JPH11134316A - Pattern recognition method and pattern recognition dictionary preparation method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern recognition dictionary preparing device for effectively preparing a pattern recognition dictionary by easily detecting a range in which an error is likely to be generated at the time of pattern recognition. SOLUTION: A pattern converting part 23 converts a feature vector in which erroneous recognition is likely to be generated into an original pattern by using a genetic algorithm. When a belonging category can be specified, this original pattern is registered as a reference pattern, and obtained as a sample for deciding a category. When the belonging category of the original pattern can not be specified, this original pattern is registered as a reject pattern being the index of pattern reject at the time of recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition technique for recognizing patterns of characters, voices, and the like, and more particularly, to a technique for creating a pattern recognition dictionary used for pattern matching.

[0002]

2. Description of the Related Art Generally, in character recognition or voice recognition, a character or voice to be subjected to pattern recognition is divided into a plurality of categories to extract a pattern vector (hereinafter, feature vector) representing each feature. The similarity between the feature vector and a reference pattern vector (hereinafter, referred to as a reference vector) representing the characteristics of each category prepared in the pattern recognition dictionary is determined using some evaluation function, for example, a Euclidean square distance, and the distance between the two vectors is determined. Is output as a recognition result. In this case, for characters, one character unit such as "A", "I", "U", etc. is defined as one category, and for speech, phonemes, phonemes, etc. constituting speech are classified into one category. Category.

Normally, a reference vector is given as an average vector of vectors (sample vectors) representing characteristics of a plurality of sample patterns existing for each category. However, characters such as numbers and hiragana tend to have various fonts depending on the writer, and there are individual differences in vocalization in speech recognition.To cope with these, multiple reference vectors are set for each category. You need to keep it.

As described above, a pattern recognition dictionary in which a plurality of reference vectors are collected for each category can be created by, for example, a learning vector quantization method (hereinafter, LVQ). Regarding LVQ, refer to the reference "T. Kohonen, G. Ba.
rna, and R.Chirisley, "Statistical pattern with neur
al networks: Benchmarking studies. ", Proceedings of
the IEEE International Conference on Neural Networ
ks, pp. I-61-68, 1988 ”, in short, a method of optimizing the vector space by performing learning while correcting errors when generating multiple reference vectors And consists of the following algorithm: (1) A plurality of reference vectors are generated by performing in-category clustering processing or the like on all sample vectors (learning vectors) used for learning, and a set of these reference vectors is used as an initial dictionary. (2) Detect a reference vector closest to one learning vector whose category is known. (3) If the detected reference vector belongs to a category different from the learning vector, that is, if the recognition vector is incorrectly recognized, the reference vector is moved away from the learning vector. On the other hand, if the detected reference vector belongs to the same category as the learning vector, the reference vector is brought closer to the learning vector. By optimizing the vector space in the initial dictionary, the above process is repeated until erroneous recognition is eliminated or a predetermined number of repetitions is reached. As described above, since the LVQ has a simple algorithm and can be expected to be effective in many cases, the LVQ is often used for creating this type of pattern recognition dictionary.

[0005]

The above-mentioned reference vector is generally generated from a large amount of data by a statistical method or an artificial neural network method including the above-mentioned LVQ. Very high cost. Further, even if a large amount of data is collected, it is not always possible to obtain data necessary for improving the performance of the pattern recognition apparatus, so that there has been a problem that the efficiency of dictionary creation cannot be improved.

In actual pattern recognition, it is known that there is a pattern which cannot be judged by humans in an intermediate part of a plurality of categories. Was difficult. Therefore, there is no clear index for rejecting (rejecting) the recognition target pattern, and there is a problem that pattern recognition cannot be performed with high accuracy.

SUMMARY OF THE INVENTION An object of the present invention is to provide an improved pattern recognition method which can solve such a conventional problem and improve the accuracy of pattern recognition. Another object of the present invention is to provide a pattern recognition dictionary creation method for efficiently creating a pattern recognition dictionary capable of improving the accuracy of pattern recognition, a pattern recognition dictionary creation apparatus suitable for implementing the method, and a general-purpose computer. An object of the present invention is to provide a recording medium suitable for creating a pattern recognition dictionary.

[0008]

A pattern recognition method according to the present invention which solves the above-mentioned problems includes a reference pattern and a pattern used for pattern matching by applying a genetic algorithm to a sample pattern in which misrecognition is likely to occur in pattern recognition. It is classified into any one of reject patterns serving as reject indicators, and rejecting the recognition target pattern when the degree of coincidence between the characteristics of the input recognition target pattern and the characteristics of the reject pattern exceeds a predetermined value. Features.

According to another pattern recognition method of the present invention, a genetic algorithm is applied to a pattern in which misrecognition is likely to occur during pattern recognition, either a reference pattern used for pattern matching or a reject pattern serving as an index of pattern rejection. If the feature difference or the feature ratio between the reference pattern most similar to the input recognition target pattern and the second similar reference pattern is equal to or less than a predetermined value, the reject pattern and the recognition target pattern Is detected, and when the detected value exceeds a predetermined value, the recognition target pattern is rejected. When the feature difference or the feature ratio exceeds a predetermined value and the detected value is equal to or less than a predetermined value, a predetermined number of reference patterns are specified as correct candidate patterns.

[0010] In the pattern recognition dictionary creating method according to the present invention for solving the above-described other problems, a sample pattern whose affiliation category is unknown (for example, a pattern similar to a plurality of reference patterns belonging to different categories) is grown by a genetic algorithm. And the step of determining the presence or absence of a category belonging to the pattern after growth. If the belonging category can be identified, the pattern after growth is used as a reference pattern for pattern matching. This method includes a step of classifying a later pattern as a reject pattern serving as an index of pattern reject, and a step of dictionary-registering feature data of each classified pattern.

In this case, the growing pattern may be converted into a visible original pattern and presented to the operator, and the pattern may be classified according to an instruction from the operator based on the presentation. For example, when the belonging category is designated, the original pattern is classified as a reference pattern used for pattern matching, and when the belonging category is not determined, the original pattern is classified as a reject pattern serving as an index of pattern rejection. The determination of the belonging category is performed each time the genetic algorithm is repeated.

A pattern recognition dictionary creating apparatus according to the present invention for solving the above-mentioned other problems can provide a pattern conversion means for growing a sample pattern whose categorization is unknown by a genetic algorithm, and can specify the categorization of the pattern after the growth. Category determining means for determining whether or not a pattern whose belonging category can be specified is classified as a reference pattern used for pattern matching, and a pattern whose belonging category cannot be specified is classified as a reject pattern serving as an index of pattern rejection. Means for registering in a predetermined area of the pattern recognition dictionary.

According to another aspect of the present invention, there is provided a recording medium for growing a sample pattern used for creating a dictionary by a genetic algorithm, determining whether a category belongs to a pattern after growth, and specifying a belonging category. If successful, a process of classifying the pattern after growth as a reference pattern used for pattern matching, and if the belonging category cannot be specified, classifying the pattern after growth as a reject pattern serving as an index of pattern rejection; and
This is a computer-readable recording medium that records a program that causes a computer to execute a process of registering feature data of each classified pattern in a dictionary.

[0014]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a configuration diagram of a pattern recognition device to which the present invention is applied. This pattern recognition device 1 is realized by a computer device.
It includes functional blocks of a pattern input unit 10, a preprocessing unit 11, a feature extraction unit 12, a collation unit 13, a pattern recognition dictionary 14, and a pattern output unit 15, which are formed by a computer device reading and executing a predetermined program. Consists of

The above program is usually incorporated as needed into an internal storage device or an external storage device accessible by the computer device, and is used as needed. Since the present invention can be implemented, the recording form may be arbitrary. For example, it may be recorded in a computer-readable form on a portable recording medium such as a CD-ROM, and may be installed in the internal storage device or the external storage device at the time of use.

The pattern input unit 10 receives an input of a pattern to be recognized (hereinafter, an input pattern) from an image input device such as a scanner. The preprocessing unit 11 performs shaping such as noise removal of the input pattern. Feature extraction unit 12
Extracts a feature vector representing the feature of the input pattern. The matching unit 13 compares the feature vector extracted by the feature extracting unit 12 with the pattern recognition dictionary 14, and outputs a matching result to the pattern output unit 15. Pattern output unit 1
5 causes the collation result to be output to an output device such as a display.

The pattern recognition dictionary 14 has a reference vector storage area 141 and a reject pattern data storage area 1.
42 are formed. Reference vector storage area 141
Is an area for storing reference vectors representative of individual categories, and reject pattern storage area 142 stores reject pattern data (feature vector of reject pattern,
The same applies hereinafter).

The pattern recognition dictionary 14 is, for example, as shown in FIG.
Can be created by the pattern recognition dictionary creating apparatus 2 having the configuration shown in FIG. Like the pattern recognition device 1, the pattern recognition dictionary creation device 2 is also formed by a computer device reading and executing a required program. An original pattern DB (database) 22 storing a plurality of original patterns in a format that can be easily checked, and a pattern conversion unit 23 that grows an input sample pattern using the original patterns in the original pattern DB 22 and converts it into a new pattern. An original pattern output unit 24 for outputting the output of the pattern conversion unit 23 to an output device such as a display device, a category determination unit 25 for determining whether a belonging category can be specified for a pattern after growth, and a pattern recognition dictionary 14. For optimizing data and registering additional data Configured with the function of the management unit 26.

The pattern converter 23 of the present embodiment converts a sample pattern into an original pattern using a genetic algorithm. The conversion procedure in the pattern conversion unit 23 is as shown in FIG. 3. First, a feature vector is extracted from the sample pattern (step S101). Are selected to form an initial group (step S102). Thereafter, feature vectors corresponding to individual original patterns of the initial group are extracted and feature comparison is performed (step S103). Specifically, the extracted feature vectors are compared with each other on a scale such as a square error, and similarity evaluation is performed. Then, a set of a pair of similar patterns is created by using a weighted probability that the original pattern determined to be higher in similarity based on the similarity evaluation becomes higher (Step S).
104). A part of the components of one similar pattern and the other similar pattern are exchanged with each other, and the pattern of the next generation is replaced by 1
A pair is created (step S105). If this pattern does not conform to the sample pattern, the processing after step S103 is repeated (step S106: No), and the processing ends when the pattern conforms to the sample pattern (step S106: Yes).

In addition, regarding the genetic algorithm,
"Looking at the characters seen by OCR" (Sakano et al., 'Image Recognition and Understanding Symposium MIRU' 96pp1313)
1318) can be referred to.

Next, a procedure for actually creating a dictionary by the pattern recognition dictionary creating apparatus 2 configured as described above will be described with reference to FIG. First, the dictionary management unit 26 is activated, and the pattern recognition dictionary 14 based on the current sample pattern is optimized by the LVQ (step S201). That is, from the initial reference vector set (initial dictionary) stored in the reference vector storage area 141, a reference vector closest to one sample vector whose category is known (this is set as a learning vector) is detected. If the detected reference vector belongs to a different category from the learning vector, that is, if it is misrecognized, the reference vector is moved away from the learning vector. On the other hand, if the detected reference vector belongs to the same category as the learning vector, it is brought closer to the learning vector. This process is repeated until erroneous recognition is eliminated or a predetermined number of repetitions is reached.

Next, a feature vector that is likely to be erroneously recognized, that is, a feature vector in the vicinity of the midpoint of a reference vector that belongs to a different category and is adjacent to each other is randomly selected (step S202).

This feature vector is converted into a pattern
To convert the original pattern (step S203), and determine whether the category belonging to the converted original pattern can be specified by the category determination unit 25 (step S204).

This determination is made by presenting the original pattern converted by the pattern conversion unit 23 to the operator through the original pattern output unit 24 and decoding the input instruction content based on the presentation. In other words, if the operator has defined a belonging category for the original pattern, the belonging category is entered. If the belonging category is reserved, the fact is input to the category determining unit 25.
Generates a determination result according to the input instruction. At this time, it may be presented to the operator for each generation of the genetic algorithm and wait for an input instruction, or may be presented to the operator at the time when a predetermined plurality of generations of processing are completed and wait for the instruction at that time. It may be. Note that the processing in the pattern conversion unit 23 is repeated as necessary.

If the category determining unit 25 can identify the category to which the original pattern belongs, the dictionary managing unit 26
The feature vector extracted from the original pattern is additionally registered in the reference vector storage area 141 together with its belonging category (step S204: Yes, S205). On the other hand, when the category to which the converted original pattern belongs cannot be specified, the feature vector of the original pattern is stored in the reject pattern data storage area 142 (step S204: No, S206). The processing from step S202 is repeated as necessary (step S20
7). Thus, the pattern recognition dictionary 14 is created.

Next, a pattern recognition method of the present invention using the pattern recognition dictionary 14 created as described above will be described with reference to FIG. The input pattern received by the pattern input unit 11 of FIG. 1 is pre-processed by the pre-processing unit 11, and then a feature vector is extracted by the feature extraction unit 12 (step S301). The matching unit 13 compares the feature vector with the reference vector stored in the reference vector storage area 141 of the pattern recognition dictionary 14, and determines the belonging category of the reference vector having the smallest distance from the feature vector, The belonging category of the reference vector having the shortest distance is specified (step S302).

Next, whether the distance difference or the distance ratio between the minimum reference vector and the reference vector having the second smallest distance is equal to or smaller than a predetermined first threshold value, that is, whether the correct answer data and the incorrect answer data It is determined whether or not it is difficult to distinguish (step S303). If it exceeds the first threshold (step S303: Yes), the category is output to the pattern output unit 15 as correct data (step S304).

On the other hand, if the distance is equal to or smaller than the first threshold value (step S303: No), the distance between the feature vector and the reject pattern data in the reject pattern data storage area 142 is calculated (step S305). It is determined whether the distance is equal to or less than the second threshold, that is, whether the input pattern is similar to the reject pattern (step S306). If the distance is equal to or less than the second threshold (step S306: No), the input pattern is immediately rejected (step S307). When the distance exceeds the second threshold, the categories of a predetermined number of reference vectors are output to the pattern output unit 15 as correct answer candidate data (step S308).

As described above, according to the present embodiment, the sample pattern is converted into the original pattern by the genetic algorithm to check whether the operator can specify the category to which the operator belongs. The pattern recognition dictionary 14 can be easily created even from a sample pattern. In addition, when the category belonging to the original pattern can be specified, this is adopted as a reference vector, so that it is not necessary to collect a large amount of data for category determination as in the past, and it is necessary to create a pattern recognition dictionary. The required work will be greatly reduced.

When the category to which the original pattern belongs cannot be specified, the feature vector of the original pattern is stored in the reject pattern storage area 142 as reject pattern data. Performance can be greatly improved.

[0031]

As is apparent from the above description, according to the present invention, it is possible to efficiently create an optimal pattern recognition dictionary having a high performance and a reject function. Also, by using this pattern recognition dictionary, the accuracy of pattern recognition can be significantly improved as compared with a conventional pattern recognition device.

[Brief description of the drawings]

FIG. 1 is a block diagram of a pattern recognition apparatus to which the present invention is applied.

FIG. 2 is a block diagram showing an embodiment of a pattern recognition dictionary creating apparatus according to the present invention.

FIG. 3 is an explanatory diagram of a procedure of a genetic algorithm according to the embodiment.

FIG. 4 is an explanatory diagram of a procedure for creating a pattern recognition dictionary according to the embodiment.

FIG. 5 is a diagram illustrating a procedure of a pattern recognition method according to the embodiment;

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 pattern recognition device 10 pattern input unit 11 preprocessing unit 12 feature extraction unit 13 collation unit 14 pattern recognition dictionary 141 reference vector storage area 142 reject pattern data storage area 15 pattern output unit 2 pattern recognition dictionary creation device 21 sample pattern input unit 22 Original pattern DB 23 Pattern conversion unit 24 Original pattern output unit 25 Category determination unit 26 Dictionary management unit

Claims (10)

[Claims] 1. A genetic pattern is applied to a sample pattern in which misrecognition is likely to occur during pattern recognition, and the sample pattern is classified into one of a reference pattern used for pattern matching and a reject pattern serving as an index of pattern rejection. A pattern recognition method for rejecting the recognition target pattern when the degree of coincidence between the obtained characteristic of the recognition target pattern and the characteristic of the reject pattern exceeds a predetermined value. 2. A method in which a genetic algorithm is applied to a pattern in which misrecognition is likely to occur during pattern recognition to classify the pattern into one of a reference pattern used for pattern matching and a reject pattern serving as an index of pattern rejection. When the feature difference or the feature ratio between the reference pattern most similar to the recognition target pattern and the second similar reference pattern is equal to or less than a predetermined value, the degree of coincidence between the reject pattern and the recognition target pattern is detected, A pattern recognition method comprising rejecting a recognition target pattern when a detection value exceeds a predetermined value. 3. The method according to claim 2, wherein when the characteristic difference or the characteristic ratio exceeds a predetermined value and the detected value is equal to or less than a predetermined value, a predetermined number of reference patterns are specified as correct candidate patterns. Pattern recognition method. 4. A process of growing a sample pattern used for creating a dictionary by a genetic algorithm, a process of determining the presence or absence of a category to which the pattern after growth belongs. If the reference pattern used for pattern matching and the belonging category cannot be specified, the process of classifying the pattern after growth as a reject pattern that is an index of pattern rejection, and the process of registering feature data of each classified pattern in a dictionary Including, how to create a pattern recognition dictionary. 5. A process of converting a sample pattern used for creating a dictionary into an original pattern that can be visually recognized by a genetic algorithm and presenting it to an operator, and, when an affiliation category is specified by the operator based on the presentation, A step of classifying the original pattern as a reference pattern used for pattern matching, and a step of classifying the original pattern as a reject pattern that is an index of pattern reject when the determination of a belonging category is reserved; a step of registering feature data of each pattern in a dictionary; A method for creating a pattern recognition dictionary. 6. The pattern recognition dictionary creating method according to claim 4, wherein whether or not the category to which the original pattern belongs is determined each time the genetic algorithm is repeated. 7. The pattern recognition dictionary creation method according to claim 4, wherein the sample pattern is a pattern similar to a plurality of reference patterns belonging to different categories. 8. A pattern conversion means for growing a sample pattern whose affiliation category is unknown by a genetic algorithm, a category determination means for judging whether or not the affiliation category of the pattern after growth can be identified, and an affiliation category can be identified. Means for classifying a pattern as a reference pattern used for pattern matching, a pattern whose belonging category cannot be specified as a reject pattern serving as an index of pattern rejection, and registering characteristic data of each pattern in a predetermined area of the pattern recognition dictionary. Pattern recognition dictionary creation device. 9. The pattern conversion unit is configured to convert the grown pattern into an original pattern in a visually recognizable format and present the converted original pattern to an operator, and the category determination unit performs input corresponding to the presentation. 9. The pattern recognition dictionary creating apparatus according to claim 8, wherein the determination is made based on the determined category determination information. 10. A process for growing a sample pattern used for creating a dictionary by a genetic algorithm, a process for determining whether or not a category belongs to a pattern after growth, and a pattern matching for the pattern after growth when the belonging category can be specified. In the case where the reference pattern to be used for the belonging category cannot be specified, the computer performs a process of classifying the pattern after growth as a reject pattern serving as an index of the pattern reject, and a process of registering feature data of each classified pattern in a dictionary. A computer-readable recording medium that records a program to be executed.