JP4820830B2 - Pattern recognition apparatus, pattern recognition program, and pattern recognition method - Google Patents

Description

  The present invention relates to a pattern recognition apparatus, a pattern recognition program, and a pattern recognition method for recognizing patterns such as characters, images, and sounds.

  Currently, many pattern recognition techniques are used to identify a category belonging to an unknown pattern input to a recognition system.

  Among them, for example, one of character pattern recognition techniques used in a character recognition system uses a recognition dictionary.

  In this type of character recognition system, character recognition is performed according to the following procedure. First, an N-dimensional feature vector representing a standard feature for each character to be recognized is extracted from a standard feature vector dictionary as a recognition dictionary and stored in a memory.

  Next, N-dimensional feature vectors are similarly extracted from the character pattern to be recognized. The degree of similarity (or dissimilarity) was calculated as the degree of similarity between this feature vector and the standard feature vector of the dictionary stored in the memory, and the degree of similarity was the largest (the degree of dissimilarity). Characters having feature vectors are output as recognition symmetry candidates (recognition solutions).

  Conventionally, as a calculation method used to measure the similarity of feature vectors, for example, a subspace method, a pseudo Bayes identification method, and the like are known.

These methods determine the eigenvalues corresponding to the eigenvalues and eigenvalues of the autocorrelation matrix, select the eigenvalues and the corresponding eigenvectors based on preset threshold values, and use them to perform pattern recognition. By using only eigenvalues and eigenvectors for pattern recognition, the amount of calculation can be reduced and components unnecessary for recognition can be removed (see Non-Patent Document 1).
Ercchioya, "Pattern recognition and subspace method", Sangyo Tosho, (1986)

  In the case of the above prior art, the eigenvector to be used is determined by obtaining the eigenvalue that falls within a preset threshold value. For this reason, differences in individual characteristics of input patterns (for example, characters) are not considered, and all eigenvectors whose eigenvalues fall within the threshold are selected as objects to be used, which adversely affects the recognition of the input pattern in the selected one. Even if an eigenvector that affects the pattern is included, the eigenvector is used for recognition, and as a result, a pattern that is different from the pattern to be recognized is included as a recognition candidate. There were limits.

  SUMMARY An advantage of some aspects of the invention is that it provides a pattern recognition apparatus, a pattern recognition program, and a pattern recognition method capable of improving a pattern recognition rate.

In order to solve the above problems, a pattern recognition apparatus according to the present invention includes an input unit that converts an input pattern to be recognized into a feature vector for use in pattern recognition, and an eigenvector of a standard pattern that is used in pattern recognition. Is stored, a feature extraction unit that obtains a plurality of eigenvectors indicating features unique to the input pattern from the feature vectors converted by the input unit, a plurality of eigenvectors obtained by the feature extraction unit, and the The first recognition unit that compares the eigenvectors stored in the recognition dictionary to recognize the input pattern, outputs a score indicating the degree of matching for each of the compared eigenvectors, and the output result of the first recognition unit based on some the score, the reliability calculating portion for determining the results of the reliability of the recognition for each eigenvector, the reliability calculation unit A vector selecting unit that selects eigenvectors to be used in the next recognition from a plurality of eigenvectors by using the obtained reliability, performs recognition of the input pattern using the eigenvectors selected by the vector selection unit, the correct answer And a second recognizing unit that outputs the score of each correct candidate used for determining a candidate or a correct answer as a result of pattern recognition .

The pattern recognition program according to the present invention includes a computer that stores an input unit for converting an input pattern to be recognized into a feature vector for use in pattern recognition, and a standard pattern eigenvector used for pattern recognition. A feature dictionary, a feature extraction unit that obtains a plurality of eigenvectors indicating features unique to the input pattern from the feature vector converted by the input unit, a plurality of eigenvectors obtained by the feature extraction unit, and the memory stored in the recognition dictionary A first recognition unit that compares the eigenvectors with each other and recognizes the input pattern and outputs a score indicating the degree of matching for each of the compared eigenvectors; and based on the score that is an output result of the first recognition unit a reliability calculating unit for obtaining the results of the reliability of the recognition for each eigenvector, determined by the reliability calculating unit A vector selecting unit that selects eigenvectors to be used in the next recognition from a plurality of eigenvectors by using the obtained reliability, have rows recognition of the input pattern using the eigenvectors selected by the vector selection unit, the correct answer It is made to function as a 2nd recognition part which outputs the said score of each correct candidate used for a candidate or a correct answer as a result of pattern recognition .

The pattern recognition method of the present invention includes a step of converting an input pattern to be recognized into a feature vector for use in pattern recognition, and a step of obtaining a plurality of eigenvectors indicating features unique to the input pattern from the converted feature vector The input pattern is recognized by comparing the plurality of eigenvectors obtained by the feature extraction unit with the eigenvector stored in the recognition dictionary in which eigenvectors of standard patterns used for pattern recognition are stored. A step of outputting a score indicating the degree of coincidence for each of the compared eigenvectors, a step of determining the reliability of the recognition result for each eigenvector based on the score that is the output result of the recognition, and the obtained reliability Is used to select the eigenvector to be used in the next recognition from multiple eigenvectors And step, have rows recognition of the input pattern using the selected eigenvectors, and having a step of outputting the score for each correct answer candidate to be used to correct the candidate or correct decision as a result of the pattern recognition .

  According to the present invention, the pattern recognition rate can be improved.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a pattern recognition apparatus according to an embodiment of the present invention.

  As shown in FIG. 1, the pattern recognition apparatus includes an input unit 21, a recognition dictionary 27, a feature extraction unit 22, a recognition unit 24 having a first recognition unit 24a and a second recognition unit 24b, a reliability calculation unit 23, A vector selection unit 25, an output unit 26, and a reliability determination information storage unit 28 are provided.

  The input unit 21 is, for example, a device such as a scanner. In addition, the input unit 21 may be a computer connected to a network line. The input unit 21 acquires a pattern (input pattern) to be recognized such as characters, images, and voices. That is, the input unit 21 converts character image data (hereinafter referred to as “character pattern”) into a feature vector for use in pattern recognition when the input pattern is, for example, a character.

  That is, the input unit 21 receives a pattern to be recognized such as a character, an image, and a sound as an input, converts it into a format suitable for recognition processing, and outputs it to the feature extraction unit 22.

  When the pattern format is not suitable for the recognition process, the input unit 21 performs conversion by a known means. For example, in the case of a character pattern, processing such as extraction of individual character regions by character extraction, size adjustment by normalization, and binarization is executed, and converted into an appropriate format for recognition.

  The data after the conversion process is output to the feature extraction unit 22. The input unit 21 gives an instruction to the feature extraction unit 22 when designing the recognition dictionary 27.

  The recognition dictionary 27 stores a standard pattern recognition vector (eigenvalue and eigenvector) of each character used for character pattern recognition.

  The feature extraction unit 22 extracts a feature component from the character pattern input from the input unit 21, and represents the character pattern in the form of an N-dimensional feature vector (N: natural number). Taking a character pattern as an example, the feature vector is a vector having N variables such as a direction component and a density feature in each part in a character area divided into a certain size, and represents a feature of the input pattern. Note that not limited to characters, any variable may be used as long as the variable can be represented by a vector.

  In other words, the feature extraction unit 22 obtains a multidimensional recognition vector (hereinafter referred to as “recognition vector”) indicating the characteristic unique to the character pattern from the feature vector of the character pattern converted by the input unit 21. That is, the feature extraction unit 22 extracts a recognition vector for comparison with the recognition vector of the recognition dictionary 27 from the feature vector.

  The recognition unit 24 performs pattern recognition in comparison with the recognition vector of the recognition dictionary 27 by using this recognition vector to obtain the distance between the recognition vectors of each character by various methods such as inner product and Euclidean distance.

  In this embodiment, the eigenvalue and eigenvector of the autocorrelation matrix of the feature vector are used as an example of the recognition vector. The eigenvector is a vector having a constant size and a given direction, and the eigenvalue determines how many times the eigenvector is given. The larger the eigenvalue, the greater the feature amount.

  It should be noted that any parameter may be used as a recognition vector as long as the concept of the present invention that extracts and uses only the optimum part for pattern recognition is not limited to eigenvalues and eigenvectors.

  The first recognition unit 24a compares the plurality of eigenvectors obtained by the feature extraction unit 22 with the eigenvectors of each character stored in the recognition dictionary 27, and calculates a score indicating the degree of matching for each eigenvector of the compared characters. Output. The output in this case means that the obtained score is stored in the memory in association with each candidate.

  The reliability calculation unit 23 obtains the reliability of the recognition result of the first candidate from the relationship between the score of the first candidate and the score of the second candidate for each eigenvector.

  The vector selection unit 25 uses, as a recognition eigenvector useful for recognizing a character pattern input from the input unit 21, an eigenvector whose reliability is equal to or higher than a preset threshold value from the reliability obtained by calculation of the reliability. elect.

  The second recognition unit 24b recognizes the character pattern using the eigenvector selected by the vector selection unit 25.

  Thus, the recognition unit 24 compares the eigenvector of the character pattern obtained from the feature extraction unit 22 with the eigenvector of each character generated by the feature extraction unit 22 and held in the recognition dictionary 27, and Determine the character to which it belongs. Note that which eigenvector is used is determined based on a selection result by the vector selection unit 25 described later.

  Moreover, the comparison between vectors, that is, the determination of which character the character pattern is, is executed using a predetermined method. For example, there is a technique in which an inner product of vectors to be compared is calculated, the calculation result is regarded as a similarity, and a character having the maximum similarity is regarded as a recognition solution. In addition, any known method may be used as long as the probability of the character to which the character pattern belongs can be expressed by a numerical value, such as the Euclidean distance.

  The output unit 26 includes various interfaces and devices connected thereto, such as a monitor connected to a VGA interface, a printer connected to a USB interface, and an external computer connected to a LAN interface. The recognition result by the recognition unit 24 is passed to the output unit 26 and output from the output unit 26.

  The recognition result includes a character (text) recognized from the input character pattern, a score used when determining reliability, a higher-ranking character that is another correct candidate, a score of each correct candidate, and the like. In addition, said item is an example and you may discard and output arbitrarily the information which a user requires suitably. The display format may be an arbitrary format determined by the user as desired. As an example, the output unit 26 displays characters that are recognition solutions recognized by the recognition unit 24.

  The reliability determination information storage unit 28 is realized by a memory or the like. The reliability determination information storage unit 28 stores a threshold value for selecting an effective recognition eigenvector from the reliability.

  First, the dictionary registration operation of the pattern recognition apparatus will be described with reference to FIG.

  When registering a character pattern for recognition in the recognition dictionary 27, the input unit 21 acquires a character pattern for dictionary design and outputs it to the feature extraction unit 22 (step S101).

  The feature extraction unit 22 calculates the feature vector and the recognition vector (eigenvalue and eigenvector) of the input dictionary design character pattern (step S102), and calculates the calculated recognition vector (eigenvalue and eigenvector) and the input character pattern. Character information (Hiragana, Katakana, Kanji, etc.) is associated with each other and stored in the recognition dictionary 27 (step S103).

  Then, the feature extraction unit 22 generates a reliability function for each eigenvector and stores it in the reliability determination information storage unit 28 (step S104).

Next, the pattern recognition operation of this pattern recognition apparatus will be described with reference to FIGS.
In this case, the input unit 21 acquires a character pattern to be recognized and outputs it to the feature extraction unit 22 (step S201 in FIG. 4).

  The feature extraction unit 22 calculates a feature vector and a plurality of eigenvectors for recognition from the input character pattern and outputs them to the recognition unit 24 (step S202).

  In the recognizing unit 24, the first recognizing unit 24a first performs first recognition by comparing each eigenvector input from the feature extracting unit 22 with the recognition vector (eigenvector) of the recognition dictionary 27, and as a recognition result. A recognition score (score) of the eigenvector for each of the obtained candidates is calculated and stored in the memory (step S203).

  If a plurality of eigenvectors obtained from one character pattern are eigenvectors A to D as shown in FIG. 3, for example, and the score is “1000 points”, in this example, the reliability determination information storage unit 28 ( For example, “900 points” is stored as the first candidate score for the eigenvector A, and “450 points” is stored as the second candidate score.

  For the eigenvector B, for example, “700 points” is stored as the first candidate score, “650 points” is stored as the second candidate score, and the like.

  For the eigenvector C, for example, “400 points” is stored as the score of the first candidate, “100 points” is stored as the score of the second candidate, and the like.

  For the eigenvector D, for example, “300 points” is stored as the first candidate score, “250 points” is stored as the second candidate score, and the like.

  When the calculation of the score for each eigenvector is completed, the reliability calculation unit 23 subsequently reads the recognition result of each eigenvector in the recognition unit 24 and is stored in the reliability determination information storage unit 28 (memory). The reliability is calculated based on the reliability function (step S204).

  For example, the reliability calculation unit 23 calculates the reliability of the recognition result of the first candidate from the relationship between the eigenvector scores of the higher candidates, and stores the reliability in the column of the reliability of the corresponding eigenvector in the reliability determination information storage unit 28.

  For example, for the eigenvector A in FIG. 3, if the score of the first candidate is “900 points” and the score of the second candidate is “450 points”, the score of the first candidate is high, and the first candidate and the second candidate Since the reliability of the first candidate is high when the score difference from the candidate is large, “0.9” is assigned as the reliability. In this case, “1.0” is the highest reliability.

  On the other hand, for the eigenvector B, when the score of the first candidate is “700 points” and the score of the second candidate is “650 points”, the score difference between the score of the first candidate and the score of the second candidate is small. Although the score is high, the reliability of the first candidate is not high. For this reason, “0.5” is given as the reliability.

  For the eigenvector C, when the score of the first candidate is “400 points” and the score of the second candidate is “100 points”, the score of the first candidate and the second candidate score are not so high. Since the score difference with the score of the two candidates is large, the reliability of the first candidate is high. For this reason, “0.8” is given as the reliability.

  Here, the process for obtaining the reliability from the calculation of the recognition score will be described in detail. The recognition unit 24 obtains a recognition result (recognition score) when a data acquisition pattern group is input using each eigenvector group.

  The reliability calculation unit 23 sets a reliability function indicating how reliable the recognition result of a certain character pattern is based on the recognition result (recognition score). That is, the recognition score calculated by the recognition unit 24 is input as to how reliable the recognition score obtained using a certain eigenvector is, and how much the identification information created based on that eigenvector is reliable (weight). Is expressed in the form of a mathematical expression that returns the reliability.

  As the reliability function, for example, as shown in FIG. 5, a step function or the like in which the reliability changes stepwise by a certain value with a certain threshold with respect to the recognition score (similarity). If the reliability can be uniquely determined based on the various data obtained from the character pattern, in addition to the reliability function, the value corresponding to the input data is held in the form of a table. A decision style may be established.

  Further, the relationship of reliability with respect to input data (what value is the reliability to be output when certain input data comes) may be arbitrarily set.

  The reliability calculation unit 23 creates a reliability function and then calculates the reliability of each eigenvector in the recognition from the reliability function and the character pattern. The obtained result is output to the vector selection unit 25.

  The vector selection unit 25 reads the reliability of each eigenvector obtained by the reliability calculation unit 23 from the reliability determination information storage unit 28, and is effective for character pattern recognition from the relationship between the reliability of each read eigenvector. The eigenvector is selected (selected), and selection information of the eigenvector is output to the recognition unit 24 (step S205).

  When selecting (selecting) eigenvectors effective for character pattern recognition, for example, as shown in FIG. 6, only eigenvector A and eigenvector C are selected from eigenvectors A to D, or as shown in FIG. A continuous range of vectors such as eigenvectors A to C may be selected from eigenvectors A to D, and the threshold value as the boundary may be determined as desired.

  In the recognition unit 24, the second recognition unit 24 b first selects a recognition vector (eigenvalue) selected from the recognition vectors (eigenvalues and eigenvectors) of the recognition dictionary 27 from the eigenvector selection information input from the feature extraction unit 22. And the corresponding eigenvector obtained from the character pattern to perform the second recognition (step S206), and the character is output to the output unit 26 as the recognition result (step S207) and stored in the memory.

  In performing the second recognition, an identification space is configured by combining a plurality of eigenvectors in consideration of the reliability of the selected eigenvector. When calculating a recognition score, the recognition scores of a plurality of selected eigenvectors A to D are totaled for one character.

Assuming that the plurality of eigenvectors are eigenvectors A to D and the recognition score of the character “A” is R (X), the recognition score (R ₀ (X)) obtained with the eigenvector A is obtained as shown in FIG. Multiply the reliability α _o for each eigenvector, multiply the recognition score (R ₁ (X)) obtained for the eigenvector B by the reliability α ₁ for each eigenvector, and obtain the recognition score (R ₂ ( X)) is multiplied by the reliability α ₂ for each eigenvector, and the sum of these values is taken as a recognition score R (X).

  Then, the recognition score is obtained for other characters, and the character having the highest value in the end is used as the recognition solution, that is, the character of the recognition result. That is, the recognition scores of a plurality of eigenvectors A to D are totaled for other characters. For example, the recognition score of the character “I” is R ′ (X), the recognition score of the character “U” is R ″ (X), etc., and the numerical values of the eigenvectors are summed.

  In addition, after obtaining a recognition solution for each eigenvector, the reliability for each eigenvector is accumulated as the recognition solution score, and the character with the highest score is used as the final recognition solution. May be applied.

As described above, according to the pattern recognition apparatus of the first embodiment, the reliability of a plurality of eigenvectors obtained in the first character recognition is obtained, and the eigenvectors to be recognized are limited by using the reliability. By selecting a recognition vector specialized for recognition of characters having eigenvectors, recognition accuracy can be improved and recognition rate of character patterns can be improved as compared with the case where all characters are recognized.
(Second Embodiment)

(Second Embodiment)
Next, a second embodiment will be described. In describing the second embodiment, the same components as those in the first embodiment are denoted by the same reference numerals, and different functions will be described.

  As shown in FIG. 9, the pattern recognition apparatus of the second embodiment is similar to the first embodiment in that an input unit 21, a feature extraction unit 22, a reliability calculation unit 23, a recognition unit 24, a vector selection unit 25, An output unit 26, a recognition dictionary 27, a reliability determination information storage unit 28, and a pattern matching unit 29 are provided.

  The feature extraction unit 22 extracts a recognition vector from the character pattern input from the input unit 21. Here, regarding the input pattern to be recognized, the recognition vector is output to the reliability calculation unit 23 as it is. The reliability calculation unit 23 receives the feature vector obtained by the feature extraction unit 22, and calculates the reliability from the received feature vector.

  The pattern matching unit 29 performs pattern matching by comparing each eigenvector input from the feature extraction unit 22 with the recognition vectors (eigenvalues and eigenvectors) of the recognition dictionary 27, and a plurality of pattern matching results are obtained. A recognition score (score) of each candidate eigenvector is calculated and output, that is, a recognition vector (eigenvalue and eigenvector) and a recognition score (score) are associated with each other and stored in a memory.

  The reliability calculation unit 23 reads the eigenvector and recognition score (score) of the character pattern stored in the memory, calculates the reliability for each eigenvector, and outputs it to the vector selection unit 25.

  In the second embodiment, as in the first embodiment, a recognition vector (eigenvalue and eigenvector) representing the degree of dispersion of the original pattern group and a recognition score are used as examples of the recognition vector.

  The reliability calculation unit 23 acquires the distribution state of eigenvalues from the recognition vector (eigenvalue and eigenvector) read from the memory.

  10 to 12, a reliability calculation table 30 (table) corresponding to the distribution state of the eigenvectors is stored in advance in the memory, and the numerical values of the eigenvalues are as shown in Table 30 of FIGS. When it is approximated or coincides with any bar graph arrangement, the corresponding reliability is obtained from the corresponding table 30 and output to the vector selection unit 25. Subsequent processing in the vector selection unit 25 is the same as that in the first embodiment. From the recognition vectors in the recognition dictionary 27, those suitable for recognition of the input pattern are selected and notified to the recognition unit 24.

  The recognition unit 24 does not perform recognition of each recognition vector, but receives a notification from the vector selection unit 25 and executes character pattern recognition processing using the selected recognition vector.

  In addition to the eigenvalues and eigenvectors, any recognition vector may be used as long as the concept of the present invention of extracting and using only the optimum part for recognition can be realized.

  In addition to the table 30 (table), any means can be used as the determining means as long as the reliability and the recognition vector used for the recognition can be uniquely selected (determined) for a specific function. Good. As a result, only those suitable for recognition of the input pattern can be selected from the recognition vectors in the recognition dictionary 27 without passing through the recognition unit 24 twice.

(Third embodiment)
Next, a third embodiment will be described. The connection relationship of the configuration of the third embodiment is the same as that of the first embodiment, and the function of each configuration will be described with reference to FIG.

  Similar to the first embodiment, the pattern recognition apparatus according to the third embodiment includes an input unit 21, a feature extraction unit 22, a reliability calculation unit 23, a recognition unit 24, a vector selection unit 25, an output unit 26, and a recognition dictionary. 27, a reliability determination information storage unit 28 is provided.

  In the third embodiment, the recognition unit 24 first performs recognition of the input character pattern using all the recognition vectors. As a result of this recognition, a character group (higher candidate character group) having a higher recognition score is output to the reliability calculation unit 23. As a result, the reliability calculation unit 23 receives the upper character group of the recognition score obtained by the recognition unit 24.

  As shown in FIG. 13, a table in which a reliability level for each recognition vector for recognition performed on a specific character group is set in advance as a list 31 in the memory.

  When the reliability calculation unit 23 receives a character group with a higher recognition score candidate, the reliability calculation unit 23 outputs the reliability for each recognition vector corresponding to the received character group from the memory list 31 to the vector selection unit 25.

  In this example, the list 31 is used. However, any other format may be used as long as the information can uniquely determine the reliability of the input, such as a function. In addition to outputting the reliability for each recognition vector to the vector selection unit 25, the reliability calculation unit 23 may send information indicating the recognition vector to be used to the input character group. In this case, the vector selection unit 25 outputs the input information as it is to the recognition unit 24.

  The recognition unit 24 first executes the first recognition using all the recognition vectors for the input character pattern, and passes the recognition result to the reliability calculation unit 23.

  According to the third embodiment, by limiting the characters to be recognized and selecting a recognition vector specialized for the recognition of the characters, the recognition accuracy is further improved as compared with the case where all the characters are to be recognized. be able to.

  In addition, this invention is not limited only to the said embodiment, You may deform | transform a component in the range which does not deviate from the summary in an implementation stage. In addition, various inventions can be configured by appropriately combining a plurality of components disclosed in the embodiment.

  For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

  Specifically, although the case where the input pattern to be recognized is a character pattern has been described in the above embodiment, other patterns such as an audio pattern, an image, and a moving image may be used.

The figure which shows the structure of the pattern recognition apparatus of 1st Embodiment and 3rd Embodiment of this invention. The flowchart which shows the dictionary registration operation | movement of this pattern recognition apparatus. The figure which shows the score and reliability for every 1st candidate provided with respect to the some eigenvector obtained from the character pattern. The flowchart which shows operation | movement of the pattern recognition apparatus of 1st Embodiment. The figure which shows an example of the step function from which a reliability changes by a fixed value with a certain threshold with respect to a recognition score. The figure which shows an example in which a useful eigenvector is selected from several eigenvectors. The figure which shows the other example from which a useful eigenvector is selected from several eigenvectors. The figure which shows the example of calculation of the authentication score about each character. The figure which shows the structure of the pattern recognition apparatus of 2nd Embodiment. The figure which shows the table | surface of the eigenvector-reliability according to pattern. The figure which shows the eigenvector-reliability table according to pattern. The figure which shows the eigenvector-reliability table according to pattern. The figure which shows the list | wrist (correspondence table | surface) of the combination of a high-order candidate, and each eigenvector reliability.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 21 ... Input part, 22 ... Feature extraction part, 23 ... Reliability calculation part, 24 ... Recognition part, 24a ... First recognition part, 24b ... Second recognition part, 25 ... Vector selection part, 26 ... Output part, 27 ... Recognizing dictionary, 28 ... reliability determination information storage unit, 29 ... pattern matching unit.

Claims (5)

An input unit that converts an input pattern to be recognized into a feature vector for use in pattern recognition;
A recognition dictionary in which eigenvectors of standard patterns used for pattern recognition are stored;
A feature extraction unit for obtaining a plurality of eigenvectors indicating features unique to the input pattern from the feature vectors converted by the input unit;
A first output unit that recognizes the input pattern by comparing a plurality of eigenvectors obtained by the feature extraction unit with eigenvectors stored in the recognition dictionary and outputs a score indicating a degree of matching for each of the compared eigenvectors. A recognition unit;
A reliability calculation unit that obtains reliability of the recognition result for each eigenvector based on the score that is an output result of the first recognition unit ;
A vector selection unit that selects an eigenvector to be used in the next recognition from a plurality of eigenvectors using the reliability obtained by the reliability calculation unit;
A second recognizing unit that recognizes the input pattern using the eigenvector selected by the vector selecting unit and outputs the correct candidate or the score of each correct candidate used for determining a correct answer as a result of pattern recognition. A pattern recognition apparatus. The reliability calculation unit includes:
The pattern recognition apparatus according to claim 1, wherein the reliability of the recognition result of the first candidate is obtained from the relationship between the score of the first candidate and the score of the second candidate for each eigenvector. The vector selection unit
The pattern recognition apparatus according to claim 1, wherein an eigenvector whose reliability is equal to or higher than a preset threshold is selected as an eigenvector used for recognition in the second recognition unit from a result of reliability calculation. Computer
An input unit that converts an input pattern to be recognized into a feature vector for use in pattern recognition;
A recognition dictionary in which eigenvectors of standard patterns used for pattern recognition are stored;
A feature extraction unit for obtaining a plurality of eigenvectors indicating features unique to the input pattern from the feature vectors converted by the input unit;
A first output unit that recognizes the input pattern by comparing a plurality of eigenvectors obtained by the feature extraction unit with eigenvectors stored in the recognition dictionary and outputs a score indicating a degree of matching for each of the compared eigenvectors. A recognition unit;
A reliability calculation unit that obtains reliability of the recognition result for each eigenvector based on the score that is an output result of the first recognition unit ;
A vector selection unit that selects an eigenvector to be used in the next recognition from a plurality of eigenvectors using the reliability obtained by the reliability calculation unit;
There line recognition of the input pattern using the selected eigenvectors by the vector selecting unit, to function as a second recognition unit which outputs the score for each correct answer candidate to be used to correct the candidate or correct decision as a result of the pattern recognition A pattern recognition program characterized by that. Converting the input pattern to be recognized into a feature vector for use in pattern recognition;
Obtaining a plurality of eigenvectors indicating features unique to the input pattern from the transformed feature vectors;
The input pattern is recognized by comparing a plurality of eigenvectors obtained by the feature extraction unit with eigenvectors stored in a recognition dictionary in which eigenvectors of standard patterns used for pattern recognition are stored. Outputting a score indicating the degree of matching for each eigenvector,
Determining the reliability of the recognition result for each eigenvector based on the score that is the output result of the recognition;
Selecting an eigenvector to be used in the next recognition from a plurality of eigenvectors using the determined reliability;
Pattern recognition method characterized by a step of outputting have line recognition of the input pattern using the selected eigenvectors, the score for each correct answer candidate to be used to correct the candidate or correct decision as a result of the pattern recognition .

Images