JP2014153763A - Pattern recognition device, pattern recognition method and pattern recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern recognition device, a pattern recognition method and a pattern recognition program which allow improvement of system performance.SOLUTION: The pattern recognition device includes an extraction section for extracting a feature vector from an input pattern, a first calculation section for calculating an identification function value for the feature vector, a second calculation section for calculating a posterior probability for each recognition class corresponding to a classification destination of the input pattern from the identification function value and a recognition section for classifying the input pattern into one of the recognition classes on the basis of the posterior probability.

Description

  Embodiments described herein relate generally to a pattern recognition apparatus, a pattern recognition method, and a pattern recognition program.

  Conventionally, a pattern recognition technique for automatically classifying input characters, sounds, images, or the like is known. In the field of such pattern recognition, a discriminant function in the form of a quadratic form is often used. These discriminant functions include the nearest neighbor determination method and the learning vector quantization of its means of realization, the composite similarity and its derivative subspace method, kernel subspace method, mutual subspace method, similarity Bayes method (modified two Order discriminant function), and spherical pseudo Bayes method.

Japanese Patent No. 2739950

  However, in the conventional pattern recognition technology, the discriminant function used does not have a probabilistic meaning. Therefore, the conventional pattern recognition technology is not compatible with other fields such as knowledge processing using a language model, and the performance of the entire system is lowered.

  The pattern recognition apparatus according to the embodiment includes an extraction unit that extracts a feature vector from an input pattern, a first calculation unit that calculates a discrimination function value for the feature vector, and a classification destination of the input pattern from the discrimination function value. A second calculation unit that calculates a posterior probability for each corresponding recognition classification; and a recognition unit that classifies the input pattern into one of the recognition classifications based on the posterior probability.

The figure which shows the structural example of the pattern recognition apparatus which concerns on 1st Embodiment. The figure which shows the structural example of the pattern recognition function which concerns on 1st Embodiment. 6 is a flowchart showing an example of a processing procedure during pattern recognition according to the first embodiment.

  Hereinafter, embodiments of a pattern recognition apparatus, a pattern recognition method, and a pattern recognition program will be described in detail with reference to the accompanying drawings.

[First Embodiment]
<Device configuration>
FIG. 1 is a diagram illustrating a configuration example of a pattern recognition apparatus 100 according to the present embodiment. As shown in FIG. 1, a pattern recognition apparatus 100 according to this embodiment includes a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, and the like. The pattern recognition apparatus 100 includes an external storage device 104, an input device 105, a display device 106, a communication IF (Interface) 107, and the like. In the pattern recognition apparatus 100 according to the present embodiment, each hardware is connected via a bus B.

  The CPU 101 is an arithmetic device that realizes overall control of the apparatus and a mounting function. The ROM 102 is, for example, a non-volatile semiconductor memory that stores programs for realizing functions, function setting data, and the like. The RAM 103 is a volatile semiconductor memory from which programs and data are read and temporarily stored. Therefore, for example, the CPU 101 reads out a program and data from the ROM 102 onto the RAM 103 and executes processing, thereby realizing control and mounting functions of the entire apparatus.

  The external storage device 104 is a non-volatile storage device such as an HDD (Hard Disk Drive) or a memory card. The external storage device 104 includes recording media such as a flexible disk (FD), a CD (Compact Disk), and a DVD (Digital Versatile Disk). The input device 105 is, for example, a numeric keypad, a touch panel, or a keyboard, and is used to input each operation signal to the pattern recognition device 100. The display device 106 is a display or the like, for example, and displays the processing result by the pattern recognition device 100. The communication IF 107 is an interface that connects the pattern recognition apparatus 100 to a predetermined data transmission path. Accordingly, the pattern recognition apparatus 100 can perform data communication with other devices via the communication IF 107.

  As described above, according to the present embodiment, a pattern recognition function for automatically classifying input signals can be provided by the above configuration.

<Pattern recognition function>
The pattern recognition function according to this embodiment will be described. The pattern recognition apparatus 100 according to the present embodiment extracts a feature vector from an input pattern. The pattern recognition apparatus 100 calculates a discrimination function value (hereinafter referred to as “discrimination function value”) for a feature vector. The pattern recognition apparatus 100 converts the discriminant function value into a posteriori probability based on a conversion parameter (hereinafter referred to as “likelihood conversion parameter”) for conversion into a probability scale having a probabilistic meaning, and sets the input pattern as a classification destination. Calculate posterior probabilities for each corresponding recognition category. The pattern recognition apparatus 100 classifies the input pattern into one of recognition classifications based on the posterior probability. The pattern recognition apparatus 100 according to the present embodiment has such a pattern recognition function.

  In the conventional pattern recognition technique, a discriminant function in a quadratic form is used, but the discriminant function used does not necessarily have a probabilistic meaning. Therefore, the conventional pattern recognition technology is not compatible with other fields such as knowledge processing using a language model, and the performance of the entire system is lowered.

  Therefore, in the pattern recognition function according to the present embodiment, focusing on the fact that any of the above discriminant functions can be regarded as an approximation of the logarithmic likelihood (natural logarithm of likelihood) of a Gaussian distribution, a handling method unified with respect to discriminant functions in general Propose. Furthermore, in the pattern recognition function according to the present embodiment, a mechanism is proposed in which convergence is guaranteed without depending on the initial value by providing a learning index based on a posteriori probability, and sequential pattern learning is performed.

  The configuration and operation of the pattern recognition function according to this embodiment will be described below. FIG. 2 is a diagram illustrating a configuration example of the pattern recognition function according to the present embodiment. As shown in FIG. 2, the pattern recognition function according to the present embodiment includes a signal input unit 11, a feature extraction unit 12, a discrimination function calculation unit (first calculation unit) 13, and a posterior probability calculation unit (second calculation unit) 14. , A recognition unit 15 and a recognition result discarding unit 16. In the present embodiment, the basic function of pattern recognition (hereinafter referred to as “basic function” for convenience) is realized by the operation of each of these functional units. The pattern recognition function according to the present embodiment includes a likelihood conversion parameter update unit (first update unit) 17 and a dictionary update unit (second update unit) 18. In the present embodiment, a pattern learning function based on the recognition result (hereinafter referred to as “learning function” for convenience) is realized by the operation of each of these functional units.

<Basic functions>
The signal input unit 11 receives an input pattern to be authenticated as an input signal. The input pattern to be authenticated includes characters, sounds, images, and the like, and the input signal is subjected to a predetermined signal process (for example, “binarization”) on the digital information and the digital information. This corresponds to preprocessed digital information.

  The feature extraction unit 12 extracts a feature vector from the input pattern. The feature extraction unit 12 converts the input signal from the signal input unit 11 into a number vector of a certain dimension. The feature extraction unit 12 performs preprocessing such as normalizing the length and quantization level on the input signal, and obtains a signal value after preprocessing. The feature extraction unit 12 performs filter processing such as a Gaussian filter and conversion processing such as Fourier transform on the pre-processed signal value to obtain a signal value after the filter / conversion processing. As a result, the feature extraction unit 12 extracts a feature vector by outputting a vector whose component is a signal value after preprocessing or a signal value after filtering / conversion processing.

  The feature vectors extracted at this time may be normalized so that the norm is 1. The feature extraction unit 12 extracts a plurality of feature vectors by extracting feature vectors for each of a plurality of input signals or adding different variations such as randomly generated noise to one input signal. It may be extracted. In addition, the feature extraction unit 12 may treat the feature vector extracted by these methods as a temporary feature vector, and may convert the feature vector into a feature vector in the reproduction kernel Hilbert space using a positive definite kernel described later.

  The discriminant function calculator 13 calculates discriminant function values for the feature vectors. The discriminant function calculator 13 calculates a discriminant function value for the recognition classification (hereinafter referred to as “category”) of the feature vector extracted by the feature extractor 12.

となる。 For example, the discriminant function calculator 13 calculates the discriminant function value f ⁽ⁱ⁾ (x) using the discriminant function based on the nearest neighbor determination method. For example, when the number of prototypes of the recognition dictionary held by the recognition dictionary holding unit 93 is 1 for each category, the identification function at this time is as follows for the category i of the input feature vector x:

It becomes.

また複合類似度法においては前記部分空間の直交補空間の正規直交基底r_ｈ＋１ ^（ｉ） ，・・・，ｒ_Ｄ ^（ｉ）及び追加のパラメータλ_１ ^（ｉ），・・・，λ_Ｄ ^（ｉ）を用いて

となる（式中の「・」は標準内積を表す）。なお、入力された特徴ベクトルは、||x||＝１と正規化されているものとする。 Further, the discriminant function calculation unit 13 may calculate the discriminant function value f ⁽ⁱ⁾ (x) using a discriminant function based on the subspace method and the composite similarity method. For example, in the subspace method, the discriminant function at this time is represented by r ₁ ⁽ⁱ⁾ ,..., R _h ^{( i)}

The composite similarity method the orthonormal basis of the orthogonal complement subspace _r ^h + 1 in ^{_{(i), ···, r D}} (i) and additional parameters ^{_{λ 1 (i), ···,}} λ D ( ^{with i)}

(“•” in the formula represents a standard inner product). It is assumed that the input feature vector is normalized as || x || = 1.

となる。 Similarly, the discriminant function calculation unit 13 may calculate the discriminant function value f ⁽ⁱ⁾ (x) using the discriminant function based on the mutual subspace method. The discrimination function at this time is, for example, when principal component analysis is performed on a plurality of input feature vectors x and the obtained input subspace is W,

It becomes.

となる。このとき正定値カーネルＫ（・，・）は、仮の特徴ベクトルｘから再生核ヒルベルト空間上への写像を与えており、この写像先を特徴ベクトルｘとして取り扱える。また、上記以外の場合も、識別関数は、一般的に値が小さいほど、カテゴリとの類似性が高くなる実数値の関数として統一的に取り扱うことができる。 Further, the discriminant function calculation unit 13 may calculate the discriminant function value f ⁽ⁱ⁾ (x) using a discriminant function based on the kernel subspace method. The discriminant function at this time is, for example, x for a temporary feature vector, r ₁ ⁽ⁱ⁾ ,..., R _N ^{(i) for} learning data, and α ₁ ⁽ⁱ⁾ ,. If α _N ⁽ⁱ⁾ ,

It becomes. At this time, the positive definite kernel K (.,.) Gives a mapping from the temporary feature vector x to the reproduction kernel Hilbert space, and can handle this mapping destination as the feature vector x. In other cases as well, the discriminant function can generally be handled uniformly as a real-valued function in which the smaller the value, the higher the similarity to the category.

  As described above, the basic function according to the present embodiment proposes a unified handling method for the discrimination function in general.

The recognition dictionary holding unit 93 corresponds to a predetermined storage area of a storage device included in the pattern recognition device 100, and is used to calculate parameters of the identification function (identification function values f ⁽ⁱ⁾ (x) for each category to be recognized. A recognition dictionary in which parameters to be used are registered. The value of the recognition dictionary is updated by a dictionary updating unit 18 described later, and the learning result is reflected. The initial value of the recognition dictionary can be determined in advance according to the type of the identification function, for example.

The posterior probability calculation unit 14 converts the discriminant function value f ⁽ⁱ⁾ (x) into the posterior probability, and calculates the posterior probability of each category. The posterior probability calculation unit 14 uses the discrimination function value f ⁽ⁱ⁾ (x) calculated by the discrimination function calculation unit 13 based on the likelihood conversion parameter held by the likelihood conversion parameter holding unit 92 as the posterior probability for each category. Convert. The likelihood conversion parameter holding unit 92 corresponds to a predetermined storage area of the storage device included in the pattern recognition apparatus 100, and holds a likelihood conversion parameter for conversion into a probability scale having a probabilistic meaning. . The likelihood conversion parameter is updated by a likelihood conversion parameter update unit 17 described later, and the learning result is reflected.

により、対数尤度ｌ^（ｉ）（ｄ）を計算する。なお、式中のｄ^（ｉ）は、識別関数計算部１３により計算された識別関数値ｆ^（ｉ）（ｘ）であり、全てのカテゴリをまとめてｄ：＝（ｄ^（１），・・・，ｄ^（ｃ））^Ｔと表せる（ｃ：カテゴリ数、Ｔ：転置）。このときの尤度変換パラメータ（ｑ^（ｉ），ｗ^（ｉ））は、カテゴリによらず、一定の値としてもよい。 The posterior probability calculation unit 14 uses the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) held by the likelihood conversion parameter holding unit 92.

To calculate the log likelihood l ⁽ⁱ⁾ (d). In addition, d ⁽ⁱ⁾ in the formula is the discriminant function value f ⁽ⁱ⁾ (x) calculated by the discriminant function calculator 13, and d: = (d ⁽¹⁾ ,. ., D ^(c) ) It can be expressed as ^T (c: number of categories, T: transpose). The likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) at this time may be constant values regardless of the category.

により、対数尤度ｌ^（ｉ）（ｄ）を計算してもよい。 In addition, the posterior probability calculation unit 14 takes w ⁽ⁱ⁾ into a c-dimensional vector value in consideration of an identification function other than the category.

Thus, the log likelihood l ⁽ⁱ⁾ (d) may be calculated.

Each of the above discriminant functions can be regarded as an approximation of log likelihood in Gaussian distribution. From this, the calculation model of the log likelihood l ⁽ⁱ⁾ (d) which the posterior probability calculation unit 14 according to the present embodiment has can be regarded as an appropriately modeled general discriminant function.

により、事後確率Ｐ^（ｉ）（ｄ）を計算する。なお、式中のＬ^（ｉ）（ｄ）は、尤度Ｌ（ｄ）＝ｅｘｐ（ｌ^（ｉ）（ｄ））である。 Next, the posterior probability calculation unit 14 uses the calculated log likelihood l ⁽ⁱ⁾ (d).

To calculate the posterior probability P ⁽ⁱ⁾ (d). Note that L ⁽ⁱ⁾ (d) in the equation is likelihood L (d) = exp (l ⁽ⁱ⁾ (d)).

により、事後確率Ｐ^（ｉ）（ｄ）を計算してもよい。 Further, the posterior probability calculation unit 14 gives the prior probability of each category in the form of p ⁽ⁱ⁾

Thus, the posterior probability P ⁽ⁱ⁾ (d) may be calculated.

を用いて、事後確率Ｐ^（ｉ）（ｄ）を計算してもよい。なお、式中のＥ^（ｉ）はカテゴリｉの分布の次元数を表し、εは微小領域の大きさを表す。これらの値は、予め定められた正の定数である。また、Ｖ_Ｅ ^（ｉ）（ε）はＥ^（ｉ）次元体積である。また、Ｅ^（ｉ），Ｖ_Ｅ ^（ｉ）（ε）は、例えば、部分空間法の場合、

となる。 In addition, the posterior probability calculation unit 14 considers the possibility that the number of dimensions of the space in which the input pattern is distributed varies from category to category, and is a value obtained by approximately integrating the log likelihood l ⁽ⁱ⁾ (d) in a small region.

May be used to calculate the posterior probability P ⁽ⁱ⁾ (d). In the equation, E ⁽ⁱ⁾ represents the number of dimensions of the distribution of category i, and ε represents the size of the minute region. These values are predetermined positive constants. V _E ⁽ⁱ⁾ (ε) is an E ⁽ⁱ⁾ dimensional volume. Further, E ⁽ⁱ⁾ and V _E ⁽ⁱ⁾ (ε) are, for example, in the case of the subspace method,

It becomes.

The recognition unit 15 classifies the input pattern into one of the categories based on the posterior probability P ⁽ⁱ⁾ (d), and recognizes the pattern. The recognizing unit 15 classifies the input pattern into the category i that maximizes the posterior probability P ⁽ⁱ⁾ (d) calculated by the posterior probability calculating unit 14. The recognition unit 15 stores the classification of the input pattern in the recognition result holding unit 91 as a recognition result. Note that the recognition result holding unit 91 corresponds to a predetermined storage area of a storage device included in the pattern recognition apparatus 100.

The recognition result discarding unit 16 verifies whether or not the recognition result (input pattern classification result) of the recognition unit 15 is appropriate. If the recognition result is inappropriate, the recognition result discarding unit 16 discards the corresponding recognition result. At this time, the recognition result discarding unit 16 verifies whether or not the category i having the maximum posterior probability P ⁽ⁱ⁾ (d) is appropriate as the recognition result. The posterior probability P ⁽ⁱ⁾ (d) calculated by the posterior probability calculation unit 14 includes the following possibilities. If the maximum value P _{max of the} posterior probability P ⁽ⁱ⁾ (d) (hereinafter referred to as “maximum posterior probability P _max ”) is not sufficiently large, it is considered that there is a high possibility of an erroneous recognition result. In addition, when the maximum likelihood L _{max of the} category or the total likelihood L _all obtained by adding the likelihoods of all categories is small, for example, a character other than the character (such as an image) is input for pattern recognition for recognizing the character. It is highly probable that a pattern other than the recognition target has been input by mistake. Note that the maximum likelihood L _max corresponds to the maximum value of the likelihood of the classification condition estimated when the input pattern is classified into categories, and the total likelihood L _all is the likelihood of the estimated classification condition. It corresponds to the total value.

により、認識結果を検証する。なお、式中のａ，ｂ，ｃ，ｄ，ｅ，ｆは予め定められた非負の定数であり、ｌ_ｍａｘ，ｌ_ａｌｌは最大尤度Ｌ_ｍａｘ，全尤度Ｌ_ａｌｌの対数である。 Therefore, in this embodiment, an index R (d) (hereinafter referred to as “discard index R (d)”) for the recognition result discarding unit 16 to discard an inappropriate recognition result is defined.

Thus, the recognition result is verified. In the formula, a, b, c, d, e, and f are predetermined non-negative constants, and l _max and l _all are logarithms of the maximum likelihood L _max and the total likelihood L _all .

  The recognition result discarding unit 16 discards an inappropriate recognition result based on the verification result. Specifically, the recognition result discarding unit 16 determines whether or not the discard index R (d) is greater than or equal to a predetermined threshold value, and verifies the recognition result. The recognition result discarding unit 16 determines that the recognition result is inappropriate when the discard index R (d) is equal to or greater than a threshold value. As a result, the recognition result discarding unit 16 accesses the recognition result holding unit 91 and discards the appropriate data by deleting the corresponding data.

As described above, in the basic function according to the present embodiment, the input pattern is recognized by the posterior probability P ⁽ⁱ⁾ (d), and the inappropriate recognition result is discarded. As a result, the basic function according to the present embodiment solves the problems of conventional pattern recognition that uses a discriminant function that does not have a probabilistic meaning (incompatibility with other fields such as knowledge processing using a language model). Can be improved. As a result, the basic function according to the present embodiment can improve the performance of the entire system.

で計算された事後確率の値Ｌ（Ｘ）を評価する。なお、ｙ_ｎはｎ番目の学習データが属するカテゴリであり、ｄ_ｎはｎ番目の識別関数値である。 《Learning function》
The likelihood conversion parameter update unit 17 updates the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) used when calculating the posterior probability P ⁽ⁱ⁾ (d). The likelihood conversion parameter updating unit 17 performs pattern learning data (hereinafter referred to as “learning data”) prepared in advance.

The value L (X) of the posterior probability calculated in (1) is evaluated. Incidentally, y _n is the category to which belongs the n-th training data, d _n is the n-th classification function value.

The likelihood conversion parameter updating unit 17 updates the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) so that the posterior probability value L (X) increases based on the evaluation result. At this time, the likelihood conversion parameter updating unit 17 accesses the likelihood conversion parameter holding unit 92 and converts the values (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) held in a predetermined storage area into the posterior probability values L ( X) is updated to a value that increases.

により、更新値を計算する。なお、δ_ｉｙｎは

で定義されるクロネッカーのデルタ記号であり、βは予め定められたパターン学習用の正の係数（以下「学習係数」という）である。また、尤度変換パラメータ（ｑ^（ｉ），ｗ^（ｉ））は、更新を繰り返すことで最適値に収束することから、初期値を無作為に定めてもよい。 The likelihood conversion parameter updating unit 17 regards the value L (X) of the posterior probability as a function of the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ), takes the logarithm, and uses the gradient method

The update value is calculated by Note that δ _iyn is

Is a Kronecker delta symbol defined by ## EQU1 ## where β is a predetermined positive coefficient for pattern learning (hereinafter referred to as “learning coefficient”). In addition, since the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) converge to the optimum values by repeating the update, the initial values may be determined randomly.

により、更新値を計算する。 The dictionary update unit 18 updates the recognition dictionary in which the identification function parameters for each category to be recognized are registered. The dictionary update unit 18 evaluates the value r ⁽ⁱ⁾ of the recognition dictionary with respect to learning data prepared in advance. Based on the evaluation result, the dictionary update unit 18 updates the value r ⁽ⁱ⁾ of the recognition dictionary so that the value L (X) of the posterior probability increases. At this time, the dictionary updating unit 18 accesses the recognition dictionary holding unit 93 and updates the value r ⁽ⁱ⁾ held in the predetermined storage area to a value that increases the value L (X) of the posterior probability. At this time, the dictionary update unit 18 is, for example, in the case of using the nearest neighbor determination method with one prototype per category,

The update value is calculated by

Thus, in the learning function according to the present embodiment, a learning index is given by the posterior probability P ⁽ⁱ⁾ (d), and pattern learning that guarantees convergence is sequentially performed without depending on the initial value. As a result, the learning function according to the present embodiment can reduce the time required for pattern learning.

  Such a pattern recognition function according to the present embodiment is realized by executing a pattern recognition program in the pattern recognition apparatus 100 and performing the above-described functional units in cooperation with each other.

  The pattern recognition program according to the present embodiment is recorded and provided in a file in an installable format or an executable format in an external storage device 104 readable by the pattern recognition device 100 (computer) as an execution environment. The pattern recognition program has a module configuration including the above functional units, and the functional units are generated on the RAM 103 when the CPU 101 reads and executes the program from the external storage device 104. The method for providing the pattern recognition program is not limited to this. For example, the pattern recognition program may be stored in a device connected to the Internet or the like and downloaded via the network via the communication IF 107. Alternatively, a method of providing a pattern recognition program by incorporating it in the ROM 102 or the like in advance may be used.

  In addition, although the example which implement | achieved the pattern recognition function based on this embodiment by execution of a pattern recognition program (software) was demonstrated above, it is not this limitation. For example, a part or all of the above functional units may be realized by mounting hardware logic (circuit or the like).

  In the following, processing at the time of executing the pattern recognition program according to the present embodiment (cooperation operation of each functional unit) will be described using a flowchart.

  FIG. 3 is a flowchart illustrating an example of a processing procedure during pattern recognition according to the present embodiment. As shown in FIG. 3, in the pattern recognition apparatus 100 according to the present embodiment, each of the functional units operates in a cooperative manner according to the following processing procedure.

  The signal input unit 11 receives an input signal of an input pattern to be authenticated (step S101).

  In response to this, the feature extraction unit 12 extracts a feature vector x from the input signal of the input pattern (step S102). At this time, the feature extraction unit 12 first performs pre-processing such as normalizing the length and quantization level on the input signal from the signal input unit 11 to obtain a pre-processed signal value. Next, the feature extraction unit 12 performs a filter process such as a Gaussian filter or a conversion process such as a Fourier transform on the pre-processed signal value to obtain a signal value after the filter / conversion process. As a result, the feature extraction unit 12 extracts a feature vector x by outputting a vector whose component is a signal value after preprocessing or a signal value after filtering / conversion processing.

In response to this, the discriminant function calculator 13 calculates the discriminant function value f ⁽ⁱ⁾ (x) for the feature vector x (step S103). At this time, the discriminant function calculation unit 13 is extracted by the feature extraction unit 12 using any discriminant function such as nearest neighbor determination method, subspace method and composite similarity method, mutual subspace method, and kernel subspace method. The discriminant function value f ⁽ⁱ⁾ (x) for the category i of the feature vector x is calculated. The discrimination function calculation unit 13 accesses the recognition dictionary holding unit 93 and calculates the discrimination function value f ⁽ⁱ⁾ (x) using the parameters of the discrimination function held by the recognition dictionary holding unit 93.

In response to this, the posterior probability calculation unit 14 converts the discriminant function value f ⁽ⁱ⁾ (x) into the posterior probability, and calculates the posterior probability P ⁽ⁱ⁾ (d) of each category (step S104). At this time, the posterior probability calculation unit 14 calculates the discriminant function value f ⁽ⁱ⁾ (x) calculated by the discriminant function calculation unit 13 based on the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) for each category. The posterior probability P ⁽ⁱ ) is converted into (d). The posterior probability calculation unit 14 first calculates the log likelihood l ⁽ⁱ⁾ (d) by a calculation formula using the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ). The posterior probability calculation unit 14 accesses the likelihood conversion parameter holding unit 92 and uses the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) held by the likelihood conversion parameter holding unit 92 to logarithmic likelihood l ⁽ⁱ⁾ Calculate (d). Next, the posterior probability calculation unit 14 calculates the posterior probability P ⁽ⁱ⁾ (d) by a calculation formula using the calculated log likelihood l ⁽ⁱ⁾ (d).

Receiving this, the recognition unit 15 classifies the input pattern into one of the categories based on the posterior probability P ⁽ⁱ⁾ (d), and recognizes the pattern (step S105). At this time, the recognizing unit 15 classifies the input pattern into the category i that maximizes the posterior probability P ⁽ⁱ⁾ (d) calculated by the posterior probability calculating unit 14. As a result, the recognition unit 15 stores the classification of the input pattern in the recognition result holding unit 91 as a recognition result.

In response to this, the recognition result discarding unit 16 determines whether or not the category i that maximizes the posterior probability P ⁽ⁱ⁾ (d) is appropriate as the recognition result (the classification destination of the input pattern) (step S106). At this time, the recognition result discarding unit 16 determines whether or not the recognition result is appropriate by determining whether or not a discard index R (d) for discarding an inappropriate recognition result is equal to or greater than a threshold value.

As a result, when the recognition result discarding unit 16 determines that the discard index R (d) is equal to or greater than the threshold and is not an appropriate recognition result (an inappropriate recognition result) (step S106: NO), the recognition result holding unit 91 By accessing and erasing the corresponding data, an inappropriate recognition result is discarded (step S107). As described above, the pattern recognition apparatus 100 according to the present embodiment recognizes an input pattern based on the posterior probability P ⁽ⁱ⁾ (d) and discards an inappropriate recognition result.

  On the other hand, when the recognition result discarding unit 16 determines that the discard index R (d) is less than the threshold and is an appropriate recognition result (step S106: YES), the recognition result is not discarded.

In response to this, the likelihood conversion parameter updating unit 17 updates the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) used when calculating the posterior probability P ⁽ⁱ⁾ (d) (step S108). . At this time, the likelihood conversion parameter updating unit 17 first evaluates the posterior probability value L (X) calculated by a predetermined calculation formula with respect to the learning data prepared in advance. Next, the likelihood conversion parameter update unit 17 updates the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) so that the value L (X) of the posterior probability increases based on the evaluation result. Specifically, the likelihood conversion parameter updating unit 17 accesses the likelihood conversion parameter holding unit 92 and uses the values (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ) held by the likelihood conversion parameter holding unit 92 after the fact. The probability value L (X) is updated to a value that increases.

  In response to this, the dictionary updating unit 18 determines whether or not to perform pattern learning based on a preset setting value (step S109).

As a result, the dictionary updating unit 18 updates the recognition dictionary (step S110) when it is determined that the pattern learning is to be executed in accordance with the setting to execute (step S109: YES). At this time, the dictionary updating unit 18 evaluates the value r ⁽ⁱ⁾ of the recognition dictionary with respect to the learning data prepared in advance. Next, the dictionary update unit 18 updates the value r ⁽ⁱ⁾ of the recognition dictionary based on the evaluation result so that the value L (X) of the posterior probability increases. Specifically, the dictionary updating unit 18 accesses the recognition dictionary holding unit 93 and changes the value r ⁽ⁱ⁾ held by the recognition dictionary holding unit 93 to a value that increases the value L (X) of the posterior probability. Update. As described above, the pattern recognition apparatus 100 according to the present embodiment sequentially gives the learning index by the posterior probability P ⁽ⁱ⁾ (d), and sequentially performs pattern learning that guarantees convergence without depending on the initial value.

  Note that if the dictionary update unit 18 determines that pattern learning is not to be executed with the setting that it is not to be executed (step S109: NO), the recognition dictionary is not updated.

<Summary>
As described above, according to the pattern recognition apparatus 100 according to the present embodiment, the feature extraction unit 12 extracts the feature vector x from the input pattern. Next, the discrimination function calculation unit 13 calculates the discrimination function value f ⁽ⁱ⁾ (x) for the feature vector x. Next, the posterior probability calculation unit 14 converts the discriminant function value f ⁽ⁱ⁾ (x) into the posterior probability based on the likelihood conversion parameters (q ⁽ⁱ⁾ , w ⁽ⁱ⁾ ), and the posterior probability P for each category. ^(I) Calculate (d). Next, the recognition unit 15 classifies the input pattern into one of the categories based on the posterior probability P ⁽ⁱ⁾ (d).

Thereby, in the pattern recognition apparatus 100 according to the present embodiment, the input pattern is recognized by the posterior probability P ⁽ⁱ⁾ (d), and an inappropriate recognition result is discarded. As a result, the pattern recognition apparatus 100 according to the present embodiment can improve the problems of conventional pattern recognition that uses a discriminant function that does not have a probabilistic meaning (incompatibility with other fields). . Specifically, the pattern recognition apparatus 100 according to the present embodiment performs pattern recognition using a posteriori probability P ⁽ⁱ⁾ (d) having a probabilistic meaning, so that integration of multiple recognitions and integrated learning is easy. Thus, the recognition accuracy can be improved and the performance of the entire system can be improved.

In the pattern recognition apparatus 100 according to the present embodiment, a learning index is given by the posterior probability P ⁽ⁱ⁾ (d), and pattern learning with guaranteed convergence is sequentially performed without depending on the initial value. Thereby, the pattern recognition apparatus 100 according to the present embodiment can reduce the time required for pattern learning.

  In the above embodiment, a configuration example in which the predetermined storage area of the storage device included in the pattern recognition device 100 corresponds to the recognition result holding unit 91, the likelihood conversion parameter holding unit 92, the recognition dictionary holding unit 93, and the like. However, this is not the case. The recognition result, likelihood conversion parameter, recognition dictionary, and the like may be held in a predetermined storage area of the external storage device 104 that can be read by the pattern recognition device 100, such as a recording medium.

  Moreover, although the pattern recognition apparatus 100 showed the structural example which has learning functions, such as the likelihood conversion parameter update part 17 and the dictionary update part 18, in the said embodiment, it is not this limitation. The learning function may be realized by a device different from the pattern recognition device 100. Specifically, the device different from the pattern recognition device 100 may include a likelihood conversion parameter update unit 17 and a dictionary update unit 18. In this case, the pattern recognition apparatus 100 performs data communication with a device including the likelihood conversion parameter update unit 17 and the dictionary update unit 18 via the communication IF 107. As a result, the basic function realized by the pattern recognition apparatus 100 and the learning function realized by a device different from the pattern recognition apparatus 100 can be linked to realize the pattern recognition function according to the present embodiment.

  Finally, although the embodiment of the present invention has been described, the above embodiment is presented as an example, and is not intended to limit the scope of the invention. The novel embodiment can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. The above-described embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalent scope thereof.

DESCRIPTION OF SYMBOLS 11 Signal input part 12 Feature extraction part 13 Discriminant function calculation part 14 A posteriori probability calculation part 15 Recognition part 16 Recognition result discard part 17 Likelihood conversion parameter update part 18 Dictionary update part 91 Recognition result holding part 92 Likelihood conversion parameter holding part 93 Recognition dictionary storage

Claims (8)

An extractor for extracting feature vectors from the input pattern;
A first calculation unit for calculating a discriminant function value for the feature vector;
A second calculation unit that calculates a posterior probability for each recognition classification corresponding to a classification destination of the input pattern from the identification function value;
A recognition unit that classifies the input pattern into one of the recognition classifications based on the posterior probability;
A pattern recognition apparatus comprising: The discriminant function value by the second calculation unit is evaluated so that the posterior probability calculated for learning data of pattern learning prepared in advance is increased based on the evaluation result. A first updating unit for updating a conversion parameter used when converting from the posterior probability to the posterior probability,
The pattern recognition apparatus according to claim 1, further comprising: Discards the recognition result of the recognition unit that classifies the input pattern into the recognition classification that maximizes the posterior probability and discards the inappropriate recognition result based on the verification result. Part
The pattern recognition apparatus according to claim 1, further comprising: The discarding unit
It is determined whether or not a discard index for discarding the inappropriate recognition result is equal to or greater than a predetermined threshold, and when the discard index is equal to or greater than the threshold, the recognition result is discarded. Item 4. The pattern recognition device according to Item 3. The discarding unit
A maximum posterior probability of the recognition classification that maximizes the posterior probability, a maximum likelihood of the recognition classification that maximizes a likelihood of a classification condition estimated when the input pattern is classified into the recognition classification, and The pattern recognition apparatus according to claim 4, wherein the discard index is calculated based on a total likelihood obtained by summing up the likelihoods of all the recognition classifications. The discriminant function value by the first calculation unit is evaluated so that the posterior probability calculated with respect to the learning data of pattern learning prepared in advance is increased based on the evaluation result. A second updating unit for updating a recognition dictionary in which parameters used for the calculation are registered,
The pattern recognition apparatus according to claim 1, further comprising: An extraction process for extracting feature vectors from the input pattern;
A first calculation step of calculating a discriminant function value for the feature vector;
A second calculation step of calculating a posterior probability for each recognition classification corresponding to the classification destination of the input pattern from the identification function value;
A recognition step of classifying the input pattern into one of the recognition classifications based on the posterior probability;
A pattern recognition method comprising: Computer
An extractor for extracting feature vectors from the input pattern;
A first calculation unit for calculating a discriminant function value for the feature vector;
A second calculation unit that calculates a posterior probability for each recognition classification corresponding to a classification destination of the input pattern from the identification function value;
A recognition unit that classifies the input pattern into one of the recognition classifications based on the posterior probability;
Pattern recognition program to make it function.

Images