JPH0654503B2 - Pattern recognition device - Google Patents

Description

TECHNICAL FIELD OF THE INVENTION The present invention relates to a pattern recognition device that can effectively improve the recognition rate.

[Technical background of the invention and its problems]

Recently, pattern recognition techniques such as character recognition, voice recognition, and drawing reading have been remarkably developed. For example, a voice word processor, a circuit drawing reading device, and the like are entering the stage of practical application. However, in pattern recognition targeting an unspecified number of users and pattern recognition in which the number of categories to be recognized is large, the recognition rate is still low, which is a major problem in pattern recognition technology.

By the way, in order to improve the recognition rate, it is known that a large number of patterns are collected and learned in advance, and a dictionary for recognition used for pattern recognition is sufficiently prepared. However, when there are a large number of categories to be recognized, for example, when thousands of kanji are to be recognized, it is practically impossible to collect sufficiently many deformed patterns of those kanji. is there. Moreover, when targeting an unspecified large number of users, the unique habit of the pattern input person has a great influence, and it is extremely difficult to sufficiently deal with the deformation of the pattern caused by this habit. There is. Further, in order to avoid this, if the pattern input is repeated and learning is performed, the pattern input person is forced to bear too much burden.

[Object of the Invention]

The present invention has been made in consideration of such circumstances, and its purpose is to collect learning patterns effectively so as to adapt to a specific user or a specific usage environment for pattern recognition. It is possible to provide a highly practical pattern recognition device that can enhance the recognition dictionary, reduce the burden on the user for learning the recognition dictionary, and can improve the recognition rate.

[Outline of Invention]

The present invention focuses on the following points. That is, what causes misrecognition during pattern recognition depends on the individuality of the user, the usage environment, and the similarity between recognition target categories, and misrecognition is concentrated in a specific category. Therefore, many categories can be covered by a normal recognition dictionary.

Therefore, in the pattern recognition device according to the present invention, the recognition means for recognizing the input pattern by collating the input pattern with the recognition dictionary, and the true category in the input pattern based on the recognition result of the recognition means Means for identifying the category obtained as the misrecognition result, and a pattern including the true category that caused the misrecognition and a pattern including the category obtained as the misrecognition result, based on each category identified by this means Is input as a learning pattern, and a means for learning / updating the recognition dictionary using each pattern input by this means.

In other words, based on the recognition result for the input pattern, the recognition dictionary is learned by using a pattern including a specific category that is likely to cause erroneous recognition or conflict, so that the performance of the recognition dictionary can be improved or enhanced. It is designed to be simple and efficient.

〔The invention's effect〕

Thus, according to the present invention, the recognition target pattern to be learned is collected based on the recognition result of the input pattern, and for example, the recognition dictionary is learned by collecting the patterns that are likely to cause misrecognition and conflicts. By collecting a small number of learning patterns, the dictionary can be easily re-created as a highly reliable recognition dictionary to improve the recognition rate. That is, it is possible to easily utilize the relatively small amount of input pattern data and easily enhance the recognition dictionary.

Example of Invention

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

FIG. 1 is a schematic configuration diagram of the embodiment apparatus. Each element that constitutes this device is different depending on its recognition target, but is basically configured as follows. The pattern input unit 1 inputs pattern data to be recognized, and the features of the pattern input via the pattern input unit 1 are extracted by the preprocessing unit 2. Recognition unit 3
Is the distance or similarity between the feature vector of the input pattern thus obtained by the preprocessing unit 2 and the recognition dictionary of a plurality of categories registered in the dictionary storage unit 4, specifically, The Mahalanobis distance and the Euclidean distance are obtained, and the composite similarity is calculated to perform recognition processing. Then, a previously known category of the recognition dictionary having the minimum distance or the maximum similarity is obtained as the recognition result of the input pattern. The recognition result thus obtained is presented to the pattern input person (user) via the display unit 5.

Here, when the pattern to be recognized is a monosyllabic voice or a word voice, the pattern input unit 1 is composed of a microphone, an amplifier that amplifies an input signal obtained through the microphone, and the like. And the preprocessing unit 2
For example, square-law detection is performed on the output that has passed through a filter bank of 16 channels, and this is further detected in the time direction by 16
It is configured to resample over a frame to obtain a total of 256 feature vectors. In the case of word speech, the feature vector is expressed as a combination of the single syllable feature vectors as described above.

When the recognition target is a handwritten character pattern or the like, the pattern input unit 1 is composed of a photoelectric converter or the like, and the preprocessing unit 2 adds a so-called blur operation to the character pattern image to It is configured to extract a feature vector. When such a character pattern or the like is to be recognized, the input pattern extracted from the feature vector should be temporarily stored in the pattern memory 6 as shown in FIG. It will be convenient if you do.

In addition, it goes without saying that the pattern input unit 1 and the preprocessing unit 2 may be configured according to the type of the pattern to be recognized, and the feature vector of the input pattern suitable for the recognition may be extracted. In addition, it goes without saying that the recognition processing based on the feature vector of the input pattern extracted in this way may appropriately use various conventionally proposed methods.

By the way, the feature of the device of the present invention is that in addition to the main body of the pattern recognition device configured as described above, the recognition dictionary stored in the dictionary storage unit 4 is used in accordance with the recognition result of the device body. The point is that learning and updating can be done. The learning input category determination unit 7 inputs the recognition result obtained by the recognition unit 3 and the instruction information of the pass / fail of the pattern input person for the recognition result, and determines the category to be learned based on this. It is a thing. For example, a confusion matrix for the input pattern is created from the information of “correct”, “wrong”, and “conflict” for the recognition result, the rank of the candidate category obtained in the recognition process and the similarity value, and the confusion matrix is created.・
The matrix determines the patterns to be learned.
Specifically, a pattern including a category that causes an erroneous recognition result and a pattern including a category obtained as the erroneous recognition result are determined as patterns for learning, and a combination of categories in which a conflict has occurred is checked. It is determined as a pattern learning target including a set of categories. Then, the input category determination unit 7 prompts the pattern input person (user) via the display unit 5 to input the pattern including the category to be learned thus obtained. The learning unit 8 learns the recognition dictionary registered in the dictionary storage unit 4 using the pattern input in response to such an input request and updates the recognition dictionary. In this learning, for example, one set or a plurality of sets of feature vectors of the input pattern are extracted, and a feature vector component effective for discriminating the two is added to the dictionary pattern from the difference between the feature vectors between the categories in which an error occurs. Done. This makes the dictionary pattern suitable for the person who inputs the pattern.

It should be noted that it is convenient to collect the learning patterns by setting the number of times of inputting the learning patterns in accordance with the degree of erroneous recognition or conflict, because the learning of the patterns can be intensively and effectively performed. Further, when there is a pattern in which misrecognition or conflict is expected in advance, the input of the pattern is requested prior to the pattern recognition processing, and the recognition dictionary is learned according to the input pattern. It is also effective to perform the above pattern recognition processing.

It should be noted that reference numeral 9 in FIG. 1 denotes a control unit that controls these series of controls.

Thus, according to the device configured as described above, from the recognition results subjected to the recognition process based on the dictionary pattern registered in advance, only the categories that are likely to cause misrecognition or conflict are selected, and the category that causes the misrecognition result is selected. It collects only a small number of pattern data consisting of patterns that include and patterns that include categories obtained as a result of misrecognition, and then learns them, which incurs a large burden on the collection of pattern data for learning. There is no. Further, since the recognition dictionary can be effectively learned with a small number of pattern data as described above, the recognition dictionary can be made highly reliable and adapted to the person who inputs the pattern.
Therefore, it is possible to effectively improve the recognition rate.

Next, a specific operation example of the device of the present invention will be described.

FIG. 3 shows the flow of processing when recognizing a monosyllabic voice pattern. For example, when implementing a pattern recognition device suitable for a specific speaker by using a device equipped with a recognition dictionary for an unspecified speaker, first, a single syllable to be input is displayed on the display unit 5 such as a CRT. Prompt for vocalization as a word consisting of syllable combinations. Then, the recognition process is performed on each of the monosyllabic words that form each of these words, and a confusion matrix is created from the recognition result. At this time, it is preferable to store the input pattern and use it for later learning. After that, a set of the category of the input pattern that causes the erroneous recognition result, the category obtained as the erroneous recognition result, and the category in which the conflict occurs are extracted from a large number of recognition results. After that, a sequence of monosyllabic voices including these categories, for example, a voicing input of a word is requested, and the feature vector of the learning target single syllable in the input voice is obtained. The recognition dictionary is learned according to the feature vector of each monosyllabic pattern thus obtained. For example, when there is a high probability that a single syllable "pa" is mistakenly recognized as "a", a word including the single syllable "pa" (for example, "ka-n-pa-
I)) and a single syllable "a" (for example, "shi-a-i") input pattern, the feature vector of the corresponding monosyllabic is extracted, and the dictionary for the monosyllabic category is learned.

As a voice recognition device for an unspecified speaker, word voice recognition of a city name or the like input via a public telephone line can be considered. In the case of such an input form, generally, there are many variations in the input pattern due to the transmission characteristics of the line, etc. When changing the recognition target vocabulary, a large amount of input word voice patterns are used for learning and a recognition dictionary is created. It is necessary to.
However, in the case of effectively learning the recognition dictionary, in the present device, dictionary learning is proceeded interactively as follows. That is, when voice is input during operation of the voice recognition system, the recognition result is transmitted to the input person via, for example, a voice synthesizer. For example, in response to a voice input "Hamamatsu", the first place of the recognition result is input by a response such as "The input is" Takamatsu "isn't it?" The recognition result is output in order from the second rank, and the recognition result is confirmed. In this case, "Yes",
Since "No" and the like are clearly different from other voices and it is easy to clearly distinguish them, they can be used as they are as information for pass / fail determination. In this way, by confirming the word voice input, the correspondence between the true category of the input pattern and the category obtained as the erroneous recognition result can be obtained. Therefore, as shown in FIG.
When operating the system, a confusion matrix of recognition results can be easily created. Therefore, it is possible to effectively learn the word dictionary by determining the category to be learned according to the confusion matrix and urging the pattern input or collecting the pattern of the category during system operation. .

Further, in the case of a continuous numeral voice recognition device for a specific speaker, it was difficult to learn the recognition dictionary because of the large number of combinations of numbers, but the procedure shown in FIG. You can do it. That is, a numerical sequence to be input for initial learning, for example, “3/6/5/3”, “3/7/5/3”, “4.
3/6/9 ”etc. are presented to the input speaker, and the recognition device sequentially recognizes the input voice using the recognition dictionary of the non-special speaker.
The recognition result is stored. At this time, it is effective to store the continuous numeric voice pattern. Thus obtained,
Using the recognition result of the continuous number voice for initial learning, the continuous number used for learning the recognition dictionary is determined as follows.
That is, the input continuous numeric voice “4,3,6,9”
Is mistakenly recognized as "4, 3, 5, 9", etc.,
Concatenated numeral voices including "6.9" are used for learning. For example, "3,8,6,9", "6,9", "2,6,9"
Etc. are presented to the speaker, the corresponding continuous voice is input, the numerical voice / pattern part to be learned is cut out, and a recognition dictionary suitable for the input speaker is created. Thus, the apparatus according to the present embodiment is advantageous because it is possible to effectively learn different numbers depending on the speaker and the habit of continuously speaking the numbers without burdening the user.

It is to be noted that such a continuous numeral voice recognition device can be configured for a specific speaker from the beginning even if it is not based on that of an unspecified speaker.

Furthermore, the present embodiment can be applied to other continuous word recognition.

By the way, although the above example gives an input pattern as a voice, in recognition of a handwritten character pattern, etc.,
The present invention can be similarly applied. In this case, for example, the sixth
Dictionary learning may be performed as shown in the figure.
That is, when the recognition target is a Chinese character pattern, the number is extremely large. Therefore, the character pattern input by handwriting is recognized and a confusion matrix of the recognition result is created. Then, the user is prompted to input a character pattern category in which erroneous recognition or conflict has occurred, or a category inferred from the erroneously recognized category, and the dictionary learning is performed using these input patterns.
For example, in general, a conflict often occurs in "knowledge, texture, job" and "7, wa, ku, fu", etc. Therefore, the recognition dictionary is learned for each of these categories.

By doing this, as in the case of the voice recognition described above,
It becomes possible to effectively learn the recognition dictionary.

Furthermore, the present invention can be applied to learning a dictionary pattern in speaker recognition and writer recognition. Also,
The present invention is also applicable to symbol / shape recognition in a handwriting drawing reading device and quality recognition for an object whose image has been input. That is, in this quality recognition, features such as shape, size, and color of the image-inputted object are extracted and classified (ranked), which in this case serves as a criterion for ranking. It is necessary to adapt the parameters to the object and learn them, and the present invention can be applied to the learning of such a recognition dictionary.

As described above, according to the present invention, based on the recognition result of the input pattern, the recognition dictionary is intensively learned only for the categories in which erroneous recognition and conflicts are likely to occur. Can be adapted to any person. Further, since the special learning is not performed for the category in which the correct recognition result is obtained, the learning effect is very high. Therefore, when there are a large number of categories to be recognized, when there are many deformations of the input pattern, and when the effect of the individuality of the pattern input person, that is, when the deformation of the input pattern is large, and when there are many similar categories For example, a great effect can be achieved in adapting the dictionary.
Then, without collecting the pattern data for all the recognition target categories, only the patterns including the minimum necessary category that easily causes misrecognition can be collected, and the dictionary learning can be performed by effectively using the small number of patterns. Therefore, the burden on the pattern provider for pattern collection can be significantly reduced.

The present invention is not limited to the above embodiment. For example, regarding the type of feature vector representing the input pattern and the pattern recognition method using the feature vector, various conventionally known methods may be appropriately used according to the application. Further, it goes without saying that the present invention can be applied to pattern recognition other than the above-mentioned examples, and in short, the present invention can be variously modified and implemented without departing from the scope of the invention.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of an embodiment apparatus of the present invention, FIG. 2 is a principal configuration diagram showing a modified example, and FIGS. 3 to 6 are views showing a flow of learning processing in the embodiment apparatus. . 1 ... Pattern input section, 2 ... Preprocessing section, 3 ... Recognition section, 4 ...
Dictionary storage unit, 5 ... Display unit, 6 ... Pattern memory, 7 ... Learning input category determination unit, 8 ... Learning unit, 9 ... Control unit.

Claims (2)

[Claims] 1. A recognition means for recognizing the input pattern by collating the input pattern with a recognition dictionary, and a true category in the input pattern and a category obtained as an erroneous recognition result based on the recognition result of the recognition means. Based on the means to identify and each category identified by this means,
Means for inputting, as a learning pattern, a pattern including a true category that has caused misrecognition and a pattern including a category obtained as a misrecognition result, and learning of the recognition dictionary by using each pattern input by this means・
A pattern recognition device comprising means for updating. 2. The category is a part or all of the constituent elements of a pattern, and the means for inputting includes means for requesting a user to input a pattern including the determined category. The pattern recognition device according to claim 1.