table of FIG. 4 when "Sun Moon Research Institute" is used as an input word, and the candidates of up to the third place are output. Among these candidates, the ones that have entered the recognition target character type table (FIG. 4) recognition target character type shown in FIG. 5 are circled. FIG. 6 (b) shows a modified candidate character table (FIG. 4) created by picking up the circled candidate characters of FIG. 6 (a), and the candidates up to the second place are It is rising. When this modified candidate character table is collated with the word dictionary shown in FIG. 5, it is most consistent with the circled “Higetsu Institute” in FIG. 6 (b), and FIG. 6 (c). The results shown in are obtained.

As described above, by automatically extracting the recognition target character type set from the word dictionary registered by the user, it is possible to eliminate the trouble of individually setting the character type definition flags as in the conventional method. There is a first feature of the invention.

The recognition dictionary is a standard pattern dictionary corresponding to a general-purpose character type set, and changes in the recognition target character type set are handled by modifying the recognition result candidate character table after recognition, and the recognition dictionary needs to be changed. do not do. This is the second feature of the present invention. As a result, it is not necessary to readjust the recognition determination parameter in response to the change in the recognition target character type set.

Example of Invention

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

FIG. 1 is a block diagram showing an embodiment of a pattern recognition device according to the present invention. The entire control is performed by the control unit 100, which is realized by an ordinary microcomputer. In the figure, the line notation showing the flow of control information from the control unit 100 is omitted.

First, the user inputs a word to be read from the terminal 10 including a video data terminal (VDT) and a keyboard using, for example, kana-kanji conversion, and registers it in the word dictionary 25 in the word registration unit 20. Next, the recognition target character type extraction unit 30 extracts different character types and creates a recognition target character type table 35.

On the other hand, the character pattern written on the form is observed by the solicitation unit 40 and converted into a digital pattern by the scanner. The image pattern 41 for one form is separated into character patterns for each character in the character cutting section 50, and pre-processing required for character recognition such as normalization is performed. The character pattern 51 cut out for each character is sent to the recognition unit 60. In the recognition unit 60, after the features used for recognition are extracted, character recognition is performed by matching with the standard pattern dictionary 65 and determination. As a result of the character recognition, a plurality of candidate character sequences 61 for each word, that is, the recognition result candidate character table in FIG. 4 is obtained and sent to the recognition result correction unit 70 together with the recognition target character type table 35 created in advance. In the recognition result correction unit 70, only the character types existing in the recognition target character type table 35 are extracted from the plurality of candidate character sequences 61, and the corrected candidate character sequence 71 is extracted.
(That is, the corrected candidate character table of FIG. 4) is created. Next, the candidate character string 71 is sent to the word matching unit 80, and the word dictionary 25 is matched with the rank difference between the word dictionary 25 and the first-ranked candidate, for example, and the word 81 having the smallest distance is output. The output word 81 is displayed on the display unit 90 and is sequentially stored in the recognition result storage unit 95.

In the embodiment described above, the observation unit 40, the character cutout unit 50, the recognition unit 60, the word matching unit 80, the display unit 90, and the recognition result storage unit 95 can be realized by known techniques.

Here, the word registration unit 20 and the recognition target character type extraction unit 3
The process of 0 will be described in detail with reference to FIG.

First, the kana character sequence input from the VDT and keyboard terminal 10 is stored in the input character buffer 11. The kana character series stored in the input buffer 11 is converted into a kanji series by the kana-kanji conversion section 21 using the kana-kanji conversion dictionary 22 and stored in the converted character buffer 23. Next, in the word registration unit 24, the character series of the words extracted from the converted character buffer 23 are sequentially checked, and after it is confirmed that they are registered in the standard pattern dictionary 65, they are registered in the word dictionary 25. If a character not registered in the standard pattern dictionary 65 appears, the fact is displayed on the VDT 10 and an instruction to change the registered word is given. Next, at 31 in the character series of the registered words, different characters are checked. This is the recognition target character type table 3 that has already been registered.
5 is compared with the sequentially extracted characters, and in the case of a newly appeared character type, the character registration unit 32 registers it in the recognition target character type table 35.

Next, the processing contents of the recognition result correction unit 70 in FIG. 1 will be described with reference to FIG.

In the candidate character sequence 61 output from the recognition unit 60, if the input word length is L and the recognition result candidate character number of the i-th character in the word is n (i), Is represented. On the other hand, the recognition target character type table is: {D (k), k = 1 ..., M} where M is the number of recognition target character types.

At this time, the recognition result correction section 70 outputs the corrected candidate character sequence 71 by the process shown in FIG. The candidate character sequence 71 corrected here Represents.

That is, first, the counter i is set to 0, the counter i is sequentially incremented, and the following processing A is repeated until the input word length L is reached.

Process A: First, the counter l is set to 0, and the counter J is used to repeat the following process B until the recognition result candidate character number n (i) of the i-th character is reached.

Process B: Using the counter k, the candidate C (i, J) is recognized as a character type table D until the number of character types for recognition M is reached.
(k) is compared or not, and if they match, the counter 1 is incremented, and a new candidate table C ′ (i,
Set D (k) to J). If k exceeds M,
The process returns to the increment process of the counter J of the process A.

In this example, as a method of inputting a word to be registered,
Kana-to-Kanji conversion using a keyboard is adopted. In addition to this, tablet input by pen touch, input by associative code, input by OCR, input by combination of voice recognition device and Kana-Kanji conversion, or online character recognition device It goes without saying that it can be realized even by using a method such as input by. Further, although different character types are extracted after the word registration, it is clear that the same effect can be obtained even before the word registration.

Further, if the characters of the registered word are not registered in the recognition dictionary, it is needless to say that instead of displaying that effect and changing the registered word, the recognition dictionary having a larger number of character types may be switched.

Needless to say, even when the recognition target is speech, it can be applied as a large-scale vocabulary word speech recognition method. In this case, it is conceivable to adopt a phoneme, a phoneme, a syllable, or the like as the recognition unit, and these are the above words.

As described above, according to this embodiment, since the read word can be registered for each user and the recognition target character type need not be individually specified, the purpose of use can be achieved without changing the recognition unit and the recognition dictionary. There is an advantage that the character reading of the word suitable for can be performed.

〔The invention's effect〕

As described above, according to the present invention, it is possible to realize the recognition in which the recognition target most suitable for the purpose of use of the user is limited, and
Since the parameter adjustment of the recognition dictionary that is delicately adjusted is unnecessary, the excellent effect that the number of steps for creating a dictionary corresponding to each recognition unit can be reduced can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a pattern recognition system according to the present invention, FIG. 2 is a flow chart showing processing of a word registration unit and a recognition target character type extraction unit in FIG. 1, and FIG. 3 is FIG. FIG. 4 is a flowchart showing the processing of the recognition result correction unit in FIG. 4, FIG. 4 is a schematic explanatory view of the processing procedure of the present invention,
FIG. 5 is an explanatory diagram of the procedure for creating the recognition target character type table in FIG. 4, and FIG. 6 is an explanatory diagram of the procedure from character recognition to word matching in FIG. 20 ... Word registration unit, 30 ... Recognition target character type extraction unit, 2
5 ... word dictionary, 35 ... recognition target character type table, 60
...... Character recognition unit, 70 ...... Recognition result correction unit, 80 ...... Word matching unit.

Claims (3)

[Claims] 1. An input means for inputting a word to be read, a word registration dictionary for registering the word input by the input means, and a character different from the already registered word among the registered words. A recognition target character type table created by extracting seeds, an observation unit that observes a character pattern written on a form and outputs an image pattern, a unit that creates a character pattern from the output of the observation unit, and the character A recognition unit that compares a pattern with a standard pattern and outputs a plurality of candidate characters, and extracts only the character types existing in the recognition target character type table from the plurality of candidate characters to create a corrected candidate character table. A recognition correction result correction unit for outputting, a word matching unit for matching the corrected candidate character table with the word dictionary and outputting a recognition result, and an output of the word matching unit are stored. A recognition result storage section, a pattern recognition apparatus characterized by comprising a display means for displaying the output of the word collating unit. 2. The pattern recognition apparatus according to claim 1, wherein the word dictionary for registering a word input from the input means stores a character subjected to kana-kanji conversion processing. Pattern recognition device. 3. The pattern recognition device according to claim 1, wherein the character pattern is output to the recognition unit via a character cutout unit that separates the output from the observation unit into character patterns one by one. A pattern recognition device.