CN105224960A

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

Info

Publication number: CN105224960A
Application number: CN201510741678.XA
Authority: CN
Inventors: 黄敏; 何楚婕; 朱启兵
Original assignee: Jiangnan University
Current assignee: Jiangnan University
Priority date: 2015-11-04
Filing date: 2015-11-04
Publication date: 2016-01-06
Anticipated expiration: 2035-11-04
Also published as: CN105224960B

Abstract

The invention discloses a method for updating a classification and recognition model of a corn seed hyperspectral image based on a clustering algorithm, comprising: calculating the spectral mean value feature in an area of interest based on collecting N hyperspectral images of all corn seeds to be identified under N bands , as characteristic parameters, input data in turn, calculate the matching between the corn seeds to be identified and the detection model; judge the matching, if the matching is unsuccessful, obtain the category of the corn seeds to be identified through chemical analysis tests, and use the obtained seeds to be identified Feature parameters and their true categories, update the training set; after completing the matching judgment of all corn seeds to be identified, use the new training set to establish a new least squares support vector machine detection model f _LS-SVM , use f _{LS- SVM} to identify the updated sample set to be identified. The invention provides a method for updating a hyperspectral image classification and recognition model of corn seeds based on a clustering algorithm, which can realize the update of the hyperspectral image classification model with good effect and high reliability.

Description

A clustering algorithm-based update method for corn seed hyperspectral image classification and recognition model

技术领域technical field

本发明涉及一种玉米种子高光谱图像分类模型的更新方法，尤其是一种基于聚类算法的玉米种子高光谱图像分类识别模型更新方法。The invention relates to a method for updating a corn seed hyperspectral image classification model, in particular to a method for updating a corn seed hyperspectral image classification recognition model based on a clustering algorithm.

背景技术Background technique

近年来，随着种子杂交技术的广泛应用，种子的品种越来越多，类间相似性越来越大，混杂现象日趋严重，这导致品种区分越来越困难，种子的纯度问题也越来越受到人们的关注。种子的纯度反映种子质量的重要参数，是评定种子等级的主要依据。传统的种子纯度检测方法存在着鉴定时间长、人员依赖性强、对种子具有破坏性等缺点，难以在实际应用中得到推广。为了提高种子检测的快速性，机器视觉技术、近红外光谱分析技术以及高光谱图像技术这类的无损检测技术得到发展。机器视觉技术主要是利用种子外在形态学信息，而近红外光谱技术是利用种子中所有有机分子含氢基团的化学特征信息。无论是机器视觉技术还是近红外光谱分析技术都只能获得种子的单一性状信息，对于某些种子来说，不同品种间的某些性状信息差异并不明显，如果仅仅依赖于这些单一性状信息，会降低种子纯度检测的准确性。相比较，高光谱图像技术可以提供包含被测对象外在形态学特征，内部结构特征和化学成分特征的所有信息，这些信息为种子纯度的准确检测提供了充分可靠的信息保证。因而高光谱图像技术在农产品无损检测中得到广泛的应用。In recent years, with the wide application of seed hybridization technology, there are more and more varieties of seeds, the similarity between classes is getting bigger and bigger, and the phenomenon of mixing is becoming more and more serious, which makes it more and more difficult to distinguish varieties and the purity of seeds. more people's attention. Seed purity reflects an important parameter of seed quality and is the main basis for evaluating seed grades. The traditional method of testing the purity of seeds has the disadvantages of long identification time, strong dependence on personnel, and destructive to seeds, so it is difficult to be popularized in practical applications. In order to improve the rapidity of seed detection, non-destructive testing technologies such as machine vision technology, near-infrared spectral analysis technology and hyperspectral image technology have been developed. Machine vision technology mainly uses the external morphological information of seeds, while near-infrared spectroscopy technology uses the chemical feature information of hydrogen-containing groups of all organic molecules in seeds. Both machine vision technology and near-infrared spectral analysis technology can only obtain single trait information of seeds. For some seeds, the difference of certain trait information between different varieties is not obvious. If only relying on these single trait information, It will reduce the accuracy of seed purity detection. In comparison, hyperspectral image technology can provide all information including the external morphological characteristics, internal structural characteristics and chemical composition characteristics of the measured object, which provides sufficient and reliable information guarantee for the accurate detection of seed purity. Therefore, hyperspectral image technology has been widely used in non-destructive testing of agricultural products.

利用高光谱图像技术进行种子纯度检测从本质上来说是个模式分类问题，其分类精度受到分类特征信息的充分性、可靠性和有效性影响。因此模型更新对于提高模型的鲁棒性和泛化能力具有重要的意义。以往的学者提出的模型更新方法选择大量样本来更新模型，这种方式虽然能达到比较高的精度，但是费时费力。因此，寻找一种能选择代表性样本的模型更新策略实现种子分类模型的更新就显得尤为重要。Seed purity detection using hyperspectral image technology is essentially a pattern classification problem, and its classification accuracy is affected by the sufficiency, reliability and effectiveness of classification feature information. Therefore, model updating is of great significance to improve the robustness and generalization ability of the model. The model update method proposed by previous scholars selects a large number of samples to update the model. Although this method can achieve relatively high accuracy, it is time-consuming and laborious. Therefore, it is particularly important to find a model update strategy that can select representative samples to update the seed classification model.

发明内容Contents of the invention

本发明的目的是克服以上技术的缺点，提供一种基于聚类算法的玉米种子高光谱图像分类识别模型更新方法，其能够实现高光谱图像分类模型更新，省时省力，快速有效，且可靠性高。The purpose of the present invention is to overcome the shortcomings of the above technologies, and provide a method for updating the classification and recognition model of corn seed hyperspectral images based on clustering algorithms, which can realize the update of hyperspectral image classification models, save time and effort, be fast and effective, and have high reliability. high.

本发明提供的技术方案，所述的聚类算法的玉米种子高光谱图像分类识别模型更新方法，具体的步骤包括：a、将待识别的玉米种子样本放置在高光谱图像采集系统中，采集并获取所有玉米种子样本在N个波段下的N个高光谱图像；b、计算感兴趣区域内的光谱均值特征，将所有玉米种子的N个波段下获得的共N个光谱均值特征作为待识别玉米种子的特征参数矩阵Y；c、将步骤b所得到的玉米种子特征参数矩阵Y，依次输入数据，计算待识别玉米种子与训练样本的匹配性；d、判断待识别玉米种子与训练样本的匹配性，若匹配不成功，则通过化学分析测试，获得待识别玉米种子的类别，利用获得的待识别种子特征参数及其真实类别，更新训练集和测试集；e、重复步骤(c-d),完成所有待识别玉米种子的匹配性判断,利用新的训练集来建立新的最小二乘支持向量机检测模型f_LS-SVM,用该模型来识别更新后的待识别样本集。The technical solution provided by the present invention, the method for updating the corn seed hyperspectral image classification and recognition model of the clustering algorithm, the specific steps include: a, placing the corn seed sample to be identified in the hyperspectral image acquisition system, collecting and Obtain N hyperspectral images of all corn seed samples under N bands; b. Calculate the spectral mean features in the region of interest, and use a total of N spectral mean features obtained under N bands of all corn seeds as the corn to be identified The characteristic parameter matrix Y of the seed; c, the corn seed characteristic parameter matrix Y obtained by step b, input data successively, calculate the matching of the corn seed to be identified and the training sample; d, judge the matching of the corn seed to be identified and the training sample property, if the matching is unsuccessful, then by chemical analysis test, obtain the category of the corn seed to be identified, utilize the obtained seed feature parameter and its true category to update the training set and test set; e, repeat step (cd), complete For the matching judgment of all corn seeds to be identified, the new training set is used to establish a new least squares support vector machine detection model f _LS-SVM , and this model is used to identify the updated sample set to be identified.

进一步的，在步骤b中，获得玉米种子的特征参数包括：Further, in step b, obtaining the characteristic parameters of corn seeds includes:

首先选择待识别玉米种子轮廓最清晰的图像对应的波段(在782.59nm处)，利用自适应阈值分割法，获得该波段下的待识别玉米种子的轮廓曲线。将该轮廓曲线投射到N个波段上，提取N个波段在该轮廓曲线内的光谱均值作为玉米种子的分类特征参数。First, select the band corresponding to the image with the clearest outline of the corn seed to be identified (at 782.59nm), and use the adaptive threshold segmentation method to obtain the outline curve of the corn seed to be identified under this band. The contour curve is projected onto N bands, and the spectral mean value of the N bands within the contour curve is extracted as the classification characteristic parameter of corn seeds.

在步骤c中判断待识别玉米种子与训练样本的匹配性的操作包括：The operation of judging the matching between the corn seed to be identified and the training sample in step c includes:

按照步骤a、b获取h类共l个玉米种子的特征参数，并利用化学分析测试，获得其类别标签,将类别标签相同的玉米种子构建为一个子训练集合，共得到h个子集合：According to steps a and b, obtain the characteristic parameters of a total of l corn seeds of class h, and use chemical analysis and testing to obtain their category labels, construct a sub-training set of corn seeds with the same category label, and obtain h sub-sets in total:

令D＝[D₁，…,D_i,…,D_h]为h类样本的训练集，为有n_i个样本的第i类样本的子集合，T＝D-D_i为剔除第i类样本的子集合。对于给定的训练样本 $x_{i}^{j} &Element; D_{i}, x_{i}^{k} &Element; D_{i},$ 和有类内距离和类间距离 Let D=[D ₁ ,...,D _i ,...,D _h ] be the training set of h class samples, is a subset of samples of type i with n _i samples, and T=DD _i is a subset of samples of type i excluded. For a given training sample $x_{i}^{j} &Element; {D.}_{i}, x_{i}^{k} &Element; {D.}_{i},$ and within-class distance and the distance between classes

${S S}_{i i}^{j j,, k k} = = | | | | {x x}_{i i}^{j j} - - {x x}_{i i}^{k k} | | | |,, {x x}_{i i}^{j j},, {x x}_{i i}^{k k} &Element; &Element; {D D.}_{i i},, j j &NotEqual; &NotEqual; k k - - - - - - ((11))$

${P P}_{i i}^{k k,, s the s} = = | | | | {x x}_{i i}^{k k} - - {x x}_{i i}^{s the s} | | | |,, {x x}_{i i}^{k k} &Element; &Element; {D D.}_{i i},, {x x}_{i i}^{s the s} &Element; &Element; {T T}_{i i} - - - - - - ((22))$

其中：1≤i≤h，||||表示2范数。Among them: 1≤i≤h, |||| represents the 2 norm.

在计算第i类的K最小类内距离平均和K最小类间距离平均后，计算一个判别指标 In calculating the K-minimum intra-class distance average of the i-th class and K-minimum inter-class distance mean After that, calculate a discriminant index

${Thr Thr}_{i i}^{k k} = = {DI DI}_{i i}^{k k} / / {DW DW}_{i i}^{k k},, i i = = 11,, ... ...,, h h,, k k = = 11,, ... ...,, {n no}_{i i} - - - - - - ((33))$

定义一个反映不同类样本间差异性的阈值Th_i：Define a threshold Th _i that reflects the difference between samples of different classes:

${Th Th}_{i i} = = \underset{k k = = 11 : : {n no}_{i i}}{m m i i n no} (({Yhr Yhr}_{i i}^{k k})) \times \times β β - - - - - - ((44))$

其中β为松弛系数。对h类训练样本，得到h个阈值(Th₁,…,Th_i,…,Th_h)。where β is the relaxation coefficient. For h class training samples, h thresholds (Th ₁ ,...,Th _i ,...,Th _h ) are obtained.

在步骤e中，建立玉米种子的最小二乘支持向量机检测模型f_LS-SVM，具体包括：In step e, the least squares support vector machine detection model f _LS-SVM of corn seed is established, specifically including:

利用步骤d中获得的h类共l个玉米种子的特征参数和类别标签，通过一对多方法，构建h个子LS-SVM分类模型。其第k(1≤k≤h)个子模型的输出表达式z^k(Y)为：Using the feature parameters and category labels of h categories of l corn seeds obtained in step d, construct h sub-LS-SVM classification models through a one-to-many method. The output expression z ^k (Y) of the kth sub-model (1≤k≤h) is:

${z z}^{k k} ((Y Y)) = = s the s g g n no [[{Σ Σ}_{j j = = 11}^{l l} {β β}_{j j} K K ((Y Y,, {Y Y}_{j j})) + + b b]] - - - - - - ((55))$

其中,sgn(·)为符号函数，Y为待识别玉米种子样本的特征参数，Y_j为训练集玉米种子样本的特征参数，β_j和b均为待定系数。当大于等于零时，sgn(·)取值为1，表示待识别样本Y属于第k类，否则不属于。Among them, sgn( ) is a symbolic function, Y is the characteristic parameter of the corn seed sample to be identified, Y _j is the characteristic parameter of the corn seed sample in the training set, and β _j and b are undetermined coefficients. when When it is greater than or equal to zero, sgn(·) takes a value of 1, indicating that the sample Y to be identified belongs to the kth class, otherwise it does not.

公式(5)中β_j和b的取值由下列形式给出，The values of β _j and b in formula (5) are given by the following forms,

$[\begin{matrix} 00 & - - D D. \\ {D D.}^{T T} & Ω Ω + + {γ γ}^{- - 11} E E. \end{matrix}] [\begin{matrix} b b \\ θ θ \end{matrix}] = = [\begin{matrix} 00 \\ I I \end{matrix}] - - - - - - ((66))$

其中，θ＝[β₁…β_j…β_l]和b为待求解参数；D＝[d₁…d_j…d_l]，当训练样本Y_j属于第k类时，d_j＝1，否则d_j＝-1。γ为为惩罚系数，E为l×l的单位方阵，I为l×1的全1列向量；Ω是l×l的方阵，其第i行第j列的元素Ω_ij＝d_id_jK(Y_i,Y_j)，核函数选取：Among them, θ=[β ₁ ... β _j ... β _l ] and b are the parameters to be solved; D = [d ₁ ... d _j ... d _l ], when the training sample Y _j belongs to the kth class, d _j = 1, Otherwise d _j =-1. γ is the penalty coefficient, E is the unit square matrix of l×l, and I is the vector of all 1 columns of l×1; Ω is the square matrix of l×l, and the element Ω _ij =d _i of the i-th row and the j-th column d _j K(Y _i ,Y _j ), kernel function Select:

$K K (({Y Y}_{i i}^{k k},, {Y Y}_{j j}^{k k})) = = exp exp ((- - \frac{| | | | {Y Y}_{i i}^{k k} - - {Y Y}_{j j}^{k k} | | {| |}^{22}}{22 {σ σ}^{22}})) - - - - - - ((77))$

高斯核函数K(Y,Y_i)的核宽度参数和惩罚系数γ通过网格搜索方法获得。The kernel width parameter and penalty coefficient γ of Gaussian kernel function K(Y,Y _i ) are obtained by grid search method.

进一步地，基于聚类算法的玉米种子高光谱图像分类识别模型更新方法，其特征在于步骤d中判断待识别玉米种子与训练样本的匹配性，具体包括：Further, the corn seed hyperspectral image classification and recognition model update method based on clustering algorithm is characterized in that judging the matching between the corn seed to be identified and the training sample in step d, specifically including:

取一个待识别样本x，假设其属于第i类，则按式(1)，(2)和(3)计算其判别指标Thr_i，如果Thr_i≥Th_i，则该待识别样本与所假设的第i类样本不匹配，此时将该待识别样本归为更新样本，并利用化学分析测试，获得其类别标签，利用获得的待识别种子特征参数及其真实类别，更新训练集；反之，不更新训练集。Take a sample x to be identified, assuming it belongs to the i-th category, then calculate its discriminant indicator Thr _i according to formulas (1), (2) and (3), if Thr _i ≥ Th _i , then the sample to be identified is consistent with the hypothesized The i-th sample does not match, at this time, the sample to be identified is classified as an update sample, and its category label is obtained by chemical analysis test, and the training set is updated by using the obtained seed feature parameters and its true category to be identified; otherwise, The training set is not updated.

附图说明Description of drawings

图1为本发明提供的基于聚类算法的玉米种子高光谱图像分类识别模型更新方法的流程图。Fig. 1 is a flowchart of a method for updating a classification and recognition model of corn seed hyperspectral images based on a clustering algorithm provided by the present invention.

具体实施方法Specific implementation method

下面结合具体附图和优选实施例对本发明做进一步说明。The present invention will be further described below in combination with specific drawings and preferred embodiments.

图1为本发明提供的基于聚类算法的玉米种子高光谱图像分类识别模型更新方法的流程图，如图1所示，其中，基于最小二乘支持向量机的模型更新具体步骤包括：a、将待识别的玉米种子样本放置在高光谱图像采集系统中，采集并获取所有玉米种子样本在N个波段下的N个高光谱图像；b、计算感兴趣区域内的光谱均值特征，将所有玉米种子的N个波段下获得的共N个光谱均值特征作为待识别玉米种子的特征参数矩阵Y；c、将步骤b所得到的玉米种子特征参数矩阵Y，依次输入数据，计算待识别玉米种子与训练样本的匹配性；d、判断待识别玉米种子与训练样本的匹配性，若匹配不成功，则通过化学分析测试，获得待识别玉米种子的类别，利用获得的待识别种子特征参数及其真实类别，更新训练集和测试集；e、重复步骤(c-d),完成所有待识别玉米种子的匹配性判断,利用新的训练集来建立新的最小二乘支持向量机检测模型f_LS-SVM,用该模型来识别更新后的待识别样本集。Fig. 1 is the flow chart of the corn seed hyperspectral image classification recognition model updating method based on clustering algorithm provided by the present invention, as shown in Fig. 1, wherein, based on the model update specific steps of least squares support vector machine comprising: a, Place the corn seed samples to be identified in the hyperspectral image acquisition system, collect and obtain N hyperspectral images of all corn seed samples in N bands; A total of N spectral mean features obtained under the N bands of the seed are used as the characteristic parameter matrix Y of the corn seed to be identified; c, the corn seed characteristic parameter matrix Y obtained in step b, is input data in turn, and the calculated corn seed to be identified and The matching of the training samples; d, judging the matching of the corn seeds to be identified and the training samples, if the matching is unsuccessful, the chemical analysis test is used to obtain the category of the corn seeds to be identified, and the obtained characteristic parameters of the seeds to be identified and their true Category, update training set and test set; e, repeat step (cd), complete the matching judgment of all corn seeds to be identified, utilize new training set to set up new least squares support vector machine detection model f _LS-SVM , Use this model to identify the updated sample set to be identified.

在步骤c中计算待识别玉米种子与训练样本的匹配性的操作包括：The operation of calculating the matching between the corn seed to be identified and the training sample in step c includes:

令D＝[D₁，…,D_i,…,D_h]为h类样本的训练集，为有n_i个样本的第i类样本的子集合，T＝D-D_i为剔除第i类样本的子集合。对于给定的训练样本 $x_{i}^{j} &Element; D_{i}, x_{i}^{k} &Element; D_{i},$ 和 $x_{i}^{s} &Element; T_{i},$ 有类内距离和类间距离 Let D=[D ₁ ,...,D _i ,...,D _h ] be the training set of h class samples, is a subset of samples of type i with n _i samples, and T=DD _i is a subset of samples of type i excluded. For a given training sample $x_{i}^{j} &Element; {D.}_{i}, x_{i}^{k} &Element; {D.}_{i},$ and $x_{i}^{the s} &Element; T_{i},$ within-class distance and the distance between classes

${Th Th}_{i i} = = \underset{k k = = 11 : : {n no}_{i i}}{m m i i n no} (({Thr Thr}_{i i}^{k k})) \times \times β β - - - - - - ((44))$

本发明的优势：提供一种基于聚类算法的玉米种子高光谱图像分类识别模型更新方法，其能够实现高光谱图像分类模型更新，省时省力，快速有效，且可靠性高。Advantages of the present invention: provide a method for updating a hyperspectral image classification and recognition model of corn seeds based on a clustering algorithm, which can realize the update of a hyperspectral image classification model, saves time and effort, is fast and effective, and has high reliability.

Claims

1. the corn seed hyperspectral image classification recognition model updating method based on clustering algorithm, it is characterized in that, comprises:

a. Place the corn seed samples to be identified in the hyperspectral image acquisition system, collect and obtain N hyperspectral images of all corn seed samples in N bands;

B, calculate the spectral mean value feature in the region of interest, and the total N spectral mean value features obtained under the N bands of all corn seeds are used as the characteristic parameter matrix Y of the corn seed to be identified;

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

d. Judging the matching between the corn seeds to be identified and the training samples, if the matching is unsuccessful, the category of the corn seeds to be identified is obtained through a chemical analysis test, and the training set and test set;

e, repeat step (cd), complete the matching judgment of all corn seeds to be identified, use the new training set to set up a new least squares support vector machine detection model f _LS-SVM , and use this model to identify the updated corn seeds Identify sample sets.

In step b, obtaining the characteristic parameter of corn seed comprises:

First, select the band corresponding to the image with the clearest outline of the corn seed to be identified (at 782.59nm), and use the adaptive threshold segmentation method to obtain the outline curve of the corn seed to be identified under this band. The profile curve is projected onto N bands, and the spectral mean value of the N bands within the profile curve is extracted as the classification characteristic parameter of corn seeds.

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

According to steps a and b, obtain the characteristic parameters of a total of l corn seeds of class h, and use chemical analysis and testing to obtain their category labels, construct a sub-training set of corn seeds with the same category label, and obtain h sub-sets in total:

Let D=[D ₁ ,...,D _i ,...,D _h ] be the training set of h class samples, is a subset of samples of type i with n _i samples, and T=DD _i is a subset of samples of type i excluded. For a given training sample and within-class distance and the distance between classes

{S S}_{i i}^{j j,, k k} = = | | | | {x x}_{i i}^{j j} - - {x x}_{i i}^{k k} | | | |,, {x x}_{i i}^{j j},, {x x}_{i i}^{k k} &Element; &Element; {D D.}_{i i},, j j &NotEqual; &NotEqual; k k - - - - - - ((11))

{P P}_{i i}^{k k,, s the s} = = | | | | {x x}_{i i}^{k k} - - {x x}_{i i}^{s the s} | | | |,, {x x}_{i i}^{k k} &Element; &Element; {D D.}_{i i},, {x x}_{i i}^{s the s} &Element; &Element; {T T}_{i i} - - - - - - ((22))

Among them: 1≤i≤h, |||| represents the 2 norm.

In calculating the K-minimum intra-class distance average of the i-th class and K-minimum inter-class distance mean After that, calculate a discriminant index

{Thr Thr}_{i i}^{k k} = = {DI DI}_{i i}^{k k} / / {DW DW}_{i i}^{k k},, i i = = 11,, ... ...,, h h,, k k = = 11,, ... ...,, {n no}_{i i} - - - - - - ((33))

Define a threshold Th _i that reflects the difference between samples of different classes:

{Th Th}_{i i} = = \underset{k k = = 11 : : {n no}_{i i}}{m m i i n no} (({Yhr Yhr}_{i i}^{k k})) \times \times β β - - - - - - ((44))

where β is the relaxation coefficient. For h class training samples, h thresholds (Th ₁ ,...,Th _i ,...,Th _h ) are obtained.

In step e, the least squares support vector machine detection model f _LS-SVM of corn seed is established, specifically including:

Using the feature parameters and category labels of h categories of l corn seeds obtained in step d, construct h sub-LS-SVM classification models through a one-to-many method. The output expression z ^k (Y) of the kth sub-model (1≤k≤h) is:

z^{k} (Y) = the s g no [Σ_{j = 1}^{l} β_{j} K (Y, Y_{j}) + b] - - - (5)

Among them, sgn( ) is a symbolic function, Y is the characteristic parameter of the corn seed sample to be identified, Y _j is the characteristic parameter of the corn seed sample in the training set, and β _j and b are undetermined coefficients. when When it is greater than or equal to zero, sgn(·) takes a value of 1, indicating that the sample Y to be identified belongs to the kth class, otherwise it does not.

The values of β _j and b in formula (5) are given by the following forms,

[\begin{matrix} 00 & - - D D. \\ {D D.}^{T T} & Ω Ω + + {γ γ}^{- - 11} E E. \end{matrix}] [\begin{matrix} b b \\ θ θ \end{matrix}] = = [\begin{matrix} 00 \\ I I \end{matrix}] - - - - - - ((66))

Among them, θ=[β ₁ ... β _j ... β _l ] and b are the parameters to be solved; D = [d ₁ ... d _j ... d _l ], when the training sample Y _j belongs to the kth class, d _j = 1, Otherwise d _j =-1. γ is the penalty coefficient, E is the unit square matrix of l×l, and I is the vector of all 1 columns of l×1; Ω is the square matrix of l×l, and the element Ω _ij =d _i of the i-th row and the j-th column d _j K(Y _i ,Y _j ), kernel function Select:

K K (({Y Y}_{i i}^{k k},, {Y Y}_{j j}^{k k})) = = exp exp ((- - \frac{| | | | {Y Y}_{i i}^{k k} - - {Y Y}_{j j}^{k k} | | {| |}^{22}}{22 {σ σ}^{22}})) - - - - - - ((77))

The kernel width parameter and penalty coefficient γ of Gaussian kernel function K(Y,Y _i ) are obtained by grid search method.

2. the corn seed hyperspectral image classification recognition model updating method based on clustering algorithm according to claim 1, is characterized in that judging the matching judgment of corn seed to be identified and training sample in the step d, specifically comprises:

Take a sample x to be identified, assuming it belongs to the i-th category, then calculate its discriminant indicator Thr _i according to formulas (1), (2) and (3), if Thr _i ≥ Th _i , then the sample to be identified is consistent with the hypothesized The i-th sample does not match, at this time, the sample to be identified is classified as an update sample, and its category label is obtained by chemical analysis test, and the training set is updated by using the obtained seed feature parameters and its true category to be identified; otherwise, The training set is not updated.