CN105224960A - Based on the corn seed classification hyperspectral imagery model of cognition update method of clustering algorithm - Google Patents

Abstract

The invention discloses the corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm, comprise: based on the N number of high spectrum image of all corn seeds to be identified of collection under N number of wave band, calculate the spectrum characteristics of mean in area-of-interest, as characteristic parameter, input data successively, calculate the matching of corn seed to be identified and detection model; Judge matching, if mate unsuccessful, then by chemical analysis test, obtain the classification of corn seed to be identified, utilize the seed characteristics parameter to be identified and true classification thereof that obtain, upgrade training set; After the matching completing all corn seeds to be identified judges, utilize new training set to set up new least square method supporting vector machine detection model f _lS-SVM, use f _lS-SVMidentify the sample set to be identified after renewal.The invention provides a kind of corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm, the renewal of classification hyperspectral imagery model can be realized, effective, and reliability is high.

Description

Based on the corn seed classification hyperspectral imagery model of cognition update method of clustering algorithm

Technical field

The present invention relates to a kind of update method of corn seed classification hyperspectral imagery model, especially a kind of corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm.

Background technology

In recent years, along with the widespread use of seed hybridization technique, the kind of seed gets more and more, and between class, similarity is increasing, mixes phenomenon and is on the rise, and it is more and more difficult that this causes kind to be distinguished, and the purity problem of seed also more and more receives the concern of people.The important parameter of the purity reflection seed quality of seed is the Main Basis of evaluation seed grade.Traditional Seed purity test method also exists qualification time length, Personnel Dependence has the shortcomings such as destructive by force, to seed, is difficult to be promoted in actual applications.In order to improve the rapidity of Seed inspection, this kind of Dynamic Non-Destruction Measurement of machine vision technique, near-infrared spectral analysis technology and hyper-spectral image technique is developed.Machine vision technique mainly utilizes seed external morphology information, and near-infrared spectrum technique is the chemical feature information utilizing all organic molecule hydric groups in seed.The single traits information that machine vision technique or near-infrared spectral analysis technology all can only obtain seed, for some seed, some trait information difference between different cultivars is also not obvious, iff depending on these single traits information, the accuracy of Seed purity test can be reduced.Compare, hyper-spectral image technique can contain measurand external morphology feature, all information of inner structural features and Chemical Composition Characteristics by providing package, and these information are that the accurate detection of seed purity provides information assurance fully reliably.Thus hyper-spectral image technique is widely used in nondestructive measuring method of the farm product.

Utilizing hyper-spectral image technique to carry out Seed purity test is essentially pattern classification problem, and its nicety of grading is subject to the adequacy of characteristic of division information, reliability and availability influence.Therefore model modification has great importance for the robustness and generalization ability improving model.The model update method that scholar in the past proposes selects great amount of samples to carry out Renewal model, although this mode can reach higher precision, wastes time and energy.Therefore, find a kind of renewal that the model modification strategy of representative sample can be selected to realize seed classification model and just seem particularly important.

Summary of the invention

The object of the invention is the shortcoming overcoming above technology, provide a kind of corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm, it can realize classification hyperspectral imagery model modification, time saving and energy saving, and fast effectively, and reliability is high.

Technical scheme provided by the invention, the corn seed classification hyperspectral imagery model of cognition update method of described clustering algorithm, concrete step comprises: a, be placed in high spectrum image acquisition system by corn seed sample to be identified, gathers and obtains the N number of high spectrum image of all corn seed samples under N number of wave band; B, the spectrum characteristics of mean calculated in area-of-interest, using common N number of spectrum characteristics of mean of obtaining under N number of wave band of all corn seeds characteristic parameter matrix Y as corn seed to be identified; C, the corn seed characteristic parameter matrix Y obtained by step b, input data successively, calculate the matching of corn seed to be identified and training sample; D, judge the matching of corn seed to be identified and training sample, if mate unsuccessful, then by chemical analysis test, obtain the classification of corn seed to be identified, utilize the seed characteristics parameter to be identified and true classification thereof that obtain, upgrade training set and test set; E, repetition step (c-d), the matching completing all corn seeds to be identified judges, utilizes new training set to set up new least square method supporting vector machine detection model f _lS-SVM, identify the sample set to be identified after renewal with this model.

Further, in stepb, the characteristic parameter obtaining corn seed comprises:

First select the corn seed profile wave band that image is corresponding the most clearly (at 782.59nm place) to be identified, utilize adaptive threshold fuzziness method, obtain the contour curve of the corn seed to be identified under this wave band.This contour curve is projected on N number of wave band, extracts the characteristic of division parameter of the spectrum average of N number of wave band in this contour curve as corn seed.

In step c, judge that the operation of the matching of corn seed to be identified and training sample comprises:

Obtain according to step a, b the characteristic parameter that h class is total to l corn seed, and utilize chemical analysis test, obtain its class label, corn seed identical for class label is configured to a son training set, obtains h subclass altogether:

Make D=[D ₁..., D _i..., D _h] be the training set of h class sample, for there being n _ithe subclass of the i-th class sample of individual sample, T=D-D _ifor rejecting the subclass of the i-th class sample.For given training sample $x_i^j\∈D_i,x_i^k\∈D_i,$ With there is inter-object distance and between class distance

$S_i^{j,k}=\left|\right|x_i^j-x_i^k\left|\right|,x_i^j,x_i^k\∈D_i,j\≠k---\left(1\right)$

$P_i^{k,s}=\left|\right|x_i^k-x_i^s\left|\right|,x_i^k\∈D_i,x_i^s\∈T_i---\left(2\right)$

Wherein: 1≤i≤h, || || represent 2 norms.

Range averaging in the K infima species of calculating i-th class average with K infima species spacing after, calculate a discriminant criterion

${\mathrm{Thr}}_i^k={\mathrm{DI}}_i^k/{\mathrm{DW}}_i^k,i=1,...,h,k=1,...,n_i---\left(3\right)$

Define the threshold value Th of otherness between a reflection inhomogeneity sample _i:

${\mathrm{Th}}_i=\underset{k=1:n_i}{min}\left({\mathrm{Yhr}}_i^k\right)\×\mathrm{\β}---\left(4\right)$

Wherein β is coefficient of relaxation.To h class training sample, obtain h threshold value (Th ₁..., Th _i..., Th _h).

In step e, set up the least square method supporting vector machine detection model f of corn seed _lS-SVM, specifically comprise:

Utilize the h class obtained in steps d to be total to characteristic parameter and the class label of l corn seed, by one-against-rest, build h sub-LS-SVM disaggregated model.The output expression formula z of its kth (1≤k≤h) individual submodel ^k(Y) be:

$z^k\left(Y\right)=sgn\[\underset{j=1}{\overset{l}{\Σ}}{\mathrm{\β}}_{j}K(Y,Y_j)+b\]---\left(5\right)$

Wherein, sgn () is sign function, and Y is the characteristic parameter of corn seed sample to be identified, Y _jfor the characteristic parameter of training set corn seed sample, β _jundetermined coefficient is with b.When when being more than or equal to zero, sgn () value is 1, represents that sample Y to be identified belongs to kth class, otherwise does not belong to.

β in formula (5) _jprovided by following form with the value of b,

$\left[\begin{array}{cc}0& -D\\ D^T& \mathrm{\Ω}+{\mathrm{\γ}}^{-1}E\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\θ}\end{array}\right]=\left[\begin{array}{c}0\\ I\end{array}\right]---\left(6\right)$

Wherein, θ=[β ₁β _jβ _l] and b be parameter to be solved; D=[d ₁d _jd _l], as training sample Y _jwhen belonging to kth class, d _j=1, otherwise d _j=-1.γ is for being penalty coefficient, and E is the unit square formation of l × l, and I is complete 1 column vector of l × 1; Ω is the square formation of l × l, the element Ω of its i-th row jth row _ij=d _id _jk (Y _i, Y _j), kernel function choose:

$K(Y_i^k,Y_j^k)=\exp (-\frac{\left|\right|Y_i^k-Y_j^k|{|}^2}{2{\mathrm{\σ}}^2})---\left(7\right)$

Gaussian kernel function K (Y, Y _i) core width parameter and penalty coefficient γ obtained by trellis search method.

Further, based on the corn seed classification hyperspectral imagery model of cognition update method of clustering algorithm, it is characterized in that the matching judging corn seed to be identified and training sample in steps d, specifically comprise:

Get a sample x to be identified, suppose that it belongs to the i-th class, then by formula (1), (2) and (3) calculate its discriminant criterion Thr _iif, Thr _i>=Th _i, then this sample to be identified does not mate with the i-th supposed class sample, now this sample to be identified is classified as more new samples, and utilizes chemical analysis test, obtain its class label, utilizes the seed characteristics parameter to be identified and true classification thereof that obtain, upgrades training set; Otherwise, do not upgrade training set.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm provided by the invention.

Specific implementation method

Below in conjunction with concrete accompanying drawing and preferred embodiment, the present invention will be further described.

Fig. 1 is the process flow diagram of the corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm provided by the invention, as shown in Figure 1, wherein, model modification concrete steps based on least square method supporting vector machine comprise: a, be placed in high spectrum image acquisition system by corn seed sample to be identified, gather and obtain the N number of high spectrum image of all corn seed samples under N number of wave band; B, the spectrum characteristics of mean calculated in area-of-interest, using common N number of spectrum characteristics of mean of obtaining under N number of wave band of all corn seeds characteristic parameter matrix Y as corn seed to be identified; C, the corn seed characteristic parameter matrix Y obtained by step b, input data successively, calculate the matching of corn seed to be identified and training sample; D, judge the matching of corn seed to be identified and training sample, if mate unsuccessful, then by chemical analysis test, obtain the classification of corn seed to be identified, utilize the seed characteristics parameter to be identified and true classification thereof that obtain, upgrade training set and test set; E, repetition step (c-d), the matching completing all corn seeds to be identified judges, utilizes new training set to set up new least square method supporting vector machine detection model f _lS-SVM, identify the sample set to be identified after renewal with this model.

Further, in stepb, the characteristic parameter obtaining corn seed comprises:

First select the corn seed profile wave band that image is corresponding the most clearly (at 782.59nm place) to be identified, utilize adaptive threshold fuzziness method, obtain the contour curve of the corn seed to be identified under this wave band.This contour curve is projected on N number of wave band, extracts the characteristic of division parameter of the spectrum average of N number of wave band in this contour curve as corn seed.

The operation calculating the matching of corn seed to be identified and training sample in step c comprises:

Obtain according to step a, b the characteristic parameter that h class is total to l corn seed, and utilize chemical analysis test, obtain its class label, corn seed identical for class label is configured to a son training set, obtains h subclass altogether:

Make D=[D ₁..., D _i..., D _h] be the training set of h class sample, for there being n _ithe subclass of the i-th class sample of individual sample, T=D-D _ifor rejecting the subclass of the i-th class sample.For given training sample $x_i^j\∈D_i,x_i^k\∈D_i,$ With $x_i^s\∈T_i,$ There is inter-object distance and between class distance

$S_i^{j,k}=\left|\right|x_i^j-x_i^k\left|\right|,x_i^j,x_i^k\∈D_i,j\≠k---\left(1\right)$

$P_i^{k,s}=\left|\right|x_i^k-x_i^s\left|\right|,x_i^k\∈D_i,x_i^s\∈T_i---\left(2\right)$

Wherein: 1≤i≤h, || || represent 2 norms.

Range averaging in the K infima species of calculating i-th class average with K infima species spacing after, calculate a discriminant criterion

${\mathrm{Thr}}_i^k={\mathrm{DI}}_i^k/{\mathrm{DW}}_i^k,i=1,...,h,k=1,...,n_i---\left(3\right)$

Define the threshold value Th of otherness between a reflection inhomogeneity sample _i:

${\mathrm{Th}}_i=\underset{k=1:n_i}{min}\left({\mathrm{Thr}}_i^k\right)\×\mathrm{\β}---\left(4\right)$

Wherein β is coefficient of relaxation.To h class training sample, obtain h threshold value (Th ₁..., Th _i..., Th _h).

In step e, set up the least square method supporting vector machine detection model f of corn seed _lS-SVM, specifically comprise:

Utilize the h class obtained in steps d to be total to characteristic parameter and the class label of l corn seed, by one-against-rest, build h sub-LS-SVM disaggregated model.The output expression formula z of its kth (1≤k≤h) individual submodel ^k(Y) be:

$z^k\left(Y\right)=sgn\[\underset{j=1}{\overset{l}{\Σ}}{\mathrm{\β}}_{j}K(Y,Y_j)+b\]---\left(5\right)$

Wherein, sgn () is sign function, and Y is the characteristic parameter of corn seed sample to be identified, Y _jfor the characteristic parameter of training set corn seed sample, β _jundetermined coefficient is with b.When when being more than or equal to zero, sgn () value is 1, represents that sample Y to be identified belongs to kth class, otherwise does not belong to.

β in formula (5) _jprovided by following form with the value of b,

$\left[\begin{array}{cc}0& -D\\ D^T& \mathrm{\Ω}+{\mathrm{\γ}}^{-1}E\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\θ}\end{array}\right]=\left[\begin{array}{c}0\\ I\end{array}\right]---\left(6\right)$

Wherein, θ=[β ₁β _jβ _l] and b be parameter to be solved; D=[d ₁d _jd _l], as training sample Y _jwhen belonging to kth class, d _j=1, otherwise d _j=-1.γ is for being penalty coefficient, and E is the unit square formation of l × l, and I is complete 1 column vector of l × 1; Ω is the square formation of l × l, the element Ω of its i-th row jth row _ij=d _id _jk (Y _i, Y _j), kernel function choose:

$K(Y_i^k,Y_j^k)=\exp (-\frac{\left|\right|Y_i^k-Y_j^k|{|}^2}{2{\mathrm{\σ}}^2})---\left(7\right)$

Gaussian kernel function K (Y, Y _i) core width parameter and penalty coefficient γ obtained by trellis search method.

Further, based on the corn seed classification hyperspectral imagery model of cognition update method of clustering algorithm, it is characterized in that the matching judging corn seed to be identified and training sample in steps d, specifically comprise:

Get a sample x to be identified, suppose that it belongs to the i-th class, then by formula (1), (2) and (3) calculate its discriminant criterion Thr _iif, Thr _i>=Th _i, then this sample to be identified does not mate with the i-th supposed class sample, now this sample to be identified is classified as more new samples, and utilizes chemical analysis test, obtain its class label, utilizes the seed characteristics parameter to be identified and true classification thereof that obtain, upgrades training set; Otherwise, do not upgrade training set.

Advantage of the present invention: provide a kind of corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm, it can realize classification hyperspectral imagery model modification, time saving and energy saving, and fast effectively, and reliability is high.

Claims (2)

1., based on the corn seed classification hyperspectral imagery model of cognition update method of clustering algorithm, it is characterized in that, comprising:

A, corn seed sample to be identified to be placed in high spectrum image acquisition system, to gather and obtain the N number of high spectrum image of all corn seed samples under N number of wave band;

B, the spectrum characteristics of mean calculated in area-of-interest, using common N number of spectrum characteristics of mean of obtaining under N number of wave band of all corn seeds characteristic parameter matrix Y as corn seed to be identified;

C, the corn seed characteristic parameter matrix Y obtained by step b, input data successively, calculate the matching of corn seed to be identified and training sample;

D, judge the matching of corn seed to be identified and training sample, if mate unsuccessful, then by chemical analysis test, obtain the classification of corn seed to be identified, utilize the seed characteristics parameter to be identified and true classification thereof that obtain, upgrade training set and test set;

E, repetition step (c-d), the matching completing all corn seeds to be identified judges, utilizes new training set to set up new least square method supporting vector machine detection model f _lS-SVM, identify the sample set to be identified after renewal with this model.

In stepb, the characteristic parameter obtaining corn seed comprises:

First select the corn seed profile wave band that image is corresponding the most clearly (at 782.59nm place) to be identified, utilize adaptive threshold fuzziness method, obtain the contour curve of the corn seed to be identified under this wave band.This contour curve is projected on N number of wave band, extracts the characteristic of division parameter of the spectrum average of N number of wave band in this contour curve as corn seed.

The operation calculating the matching of corn seed to be identified and training sample in step c comprises:

Obtain according to step a, b the characteristic parameter that h class is total to l corn seed, and utilize chemical analysis test, obtain its class label, corn seed identical for class label is configured to a son training set, obtains h subclass altogether:

Make D=[D ₁..., D _i..., D _h] be the training set of h class sample, for there being n _ithe subclass of the i-th class sample of individual sample, T=D-D _ifor rejecting the subclass of the i-th class sample.For given training sample with there is inter-object distance and between class distance

$S_i^{j,k}=\left|\right|x_i^j-x_i^k\left|\right|,x_i^j,x_i^k\∈D_i,j\≠k---\left(1\right)$

$P_i^{k,s}=\left|\right|x_i^k-x_i^s\left|\right|,x_i^k\∈D_i,x_i^s\∈T_i---\left(2\right)$

Wherein: 1≤i≤h, || || represent 2 norms.

Range averaging in the K infima species of calculating i-th class average with K infima species spacing after, calculate a discriminant criterion

${\mathrm{Thr}}_i^k={\mathrm{DI}}_i^k/{\mathrm{DW}}_i^k,i=1,...,h,k=1,...,n_i---\left(3\right)$

Define the threshold value Th of otherness between a reflection inhomogeneity sample _i:

${\mathrm{Th}}_i=\underset{k=1:n_i}{min}\left({\mathrm{Yhr}}_i^k\right)\×\mathrm{\β}---\left(4\right)$

Wherein β is coefficient of relaxation.To h class training sample, obtain h threshold value (Th ₁..., Th _i..., Th _h).

In step e, set up the least square method supporting vector machine detection model f of corn seed _lS-SVM, specifically comprise:

Utilize the h class obtained in steps d to be total to characteristic parameter and the class label of l corn seed, by one-against-rest, build h sub-LS-SVM disaggregated model.The output expression formula z of its kth (1≤k≤h) individual submodel ^k(Y) be:

$z^k\left(Y\right)=sgn\[\underset{j=1}{\overset{l}{\Σ}}{\mathrm{\β}}_{j}K(Y,Y_j)+b\]---\left(5\right)$ Wherein, sgn () is sign function, and Y is the characteristic parameter of corn seed sample to be identified, Y _jfor the characteristic parameter of training set corn seed sample, β _jundetermined coefficient is with b.When when being more than or equal to zero, sgn () value is 1, represents that sample Y to be identified belongs to kth class, otherwise does not belong to.

β in formula (5) _jprovided by following form with the value of b,

$\left[\begin{array}{cc}0& -D\\ D^T& \mathrm{\Ω}+{\mathrm{\γ}}^{-1}E\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\θ}\end{array}\right]=\left[\begin{array}{c}0\\ I\end{array}\right]---\left(6\right)$

Wherein, θ=[β ₁β _jβ _l] and b be parameter to be solved; D=[d ₁d _jd _l], as training sample Y _jwhen belonging to kth class, d _j=1, otherwise d _j=-1.γ is for being penalty coefficient, and E is the unit square formation of l × l, and I is complete 1 column vector of l × 1; Ω is the square formation of l × l, the element Ω of its i-th row jth row _ij=d _id _jk (Y _i, Y _j), kernel function choose:

$K(Y_i^k,Y_j^k)=\exp (-\frac{\left|\right|Y_i^k-Y_j^k|{|}^2}{2{\mathrm{\σ}}^2})---\left(7\right)$

Gaussian kernel function K (Y, Y _i) core width parameter and penalty coefficient γ obtained by trellis search method.

2. the corn seed classification hyperspectral imagery model of cognition update method based on clustering algorithm according to claim 1, is characterized in that the matching judging corn seed to be identified and training sample in steps d judges, specifically comprises:

Get a sample x to be identified, suppose that it belongs to the i-th class, then by formula (1), (2) and (3) calculate its discriminant criterion Thr _iif, Thr _i>=Th _i, then this sample to be identified does not mate with the i-th supposed class sample, now this sample to be identified is classified as more new samples, and utilizes chemical analysis test, obtain its class label, utilizes the seed characteristics parameter to be identified and true classification thereof that obtain, upgrades training set; Otherwise, do not upgrade training set.