CN104714237A - Fish identification method with multi-feature and multidirectional data fused - Google Patents

Abstract

The invention relates to the field of acoustic fish identification, in particular to a fish identification method with multi-feature and multidirectional data fused. The method comprises the steps that acoustical signals are emitted to the underwater, and fish body multidirectional acoustic scattering signals are acquired; the acquired multidirectional acoustic scattering signals are normalized and filtered; the preprocessed signals are subjected to multi-feature extraction; the preprocessed multidirectional acoustic scattering data are subjected to orthogonal transformation, envelopes are extracted, the wavelet packet coefficient singular value features, the time domain mass center features and the frequency domain mass center features of envelope information are extracted, and feature fusion and dimension reduction processing are conducted. The multidirectional data acquiring method is simple and easy to implement. Based on the extracted multiple features, the multidirectional acoustical scattering features are subjected to collaboration fusion, the fusion degree is high, fusion is compact, and the problems that identification is not clear, and correct identification even cannot be achieved when only single-dimension acoustical scattering information is classified can be effectively solved.

Description

A kind of multiple features and multi-faceted data fusion fish identification method

Technical field

What the present invention relates to is a kind of acoustics fish identification field, is specifically related to a kind of multiple features and multi-faceted data fusion fish identification method.

Background technology

Patent of invention " a kind of fish identification method based on wavelet packet multi-scale information entropy and system in substantive examination, publication number: 103308919A " describe a kind of fish identification method, the technical characterstic of the method is to carry out fish identification based on the single features of single bearing data, but fish is a complicated sound scatterer, the orientation of fish body and air bladder, size, shape has a strong impact on the acoustic scattering feature of fish, thus make the acoustic scattering signal of fish also become very complicated, extract from sophisticated signal and can reflect that the feature of fish essence sound scattering characteristics is particularly important in fish identification.The scattering properties that fish shows in different acoustic scattering orientation is different, and therefore single orientation detection will certainly lose the scattered information of fish.And fish is free-swimming in water, the concrete orientation of unpredictable fish in wave beam during detection, therefore the sound scattering characteristics obtaining fish under actual environment shows certain randomness in scattering orientation, causes the accuracy of identification problem that is poor, that even can not correctly identify that to there is the orientation difference due to fish and cause in fish identifying.

Summary of the invention

The object of the invention is to solve the high precision fish identification problem based on acoustic data, and a kind of multiple features and multi-faceted data fusion fish identification method are provided.

The object of the present invention is achieved like this:

Multiple features and multi-faceted data fusion fish identification method, comprise the steps:

(1) to underwater emission acoustical signal, the multi-faceted acoustic scattering signal of fish body is obtained;

(2) to obtain multi-faceted acoustic scattering signal be normalized, filtering process;

(3) multi-feature extraction is carried out to pretreated signal: orthogonal transformation is carried out to pretreated multi-faceted acoustic scattering data, extract envelope, wavelet packet coefficient singular value features, time domain centroid feature, frequency domain centroid feature are extracted to envelope information, carries out Fusion Features and dimension-reduction treatment;

(4) characteristic quantity is input to support vector machine classifier and carries out fish identification: by support vector machine classifier, the class result of decision is expressed as posterior probability, utilize the decision probability of the decision probability in each orientation to all the other orientation to be weighted simultaneously, all decision packages cooperate mutually, complete fish identification.

Concrete grammar envelope information being extracted to wavelet packet coefficient singular value features is: obtain wavelet packet coefficient by WAVELET PACKET DECOMPOSITION and reconstruct, formed wavelet packet coefficient matrix, obtain wavelet packet coefficient singular value:

${\mathrm{SD}}_{m,n}=\sqrt{{\mathrm{\λ}}_{m,n}}$

D _mfor the matrix of wavelet package reconstruction coefficient composition, λ _m,nfor d _m ^hd _mnonzero eigenvalue, SD _m,nfor characteristic quantity, m is take over party's item, and n is Characteristic Number.

Concrete grammar envelope information being extracted to time domain centroid feature is:

(3.1.1) whole segment signal time domain barycenter TC is calculated ₁₁, obtain two subband [0, TC of ground floor ₁₁] and [TC ₁₁, T _max], signal time length is 0-T _max;

(3.1.2) on the basis of ground floor segmentation, calculate the time domain barycenter of each subband respectively, be respectively TC ₂₁and TC ₂₂, obtain three subband [0, TC of the second layer ₂₁], [TC ₂₁, TC ₂₂], [TC ₂₂, T _max];

(3.1.3) in i-th layer of each subband, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, be generally advisable with 3 ~ 6 layers, then character pair number is 3 ~ 6.

Extract frequency domain centroid feature to envelope information to comprise:

(3.2.1) the frequency domain barycenter SC of frequency range is calculated ₁₁, obtain two subband [0, SC of ground floor ₁₁] and [SC ₁₁, f _max], the scope that desirable signal energy is concentrated is as original analysis frequency range, and signal frequency range is 0-f _max;

(3.2.2) on the basis of ground floor segmentation, calculate the frequency domain barycenter of each subband respectively, be respectively SC ₂₁and SC ₂₂, obtain three subband [0, SC of the second layer ₂₁], [SC ₂₁, SC ₂₂], [SC ₂₂, f _max];

(3.2.3) by that analogy, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, be generally advisable with 3 ~ 6 layers in each subband of jth layer, then character pair number is 3 ~ 6.

Fusion Features and dimension-reduction treatment refer to three kinds of features to combine, composition fusion feature vector, and carry out Feature Dimension Reduction process by Fisher method of discrimination.

In step (4), support vector machine exports and is:

f(T _m)＝wφ(T _m)+b

Wherein, f (x) is for exporting lineoid, and w is weight vector, and b is bias amount, and φ is Nonlinear Mapping, T _mrepresentative feature, m is orientation number;

Posterior probability estimation:

$p\left(c\right|T_m)=\frac{\exp (\mathrm{w\φ}\left(T_m\right)+b)}{{\mathrm{\Σ}}_{c=1}^{C_l}\exp (\mathrm{w\φ}\left(T_m\right)+b)}$

Wherein c is category label, C _lfor total class quantity;

The posterior probability vector that each orientation exports is:

$p_m=(p\left(c_1\right|T_m),...,p\left(c_{C_l}\right|T_m))$

The probability vector p that orientation m exports _mbe sent to other decision packages, receive the probability vector p that other decision packages export simultaneously ₁..., p _m-1, p _m+1... p _m, be independent distribution, then the probability-weighted of orientation m:

$p_{\mathrm{weig}}\left(c_k\right|t_m)=\underset{i\≠j}{\overset{A}{\mathrm{\Π}}}{p}_m\left(c_k\right|T_m)$

K represents the label of a certain class, t _mcharacteristic vector space after representative reconfigures;

After obtaining decision probability vector and probability-weighted vector, for avoiding introducing new sorter again, combine, definition slope K _sL, then K _sLbe expressed as:

$K_{\mathrm{SL}}=\frac{p\left(c_1\right|T)-p\left(c_2\right|T)}{p\left(c_1\right|T)}$

Wherein

$p\left(c_1\right|T)\≥p\left(c_2\right|T)\≥...p\left(c_{C_l}\right|T)$

Then decision probability vector is respectively with probability-weighted vector weight:

$W_{\mathrm{init}}=\frac{K_{\mathrm{SL}}^{\mathrm{init}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}},W_{\mathrm{weig}}=\frac{K_{\mathrm{SL}}^{\mathrm{weig}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}}$

According to shared weight, calculate the final decision probability in each orientation:

$p_m^f\left(c\right|T)=p_m\left(c\right|T_m)*W_{\mathrm{init}}+p_{\mathrm{weig}}\left(c\right|T_m)*W_{\mathrm{weig}}$

Wherein T is total characteristic vector space;

Final Classification and Identification result then for each class is:

$P\left(c\right|T)=\underset{m=1}{\overset{M}{\mathrm{\Π}}}{p}_m^f\left(c\right|T)$

Wherein M is orientation quantity;

For the feature being input to support vector machine classifier, the probability of which class greatly then expresses support for vector machine classifier and the feature of current input is identified as this class.

Beneficial effect of the present invention is:

(1) the multi-faceted acoustic scattering data based on fish are extracted wavelet packet coefficient singular value, time domain barycenter, frequency domain barycenter three kinds of features, multiple features reflects the acoustics essential characteristic of fish, describe different fingerling at time domain and frequency domain distribution property difference, and can effective supplement be carried out between these features, these features are carried out merging and improves the comprehensive of characteristic information, simultaneously, for improving feature stability, reducing identifying complexity, adopt Fisher method of discrimination to carry out Feature Dimension Reduction process, remain the more effective feature of fish identification.Multiple features can effectively solve based on the low problem of the discrimination of single features.

(2) multi-faceted data capture method is simple, is easy to realize; Based on the multiple features of said extracted, multi-faceted acoustic scattering feature is carried out cooperation and is merged by the present invention, merges degree high and tight, effectively can solve when a folk prescription position acoustic scattering information is classified and identify problem that is unclear, that even can not correctly identify.

Accompanying drawing explanation

Fig. 1 is multiple features and multi-faceted data fusion fish identification method structured flowchart

Fig. 2 is multi-faceted acoustic scattering data capture method schematic diagram;

Fig. 3 is multiple features feature extracting method structured flowchart;

Fig. 4 is that multi-faceted acoustic scattering collaboration data merges fish identification method structured flowchart

Embodiment

Below in conjunction with accompanying drawing, the invention will be described further with enforcement:

Step (01) utilizes a single beam transmitting transducer to launch acoustical signal to fish body, utilizes M nautical receiving set to receive fish volume scattering acoustical signal M different azimuth, obtains multi-faceted fish volume scattering data, M=6 ~ 9 simultaneously.

The process such as step (02) is normalized the multi-faceted acoustic scattering signal obtained, filtering;

Step (03) extracts wavelet packet coefficient singular value, time domain centroid feature, and carries out Fusion Features, and concrete grammar is:

Step (03) obtains wavelet packet coefficient singular value by formula (1):

${\mathrm{SD}}_{m,n}=\sqrt{{\mathrm{\λ}}_{m,n}}---\left(1\right)$

If d _mfor the matrix of wavelet package reconstruction coefficient composition, then λ _m,nfor d _m ^hd _mnonzero eigenvalue, SD _m,nfor characteristic quantity, m is take over party's item, and n is Characteristic Number.

Wavelet packet coefficient singular value features is:

T _sd,m＝(SD _m,1,SD _m,2,…,SD _m,n) (2)

Time domain barycenter is solved by method below:

(1) acoustic scattering signal time domain barycenter TC is calculated ₁₁, obtain two subband [0, TC of ground floor ₁₁] and [TC ₁₁, T _max];

(2) on the basis of ground floor segmentation, calculate the time domain barycenter of each subband respectively, be respectively TC ₂₁and TC ₂₂, obtain three subband [0, TC of the second layer ₂₁], [TC ₂₁, TC ₂₂], [TC ₂₂, T _max];

(3) by that analogy, in i-th layer of each subband, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, i=3 ~ 6.

According to the strategy of (1)-(3), each time domain barycenter method for solving is:

${\mathrm{TC}}_m=\frac{\underset{0}{\overset{L-1}{\mathrm{\Σ}}}n{n_m}^2\left(l\right)}{\underset{0}{\overset{L-1}{\mathrm{\Σ}}}{x_m}^2\left(l\right)}---\left(3\right)$

Wherein, L is the signal length of each subband, the amplitude that x (l) is signal.

Time domain centroid feature is:

T _tc,m＝(TC _m,1,TC _m,2,…,TC _m,n) (4)

Frequency domain barycenter is solved by method below:

(1) the frequency domain barycenter SC of frequency range is calculated ₁₁, obtain two subband [0, SC of ground floor ₁₁] and [SC ₁₁, f _max], the scope that desirable signal energy is concentrated is as original analysis frequency range;

(2) on the basis of ground floor segmentation, calculate the frequency domain barycenter of each subband respectively, be respectively SC ₂₁and SC ₂₂, obtain three subband [0, SC of the second layer ₂₁], [SC ₂₁, SC ₂₂], [SC ₂₂, f _max];

(3) by that analogy, in each subband of jth layer, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, j=3 ~ 6.

According to the strategy of (1)-(3), each frequency domain barycenter method for solving is:

${\mathrm{SC}}_m=\frac{{\∫}_0^{f_{\max }}f{E}_m\left(f\right)\mathrm{df}}{{\∫}_0^{f_{\max }}{E}_m\left(f\right)\mathrm{df}}---\left(5\right)$

Wherein: f is signal frequency; E (f) is the spectrum energy of respective frequencies after continued time domain signal x (t) Fourier transform.

Frequency domain centroid feature is:

T _sc,m＝(SC _m,1,SC _m,2,…,SC _m,n) (6)

Fusion feature is:

T _m＝(T _sd,m,T _tc,m,T _sc,m) (7)

On the basis of merging, carry out Feature Dimension Reduction by formula (8):

$D_T=\frac{{|{\mathrm{\μ}}_x-{\mathrm{\μ}}_y|}^2}{{\mathrm{\σ}}_x^2+{\mathrm{\σ}}_y^2}---\left(8\right)$

Wherein x, y represent classification, wherein μ _x, μ _y, be respectively class x, the mean and variance of y character pair amount.

Step (04) adopts support vector machine classifier to carry out fish identification, and support vector machine exports and is:

f(T _m)＝wφ(T _m)+b (9)

Wherein, f (x) is for exporting lineoid, and w is weight vector, and b is bias amount, and φ is Nonlinear Mapping, T _mrepresentative feature, m is orientation number.

Described step (04) obtains posterior probability estimation by formula (10):

$p\left(c\right|T_m)=\frac{\exp (\mathrm{w\φ}\left(T_m\right)+b)}{{\mathrm{\Σ}}_{c=1}^{C_l}\exp (\mathrm{w\φ}\left(T_m\right)+b)}---\left(10\right)$

Wherein c is category label, C _lfor total class quantity.

The posterior probability vector that each orientation exports is:

$p_m=(p\left(c_1\right|T_m),...,p\left(c_{C_l}\right|T_m))---\left(11\right)$

The probability vector p that orientation m exports _mbe sent to other decision packages, receive the probability vector p that other decision packages export simultaneously ₁..., p _m-1, p _m+1... p _m, suppose that these probability are independent distribution, then the probability-weighted of orientation m:

$p_{\mathrm{weig}}\left(c_k\right|t_m)=\underset{i\≠j}{\overset{A}{\mathrm{\Π}}}{p}_m\left(c_k\right|T_m)---\left(12\right)$

Wherein k represents the label of a certain class, t _mfeature space after representative reconfigures.

After obtaining decision probability vector and probability-weighted vector, for avoiding introducing new sorter again, formula (13) method is adopted to combine, definition slope K _sL, then K _sLbe expressed as:

$K_{\mathrm{SL}}=\frac{p\left(c_1\right|T)-p\left(c_2\right|T)}{p\left(c_1\right|T)}---\left(13\right)$

Wherein

$p\left(c_1\right|T)\≥p\left(c_2\right|T)\≥...p\left(c_{C_l}\right|T)---\left(14\right)$

Then decision probability vector is respectively with probability-weighted vector weight:

$W_{\mathrm{init}}=\frac{K_{\mathrm{SL}}^{\mathrm{init}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}},W_{\mathrm{weig}}=\frac{K_{\mathrm{SL}}^{\mathrm{weig}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}}---\left(15\right)$

According to shared weight, calculate the final decision probability in each orientation:

$p_m^f\left(c\right|T)=p_m\left(c\right|T_m)*W_{\mathrm{init}}+p_{\mathrm{weig}}\left(c\right|T_m)*W_{\mathrm{weig}}---\left(16\right)$

Final Classification and Identification result then for each class is:

$P\left(c\right|T)=\underset{m=1}{\overset{M}{\mathrm{\Π}}}{p}_m^f\left(c\right|T)---\left(17\right)$

For the feature being input to support vector machine classifier, the probability of which class greatly then expresses support for vector machine classifier and the feature of current input is identified as this class.

Fig. 2 is multi-faceted acoustic scattering data capture method schematic diagram, and wherein filled circles represents transmitting transducer, and open circles represents receiving transducer, and transmitting transducer of the present invention is phase array transducer, and receiving transducer is standard hydrophone.

Extract in the present invention and obtain wavelet packet coefficient singular value, time domain barycenter, frequency domain centroid feature, multiple features feature extraction flow process as shown in Figure 3.Feature extraction mainly comprises several steps such as envelope information extraction, feature extraction, Fusion Features, Feature Dimension Reduction, wavelet packet coefficient singular value characterizes the joint distribution characteristic of fish on time-frequency domain, time domain barycenter then illustrates the power distribution properties of fish in time domain, and frequency domain barycenter then illustrates the power distribution properties of fish on frequency domain.

Specific implementation method is:

Step (01) utilizes a single beam phased transducer to launch acoustical signal to fish body, utilizes M nautical receiving set to receive fish body acoustic scattering signal M different azimuth, obtains multi-faceted fish volume scattering data, M=6 ~ 9 simultaneously;

The process such as step (02) is normalized the multi-faceted scattered signal obtained, filtering;

Step (03) extracts wavelet packet coefficient singular value features, time domain centroid feature, frequency domain centroid feature, obtains the fusion feature of three, and carries out Feature Dimension Reduction process;

As shown in Figure 4, the corresponding decision package of the proper vector in each orientation in step (04), decision package mainly comprises the processing procedure to acoustic scattering data such as sorter, probability weight.For orientation m, first extraction feature is carried out to acoustic scattering data, and by support vector machine classifier, the class result of decision is expressed as posterior probability, utilize the decision probability of the decision probability of orientation m to all the other each orientation to be weighted simultaneously, all decision packages cooperate mutually, complete the prediction of unknown fingerling.

Compared with prior art, advantage of the present invention is:

(1) the multi-faceted acoustic scattering data based on fish are extracted wavelet packet coefficient singular value, time domain barycenter, frequency domain barycenter three kinds of features, multiple features reflects the acoustics essential characteristic of fish, describe different fingerling at time domain and frequency domain distribution property difference, and can effective supplement be carried out between these features, these features are carried out merging and improves the comprehensive of characteristic information, simultaneously, for improving feature stability, reducing identifying complexity, adopt Fisher method of discrimination to carry out Feature Dimension Reduction process, remain the more effective feature of fish identification.Multiple features can effectively solve based on the low problem of the discrimination of single features.

(2) multi-faceted data capture method is simple, is easy to realize; Based on the multiple features of said extracted, multi-faceted acoustic scattering feature is carried out cooperation and is merged by the present invention, merges degree high and tight, effectively can solve when a folk prescription position acoustic scattering information is classified and identify problem that is unclear, that even can not correctly identify.

Claims (6)

1. multiple features and a multi-faceted data fusion fish identification method, is characterized in that, comprise the steps:

(1) to underwater emission acoustical signal, the multi-faceted acoustic scattering signal of fish body is obtained;

(2) to obtain multi-faceted acoustic scattering signal be normalized, filtering process;

(3) multi-feature extraction is carried out to pretreated signal: orthogonal transformation is carried out to pretreated multi-faceted acoustic scattering data, extract envelope, wavelet packet coefficient singular value features, time domain centroid feature, frequency domain centroid feature are extracted to envelope information, carries out Fusion Features and dimension-reduction treatment;

(4) characteristic quantity is input to support vector machine classifier and carries out fish identification: by support vector machine classifier, the class result of decision is expressed as posterior probability, utilize the decision probability of the decision probability in each orientation to all the other orientation to be weighted simultaneously, all decision packages cooperate mutually, complete fish identification.

2. a kind of multiple features according to claim 1 and multi-faceted data fusion fish identification method, it is characterized in that: described to the concrete grammar of envelope information extraction wavelet packet coefficient singular value features is: obtain wavelet packet coefficient by WAVELET PACKET DECOMPOSITION and reconstruct, formed wavelet packet coefficient matrix, obtained wavelet packet coefficient singular value:

${\mathrm{SD}}_{m,n}\sqrt{{\mathrm{\λ}}_{m,n}}$

D _mfor the matrix of wavelet package reconstruction coefficient composition, λ _m,nfor d _m ^hd _mnonzero eigenvalue, SD _m,nfor characteristic quantity, m is take over party's item, and n is Characteristic Number.

3. a kind of multiple features according to claim 1 and multi-faceted data fusion fish identification method, is characterized in that: the described concrete grammar to envelope information extraction time domain centroid feature is:

(3.1.1) whole segment signal time domain barycenter TC is calculated ₁₁, obtain two subband [0, TC of ground floor ₁₁] and [TC ₁₁, T _max], signal time length is 0-T _max;

(3.1.2) on the basis of ground floor segmentation, calculate the time domain barycenter of each subband respectively, be respectively TC ₂₁and TC ₂₂, obtain three subband [0, TC of the second layer ₂₁], [TC ₂₁, TC ₂₂], [TC ₂₂, T _max];

(3.1.3) in i-th layer of each subband, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, be generally advisable with 3 ~ 6 layers, then character pair number is 3 ~ 6.

4. a kind of multiple features according to claim 1 and multi-faceted data fusion fish identification method, is characterized in that: described comprises envelope information extraction frequency domain centroid feature:

(3.2.1) the frequency domain barycenter SC of frequency range is calculated ₁₁, obtain two subband [0, SC of ground floor ₁₁] and [SC ₁₁, f _max], the scope that desirable signal energy is concentrated is as original analysis frequency range, and signal frequency range is 0-f _max;

(3.2.2) on the basis of ground floor segmentation, calculate the frequency domain barycenter of each subband respectively, be respectively SC ₂₁and SC ₂₂, obtain three subband [0, SC of the second layer ₂₁], [SC ₂₁, SC ₂₂], [SC ₂₂, f _max];

(3.2.3) by that analogy, calculate time domain barycenter, and it can be used as the foundation that lower one deck divides, be generally advisable with 3 ~ 6 layers in each subband of jth layer, then character pair number is 3 ~ 6.

5. a kind of multiple features according to claim 1 and multi-faceted data fusion fish identification method, it is characterized in that: described Fusion Features and dimension-reduction treatment refer to three kinds of features to combine, composition fusion feature vector, and carry out Feature Dimension Reduction process by Fisher method of discrimination.

6. a kind of multiple features according to claim 1 and multi-faceted data fusion fish identification method, is characterized in that: in described step (4), support vector machine exports and is:

f(T _m)＝wφ(T _m)+b

Wherein, f (x) is for exporting lineoid, and w is weight vector, and b is bias amount, and φ is Nonlinear Mapping, T _mrepresentative feature, m is orientation number;

Posterior probability estimation:

$p\left(c\right|T_m)=\frac{\exp (\mathrm{w\φ}\left(T_m\right)+b)}{{\mathrm{\Σ}}_{c=1}^{C_l}\exp (\mathrm{w\φ}\left(T_m\right)+b)}$

Wherein c is category label, C _lfor total class quantity;

The posterior probability vector that each orientation exports is:

$p_m=(p\left(c_1\right|T_m),...,p\left(c_{C_l}\right|T_m))$

The probability vector p that orientation m exports _mbe sent to other decision packages, receive the probability vector p that other decision packages export simultaneously ₁..., p _m-1, p _m+1... p _m, be independent distribution, then the probability-weighted of orientation m:

$p_{\mathrm{weig}}\left(c_k\right|t_m)=\underset{i\≠j}{\overset{A}{\mathrm{\Π}}}{p}_m\left(c_k\right|T_m)$

K represents the label of a certain class, t _mcharacteristic vector space after representative reconfigures;

After obtaining decision probability vector and probability-weighted vector, for avoiding introducing new sorter again, combine, definition slope K _sL, then K _sLbe expressed as:

$K_{\mathrm{SL}}=\frac{p\left(c_1\right|T)-p\left(c_2\right|T)}{p\left(c_1\right|T)}$

Wherein

p(c ₁|T)≥p(c ₂|T)≥…p(c _Cl|T)

Then decision probability vector is respectively with probability-weighted vector weight:

$W_{\mathrm{init}}=\frac{K_{\mathrm{SL}}^{\mathrm{init}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}},W_{\mathrm{weig}}=\frac{K_{\mathrm{SL}}^{\mathrm{weig}}}{K_{\mathrm{SL}}^{\mathrm{init}}+K_{\mathrm{SL}}^{\mathrm{weig}}}$

According to shared weight, calculate the final decision probability in each orientation:

$p_m^f\left(c\right|T)=p_m\left(c\right|T_m)*W_{\mathrm{init}}+W_{\mathrm{init}}+p_{\mathrm{weig}}\left(c\right|T_m)*W_{\mathrm{weig}}$

Wherein T is total characteristic vector space;

Final Classification and Identification result then for each class is:

$P\left(c\right|T)=\underset{m=1}{\overset{M}{\mathrm{\Π}}}{p}_m^f\left(c\right|T)$

Wherein M is orientation quantity;

For the feature being input to support vector machine classifier, the probability of which class greatly then expresses support for vector machine classifier and the feature of current input is identified as this class.