JP6313062B2 - Pattern recognition device, pattern recognition method and program - Google Patents

Description

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

  In the field of pattern recognition, hidden Markov models (HMMs) and conditional random fields are used as methods for recognizing input signals whose recognition unit boundaries are not clear, such as speech signals and character string images. Many of its derivatives are used. This method has the disadvantage that it can simultaneously perform the determination and recognition of the separation of the recognition target, but requires a lot of calculation time to collate the internal state model with the feature vector. For this reason, it is desired to provide a new technique capable of performing highly accurate recognition in a short time with respect to an input signal whose recognition unit break is not clear.

C. M. Bishop, “Pattern Recognition and Machine Learning (Up / Down)” (translated by Noboru Murata), Springer Japan, 2007 F. Camastra et al. “Machine Learning for Audio, Image and Video Analysis: Theory and Applications”, Springer-Verlag, 2007 Daiya Takamura et al., “Introduction to Machine Learning for Language Processing” (Natural Language Processing Series 1), Corona, 2010

  The problem to be solved by the present invention is to provide a pattern recognition apparatus, a pattern recognition method, and a program that can perform high-accuracy recognition in a short time for an input signal whose recognition unit separation is not clear.

The pattern recognition apparatus according to the embodiment performs pattern recognition of an input signal by converting the input signal into a feature vector and comparing the feature vector with a recognition dictionary. The recognition dictionary is a dictionary subspace base vector representing a dictionary subspace which is a subspace of the feature vector space, and a similarity calculated from the feature vector and the dictionary subspace is converted into likelihood. A plurality of stochastic parameters. The pattern recognition apparatus includes a recognition unit. The recognizing unit calculates the similarity by a quadratic polynomial of the inner product value of the feature vector and the dictionary subspace basis vector, and calculates the likelihood by an exponential function of a linear sum of the similarity and the probability parameter. To do. The recognition dictionary is learned by an expected value maximization method using a constraint condition between the plurality of probabilistic parameters.

FIG. 1 is a conceptual diagram illustrating the replacement of an HMM state model with a probabilistic subspace model. FIG. 2 is a block diagram illustrating a functional configuration of the pattern recognition apparatus according to the embodiment. FIG. 3 is a flowchart illustrating an example of a processing procedure performed by the pattern recognition apparatus according to the embodiment. FIG. 4 is a block diagram illustrating a hardware configuration example of the pattern recognition apparatus according to the embodiment.

  Hereinafter, a pattern recognition apparatus, a pattern recognition method, and a program according to embodiments will be described with reference to the drawings.

First, the basic concept of this embodiment will be described. In the present embodiment, in order to enable recognition with high accuracy in a short time for an input signal whose recognition unit break is not clear, a large amount of calculation time is required in the conventional method for simultaneously performing recognition determination and recognition of a recognition target. Consider the replacement of the internal state model and feature vector matching operations that were required by the subspace method and its similarity calculation. The subspace method and its derivatives are known as methods used for recognition of a single feature vector (see Reference 1 below), and have an advantage that high recognition accuracy can be obtained compared to processing time. .
<Reference Document 1> E. Oya, "Pattern Recognition and Subspace Method" (Hidemitsu Ogawa, Makoto Sato), Sangyo Tosho, 1986

Similarity calculation by the subspace method or its derivative form has the same purpose as the state model and feature vector matching operation in the conventional method that simultaneously performs recognition and delimitation of recognition targets. It can be regarded as an approximation (see Reference 2 below).
<Reference 2> Yoshiaki Kurosawa, “Subspace Method Derived from Spherical Gaussian Distribution” Theory of Science (D-2), J81-D2 (6), pp.1205-1212, 1998

ただし、式（１）におけるＰ’は、正規直交する辞書部分空間基底ベクトルｕ_１，・・・，ｕ_ｋから下記式（２）で計算される行列であり、ｑ，ｗは確率化パラメータである。

また、このとき、下記式（３）で示されるｓは、特徴ベクトルｘの類似度である。

以下では、式（１）の計算を行うモデルを「確率的部分空間モデル」と呼ぶ。 Therefore, a model for calculating the likelihood L (x) of the feature vector x as shown in the following equation (1) by introducing a probability parameter that converts the similarity in the former into a probability scale such as the likelihood used in the latter is considered. .

However, P ′ in the equation (1) is a matrix calculated by the following equation (2) from the orthonormal dictionary subspace basis vectors u ₁ ,..., U _k , and q and w are the stochastic parameters. is there.

At this time, s represented by the following formula (3) is the similarity of the feature vector x.

Hereinafter, the model that performs the calculation of Expression (1) is referred to as a “probabilistic subspace model”.

  Here, an example of replacing the hidden Markov model (HMM) state model with a probabilistic subspace model will be described with reference to FIG. The HMM is constituted by a plurality of states as schematically shown in FIG. The input of the HMM is a series of feature vectors. Each state is a statistical model of a single feature vector, and usually a mixed Gaussian distribution model (GMM: Gaussian Mixture Model) as schematically shown in FIG. 1B is used (see Non-Patent Document 1). ). When learning the HMM, the states and the parameters between the states are learned independently.

  A GMM used as an HMM state model has a large amount of calculation for recognition accuracy as a single feature vector model. Therefore, the state model of the HMM is replaced with a probabilistic subspace model as schematically shown in FIG. Since the probabilistic subspace model performs an operation in a subspace having a smaller number of dimensions than the dimension number of the feature vector, the amount of calculation is smaller than that of the GMM. Therefore, as shown in FIG. 1D, the HMM state model is replaced with the probabilistic subspace model from the GMM, and the feature vector matching operation is performed, thereby enabling high-precision recognition in a short time. Become.

  At this time, in the recognition, the Viterbi algorithm for calculating the likelihood of the HMM does not depend on the likelihood calculation method in each state model (see Non-Patent Document 1), so the HMM state model is changed to a probabilistic subspace model. Even if it is replaced, the Viterbi algorithm can be used as it is.

  On the other hand, at the time of learning a probabilistic subspace model, if this is performed by the expected value maximization method (EM method) as in the Baum-Welch algorithm (see Non-Patent Document 1) used for HMM learning, Since the burden ratio to the state (see Non-Patent Document 1) does not depend on the form of the state model, it can be calculated in the same manner as the Baum-Welch algorithm.

ＥＭ法においては、式（４）で表される対数尤度を最大化するようにパラメータを最大化するが、Ｐ’については主成分分析（非特許文献１を参照）と同様であり、下記式（５）を対角化し、上位ｋ個の固有値に対応する固有ベクトルを順番にとって、辞書部分空間基底ベクトルｕ_１，・・・，ｕ_ｋとすればよい。

Therefore, the learning data _x 1, ..., with _{x N,} the probability parameter q, w and subspace basis vector _u 1 is a parameter of the stochastic subspace model, ..., updates _{u k} Think about it. At this time, _{assuming that} the burden ratio of x ₁ ,..., X _n is γ ₁ ,..., Γ _n , the log likelihood of all learning data assumes the independence of the learning data, and the following equation (4) It can be expressed as

In the EM method, the parameters are maximized so as to maximize the log likelihood represented by the equation (4), but P ′ is the same as in the principal component analysis (see Non-Patent Document 1). The equation (5) is diagonalized, and the eigenvectors corresponding to the top k eigenvalues are taken in order as the dictionary subspace basis vectors u ₁ ,..., U _k .

  However, there is a problem with the stochastic parameters q and w. As will be apparent from the form of Equation (6) described later, the data likelihood L is monotonous with respect to q and w, and it is impossible to maximize the data likelihood L with respect to the probability parameters q and w. In fact, if w is made small and q is made large, L can be arbitrarily increased, which is inappropriate as a recognition model.

  Therefore, in this embodiment, it is possible to appropriately learn the probabilistic subspace model by introducing an appropriate constraint condition f (q, w) = 0 between the probabilistic parameters q and w. As a result, a pattern recognition apparatus capable of performing high-accuracy recognition in a short time for an input signal whose recognition unit break is not obvious by a novel method in which an HMM state model or the like is replaced with a probabilistic subspace model. Make it feasible.

  FIG. 2 is a block diagram showing a functional configuration of the pattern recognition apparatus of the present embodiment. As shown in FIG. 1, the pattern recognition apparatus of this embodiment includes a signal input unit 1, a feature extraction unit 2, a recognition unit 3, and a dictionary update unit 4.

  The signal input unit 1 receives an input of a signal to be recognized. Signals to be recognized are, for example, characters and character strings represented as images, other images, audio signals represented as waveforms, various sensor signals, and the like. These digital information, or two as necessary. Digital information subjected to preprocessing such as valuation is input to the signal input unit 1.

The feature extraction unit 2 converts the signal input to the signal input unit 1 into a set of feature vectors having a certain number of dimensions. Specifically, the feature extraction unit 2 first extracts a partial signal in the window range by applying a window to the signal input to the signal input unit 1. Next, the feature extraction unit 2 performs preprocessing such as normalizing the length and quantization level for each of the extracted signal portions. Then, the feature extraction unit 2 outputs a feature vector whose component is a value after the pre-processing, or a value after further performing a filtering process such as a Gaussian filter or a transformation process such as a Fourier transform on the pre-processed signal. Then, a set of feature vectors corresponding to the signal input to the signal input unit 1 is generated. As a specific example, the technique described in Reference Document 3 below can be used.
<Reference Document 3> J. A. Rodriguez and F. Perronin, “Local Gradient Histogram Features for Word Spotting in Unconstrained Handwritten Documents”, Proc. ICFHR2008, 2008

  The recognition unit 3 evaluates the set of feature vectors generated by the feature extraction unit 2 using the recognition dictionary 10 and outputs a recognition result representing the class or set of classes to which the signal input to the signal input unit 1 belongs. To do.

The recognition dictionary 10 is a database including models corresponding to the respective classes handled by the pattern recognition apparatus of the present embodiment as signal classification destinations, and is held inside or outside the pattern recognition apparatus of the present embodiment. Each class model held by the recognition dictionary 10 is composed of a plurality of states like the HMM, and each state is the above-described stochastic subspace model. That is, the recognition dictionary 10 holds dictionary subspace basis vectors u ₁ ,..., U _k corresponding to each state of the model for each class, and the stochastic parameters q, w. The dictionary subspace basis vectors u ₁ ,..., U _k are parameters that represent a dictionary subspace having a dimension smaller than the number of dimensions of the feature vector, and the stochastic parameters q and w are the feature vector and the dictionary subspace. This is a parameter for converting the similarity calculated from the above into likelihood.

The recognition unit 3 combines the models included in the recognition dictionary 10 to search for an optimum correspondence with the set of feature vectors generated by the feature extraction unit 2, and outputs a set of model labels. At this time, the recognition unit 3 performs the feature vector and the dictionary subspace basis vector u ₁ ,... With respect to one or a plurality of feature vectors in the set of feature vectors in each state of each model included in the recognition dictionary 10. .., U _k calculates the similarity with a quadratic polynomial of the inner product value, and calculates the likelihood with an exponential function of the linear sum of the similarity and the probabilistic parameters q and w. Then, a combination of models that maximizes the data likelihood L as a whole is selected, and the label set is output.

このとき、モデル列がＭ_１，・・・，Ｍ_Ｔとなる確率Ｐ（Ｍ_１，・・・，Ｍ_Ｔ）は、Ｎグラム（下記の参考文献４を参照）などの確率的言語モデルによって決定することができ、通常はバイグラムを用いてＢａｕｍ−Ｗｅｌｃｈアルゴリズムにより学習される（非特許文献１を参照）。
＜参考文献４＞北研二、「確率的言語モデル」（言語と計算５）、東京大学出版会、1999年 Data likelihood L, the element _x 1 of the set of feature vectors, ..., _{x T} dictionary subspace _U 1, respectively, ..., model _M 1 with _{U t,} ..., corresponding to the _{M T} Is obtained as the following formula (6).

In this case, the model train _M 1, · · ·, probability _P to be _{M T (M 1, ···,} M T) , depending on probabilistic language model, such as N-gram (see reference 4 below) Usually, it is learned by the Baum-Welch algorithm using bigram (see Non-Patent Document 1).
<Reference 4> Kenji Kita, "Probabilistic Language Model" (Language and Calculation 5), University of Tokyo Press, 1999

  The dictionary update unit 4 updates the recognition dictionary 10 using a set of feature vectors generated from the input signal after the processing by the recognition unit 3 is completed. At this time, the dictionary updating unit 4 learns the recognition dictionary 10 by the expected value maximization method using the constraint condition f (q, w) = 0 between the probabilistic parameters q and w. Hereinafter, a specific example of a method for updating the recognition dictionary 10 will be described.

_X 1 input feature vector to a state model, ..., and _{x N, 1} its burden rate gamma, ..., when written as gamma _N, firstly, calculates the following equation (7), K top k u ₁ eigenvectors corresponding to _eigenvalues, ..., as u _k, updating the dictionary subspace basis vector in the state model.

ただし、μは認識時の類似度を１から引いたものの重み付き平均であり、下記式（９）のように表される。

Then, for the constraint condition f (q, w) = 0 between the stochastic parameters q and w, the solution of the equation shown in the following equation (8) is set to q, and the solution is further set to f (q, w) = 0. Let w be the solution obtained by substitution.

However, μ is a weighted average obtained by subtracting the similarity at the time of recognition from 1, and is expressed as the following formula (9).

As a constraint condition between the probabilistic parameters q and w, for example, a condition that the effective dimension of the state model expressed by the following formula (10) is kept constant, that is, a condition expressed by the following formula (11) is given. It is done.

さらに、ｗ＝Ｅ／２μを上記式（１０）に代入して、下記式（１３）が得られる。これらの値で、確率化パラメータｑ，ｗを更新すればよい。

At this time, since the above equation (8) becomes the following equation (12), that is, E / 2w = μ, this is solved to obtain w = E / 2μ.

Further, by substituting w = E / 2μ into the above equation (10), the following equation (13) is obtained. The probability parameters q and w may be updated with these values.

The dictionary update unit 4 updates the dictionary subspace basis vectors u ₁ ,..., U _k as described above and the probability parameters q and w for each state model used for recognition by the recognition unit 3. Update. Thereby, whenever the recognition is performed, the recognition dictionary 10 is automatically learned, and the recognition accuracy is improved.

  Next, an outline of processing by the pattern recognition apparatus of the present embodiment will be described with reference to FIG. FIG. 3 is a flowchart illustrating an example of a processing procedure performed by the pattern recognition apparatus according to the present embodiment.

  First, the signal input unit 1 receives an input of a signal to be recognized (step S101). The signal input to the signal input unit 1 is passed to the feature extraction unit 2.

  Next, the feature extraction unit 2 receives the signal input in step S101 from the signal input unit 1, and generates a set of feature vectors from this signal by the method described above (step S102). A set of feature vectors generated by the feature extraction unit 2 is passed to the recognition unit 3.

  Next, the recognition unit 3 receives the set of feature vectors generated in step S102 from the feature extraction unit 2, evaluates the set of feature vectors using the recognition dictionary 10, and the signal input in step S101 belongs. A recognition result representing a class or a set of classes is output (step S103). At this time, the recognition unit 3 performs the above-described feature vector similarity calculation and likelihood calculation in each state of each model included in the recognition dictionary 10, and the data likelihood L as a whole is maximized. Select a combination of models and output the label set. After this recognition processing, the set of feature vectors input to the recognition unit 3 and the label set output from the recognition unit 3 are passed to the dictionary update unit 4.

  Next, the dictionary update unit 4 receives the set of feature vectors input to the recognition unit 3 and the label set output from the recognition unit 3, and the above-described method for each state model used for recognition by the recognition unit 3 Thus, the dictionary subspace basis vector and the stochastic parameter are updated (step S104). At this time, in particular, for updating the probability parameter, the constraint condition between the probability parameters described above is used.

  As described above, as described with specific examples, the pattern recognition apparatus according to the present embodiment uses the above-described stochastic subspace model as a state model of a model for each class. The recognition dictionary 10 holds dictionary subspace basis vectors and probability parameters corresponding to each state model. Then, the recognizing unit 3 calculates a similarity to the set of feature vectors generated from the input signal by using a quadratic polynomial of the inner product value of each feature vector and the dictionary subspace basis vector, and the obtained similarity The likelihood is calculated by the exponential function of the linear sum of the probability parameter and the model combination that maximizes the overall data likelihood is selected, and the label set is output as the recognition result. When the processing by the recognition unit 3 is finished, the dictionary update unit 4 updates the recognition dictionary 10 using a set of feature vectors generated from the input signal. At this time, the dictionary update unit 4 learns the recognition dictionary 10 by the expected value maximization method using the constraint condition between the probabilistic parameters. Therefore, according to the pattern recognition apparatus of this embodiment, high-accuracy recognition can be performed in a short time for an input signal whose recognition unit break is not clear.

In the above description, the condition that the effective dimension of the state model is kept constant is used as the constraint condition between the probabilistic parameters. However, the usable constraint condition is not limited to this. For example, a condition that the following expression (14) is kept constant in each state of the model, that is, a condition represented by the following expression (15) may be used.

At this time, since the above equation (8) becomes the following equation (16), the following equation (17) is obtained by substituting this into the above equation (14). The probability parameters q and w may be updated with these values.

  In the above description, an example in which the state model of the HMM is replaced with the probabilistic subspace model represented by Expression (1) is assumed, but the present invention is not limited to this. In another method of performing recognition determination and recognition at the same time, a model that takes time to calculate likelihood may be replaced with a probabilistic subspace model. Further, instead of the probabilistic subspace model represented by the equation (1), a model having the same function, that is, another model for calculating the similarity by the subspace method and calculating the likelihood from the similarity is used. You may do it.

  In the above description, an example in which the dictionary updating unit 4 is provided inside the pattern recognition device is assumed. However, the dictionary updating unit 4 may be provided outside the pattern recognition device. In this case, the dictionary update unit 4 provided outside the pattern recognition device performs the update process of the recognition dictionary 10 described above while communicating with the pattern recognition device, for example.

  As shown in FIG. 4, for example, the pattern recognition apparatus of this embodiment includes a processor such as a CPU (Central Processing Unit) 101, a storage device such as a ROM (Read Only Memory) 102 and a RAM (Random Access Memory) 103, an HDD ( (Hard Disk Drive) 104 or the like, a communication I / F 105 that communicates by connecting to a network, a bus 106 that connects each part, and the like can be employed. . In this case, each functional component described above can be realized by executing a predetermined pattern recognition program on the computer.

  This pattern recognition program is a file in an installable or executable format and is a CD-ROM (Compact Disk Read Only Memory), flexible disk (FD), CD-R (Compact Disk Recordable), DVD (Digital Versatile Disc). The program is recorded on a computer-readable recording medium such as a computer program product.

  Further, the pattern recognition program may be provided by being stored on another computer connected to a network such as the Internet and downloaded via the network. The pattern recognition program may be provided or distributed via a network such as the Internet.

  Further, the pattern recognition program may be provided by being incorporated in advance in the ROM 102 or the like.

  This pattern recognition program has a module configuration including each processing unit (the signal input unit 1, the feature extraction unit 2, the recognition unit 3, and the dictionary update unit 4) of the pattern recognition apparatus according to the present embodiment. As the hardware, for example, the CPU 101 (processor) reads out the program from the recording medium and executes the program, whereby each processing unit described above is loaded onto the RAM 103 (main memory), and each processing unit described above is loaded into the RAM 103 (main memory). ) To be generated on top. Note that the pattern recognition apparatus of the present embodiment realizes part or all of the above-described processing units using dedicated hardware such as an application specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). It is also possible.

  As mentioned above, although embodiment of this invention was described, embodiment described here is shown as an example and is not intending limiting the range of invention. The novel embodiments described herein can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. The embodiments and modifications described herein are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

1 the signal input unit 2 feature extraction section 3 recognition section 4 dictionary update unit 10 a recognition dictionary _{u 1,} ···, u _k subspace basis vector q, w probability parameter

Claims (6)

A pattern recognition device that performs pattern recognition of an input signal by converting the input signal into a feature vector and collating the feature vector with a recognition dictionary,
The recognition dictionary is a dictionary subspace base vector representing a dictionary subspace which is a subspace of the feature vector space, and a similarity calculated from the feature vector and the dictionary subspace is converted into likelihood. A plurality of stochastic parameters,
A recognizing unit that calculates the similarity by a quadratic polynomial of the inner product value of the feature vector and the dictionary subspace basis vector, and calculates the likelihood by an exponential function of a linear sum of the similarity and the probability parameter Prepared,
The pattern recognition apparatus, wherein the recognition dictionary is learned by an expected value maximization method using a constraint condition between the plurality of probability parameters.   The pattern recognition apparatus according to claim 1, further comprising a dictionary update unit that learns the recognition dictionary by an expected value maximization method using a constraint condition between the probability parameters. The recognition dictionary has a model composed of a plurality of states for each class, the dictionary subspace basis vector corresponding to each of the states of the model, and the probability parameter,
The said recognition part performs the calculation of the said similarity degree, and the calculation of the said likelihood with respect to one or more of the said feature vectors in each state of the said model, The Claim 1 or 2 characterized by the above-mentioned. Pattern recognition device.   The pattern recognition apparatus according to claim 3, wherein the constraint condition is a condition that an effective dimension of the state corresponding to the plurality of probability parameters is kept constant. A pattern recognition method executed in a pattern recognition apparatus that performs pattern recognition of an input signal by converting an input signal into a feature vector and collating the feature vector with a recognition dictionary,
The recognition dictionary is a dictionary subspace base vector representing a dictionary subspace which is a subspace of the feature vector space, and a similarity calculated from the feature vector and the dictionary subspace is converted into likelihood. A plurality of stochastic parameters,
The pattern recognition device calculates the similarity by a second order polynomial of the inner product value of the feature vector and the dictionary subspace basis vector;
The pattern recognition device calculating the likelihood by an exponential function of a linear sum of the similarity and the probability parameter; and
The pattern recognition method, wherein the recognition dictionary is learned by an expected value maximization method using a constraint condition between the plurality of probability parameters. A program for causing a computer to function as a pattern recognition device that performs pattern recognition of an input signal by converting the input signal into a feature vector and collating the feature vector with a recognition dictionary,
The recognition dictionary is a dictionary subspace base vector representing a dictionary subspace which is a subspace of the feature vector space, and a similarity calculated from the feature vector and the dictionary subspace is converted into likelihood. A plurality of stochastic parameters,
In the computer,
A function of calculating the degree of similarity by a quadratic polynomial of the inner product value of the feature vector and the dictionary subspace basis vector;
A function for calculating the likelihood by an exponential function of a linear sum of the similarity and the probability parameter;
The recognition dictionary is learned by an expected value maximization method using a constraint condition between the plurality of stochastic parameters.

Images