CN106203452A - Corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis - Google Patents

Abstract

The present invention relates to a kind of corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis, it is characterized in that, including: a, utilize high spectrum image acquisition system collection and obtain the high spectrum image of corn seed；B, utilize Threshold segmentation to obtain the contour curve of corn seed, ask for spectrum characteristics of mean and the textural characteristics of corn seed under profile, and carry out characteristic binding and obtain the characteristic parameter X of corn seed；C, based on MLDA algorithm characteristic parameter X carried out Feature Conversion (feature extraction and waveband selection)；D, set up forecast model, it is thus achieved that the evaluation result of this feature conversion method.The present invention is by carrying out Feature Conversion to corn seed high spectrum image, and it is capable of the Feature Conversion under the conditions of multiple features, simple to operate, the most effectively, and has the advantages such as higher robustness.

Description

Corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis

Technical field

The present invention relates to a kind of corn seed high spectrum image multiple features conversion method, especially one sentence based on polyteny The corn seed high spectrum image multiple features conversion method do not analyzed.

Background technology

Semen Maydis is the cereal crops and main livestock feed source that Gross World Product is the highest, is also commercial production simultaneously In important source material, thus for subtracting oligospermous mixing, it is ensured that being smoothed out of agricultural production has important value.Mesh Before, the external sort of corn seed is detected the more ripe skills such as main employing visible images detection and infrared image detection Art, but these detection techniques cannot obtain corn seed inside quality effective information.Near-infrared spectrum technique can be the most real-time The inside quality of Non-Destructive Testing seed, and near-infrared spectrum technique provides only the detection of a zonule individual to inspection.Due to The quality of seed spatially there are differences, and there is also certain limitation in this way.Hyper-spectral image technique is a kind of wide The quality of agricultural product lossless detection method of general application, it is possible to obtain the advantages such as the geometric state feature of seed and spectral signature, Seed variety identification is widely used.

The variety ecotype of seed is essentially a pattern recognition problem, and the acquisition of fully excavating of characteristic of division is Identify model accuracy and the guarantee of robustness.Although hyper-spectral image technique can obtain external morphology feature and the inside of seed Chemical Composition Characteristics, but in existing high spectrum image identification, be mostly to utilize single features (such as spectral signature etc.), exist The probability of characteristic of division information dropout.More sufficient characteristic information is conducive to the study of disaggregated model, thus improves model Identification ability.But, too much feature learning also can also exist the risk of over-fitting, high spectrum image wave band Numerous, Simultaneously containing substantial amounts of redundancy between wave band, not only affect the precision of identification, and be unfavorable for that hyper-spectral image technique exists The real-time of seed quality detection application on site, thus be necessary data are carried out dimension-reduction treatment.

Feature extraction and waveband selection are two kinds of main paties of high-spectral data dimensionality reduction.Feature extraction is by primitive character Be converted to one group of feature dimension reduction method with obvious physical significance, statistical significance or core, such as principal component analysis (Principal Component Analysis, PCA), linear discriminant analysis (Linear discriminant analysis, LDA) etc..Wave band Selection is the process found and can represent raw spectroscopic data wave band, reduces and optimizes feature space with this and do not change original Spectral information, as calculated without information variable method of elimination (Uninformative Variable Eliminate, UVE), successive projection Method (Successive Projections Algorithm, SPA) etc..The dimension reduction method of feature extraction and waveband selection by It is widely applied in the Non-Destructive Testing of quality of agricultural product.On the one hand, what these methods utilized is the feature favourable for classification, Comprise more for classification useless feature time, these feature extracting method effects are difficult to ensure that.On the other hand, many due to these Mostly characteristic wave bands system of selection is to carry out waveband selection first with average spectral signature, then the wave band of selection is applied to other Feature, causes other characteristic informations are considered deficiency, it is impossible to guarantee the optimum wave band suitability for other features of selection.? The most fully extract the useful information of high spectrum image, and meet the reality of high spectrum image Dynamic Non-Destruction Measurement application on site simultaneously The space improved is still suffered from time property and high-class precision.

Summary of the invention

It is an object of the invention to overcome the deficiencies in the prior art, it is provided that a kind of jade based on polyteny discriminant analysis Rice seed high spectrum image multiple features conversion method, it is capable of feature choosing under the conditions of corn seed high spectrum image multiple features Select, simple to operate, the most effectively, and there is the advantages such as higher robustness.

The technical scheme provided according to the present invention, described corn seed high spectrum image based on polyteny discriminant analysis is many Feature Conversion method, is characterized in that, including:

A, Num grain corn seed sample to be identified is placed in high spectrum image acquisition system, gathers and obtain this M Grain corn seed sample L high spectrum image under L wave band, L is natural number；

B, utilize Threshold segmentation to obtain the contour curve of corn seed, extract the average spectral signature of corn seed under profile And textural characteristics, common L the average spectral signature obtained under L wave band and L textural characteristics are carried out Feature Fusion, will merge Feature is as the characteristic parameter X of corn seed to be identified；

C, utilize the corn seed characteristic parameter X obtained by step b, based on polyteny discriminant analysis (MLDA) algorithm to jade Rice seed characteristics parameter X carries out feature extraction, and obtains the fusion feature matrix after feature extractionCalculate under each wave band Primitive character contribution rate W to first principal component, set threshold value, selects the wave band corresponding to bigger W value as optimum wave band, The optimum Band Set Θ of output, is applied to the fusion feature matrix after feature extraction by optimum Band Set ΘObtain feature Feature S after conversion；

D, set up forecast model, the feature after conversion is evaluated.

Further, in described step b, it is thus achieved that the concrete steps of the characteristic parameter X of corn seed include:

The average spectral signature of corn seed high spectrum image and textural characteristics are respectively

(wherein, X¹Table Show average spectral signature；X²～X¹⁴Represent textural characteristics, and X²～X⁶It is expressed as single order data statistics in textural characteristics: Energy (energy), entropy (entropy), the degree of bias (skew), standard deviation (standard deviation), moment (moment) are special Levy；X⁷～X¹⁴It is respectively second order data statistics contrast average on 0 °, 45 °, 90 ° and 135 ° of four directions in textural characteristics (contrast mean), contrast standard deviation (contrast standard deviation), dependency average (correlation mean), correlation criterion poor (correlation standard deviation), average energy value (energy mean), energy scale poor (energy standard deviation), homogeneity degree average (homogenous Mean) and homogeneity degree standard deviation (homogenous standard deviation) feature), r_i ^jIt is the of i-th corn seed J feature, Num sample constitutes the matrix of a L × Num under L wave band；First feature is normalized, makes Fusion feature reaches the concordance on the order of magnitude；Again each sample is carried out under pth wave band respectively feature combination ( For jth feature), obtain 14 features Num sample under L wave band Fusion feature matrix X ∈ { K^{14×14×L×Num}, wherein, p=1,2 ..., L.

Further, in described step c, it is thus achieved that the concrete steps of the characteristic parameter converting characteristic of corn seed include:

Best projection vector is calculated according to MLDABy 14 × 14 × L × Num dimension The original fusion eigenspace projection of degree is to J₁×J₂×J₃×J₄Feature space, the eigenmatrix after being projectedWherein, Y_mComputing formula be:

$Y_m=X\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}---\left(1\right);$

The projecting direction U of different n-mode⁽ⁿ⁾Computing formula is:

$\begin{array}{c}{U}^{\left(n\right)}{|}_{n=1}^4=\underset{U^{\left(1\right)},U^{\left(2\right)},U^{\left(3\right)},U^{\left(4\right)}}{\arg \max }\frac{{\Σ}_{c=1}^C{n}_c\left|\right|({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X})\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}|{|}^2}{{\Σ}_{i=1}^{Num}{\Σ}_{c=1}^C\left|\right|(X_i-{\stackrel{\‾}{X}}_c)\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}|{|}^2}\\ =\underset{U^{{\left(n\right)}^T}{U}^{\left(n\right)}=I}{\arg \max }\frac{trace(U^{{\left(n\right)}^T}\·S_{B\left(n\right)}\·U^{{\left(n\right)}^T})}{trace(U^{{\left(n\right)}^T}\·S_{W\left(n\right)}\·U^{{\left(n\right)}^T})}\end{array}---\left(2\right);$

In formula (2), total C class sample, n_cIt is the sample number of c class,It is the meansigma methods of c class samples fusion feature,For the meansigma methods of whole samples fusion features, X_iRepresent i-th sample； WithIt is expressed as discrete side in the class according to the expansion of l-mode tensor Dispersion variance between difference and class,Wherein, U⁽³⁾It is set to the Orthogonal Units of L × L Matrix, U⁽⁴⁾Unit matrix for Num × Num；

According to main constituent and primitive character in linear discriminant analysis (Linear Discriminant Analysis, LDA) Between meet linear relationship t=WXⁱThought, be generalized in MLDA, striked first principal component t₁With original fusion feature X ∈{K^{14×14×L×Num}Also linear relationship, U are met⁽³⁾The characteristic information that can be reflected as under each wave band of the first row to the first main one-tenth Contribution rate W divided, value corresponding for W is the biggest, then the characteristic information under this wave band is the biggest to the contribution rate of main constituent, selects with this Excellent Band Set Θ；Optimum Band Set is applied to the fusion feature matrix after feature selection, obtains the square after Feature Conversion Battle array

Further, in described step c, first MLDA algorithm carries out feature extraction to fusion feature, after obtaining feature extraction Matrix, according to the contribution rate on first principal component of the characteristic information under each wave band as the foundation of waveband selection, will relatively Big wave band corresponding to contribution rate is as optimum wave band, and is applied to the matrix after feature extraction and obtains converting characteristic, specifically wraps Include:

S1, the overall sample average of calculatingC class sample averageInitialize and throw Shadow matrixAccording to formula (2) and formula (3), fusion feature is carried out feature selection；

S2, obtain the t time S by optimization iterative method_W(n) ^(t)And S_B(n) ^(t)L-mode tensor launch be respectively as follows:

$\begin{array}{c}{S_{B\left(n\right)}}^{\left(t\right)}={\Σ}_{c=1}^C{n}_c{({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X})}_{\left(n\right)}\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^{{(t-1)}^T}\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^{(t-1)}\·{({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X})}_{\left(n\right)}^T\\ S_{W\left(n\right)}={\Σ}_{i=1}^{Num}{\Σ}_{c=1}^C(X_i-{\stackrel{\‾}{X}}_c)\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^T\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}\·{({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X})}_{\left(n\right)}^T\end{array}---\left(3\right);$

S3, according to generalized eigenvalue decomposition calculate (S_W(n) ^(t))^-1S_B(n) ^(t)=U Λ U^TCorresponding to middle maximum generalized characteristic value Unit generalized eigenvector

S4, when reach set threshold valueOr during maximum iteration time K, then stop iteration, output Best projection direction vector U⁽ⁿ⁾With the fusion feature matrix after feature selectionOtherwise return step S2；

S5, the primitive character calculated under each wave band contribution rate W to first principal component, set threshold value, select bigger W value Corresponding wave band is as optimum wave band, output optimum Band Set Θ；

S6, optimum Band Set Θ is applied to the eigenmatrix after feature selectionObtain converting characteristic S.

Corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis of the present invention, by right Corn seed high spectrum image carries out Feature Conversion, and it is capable of Feature Conversion under the conditions of multiple features, simple to operate, quickly has Effect, and there is the advantages such as higher robustness.

Accompanying drawing explanation

Fig. 1 is image segmentation and the characteristic extraction procedure schematic diagram of corn seed high spectrum image.

Fig. 2 is corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis of the present invention Schematic flow sheet.

Detailed description of the invention

Below in conjunction with concrete accompanying drawing, the invention will be further described.

Corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis of the present invention, including:

A, as shown in Figure 1: first select wave band that corn seed profile image the most clearly to be identified is corresponding ( At 700.1nm), utilize adaptive threshold fuzziness method, it is thus achieved that the contour curve of the corn seed to be identified under this wave band.This is taken turns On wide CURVE PROJECTION to L wave band, extract average spectral signature and the textural characteristics of corn seed contour images under L wave band (first order statistic and second-order statistic), L is natural number；

B, as shown in Figure 2: be normalized average spectral signature and textural characteristics, eliminates between feature in quantity Difference in level；

C, the average spectral signature of corn seed high spectrum image and textural characteristics are respectively

(wherein, X¹Table Show average spectral signature；X²～X¹⁴Represent textural characteristics, and X²～X⁶It is expressed as single order data statistics in textural characteristics: Energy (energy), entropy (entropy), the degree of bias (skew), standard deviation (standard deviation), moment (moment) are special Levy；X⁷～X¹⁴It is respectively second order data statistics contrast average on 0 °, 45 °, 90 ° and 135 ° of four directions in textural characteristics (contrast mean), contrast standard deviation (contrast standard deviation), dependency average (correlation mean), correlation criterion poor (correlation standard deviation), average energy value (energy mean), energy scale poor (energy standard deviation), homogeneity degree average (homogenous Mean) and homogeneity degree standard deviation (homogenous standard deviation) feature),It is the of i-th corn seed J feature, Num sample constitutes the matrix of a L × Num under L wave band；First feature is normalized, makes Fusion feature reaches the concordance on the order of magnitude；Again each sample is carried out under pth wave band respectively feature combination ( For jth feature), obtain 14 features Num sample under L wave band Fusion feature matrix X ∈ { K^{14×14×L×Num}, wherein, p=1,2 ..., L；

D, according to MLDA (multilinear discriminant analysis, Multiple Linear Discrminant Analysis) find There is the thought of the projecting direction of divergence (generalized Rayleigh quaotient) in maximum class scatter and infima species and calculate best projection vector Good projection vectorThe original fusion feature space of 14 × 14 × L × Num dimension is thrown Shadow is to J₁×J₂×J₃×J₄Feature space, the eigenmatrix after being projectedIts In, Y_mComputing formula be:

$Y_m=X\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}---\left(1\right);$

Assume total C class sample, n_cIt is the sample number of c class,It is the meansigma methods of c class samples fusion feature,For The all meansigma methods of samples fusion feature, X_iRepresent i-th sample.The projecting direction U of different n-mode⁽ⁿ⁾Computing formula is:

$\begin{array}{c}{U}^{\left(n\right)}{|}_{n=1}^4=\underset{U^{\left(1\right)},U^{\left(2\right)},U^{\left(3\right)},U^{\left(4\right)}}{\arg \max }\frac{{\Σ}_{c=1}^C{n}_c\left|\right|({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X})\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}|{|}^2}{{\Σ}_{i=1}^{Num}{\Σ}_{c=1}^C\left|\right|(X_i-{\stackrel{\‾}{X}}_c)\×{U_1}^{{\left(1\right)}^T}\×{U_2}^{{\left(2\right)}^T}\×{U_3}^{{\left(3\right)}^T}\×{U_4}^{{\left(4\right)}^T}|{|}^2}\\ =\underset{U^{{\left(n\right)}^T}{U}^{\left(n\right)}=I}{\arg \max }\frac{trace(U^{{\left(n\right)}^T}\·S_{B\left(n\right)}\·U^{{\left(n\right)}^T})}{trace(U^{{\left(n\right)}^T}\·S_{W\left(n\right)}\·U^{{\left(n\right)}^T})}\end{array}---\left(2\right);$

In formula (2),With It is expressed as dispersion variance between dispersion variance and class in the class launched according to l-mode tensor,

Best projection vector U^(l)Calculating matrix S can be converted into_W(l) ^-1S_B(l)Ask for the feature corresponding to eigenvalue of maximum The problem of vector, calculates U by k rank optimization iterative method^(l).Wherein, U⁽³⁾It is set to the orthogonal matrices of L × L, U⁽⁴⁾ For the unit matrix of Num × Num, thus realize the dimensionality reduction of feature is not changed the number of wave band.

E, according to main constituent in linear discriminant analysis (Linear Discriminant Analysis, LDA) and original spy Linear relationship t=WX is met between levyingⁱThought, be generalized in MLDA, striked first principal component t₁With original fusion feature X∈{K^{14×14×L×Num}Also linear relationship, U are met⁽³⁾The characteristic information that can be reflected as under each wave band of the first row main to first Contribution rate W of composition, value corresponding for W is the biggest, then the characteristic information under this wave band is the biggest to the contribution rate of main constituent, selects with this Optimum Band Set Θ；

F, optimum Band Set is applied to the fusion feature matrix after feature selection, obtains the matrix after Feature Conversion

G, to after Feature Conversion feature combine set up Classification and Identification model, be evaluated according to nicety of grading.

Claims (4)

1. a corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis, is characterized in that, including:

A, Num grain corn seed sample to be identified is placed in high spectrum image acquisition system, gathers and obtain this M grain beautiful Rice seed specimen L high spectrum image under L wave band, L is natural number；

B, utilize Threshold segmentation to obtain the contour curve of corn seed, extract average spectral signature and the stricture of vagina of corn seed under profile Reason feature, carries out Feature Fusion, by fusion feature by common L the average spectral signature and L the textural characteristics that obtain under L wave band Characteristic parameter X as corn seed to be identified；

C, utilize the corn seed characteristic parameter X obtained by step b, based on polyteny discriminant analysis (MLDA) algorithm to maize seed Subcharacter parameter X carries out feature extraction, and obtains the fusion feature matrix after feature extractionCalculate under each wave band is former Beginning feature contribution rate W to first principal component, sets threshold value, selects the wave band corresponding to bigger W value as optimum wave band, output Optimum Band Set Θ, is applied to the fusion feature matrix after feature extraction by optimum Band Set ΘAfter being changed Feature S；

D, set up forecast model, the feature after conversion is evaluated.

2. corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis as claimed in claim 1, It is characterized in that: in described step b, it is thus achieved that the concrete steps of the characteristic parameter X of corn seed include:

The average spectral signature of corn seed high spectrum image and textural characteristics are

Wherein, X¹Represent Average spectral signature；X²～X¹⁴Represent textural characteristics, r_i ^jBeing the jth feature of i-th corn seed, Num sample is at L wave band The matrix of lower one L × Num of composition；First being normalized feature, make that fusion feature reaches on the order of magnitude is consistent Property；Again each sample is carried out feature combination under pth wave band respectively For jth feature), obtain 14 features fusion feature matrix X ∈ { K of Num sample under L wave band^{14×14×L×Num}, Wherein, p=1,2 ..., L.

3. corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis as claimed in claim 1, It is characterized in that: in described step c, it is thus achieved that the concrete steps of the characteristic parameter converting characteristic of corn seed include:

Best projection vector is calculated according to MLDABy 14 × 14 × L × Num dimension Original fusion eigenspace projection is to J after feature extraction₁×J₂×J₃×J₄The feature space of dimension, the feature after being projected MatrixWherein, Y_mComputing formula be:

The projecting direction U of different n-mode⁽ⁿ⁾Computing formula is:

In formula (2), total C class sample, n_cIt is the sample number of c class,It is the meansigma methods of c class samples fusion feature,For The all meansigma methods of samples fusion feature, X_iRepresent i-th sample；WithBe expressed as according to l-mode tensor launch class in discrete Dispersion variance between variance and class, andWherein, U⁽³⁾It is set to L × L's Orthogonal matrices, U⁽⁴⁾Unit matrix for Num × Num；

According to striked first principal component t₁With original fusion feature X ∈ { K^{14×14×L×Num}Meet linear relation, obtain every The characteristic information under individual wave band contribution rate W to first principal component, value corresponding for W is the biggest, then the characteristic information pair under this wave band The contribution rate of main constituent is the biggest, selects optimum Band Set Θ with this；Melting after optimum Band Set is applied to feature selection Close eigenmatrix, obtain the matrix after Feature Conversion

4. corn seed high spectrum image multiple features conversion method based on polyteny discriminant analysis as claimed in claim 3, It is characterized in that: in described step c, first MLDA algorithm carries out feature extraction to fusion feature, obtain the matrix after feature extraction, According to the contribution rate on first principal component of the characteristic information under each wave band as the foundation of waveband selection, will bigger contribution rate Corresponding wave band is as optimum wave band, and is applied to the matrix after feature extraction and obtains converting characteristic, specifically includes:

S1, the overall sample average of calculatingC class sample averageInitial projection square Battle arrayAccording to formula (2) and formula (3), fusion feature is carried out feature selection；

S2, obtain the t time S by optimization iterative method_W(n) ^(t)And S_B(n) ^(t)L-mode tensor launch be respectively as follows:

$\begin{array}{c}{S_{B\left(n\right)}}^{\left(t\right)}={\mathrm{\Σ}}_{c=1}^C{n}_c{\left({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X}\right)}_{\left(n\right)}\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^{{\left(t-1\right)}^T}\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^{\left(t-1\right)}\·{\left({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X}\right)}_{\left(n\right)}^T\\ S_{W\left(n\right)}={\mathrm{\Σ}}_{i=1}^{Num}{\mathrm{\Σ}}_{c=1}^C\left(X_i-{\stackrel{\‾}{X}}_c\right)\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}^T\·{\mathrm{\Φ}}_{\stackrel{\‾}{n}}\·{\left({\stackrel{\‾}{X}}_c-\stackrel{\‾}{X}\right)}_{\left(n\right)}^T\end{array}---\left(3\right);$

S3, according to generalized eigenvalue decomposition calculateUnit corresponding to middle maximum generalized characteristic value Generalized eigenvector

S4, when reach set threshold valueOr during maximum iteration time K, then stopping iteration, output is optimal Projecting direction vector U⁽ⁿ⁾With the fusion feature matrix after feature selectionOtherwise return step S2；

S5, the primitive character calculated under each wave band contribution rate W to first principal component, set threshold value, selects bigger W value institute right The wave band answered is as optimum wave band, output optimum Band Set Θ；

S6, optimum Band Set Θ is applied to the eigenmatrix after feature selectionObtain converting characteristic S.