CN108596123A - A kind of fruit hardness detection method and device based on hyperspectral analysis - Google Patents
A kind of fruit hardness detection method and device based on hyperspectral analysis Download PDFInfo
- Publication number
- CN108596123A CN108596123A CN201810399828.7A CN201810399828A CN108596123A CN 108596123 A CN108596123 A CN 108596123A CN 201810399828 A CN201810399828 A CN 201810399828A CN 108596123 A CN108596123 A CN 108596123A
- Authority
- CN
- China
- Prior art keywords
- fruit
- wave band
- matrix
- sample
- pls
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06395—Quality analysis or management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Economics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Strategic Management (AREA)
- Evolutionary Computation (AREA)
- Entrepreneurship & Innovation (AREA)
- Educational Administration (AREA)
- Development Economics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Evolutionary Biology (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a kind of fruit hardness detection method and device based on hyperspectral analysis, collect fruit body surface EO-1 hyperion dispersion image first, extract characteristics of image, obtain several fruit samples, sample set division is carried out to several fruit samples, calibration set is obtained, screens the optimal wave band of calibration set, using optimal wave band as sample predictions value, fruit hardness BP neural network Fusion Model is established.The present invention carries out feature extraction using Lorentz distribution function fitting process to the EO-1 hyperion dispersion image of acquisition, and no information variable elimination algorithm, PLS Projection Analysis algorithm and semi-supervised neighbour's propagation clustering algorithm is respectively adopted and screens optimal wave band, Fusion Model is established using optimal wave band, is realized to fruit Hardness Prediction.
Description
Technical field
The present invention relates to fruit hardness determination fields, and in particular to a kind of fruit hardness determination side based on hyperspectral analysis
Method and device.
Background technology
Since EO-1 hyperion can be detected the hardness of apple internal, thus EO-1 hyperion is current for apple quality sorting
Compare popular research direction.Hyperspectral technique is that the spectral information progress of irreflexive light occurs to being irradiated to apple surface
One technology of analysis, but spectral information amount is huge, and mainstream algorithm model is complicated, and processing speed is slow, and pole is required to calculating speed
Height, thus be also not used for being commercialized, only exist in laboratory test state.EO-1 hyperion is for most important two of apple quality detection
Aspect is the foundation of the selection and prediction model of optimal wave band, and the selection of optimal wave band can find criterion function in a large amount of wave bands
Optimal wave band, the algorithm mainly used at present are genetic algorithm, which retained most with the methods of selection variation mutation
Excellent wave band eliminates poor quality wave band.At present there are mainly two types of the methods of the apple internal information prediction model foundation of comparative maturity, the
A kind of partial least squares algorithm, the algorithm are a kind of common mathematical optimization techniques, it is looked for by minimizing the quadratic sum of error
To the optimal function matching of one group of data, but since the algorithm is common linear model, come to water only with linear model
Fruit inside quality parameter prediction will certainly cause built prediction model performance relatively low；Second is neural network, which is mould
A kind of algorithm of the structure and function of apery cranial nerve cell, it is more superior than the first algorithm in terms of predicting fruit hardness, but
Its pace of learning is slow, influences detection speed.
Lot of domestic and international mechanism is all doing the detection about agricultural product quality using EO-1 hyperion at present：At home, in hardness
This aspect, Zhao Jiewen etc. correct hardness model using support vector machines and Partial Least Squares, and prediction related coefficient reaches
0.681；In recent years external present Research approximately as:Singh etc. obtains degree of precision using hyper-spectral image technique
Potato moisture offset minimum binary prediction model.The application high light spectrum image-forming technology such as Lu is by the scattering of light to apple
Hardness and soluble solids content (SSC) prediction are studied, research shows that containing there are three the Lorentz letters of parameter after modification
Number (scattering peak value, width and the gradient) can provide good prediction；The prediction result of the hardness of fruit is 0.89.The profits such as Huang
With layering evolution algorithm combination Subspace Decomposition to ' the optimal wave of Golden Delicious' apple sample EO-1 hyperion dispersion images
Section selection, establishes apple hardness and the offset minimum binary prediction model of pol.It is excellent to have selected 17 wave bands for establishing hardness
Prediction model, the prediction related coefficient for having obtained apple hardness are 0.857.The result shows that establishing apple after waveband selection
The prediction model of hardness than all band when function admirable.Peng etc. has studied to be established firmly using single Lorentz fit parameter
The prediction model of degree and soluble dissolved solids content, the wherein prediction related coefficient of hardness are 0.89, soluble soluble solids it is pre-
It is 0.88 to survey related coefficient.At image and ANN Technology designs, feedforward is reversed passes for the research and utilizations such as ElMasry, G. EO-1 hyperions institute
ANN model is broadcast to select influence of the wavelength detection freezing injury to apple hardness.Experiment measures the related coefficient between hardness number
It is 0.93.
In conclusion in the prior art, on quickly and effectively establishing rigidity prediction model this problem, still lack effective
Solution.
Invention content
In order to overcome the above-mentioned deficiencies of the prior art, the fruit hardness inspection based on hyperspectral analysis that the present invention provides a kind of
Method and device is surveyed, with Lorentz distribution function (MLD) fitting process, without information variable elimination algorithm (UVE), Kennard-
Stone algorithms etc. predict apple hardness solve relatively low, pre- in the presence of built prediction model performance in fruit hardness determination
The problems such as performing poor in terms of surveying pol.
The technical solution adopted in the present invention is：
A kind of fruit hardness detection method based on hyperspectral analysis, this approach includes the following steps：
(1) collect fruit body surface EO-1 hyperion dispersion image；
(2) characteristics of image is extracted, several fruit internal message samples are obtained；
(3) sample set division is carried out to several fruit internal message samples, obtains calibration set；
(4) the optimal wave band of calibration set is screened；
(5) using optimal wave band as sample predictions value, fruit hardness BP neural network Fusion Model is established.
Fruit hardness detection method based on hyperspectral analysis as described above, in the step (1), collect fruit ontology
The step of surface EO-1 hyperion dispersion image includes：
Collect fruit ontology scattering light, and the scattering light of acquisition is scatter according to the difference of wave band, is disperseed
Optical information, and CCD camera is projected, two dimensional image is formed, horizontal axis represents spatial information, and the longitudinal axis represents spectral information.
The fruit hardness detection method of hyperspectral analysis as described above in the step (2), extracts the step of characteristics of image
Suddenly include：
Using the scattering curve of the Lorentzian fit-spectra of ternary, the asymptotic value of light scatter intensity maximum value is added
Scattering curve is fitted perfect, multiple data points is obtained, using data point as fruit internal message sample.
Fruit hardness detection method based on hyperspectral analysis as described above, in the step (3), to several fruit
Internal information sample carry out sample set division the step of include：
The distance between any two fruit internal message sample in all fruit internal message samples is calculated, is therefrom selected
Apart from maximum two fruit internal message samples；
It calculates separately between all remaining fruit internal message samples and selected two fruit internal message samples
Distance；
The shortest distance between each remaining fruit internal message sample of selection and selected fruit internal message sample,
And be ranked up, it selects the sample corresponding to maximum distance as calibration samples, recycles successively, until selecting multiple correction samples
This, constitutes calibration set.
In the step (4), no letter is respectively adopted in fruit hardness detection method based on hyperspectral analysis as described above
Cease the optimal wave band of variable elimination algorithm, PLS Projection Analysis algorithm and semi-supervised neighbour's propagation clustering algorithm screening calibration set.
Fruit hardness detection method based on hyperspectral analysis as described above, it is described using no information variable elimination algorithm
Screen calibration set optimal wave band the step of include：
Spectrum matrix X and concentration of component matrix Y to be measured to calibration set carry out PLS recurrence, and choose best main cause subnumber；
Stochastic variable matrix R is set up, is combined spectrum battle array X and stochastic variable matrix R to obtain matrix XR；
PLS recurrence is carried out to stochastic variable matrix R and concentration matrix Y, obtains the square being combined by n PLS regression coefficient
Battle array B；
The standard deviation and average value on the column direction of matrix B are calculated separately out, average value and standard deviation are divided by, obtained
To parameter C；
The maximum value C of parameter C is obtained in section [m+1,2m]max；
It is rejected in section [1, m] and is less than C in matrix XmaxThe corresponding variable of parameter, remaining variable is formed into new square
Battle array XUVE, new matrix XUVEMiddle element is then the optimal wave band filtered out.
Fruit hardness detection method based on hyperspectral analysis as described above, it is described to be sieved using PLS Projection Analysis algorithms
The step of optimal wave band for selecting calibration set includes：
What spectrum matrix X to calibration set and concentration of component vector y to be measured carried out column direction goes equalization, obtains Δ X and Δ
Y, and PLS recurrence is carried out to Δ X and Δ y；
The standard deviation for calculating Δ X each columns, by the standard deviation of Δ X each columns and PLS regression coefficients specific band recurrence because
The multiplication of son simultaneously seeks absolute value, obtains the PLS projection coefficients of all calibration samples；
The PLS projection coefficients of each calibration samples are normalized, PLS projection coefficient vectors are obtained；
The wave band corresponding to the PLS averaging projections coefficient for being more than given threshold in PLS projection coefficient vectors is chosen as most
Excellent wave band.
Fruit hardness detection method based on hyperspectral analysis as described above, wherein described to be passed using semi-supervised neighbour
Broadcasting the step of clustering algorithm screens the optimal wave band of calibration set includes：
Similitude between arbitrary two wave band is sought using Kullback-Leibler function of degree of disagreement, builds similarity moment
Battle array S；
Fruit hardness actual value is measured, similarity matrix is constrained using the fruit hardness actual value as label information
Adjustment；
If matrix [r (i, k)], data point i information is sent to candidate representative point k, indicate representative points of the k as i
Matching degree；If matrix [a (i, k)], the information of candidate representative point k is sent to data point i, indicate that i selects k as class
Represent the matching degree of point；And it is set up for each data point i and is biased to parameter P (i), i.e. similarity matrix S diagonal line values s (k, k)；
A (i, k)=0 is initialized, similar matrix is calculated and is biased to parameter P (i), i.e. s (k, k)；
Damping factor λ is introduced, iteration is updated several times to matrix [r (i, k)] and [a (i, k)]；
Matrix [r (i, k)] and [a (i, k)] are added, decision matrix E is obtained；
Judge whether the diagonal line value E (k, k) of decision matrix E is more than zero, is represented as class if more than will then put accordingly
Point, recycles successively, until finding out the representative point of all data, obtains final cluster result, which is to screen
The optimal wave band gone out.
Fruit hardness detection method based on hyperspectral analysis as described above in the step (5), establishes fruit hardness
The step of BP neural network Fusion Model includes：
By the optimal wave band screened by semi-supervised neighbour's propagation clustering and by the screening of no information variable elimination algorithm
Optimal wave band blends, and obtains SAP-UVE sample predictions values；
By the optimal wave band screened by no information variable elimination algorithm with by PLS Projection Analysis algorithms screening it is optimal
Wave band blends, and obtains UVE-PLS sample predictions values；
SAP-UVE sample predictions value and UVE-PLS sample predictions values are merged, fruit is established using BP neural network
Hardness BP neural network Fusion Model.
The invention also provides a kind of fruit rigidity detection device based on hyperspectral analysis, which includes Image Acquisition
Device and processor；
Described image harvester is configured as the inside that transmitting light beam is irradiated into fruit ontology and occurs to dissipate with interior tissue
It penetrates, collection of scattered light line, the scattering light of acquisition is scatter according to the difference of wave band, the optical information disperseed, and project
To CCD camera, two dimensional image is formed；
The processor is configured as extraction characteristics of image, several fruit internal message samples is obtained, to several water
Fruit internal information sample carries out sample set division, obtains calibration set, the optimal wave band of calibration set is screened, using optimal wave band as sample
This predicted value establishes fruit hardness BP neural network Fusion Model.
Compared with prior art, the beneficial effects of the invention are as follows：
(1) present invention carries out feature extraction using Lorentz distribution function fitting process to the EO-1 hyperion dispersion image of acquisition,
And it is optimal that no information variable elimination algorithm, PLS Projection Analysis algorithm and the screening of semi-supervised neighbour's propagation clustering algorithm is respectively adopted
Wave band is established Fusion Model using optimal wave band, is realized to fruit Hardness Prediction；
(2) present invention carries out sample set using Kennard-Stone algorithms to the numerous fruits internal information sample of collection
It divides, it may be determined that the edge samples of predetermined number in sample, and carried out according to the opposite Euclidean distance between sample spectrum data
Sample divides, and makes calibration set sample distribution evenly, and it is more reasonable that sample set divides.
Description of the drawings
The accompanying drawings which form a part of this application are used for providing further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation do not constitute the improper restriction to the application for explaining the application.
Fig. 1 is the fruit hardness detection method flow chart disclosed by the embodiments of the present invention based on hyperspectral analysis.
Specific implementation mode
The invention will be further described with embodiment below in conjunction with the accompanying drawings.
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific implementation mode, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singulative
It is also intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or combination thereof.
As background technology is introduced, it is inclined to there is built prediction model performance for fruit hardness determination in the prior art
The deficiencies of performing poor in terms of low, prediction pol, in order to solve technical problem as above, present applicant proposes one kind being based on bloom
The fruit hardness detection method and device of spectrum analysis.
In a kind of typical embodiment of the application, as shown in Figure 1, providing a kind of fruit based on hyperspectral analysis
Hardness detection method is carried out sample set and drawn by the body interior EO-1 hyperion dispersion image that collects fruit to image characteristics extraction
Point, optimal wave band is chosen, the fruit hardness BP neural network Fusion Model based on UVE and SAP predicts fruit hardness.
In order to make those skilled in the art be best understood from the present invention, a more detailed embodiment is set forth below,
In the present embodiment take apple as an example, it is proposed that a kind of apple hardness detection method based on hyperspectral analysis, this method are specific
Include the following steps：
Step 1：Acquire apple EO-1 hyperion dispersion image.
When light beam is irradiated to apple surface, diffusing reflection can occur with it, the optical wavelength difference of reflection carries to obtain information
Difference reflects the information such as the inner hardness sugariness of apple.
For cylindrical light beam, light beam is irradiated to be scattered the light source that the present embodiment uses into apple internal and interior tissue,
Scattering light is acquired by top side camera, and it is straight that the distance at wherein Hyperspectral imager scan line to incident center is set as light beam
Diameter, in this case, it is ensured that camera center will not generate saturated phenomenon because signal strength is excessively high.Image light spectrometer is adopted
With prism-grating-prism structure, under the premise of not destroying light original spatial information, by the light of acquisition according to the difference of wavelength
It spreads out, the optical information of dispersion is finally projected into CCD camera, form two dimensional image, horizontal axis represents spatial information, the longitudinal axis
Represent spectral information.
Step 2：Extract characteristics of image.
Image characteristics extraction is realized with Lorentz distribution function (MLD) fitting process.
First, with the scattering curve of the Lorentzian simulated spectra of ternary, formula is：
Wherein, R is light scatter intensity values, and b is scattering curve extreme value, and c is the scattering width at half extreme value, and d is in half-shadow
Slope at value, z are scattering distance.The spectral dispersion curve being fitted with tri- variables of b, c, d under some wave band has been achieved.
So each sample can be described with 3*101 data point.
In view of spectral dispersion curve is also influenced by light scatter intensity, then being added what variable a carried out curve fitting
Perfect, formula is
A is the asymptotic value of light scatter intensity maximum value, then final apple internal message sample can be by 101 × 4 numbers
Strong point is described.
Step 3：Divide sample set.
Sample set division is carried out to numerous apple internal message samples of collection using Kennard-Stone algorithms.The calculation
Method can determine the edge samples of predetermined number in sample, and carry out sample according to the opposite Euclidean distance between sample spectrum data
It divides, makes calibration set sample distribution evenly, it is more reasonable that sample set divides.
If matrix X, behavior sample are classified as sample variable, sample set division is carried out to it.
Step 3.1：Distance between any two in all samples is calculated, is therefrom selected apart from maximum two samples, formula
For
Wherein, DijIndicate sample xiWith sample xjThe distance between, xivIndicate the sample variable of i-th of sample, xjvIt indicates
The sample variable of j-th of sample, k are variable number.
Step 3.2：It calculates separately and removes two selected samples remaining sample and two above in all samples
The distance between sample.
Step 3.3：For remaining each sample, selection and the shortest distance between sampling sheet, and to by these most
Short distance is ranked up, and selects the sample corresponding to longest distance, and as calibration samples, formula is
di(M)=min { Dt1,Di2,...,DiM}
di(M+1)=max { di(M)}
Step 3.4：Step 3.3 is repeated, until selecting enough calibration samples, constitutes calibration set.
Step 4：Screen optimal wave band.
Due to, containing a large amount of useless or redundancy wavelength informations, to be carried out most before carrying out PLS modelings in calibration sample
Excellent waveband selection is established model using the maximum wave band of information content as the input variable of the prediction model of PLS apple hardness, is allowed to
The information contained by its all band is represented to the greatest extent.
In the present embodiment, using no information variable elimination algorithm (UVE), PLS Projection Analysis algorithm and semi-supervised neighbour
Propagation clustering (SAP) respectively screens the optimal wave band of calibration set, and specific implementation is as follows.
Step 4.1：Without information variable elimination algorithm (UVE)
Optimal wave band Select to use puts forward no information variable elimination algorithm (UVE) by Centner et al..Without information variable
Elimination algorithm can reduce the variable number for including in the PLS models finally established, and improve PLS models, reduce the complexity of model.
Step 4.1.1：Spectrum matrix X (m × n) and concentration of component matrix Y (m × 1) to be measured to calibration set carry out PLS and return
Return, and choose best main cause subnumber, wherein m is sample size, and n is wave band quantity；
Step 4.1.2：Stochastic variable matrix R (m × n) is set, spectrum battle array X and stochastic variable matrix R are combined
To matrix XR (m × 2n)；
Step 4.1.3：PLS recurrence is carried out to matrix R and Y, the matrix B that obtains being combined by n PLS regression coefficient (m ×
2n)；
Step 4.1.4：Calculate separately out the standard deviation S (b) and average value mean on the column direction of matrix B (m × 2n)
(b), parameter C and by formula (1) is calculatedi；
In formula, s (bi) indicate the standard deviation of regression coefficient vector b；mean(bi) indicate the average value of regression coefficient b.
Step 4.1.5：The maximum value that parameter C is obtained at section [m+1,2m] is Cmax=max (abs (C))；
Step 4.1.6：C in matrix X is removed in section [1, m]i＜ CmaxCorresponding variable, then remaining variable composition
Matrix be the new matrix X that is obtained through UVE algorithmsUVE, new matrix XUVEMiddle element is then the optimal wave band filtered out.
The optimal wave band that sample set is filtered out using no information variable elimination algorithm obtains new matrix XUVE, the new matrix
XUVEMiddle element is the optimal wave band filtered out through UVE algorithms.
Step 4.2：PLS Projection Analysis algorithms
Another algorithm of optimal wave band selected is PLS Projection Analysis algorithms.Its advantage is that considering a certain feature wave
The joint effect of the variable quantity and projection components of the spectrum PLS regression coefficients at the wave band of long spectroscopic data.
Steps are as follows for specific implementation：
Step 4.2.1：To the spectrum matrix X (m × n) of calibration set and concentration of component vector y (m × 1) to be measured into ranks side
To go equalization, obtain Δ X and Δ y；PLS recurrence is carried out to Δ X and Δ y, because Δ y is a dimensional vector, processing is single
Target variable problem, using PLS1 algorithms；
Step 4.2.2：The standard deviation of Δ X each columns is sought according to formula (2).Since Δ X is zero per column mean, can obtain
It arrives：
In formula, m is sample size, and n is wave band quantity, Δ Xi,jFor the i-th row jth column element of Δ X；qjIt is arranged for Δ X jth
Standard deviation；
Step 4.2.3：PLS projection coefficients are calculated according to formula (3)：
rj=| qj×fj| (3)
Wherein, fjFor PLS regression coefficients jth wave band regression vectors；rjFor the PLS projection coefficients of jth wave band；
Step 4.2.4：To rj(j=1,2 ..., m) is normalized, you can obtains PLS projection coefficient vectors R：
In formula, rjIndicate the PLS projection coefficients of jth wave band；Indicate the PLS averaging projections coefficient of jth wave band；
Finally, according to PLS averaging projections coefficient curve, the wave band more than the threshold value can be selected by the threshold value of a certain setting
Spectroscopic data modeled, to be optimal the purpose of waveband selection.
Step 4.3：Semi-supervised neighbour's propagation clustering (SAP)
Since neighbour's propagation clustering (AP) is to carry out data analysis in the case of no any constraints, in order to obtain
Better cluster result introduces semi-supervised neighbour's propagation clustering (SAP), is used as about by the way that information or label known to part is added
Beam is adjusted to form similarity matrix to having matrix.Similarity letter is sought using Kullback-Leibler function of degree of disagreement
Number, and then seek similarity matrix.Steps are as follows for its specific implementation：
Step 4.3.1：Similitude between arbitrary two wave band is sought using Kullback-Leibler function of degree of disagreement, into
And similarity matrix is obtained, the similitude between arbitrary two wave band is described with following formula：
S (i, k)=1-KL (i, k)
In formula, s (i, k) indicates the similarity set between arbitrary two data points i, k；KL (i, k) indicates luxuriant divergence letter
Number；
Step 4.3.2：Apple hardness actual value is measured with external force damage method；
Step 4.3.3：The hardness number of the actual measurement of apple internal message sample obtains 4.3.1 as label information similar
Degree matrix carries out constraint adjustment, and formula is as follows：
Wherein, Ym×lThat indicate is the measured value of each sample, matrix Pm×nIt is each apple internal message sample each
Predicted value under wave band；K indicates a weight coefficient；
Step 4.3.4：If matrix [r (i, k)], sends the information of data point i to candidate representative point k, indicate k conducts
The matching degree of the representative point of i；If matrix [a (i, k)], sends the information of candidate representative point k to data point i, indicate i
Select matching degrees of the k as representative point；It is set up for each data point i and is biased to parameter P (i), i.e. similarity matrix S diagonal lines
Value s (k, k),
Step 4.3.5：A (i, k)=0 is initialized, similar matrix is calculated and is biased to parameter P (i) i.e. s (k, k), formula is such as
Under：
Wherein, α ＞ 0 obtain different number of optimal wave band for controlling, and under normal circumstances, α values are bigger, select
Optimal wave band number is more, and α values are smaller, and the optimal wave band number selected is fewer；XkWhat is indicated is the kth train wave section of spectrum matrix X.
Step 4.2.6：Damping factor λ is introduced, according to following formula
Update iteration n times matrix [r (i, k)] and [a (i, k)]；
Step 4.3.7：Matrix a is added with r to obtain decision matrix E=r+a；
Step 4.3.8：Judge whether the value of E (k, k) is more than 0, is exactly representative point if more than then corresponding point；
Step 4.3.9：The representative point for finding out all data points can be obtained final cluster result, and cluster result is
The optimal wave band filtered out.
Step 5：Establish apple hardness BP neural network Fusion Model.
In order to eliminate a kind of error that optimal waveband selection is brought and a kind of algorithm to establishing the limitation of model, use
The mode of two kinds of algorithm fusions of SAP-UVE and UVE-PLS carries out the foundation of Fusion Model, by SAP-UVE and UVE-PLS models
Two input nodes of the sample Hardness Prediction value as Fusion Model.Then the output of Fusion Model is to carry out nerve to sample data
After network training, the final calibration set Hardness Prediction value of generation.The structure of the apple hardness BP neural network Fusion Model
Method is specially：
The optimal wave band of semi-supervised neighbour's propagation clustering (SAP) screening will be passed through and by no information variable elimination algorithm
(UVE) the optimal wave band screened blends, and obtains SAP-UVE sample predictions values；
By the optimal wave band screened by no information variable elimination algorithm (UVE) and process PLS Projection Analysis algorithm (PLS)
The optimal wave band of screening blends, and obtains UVE-PLS sample predictions values；
SAP-UVE sample predictions value and UVE-PLS sample predictions values are merged, apple is established using BP neural network
Hardness BP neural network Fusion Model.
In the typical embodiment of another kind of the application, a kind of apple hardness detection based on hyperspectral analysis is provided
Device, the device include image collecting device and processor.
Described image harvester is configured as the inside that transmitting light beam is irradiated into apple ontology and occurs to dissipate with interior tissue
It penetrates, collection of scattered light line, the scattering light of acquisition is scatter according to the difference of wave band, the optical information disperseed, and project
To CCD camera, two dimensional image is formed.
In the present embodiment, described image harvester includes beam emissions module, video camera, image light spectrometer and CCD
Camera, the beam emissions module transmitting light beam is irradiated to be scattered into apple internal and interior tissue, by being arranged in apple
The scattering light of acquisition scatter according to the difference of wavelength, obtains by the camera acquisition scattering light at the top of fruit, image light spectrometer
To the optical information of dispersion, and CCD camera is projected, forms two dimensional image.
The processor is configured as extraction characteristics of image, several apple internal message samples is obtained, to several apples
Fruit internal information sample carries out sample set division, obtains calibration set, the optimal wave band of calibration set is screened, using optimal wave band as sample
This predicted value establishes apple hardness BP neural network Fusion Model.
It can be seen from the above description that the application the above embodiments realize following technique effect：
(1) present invention carries out feature using Lorentz distribution function (MLD) fitting process to the EO-1 hyperion dispersion image of acquisition
Extraction, and no information variable elimination algorithm, PLS Projection Analysis algorithm and the screening of semi-supervised neighbour's propagation clustering algorithm is respectively adopted
Optimal wave band establishes Fusion Model using optimal wave band, realizes and predict apple hardness；
(2) present invention carries out sample set division using Kennard-Stone algorithms to numerous apple samples of collection, can be with
It determines the edge samples of predetermined number in sample, and sample division is carried out according to the opposite Euclidean distance between sample spectrum data,
Make calibration set sample distribution evenly, it is more reasonable that sample set divides.
Above-mentioned, although the foregoing specific embodiments of the present invention is described with reference to the accompanying drawings, not protects model to the present invention
The limitation enclosed, those skilled in the art should understand that, based on the technical solutions of the present invention, those skilled in the art are not
Need to make the creative labor the various modifications or changes that can be made still within protection scope of the present invention.
Claims (10)
1. a kind of fruit hardness detection method based on hyperspectral analysis, characterized in that include the following steps：
(1) collect fruit body surface EO-1 hyperion dispersion image；
(2) characteristics of image is extracted, several fruit internal message samples are obtained；
(3) sample set division is carried out to several fruit internal message samples, obtains calibration set；
(4) the optimal wave band of calibration set is screened；
(5) using optimal wave band as sample predictions value, fruit hardness BP neural network Fusion Model is established.
2. the fruit hardness detection method according to claim 1 based on hyperspectral analysis, characterized in that the step
(1) in, collect fruit body surface EO-1 hyperion dispersion image the step of include：
Collect fruit ontology scattering light, and the scattering light of acquisition is scatter according to the difference of wave band, the light letter disperseed
Breath, and CCD camera is projected, two dimensional image is formed, horizontal axis represents spatial information, and the longitudinal axis represents spectral information.
3. the fruit hardness detection method according to claim 1 based on hyperspectral analysis, characterized in that the step
(2) in, extract characteristics of image the step of include：
Using the scattering curve of the Lorentzian fit-spectra of ternary, the asymptotic value of light scatter intensity maximum value is added to dissipating
It penetrates curve and is fitted perfect, multiple data points are obtained, using data point as fruit internal message sample.
4. the fruit hardness detection method according to claim 1 based on hyperspectral analysis, characterized in that the step
(3) in, include to the step of several fruit internal message samples progress sample set division：
The distance between any two fruit internal message sample in all fruit internal message samples is calculated, distance is therefrom selected
Maximum two fruit internal message samples；
Calculate separately all remaining fruit internal message samples and selected the distance between two fruit internal message samples；
The shortest distance between each remaining fruit internal message sample of selection and selected fruit internal message sample, goes forward side by side
Row sequence, selects the sample corresponding to maximum distance as calibration samples, recycles successively, until multiple calibration samples are selected,
Constitute calibration set.
5. the fruit hardness detection method according to claim 1 based on hyperspectral analysis, characterized in that the step
(4) in, no information variable elimination algorithm, PLS Projection Analysis algorithm and the screening of semi-supervised neighbour's propagation clustering algorithm is respectively adopted
The optimal wave band of calibration set.
6. the fruit hardness detection method according to claim 5 based on hyperspectral analysis, characterized in that described to use nothing
Information variable elimination algorithm screen calibration set optimal wave band the step of include：
Spectrum matrix X and concentration of component matrix Y to be measured to calibration set carry out PLS recurrence, and choose best main cause subnumber；
Stochastic variable matrix R is set up, is combined spectrum battle array X and stochastic variable matrix R to obtain matrix XR；
PLS recurrence is carried out to stochastic variable matrix R and concentration matrix Y, obtains the matrix B being combined by n PLS regression coefficient；
The standard deviation and average value on the column direction of matrix B are calculated separately out, average value and standard deviation are divided by, joined
Number C；
The maximum value C of parameter C is obtained in section [m+1,2m]max；
It is rejected in section [1, m] and is less than C in matrix XmaxThe corresponding variable of parameter, remaining variable is formed into new matrix
XUVE, new matrix XUVEMiddle element is then the optimal wave band filtered out.
7. the fruit hardness detection method according to claim 5 based on hyperspectral analysis, characterized in that the use
PLS Projection Analysis algorithms screen calibration set optimal wave band the step of include：
What spectrum matrix X to calibration set and concentration of component vector y to be measured carried out column direction goes equalization, obtains Δ X and Δ y, and
PLS recurrence is carried out to Δ X and Δ y；
The standard deviation that calculates Δ X each columns, by the standard deviation of Δ X each columns and PLS regression coefficients specific band regression vectors
It is multiplied and seeks absolute value, obtain the PLS projection coefficients of all calibration samples；
The PLS projection coefficients of each calibration samples are normalized, PLS projection coefficient vectors are obtained；
The wave band corresponding to the PLS averaging projections coefficient for being more than given threshold in PLS projection coefficient vectors is chosen as optimal wave
Section.
8. the fruit hardness detection method according to claim 5 based on hyperspectral analysis, characterized in that described using half
Supervising the step of neighbour's propagation clustering algorithm screens the optimal wave band of calibration set includes：
Similitude between arbitrary two wave band, structure similarity matrix S are sought using Kullback-Leibler function of degree of disagreement；
Fruit hardness actual value is measured, constraint tune is carried out to similarity matrix using the fruit hardness actual value as label information
It is whole；
If matrix [r (i, k)], data point i information is sent to candidate representative point k, indicate symbols of the k as the representative point of i
Conjunction degree；If matrix [a (i, k)], the information of candidate representative point k is sent to data point i, indicate that i selects k to be represented as class
The matching degree of point；And it is set up for each data point i and is biased to parameter P (i), i.e. similarity matrix S diagonal line values s (k, k)；
A (i, k)=0 is initialized, similar matrix is calculated and is biased to parameter P (i), i.e. s (k, k)；
Damping factor λ is introduced, iteration is updated several times to matrix [r (i, k)] and [a (i, k)]；
Matrix [r (i, k)] and [a (i, k)] are added, decision matrix E is obtained；
Judge whether the diagonal line value E (k, k) of decision matrix E is more than zero, if more than will then put accordingly as representative point, according to
Secondary cycle obtains final cluster result until finding out the representative point of all data, which is to filter out most
Excellent wave band.
9. the fruit hardness detection method according to claim 5 based on hyperspectral analysis, characterized in that the step
(5) in, the step of establishing fruit hardness BP neural network Fusion Model, includes：
By by semi-supervised neighbour's propagation clustering screening optimal wave band with screened by no information variable elimination algorithm it is optimal
Wave band blends, and obtains SAP-UVE sample predictions values；
By the optimal wave band screened by no information variable elimination algorithm and the optimal wave band by the screening of PLS Projection Analysis algorithms
It blends, obtains UVE-PLS sample predictions values；
SAP-UVE sample predictions value and UVE-PLS sample predictions values are merged, fruit hardness is established using BP neural network
BP neural network Fusion Model.
10. a kind of fruit rigidity detection device based on hyperspectral analysis, characterized in that including image collecting device and processing
Device；
Described image harvester is configured as the inside that transmitting light beam is irradiated into fruit ontology and is scattered with interior tissue,
The scattering light of acquisition scatter by collection of scattered light line according to the difference of wave band, the optical information disperseed, and projects
CCD camera forms two dimensional image；
The processor is configured as extraction characteristics of image, several fruit internal message samples is obtained, in several fruit
Portion's message sample carries out sample set division, obtains calibration set, screens the optimal wave band of calibration set, and optimal wave band is pre- as sample
Measured value establishes fruit hardness BP neural network Fusion Model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810399828.7A CN108596123A (en) | 2018-04-28 | 2018-04-28 | A kind of fruit hardness detection method and device based on hyperspectral analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810399828.7A CN108596123A (en) | 2018-04-28 | 2018-04-28 | A kind of fruit hardness detection method and device based on hyperspectral analysis |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108596123A true CN108596123A (en) | 2018-09-28 |
Family
ID=63619147
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810399828.7A Pending CN108596123A (en) | 2018-04-28 | 2018-04-28 | A kind of fruit hardness detection method and device based on hyperspectral analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108596123A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112465077A (en) * | 2021-02-02 | 2021-03-09 | 中国人民解放军国防科技大学 | Fruit sugar degree nondestructive detection method and device, computer equipment and storage medium |
| CN113686823A (en) * | 2021-08-27 | 2021-11-23 | 西安石油大学 | Water body nitrite content estimation method based on transmission spectrum and PLS-Elman neural network |
| CN114002167A (en) * | 2021-11-02 | 2022-02-01 | 浙江大学 | Method for updating fruit spectral analysis model through deep learning |
| CN114112932A (en) * | 2021-11-08 | 2022-03-01 | 南京林业大学 | Hyperspectral detection method and sorting equipment for maturity of oil-tea camellia fruits based on deep learning |
| CN115308135A (en) * | 2022-08-06 | 2022-11-08 | 福州大学 | Spectrum imaging-based marine dinoflagellate cell concentration detection variable selection method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101403689A (en) * | 2008-11-20 | 2009-04-08 | 北京航空航天大学 | Nondestructive detection method for physiological index of plant leaf |
| CN106644983A (en) * | 2016-12-28 | 2017-05-10 | 浙江大学 | Spectrum wavelength selection method based on PLS-VIP-ACO algorithm |
-
2018
- 2018-04-28 CN CN201810399828.7A patent/CN108596123A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101403689A (en) * | 2008-11-20 | 2009-04-08 | 北京航空航天大学 | Nondestructive detection method for physiological index of plant leaf |
| CN106644983A (en) * | 2016-12-28 | 2017-05-10 | 浙江大学 | Spectrum wavelength selection method based on PLS-VIP-ACO algorithm |
Non-Patent Citations (1)
| Title |
|---|
| 王爽: "基于高光谱散射图像的苹果内部品质预测建模", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112465077A (en) * | 2021-02-02 | 2021-03-09 | 中国人民解放军国防科技大学 | Fruit sugar degree nondestructive detection method and device, computer equipment and storage medium |
| CN113686823A (en) * | 2021-08-27 | 2021-11-23 | 西安石油大学 | Water body nitrite content estimation method based on transmission spectrum and PLS-Elman neural network |
| CN113686823B (en) * | 2021-08-27 | 2024-01-23 | 西安石油大学 | Water nitrite content estimation method based on transmission spectrum and PLS-Elman neural network |
| CN114002167A (en) * | 2021-11-02 | 2022-02-01 | 浙江大学 | Method for updating fruit spectral analysis model through deep learning |
| CN114112932A (en) * | 2021-11-08 | 2022-03-01 | 南京林业大学 | Hyperspectral detection method and sorting equipment for maturity of oil-tea camellia fruits based on deep learning |
| CN115308135A (en) * | 2022-08-06 | 2022-11-08 | 福州大学 | Spectrum imaging-based marine dinoflagellate cell concentration detection variable selection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108596123A (en) | A kind of fruit hardness detection method and device based on hyperspectral analysis | |
| CN110516596B (en) | Octave convolution-based spatial spectrum attention hyperspectral image classification method | |
| Zou et al. | Nondestructive identification of coal and gangue via near-infrared spectroscopy based on improved broad learning | |
| Suryawati et al. | Deep structured convolutional neural network for tomato diseases detection | |
| CN109100323B (en) | Nondestructive quantitative evaluation method for transmission spectrum of apple water core disease | |
| Cassata et al. | The VIMOS Ultra-Deep Survey (VUDS): fast increase in the fraction of strong Lyman-α emitters from z= 2 to z= 6 | |
| CN107271375A (en) | A kind of high spectral image detecting method of quality of mutton index | |
| CN107301380A (en) | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again | |
| CN106604229A (en) | Indoor positioning method based on manifold learning and improved support vector machine | |
| CN107421895A (en) | A kind of water quality parameter retrieving concentration method and apparatus of multiband optimum organization | |
| CN110849828A (en) | Saffron crocus classification method based on hyperspectral image technology | |
| CN109030378A (en) | Japonica rice canopy chlorophyll content inverse model approach based on PSO-ELM | |
| CN104751179A (en) | Multi-target high spectral remote sensing image wave band selection method based on game theory | |
| CN105158200A (en) | Modeling method capable of improving accuracy of qualitative near-infrared spectroscopic analysis | |
| CN103278467A (en) | Rapid nondestructive high-accuracy method with for identifying abundance degree of nitrogen element in plant leaf | |
| CN109766926A (en) | A kind of radar emitter signal intrapulse feature comprehensive estimation method and system | |
| CN110363218A (en) | A kind of embryo's noninvasively estimating method and device | |
| WO2023207453A1 (en) | Traditional chinese medicine ingredient analysis method and system based on spectral clustering | |
| CN114660040A (en) | Method, device, medium and equipment for identifying microbial unicell species | |
| Wiering et al. | Lidar and RGB image analysis to predict hairy vetch biomass in breeding nurseries | |
| CN114973024A (en) | Method and system for monitoring and evaluating comprehensive evaluation index of mechanical cotton picking and defoliating effect | |
| Yu et al. | Hyperspectral technique combined with deep learning algorithm for prediction of phenotyping traits in lettuce | |
| Liu et al. | Detection of Apple Taste Information Using Model Based on Hyperspectral Imaging and Electronic Tongue Data. | |
| Bagherian et al. | Phenotyping agronomic and physiological traits in peanut under mid‐season drought stress using UAV‐based hyperspectral imaging and machine learning | |
| CN108663334A (en) | The method for finding soil nutrient spectral signature wavelength based on multiple Classifiers Combination |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180928 |