CN104077382A - Method for improving GDM (Global Data Manager) feature selection of audio classifier - Google Patents

Abstract

The invention discloses a method for improving GDM (Global Data Manager) feature selection of an audio classifier. The method comprises the following steps: training the Gaussian mixed model Gfc of each feature f specific to each audio type c, wherein c is between 1 and C; selecting a first feature f1, calculating the degree of separation between every two types specific to each feature, and selecting a first feature under which the average degree of separation among all the types is maximum; after the first feature f1 is selected, removing the feature f1 from a feature set to be selected, and finding out a type c1 and a type c2 between which the degree of separation is minimum in each type corresponding to f1; selecting a second feature f2, and selecting a feature f2 under which the degree of separation between the type c1 and the type c2 found in the third step is maximum; after the second feature f2 is selected, removing the feature f2 from the feature set to be selected, and constructing a feature vector by using the selected features f1 and f2 to obtain a final selected feature. Through a selected feature subset, types which are easily mixed up can be distinguished optimally, thereby increasing the overall classification accuracy of the classifier.

Description

A kind of for improving the GDM feature selection approach of audio classifiers

Technical field

The invention belongs to audio feature extraction field, relate in particular to a kind of for improving the GDM feature selection approach of audio classifiers.

Background technology

Audio frequency characteristics is another key factor that affects audio classifiers performance.One section of raw audio streams itself is only that a kind of non-semantic symbol represents and non-structured binary stream, and except containing the limited information such as sample frequency, quantified precision and coding method, itself does not comprise clear and definite structural information and semantic information.

People's ear has extremely strong resolution characteristic, and a given section audio stream not only can be told the type of audio frequency immediately, can also tell the sound characteristic of the very difficult descriptions such as speaker's mood in audio frequency and the keynote of music (excited or constrain etc.).Make computing machine can possess the function of as people's ear, audio frequency being classified and being identified, first need from a series of two-value symbol, to change audio stream into the characteristic parameter that can reflect difference between different audio types, i.e. feature extraction.Feature extraction is the basis of various classification problems.

According to the character of particular problem and specific area, select to have the feature of obvious separating capacity, be a part very crucial in design category process.The in the situation that of limited training sample, we wish that the characteristic Design with the least possible has the sorter of good versatility.

The average dispersion degree of the algorithm that traditional characteristic is selected based on making all categories maximum (GMM based Mean Separability Maximization is called for short GMSM) criterion, the performance of this feature selecting algorithm is easy to be subject to easily sub-category impact.And in fact, the performance of multiclass audio classifiers is except being subject to easily sub-category impact, it is more the impact that is subject to easily to obscure classification, improve the performance of sorter, improving the nicety of grading of easily obscuring between classification is the key point of problem, therefore,, during feature selecting, should choose those and make easily to obscure the feature that classification is more easily distinguished.

Summary of the invention

The object of the present invention is to provide a kind ofly for improving the GDM feature selection approach of audio classifiers, be intended to improve and easily obscure the nicety of grading between classification.

The present invention is achieved in that a kind of as follows for improving the concrete steps of GDM feature selection approach of audio classifiers:

Step 1, training pattern, to each audio types c, c ∈ [1, C], trains the gauss hybrid models G of every kind of feature f _fc;

Step 2, select first feature f ₁, to each feature, calculate the degree of separation between every two classifications, select the satisfied average divided degree maximum making between all categories of first feature, that is:

$f_1=\arg \underset{f\∈[1,F]}{\max }\underset{i=1}{\overset{C}{\mathrm{\Σ}}}\underset{j=1}{\overset{C}{\mathrm{\Σ}}}{S}_f(G_{\mathrm{fi}},G_{\mathrm{fj}});$

Step 3, selecting first feature f ₁after, from characteristic set to be selected, remove feature f ₁, and find out f ₁two class c of degree of separation minimum in each corresponding classification ₁and c ₂, that is:

$(c_1,c_2)=\arg \underset{i,j}{\min }{S}_{f_1}(G_{f_{1}i},G_{f_{1}j});$

Step 4, second feature f of selection ₂, selection makes two class c that find out in step 3 ₁and c ₂the feature f of degree of separation maximum ₂, that is:

$f_2=\arg \underset{f}{\max }{S}_{f_1}(G_{{\mathrm{fc}}_1},G_{{\mathrm{fc}}_2});$

Step 5, selecting second feature f ₂after, from characteristic set to be selected, remove feature f ₂, and with the feature f selecting ₁and f ₂form an eigenvector, iterative step three-step 4, when iteration, replaces respectively two formula in step 3 and step 4 with formula below, that is:

$\{c_l,c_{l+1}\}=\arg \underset{i,j}{\min }\underset{m=1}{\overset{l-1}{\mathrm{\Σ}}}{S}_{f_m}(i,j)$

$f_l=\arg \underset{f}{\max }{S}_{f_{l-1}}(G_{f_{l-1}{c}_l},G_{f_{l-1}{c}_{l+1}})$

Wherein, l represents the number of times of iteration;

Judge whether to meet iteration cut-off condition, if l<L returns to step 3, otherwise, stop iteration, what obtain selecting is characterized as: f ₁, f ₂..., f _l.

Further, for improving the problem of the GDM feature selection approach of audio classifiers, be described below:

Suppose to have C audio types, F kind feature, therefrom select L subcharacter, first, for each feature f, trains the gauss hybrid models GMM of each classification _fc, c ∈ [1, C], the probability density function that c class Gaussian Mixture distributes is:

$p\left(X\right|{\mathrm{\&Theta;}}_c)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&pi;}}_{i}p\left(X\right|{\mathrm{\&theta;}}_i)$

Wherein, K represents the number of mixed components, Θ _c=(π ₁..., π _k, θ ₁..., θ _k), the parameter of expression model, π _i, the weight of i mixed components of expression, meets constraint condition: θ _i={ μ _i, Σ _i, the parameter of i mixed components of expression;

P (X| θ _i) be each gaussian component, its expression-form is as follows:

$p\left(X\right|{\mathrm{\&theta;}}_i)=\frac{1}{\sqrt{{\left(2\mathrm{\&pi;}\right)}^{D/2}{\left|{\mathrm{\&Sigma;}}_i\right|}^{1/2}}}\exp (-\frac{1}{2}{(X-{\mathrm{\&mu;}}_i)}^T{{\mathrm{\&Sigma;}}_i}^{-1}(X-{\mathrm{\&mu;}}_i))$

Wherein, μ _ifor the mean value vector of D dimension, represent the average of gaussian component;

Σ _icovariance matrix for D * D;

The degree of separation (Separability) defining between two classifications is:

S _f(GMM _fk,GMM _fl)＝dis(GMM _fk,GMM _fl)

Dis () represents the distance between two gauss hybrid models, adopts improved symmetry distance tolerance K-L2 distance, and computing formula is as follows:

$d(i,j)=\frac{1}{2}{({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)}^T({\mathrm{\&Sigma;}}_i^{-1}+{\mathrm{\&Sigma;}}_j^{-1})({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)+\frac{\left|{\mathrm{\&Sigma;}}_j\right|}{\left|{\mathrm{\&Sigma;}}_i\right|}+\frac{\left|{\mathrm{\&Sigma;}}_i\right|}{\left|{\mathrm{\&Sigma;}}_j\right|}$

effect gathers

Of the present inventionly for improving character subset that the GDM feature selection approach of audio classifiers selects, can make the class discrimination the most easily obscured best, can improve the nicety of grading of sorter integral body.

Accompanying drawing explanation

Fig. 1 be the embodiment of the present invention provide for improving the GDM feature selection approach process flow diagram of audio classifiers.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.

Of the present inventionly for improving the problem of the GDM feature selection approach of audio classifiers, be described below:

Suppose to have C audio types, F kind feature, therefrom select L subcharacter, first, for each feature f, trains the gauss hybrid models GMM of each classification _fc, c ∈ [1, C], the probability density function that c class Gaussian Mixture distributes is:

$p\left(X\right|{\mathrm{\&Theta;}}_c)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&pi;}}_{i}p\left(X\right|{\mathrm{\&theta;}}_i)$

Wherein, K represents the number of mixed components, Θ _c=(π ₁..., π _k, θ ₁..., θ _k), the parameter of expression model, π _i, the weight of i mixed components of expression, meets constraint condition: θ _i={ μ _i, Σ _i, the parameter of i mixed components of expression;

P (X| θ _i) be each gaussian component, its expression-form is as follows:

$p\left(X\right|{\mathrm{\&theta;}}_i)=\frac{1}{\sqrt{{\left(2\mathrm{\&pi;}\right)}^{D/2}{\left|{\mathrm{\&Sigma;}}_i\right|}^{1/2}}}\exp (-\frac{1}{2}{(X-{\mathrm{\&mu;}}_i)}^T{{\mathrm{\&Sigma;}}_i}^{-1}(X-{\mathrm{\&mu;}}_i))$

Wherein, μ _ifor the mean value vector of D dimension, represent the average of gaussian component;

Σ _icovariance matrix for D * D, has various ways, can be complete matrix, Block diagonal matrix or diagonal matrix.For ease of computing, conventionally suppose between each dimensional feature separate, i.e. Σ _iit is diagonal matrix.

The degree of separation (Separability) defining between two classifications is:

S _f(GMM _fk,GMM _fl)＝dis(GMM _fk,GMM _fl)

Dis () represents the distance between two gauss hybrid models, and its measurement criterion has a variety of.As shown in table 1, A and B represent respectively two gauss hybrid models, and d (i, j) represents the distance between two gaussian component, and its conventional tolerance has Euclidean distance, mahalanobis distance, K-L distance and Bhattachyaryya distance etc.Wherein, between gauss hybrid models, in class divergence range formula, dis (A, B) can be any one in these four kinds of distances.

Table 1 common distance measurement criterion

In above table, K-L distance is asymmetrical, and for ease of calculating, the present embodiment adopts improved symmetry distance tolerance K-L2 distance, and computing formula is as follows:

$d(i,j)=\frac{1}{2}{({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)}^T({\mathrm{\&Sigma;}}_i^{-1}+{\mathrm{\&Sigma;}}_j^{-1})({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)+\frac{\left|{\mathrm{\&Sigma;}}_j\right|}{\left|{\mathrm{\&Sigma;}}_i\right|}+\frac{\left|{\mathrm{\&Sigma;}}_i\right|}{\left|{\mathrm{\&Sigma;}}_j\right|}$

By formula S _f(GMM _fk, GMM _fl)=dis (GMM _fk, GMM _fl) computing formula of degree of separation can find out, from two gauss hybrid models away from must be more, its degree of separation is larger, corresponding, and two classifications are also more easily distinguished.For easy differentiation classification, its classification accuracy is higher, therefore, when feature selecting, needn't give too many consideration.On the contrary, two its degree of separation of gauss hybrid models close to must be are more less, mutually deserved, and two such classifications are more easily obscured.For the classification of easily obscuring, its classification accuracy is very low, causes declining to a great extent of whole classifier performance.Therefore, during feature selecting, should choose those and can make easily to obscure the feature that classification degree of separation is large, so just can make final sorter improve the classification accuracy of all classifications.In other words, when selecting feature, first to find those classifications of easily obscuring, then select to make these easily to obscure the feature that classification is easily distinguished.

Fig. 1 shows of the present invention for improving the flow process of the GDM feature selection approach of audio classifiers, and as shown in the figure, the present invention is achieved in that a kind of as follows for improving the concrete steps of GDM feature selection approach of audio classifiers:

S101: training pattern, to each audio types c, c ∈ [1, C], trains the gauss hybrid models G of every kind of feature f _fc;

S102: select first feature f ₁, to each feature, calculate the degree of separation between every two classifications, select the satisfied average divided degree maximum making between all categories of first feature, that is:

$f_1=\arg \underset{f\&Element;[1,F]}{\max }\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{C}{\mathrm{\&Sigma;}}}{S}_f(G_{\mathrm{fi}},G_{\mathrm{fj}});$

S103: selecting first feature f ₁after, from characteristic set to be selected, remove feature f ₁, and find out f ₁two class c of degree of separation minimum in each corresponding classification ₁and c ₂, that is:

$(c_1,c_2)=\arg \underset{i,j}{\min }{S}_{f_1}(G_{f_{1}i},G_{f_{1}j});$

S104: select second feature f ₂, selection makes two class c that find out in step 3 ₁and c ₂the feature f of degree of separation maximum ₂, that is:

$f_2=\arg \underset{f}{\max }{S}_{f_1}(G_{{\mathrm{fc}}_1},G_{{\mathrm{fc}}_2});$

S105: selecting second feature f ₂after, from characteristic set to be selected, remove feature f ₂, and with the feature f selecting ₁and f ₂form an eigenvector, iterative step three-step 4, when iteration, replaces respectively two formula in step 3 and step 4 with formula below, that is:

$\{c_l,c_{l+1}\}=\arg \underset{i,j}{\min }\underset{m=1}{\overset{l-1}{\mathrm{\&Sigma;}}}{S}_{f_m}(i,j)$

$f_l=\arg \underset{f}{\max }{S}_{f_{l-1}}(G_{f_{l-1}{c}_l},G_{f_{l-1}{c}_{l+1}})$

Wherein, l represents the number of times of iteration;

Judge whether to meet iteration cut-off condition, if l<L returns to step 3, otherwise, stop iteration, what obtain selecting is characterized as: f ₁, f ₂..., f _l.

Although above-mentioned, by reference to the accompanying drawings the specific embodiment of the present invention is described; but be not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various modifications that performing creative labour can make or distortion still within protection scope of the present invention.

Claims (2)

1. for improving a GDM feature selection approach for audio classifiers, it is characterized in that, described as follows for improving the concrete steps of GDM feature selection approach of audio classifiers:

Step 1, training pattern, to each audio types c, c ∈ [1, C], trains the gauss hybrid models G of every kind of feature f _fc;

Step 2, select first feature f1, to each feature, calculate the degree of separation between every two classifications, the average divided degree of selecting first feature to meet to make between all categories is maximum, that is:

$f_1=\arg \underset{f\&Element;[1,F]}{\max }\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{C}{\mathrm{\&Sigma;}}}{S}_f(G_{\mathrm{fi}},G_{\mathrm{fj}});$

Step 3, selecting first feature f ₁after, from characteristic set to be selected, remove feature f ₁, and find out f ₁two class c of degree of separation minimum in each corresponding classification ₁and c ₂, that is:

$(c_1,c_2)=\arg \underset{i,j}{\min }{S}_{f_1}(G_{f_{1}i},G_{f_{1}j});$

Step 4, second feature f of selection ₂, selection makes two class c that find out in step 3 ₁and c ₂the feature f of degree of separation maximum ₂, that is:

$f_2=\arg \underset{f}{\max }{S}_{f_1}(G_{{\mathrm{fc}}_1},G_{{\mathrm{fc}}_2});$

Step 5, selecting second feature f ₂after, from characteristic set to be selected, remove feature f ₂, and with the feature f selecting ₁and f ₂form an eigenvector, iterative step three-step 4, when iteration, replaces respectively two formula in step 3 and step 4 with formula below, that is:

$\{c_l,c_{l+1}\}=\arg \underset{i,j}{\min }\underset{m=1}{\overset{l-1}{\mathrm{\&Sigma;}}}{S}_{f_m}(i,j)$

$f_l=\arg \underset{f}{\max }{S}_{f_{l-1}}(G_{f_{l-1}{c}_l},G_{f_{l-1}{c}_{l+1}})$

Wherein, l represents the number of times of iteration;

Judge whether to meet iteration cut-off condition, if l<L returns to step 3, otherwise, stop iteration, what obtain selecting is characterized as: f ₁, f ₂..., f _l.

2. as claimed in claim 1ly for improving the GDM feature selection approach of audio classifiers, it is characterized in that, for improving the problem of the GDM feature selection approach of audio classifiers, be described below:

Suppose to have C audio types, F kind feature, therefrom select L subcharacter, first, for each feature f, trains the gauss hybrid models GMM of each classification _fc, c ∈ [1, C], the probability density function that c class Gaussian Mixture distributes is:

$p\left(X\right|{\mathrm{\&Theta;}}_c)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&pi;}}_{i}p\left(X\right|{\mathrm{\&theta;}}_i)$

Wherein, K represents the number of mixed components, Θ _c=(π ₁..., π _k, θ ₁..., θ _k), the parameter of expression model, π _i, the weight of i mixed components of expression, meets constraint condition: θ _i={ μ _i, Σ _i, the parameter of i mixed components of expression;

P (X| θ _i) be each gaussian component, its expression-form is as follows:

$p\left(X\right|{\mathrm{\&theta;}}_i)=\frac{1}{\sqrt{{\left(2\mathrm{\&pi;}\right)}^{D/2}{\left|{\mathrm{\&Sigma;}}_i\right|}^{1/2}}}\exp (-\frac{1}{2}{(X-{\mathrm{\&mu;}}_i)}^T{{\mathrm{\&Sigma;}}_i}^{-1}(X-{\mathrm{\&mu;}}_i))$

Wherein, μ _ifor the mean value vector of D dimension, represent the average of gaussian component;

Σ _icovariance matrix for D * D;

The degree of separation (Separability) defining between two classifications is:

S _f(GMM _fk,GMM _fl)＝dis(GMM _fk,GMM _fl)

Dis () represents the distance between two gauss hybrid models, adopts improved symmetry distance tolerance K-L2 distance, and computing formula is as follows:

$d(i,j)=\frac{1}{2}{({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)}^T({\mathrm{\&Sigma;}}_i^{-1}+{\mathrm{\&Sigma;}}_j^{-1})({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)+\frac{\left|{\mathrm{\&Sigma;}}_j\right|}{\left|{\mathrm{\&Sigma;}}_i\right|}+\frac{\left|{\mathrm{\&Sigma;}}_i\right|}{\left|{\mathrm{\&Sigma;}}_j\right|}.$