CN101226146A - Prediction method for apple hardness based on multiple optical spectrum dispersion image - Google Patents

Abstract

The invention relates to an apple rigidity prediction model refresh method based on multi-spectrum scattering image, belonging to the technical field of agricultural product non-destructive quality determination technology. The invention uses different regression methods to refresh prior model to predict one new sample group, therefore, the invention provides a regression least square method that adds new sample to refresh prior prediction models and provides a sample adding method for obtaining better fruit rigidity prediction result via which the refreshed model can effectively predict fruit rigidity. The invention uses the apple sample groups collected at one time and provided with different storage times to test the apple rigidity prediction effects of the model, which needs small new samples. The invention further compares four sample suspension methods that refresh prediction models, while the result presents that the inventive model refresh method combined with multi-spectrum scattering technique can real-time predict the rigidity of apples from different harvest reasons and areas.

Description

Prediction method for apple hardness based on multiple optical spectrum dispersion image

Technical field

The invention belongs to agricultural product Non-Destructive Testing field, particularly a kind of prediction method for apple hardness based on multiple optical spectrum dispersion image.

Background technology

Hardness is an important indicator of apple internal attribute, and this index can be along with such as climate condition, cultural difference, harvest time or degree of ripeness, and receive post-processing approach and actual storage mode etc. differently have very big difference.The instrumentation style hardness measurement method of standard can be damaged the fruit sample in test process, therefore, can not adopt this method to sorted fruits and classification.The research of new Dynamic Non-Destruction Measurement can overcome the shortcoming of classic method, and bigger using value is arranged.Having a large amount of research aspect the fruit hardness Dynamic Non-Destruction Measurement.The multiple optical spectrum dispersion image technology is one of non-destructive fruit hardness measurement technology.Studies show that multispectral scattering technology is useful in fruit hardness context of detection, and be better than near-infrared spectral analysis technology.

Summary of the invention

The objective of the invention is to propose a kind of prediction method for apple hardness based on multiple optical spectrum dispersion image.It is characterized in that, adopt different homing methods to be used to predict one group of sample.Promptly developed and a kind ofly upgraded existing least square method forecast model by increasing new samples; And propose a kind ofly can obtain fruit Hardness Prediction result's sample increase method preferably, make new model can effectively predict fruit hardness.Gather with the same time at last but have different two groups of different apple sample collection receiving the back storage time and test the Hardness Prediction effect of this model, and the fruit Hardness Prediction result of Forecasting Methodology compares with adopting in the past apple.Concrete steps are as follows

(1) the apple sample material chooses

Before on-test, need choose suitable apple sample, and place it under the room temperature at least 15 hours, during test apple sample is divided into two groups: group I and group II.

(2) data acquisition and processing

The apple sample of I of collection group respectively of multispectral image system that data acquisition is to use is pocket, be equipped with liquid crystal tunable filter (LCTF) and group II obtains its spectral dispersion image, obtain each apple sample at 8 wavelength (650,680,700,740,820,880,910, and 990nm) 8 width of cloth dispersion images under.

Data processing uses the Lorentz distribution function of revising (MLD) to calculate the profile parameters of each dispersion image

$R=a_1+\frac{a_2}{1+{\left(\frac{\left|z\right|}{a_3}\right)}^{a_4}}---\left(1\right)$

In the following formula, R is the optical density of representing with the CCD grey; Z is the scattering distance; a ₁Be optical density asymmetry value, a ₂Be scattering profile peak value; a ₃It is the full scattering width (FWHM) of half peak-peak correspondence; a ₄Be the scattering slope around the FWHM.

After having gathered the spectral dispersion image, the MT hardness-testing device of use standard is measured the hardness number of each fruit, in the zone around each fruit equator identical, be that the probe of 11mm thrusts peeling fruit 9.0mm and measures its hardness number with diameter with the speed of fall of 2mm/s with images acquired.

(3) step of regressive prediction model

Fruit hardness regressive prediction model comprises following 5 steps altogether:

1) be divided into two independently sample sets by group I and group II, sample according to hardness is worth big or small descending sort.In per four apples first three is used for calibration, and the 4th is used for checking.This process finally makes from group I and group II and extracts two checking collection and two calibration collection respectively.

2) under different wavelength, adopt nonlinear regression analysis fitted figure astigmatism to penetrate profile to each sample, thus their MLD parameter in definite equation one.Can obtain a cover totally 32 MLD parameters (4 parameter *, 8 wavelength) from each sample.Further, with the MLD parameter of fruit sample divided by with Teflon normative reference corresponding parameters, obtain the influence that light source fluctuation brings the process of the test of dispersion image to proofread and correct from apple sample.

3) adopt multiple linear regression (MLR) calculating group I calibration collection and group II calibration to concentrate sample MT hardness and MLD hardness r value and calibration criterion poor (SEC) between the two, adopt the MLR method to set up the 8 wavelength forecast models of group I and group II respectively, then it is verified with their checking set pairs separately respectively.Checking is the result compare as obtaining the verification of model result with follow-up described new model update method.The fruit rigidity prediction model of group I can be represented with equation 2:

$F={\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0+c_1{a}_1+c_2{a}_2+\&CenterDot;\&CenterDot;\&CenterDot;+c_n{a}_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j---\left(2\right)$

Wherein, the MT prediction hardness of F for representing with N; a _j(j=1,2 ... n; N=32) be the MLD parameter; Subscript j represents each parameter; c _jFor each wavelength down with 4 MLD parameters in the regression coefficient that interrelates of each parameter; c ₀For being the intercept of unit with N.

4) use the sample correction chosen from group II calibration collection by the group I calibration collection forecast model of building, then verify this more new model with group II checking collection, update algorithm is derived by following process.

The regression coefficient c of forecast model (equation 1) will become c behind the adding new samples _j+ Δ c _j, can be expressed as:

$F+\mathrm{\&Delta;F}=({\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0+{\mathrm{\&Delta;c}}_0)+(c_1+{\mathrm{\&Delta;c}}_1)a_1+({\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">c</figure-callout>}}_2+{\mathrm{\&Delta;c}}_2)a_2+\&CenterDot;\&CenterDot;\&CenterDot;+\left(c_n{\mathrm{\&Delta;c}}_n\right)a_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j+\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&Delta;c}}_j{a}_j---\left(3\right)$

Wherein Δ F is the corrected value of prediction hardness, supposes that the MLD parameter of new samples and MT hardness are:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{2n}& f_2\end{array}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{m1}& a_{\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">m</figure-callout>}2}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)---\left(4\right)$

F wherein _iBe the new MT hardness that adds sample; (i 1,2 ... m) represent each sample; M is the new quantity that adds sample.The model that upgrades must guarantee newly to add the predicted value F of sample _i+ Δ F _i) and with reference to f _iBetween minimum difference, promptly utilize the least square of equation (5) to ask for the Δ c that upgrades forecast model _jValue.

$\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(F_i+{\mathrm{\&Delta;F}}_i-f_i)}^2---\left(5\right)$

According to least square method, can derive following equation to the single order differential of Δ cj by asking for:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& ...& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& ...& \begin{array}{cc}{a}_{2\mathrm{<figure-callout\; id="2"\; label="n\; f"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">n</figure-callout>}}& f_{}{}_2\end{array}\\ ...& ...& ...& ...\\ a_{m1}& a_{\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">m</figure-callout>}2}& ...& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)\left(\begin{array}{c}{\mathrm{\&Delta;<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0\\ {\mathrm{\&Delta;c}}_1\\ ...\\ {\mathrm{\&Delta;c}}_n\end{array}\right)=\left(\begin{array}{c}{t}_0\\ t_1\\ ...\\ t_n\end{array}\right)---\left(6\right)$

$S_{\mathrm{kj}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{a}_{\mathrm{ik}},t_k=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}(f_i-F_i)a_{\mathrm{ik}},a_{i0}=1,k=\mathrm{0,1},\&CenterDot;\&CenterDot;\&CenterDot;,n$

Regression coefficient c behind the adding new samples _jCorrection value delta c _jCan derive by equation 6.With Δ c _jGo into equation 3 and can derive the renewal forecast model.

5) select sample that following four kinds of methods are arranged from group II calibration collection: a) to select at random, b) only choose from ' soft ' fruit group; C) only choose from ' firmly ' fruit group; D) to choose from ' soft ' fruit group and ' firmly ' fruit group simultaneously. sample standard deviations all the group II are pressed descending sort according to the MT hardness number.' soft ' fruit sample since 1 by choosing to front sequence; ' firmly ' fruit sample from last sample by backward the order choose; To join respectively by the sample that above four kinds of systems of selection are chosen by group I calibration collection gained forecast model, to upgrade this forecast model.At last, verify more new model of institute with group II checking collection.

The invention has the beneficial effects as follows and adopt different homing methods to upgrade existing model to be used to predict new one group of sample.Promptly developed a kind of by increasing the recurrence least square method that new samples upgrades existing forecast model; And propose a kind ofly can obtain fruit Hardness Prediction result's sample increase method preferably, make model after the renewal can effectively predict the hardness of fruit.Gather with the same time at last but have different two groups of different apple sample collection receiving the back storage time and test the Hardness Prediction effect of this model apple, and with adopting in the past the fruit Hardness Prediction result of Forecasting Methodology compares, the model update method of this paper can be used for the hardness from the apple in different harvest seasons and different places is carried out real-time estimate together with multispectral scattering technology.

Description of drawings

Fig. 1 is the original radially dispersion image of an apple profile diagram, and Fig. 1 (a) is a representational apple raw scattered image.Fig. 1 (b) is an one-dimensional scattering profile diagram that extracts from dispersion image shown in Fig. 1 (a).

The relation of the MT hardness of Fig. 2 group I and estimation hardness: (a) the calibration collection of group I; (b) the checking collection of group I.

The relation of the MT hardness of Fig. 3 II and estimation hardness: (a) the checking collection of calibration collection (b) the group II of group II.

Fig. 4 is used to organize the checking collection of II by group forecast model that I builds.

Fig. 5 uses the related coefficient of four kinds of different sample adding methods with the new samples adding group I fruit calibration set established model prediction group II of institute apple hardness.

Embodiment

The present invention proposes a kind of apple hardness forecast model update method based on multiple optical spectrum dispersion image.Adopt different homing methods to upgrade existing model to be used to predict new one group of sample.Promptly developed a kind of by increasing the recurrence least square method that new samples upgrades existing forecast model; And propose a kind ofly can obtain fruit Hardness Prediction result's sample increase method preferably, make the model after the renewal can effectively predict fruit hardness.

At first use the apple sample of I of collection group respectively of multispectral image system pocket, that liquid crystal tunable filter (LCTF) is housed and group II to obtain its spectral dispersion image, obtain each apple sample at 8 wavelength (650,680,700,740,820,880,910, and 990nm) 8 width of cloth dispersion images (as shown in Figure 1) at each wavelength under, Fig. 1 (a) is a representational apple raw scattered image.Fig. 1 (b) is an one-dimensional scattering profile diagram that extracts from dispersion image shown in Fig. 1 (a).In addition, corresponding to per 100 apples, a white polytetrafluoroethylpipe disk is arranged all, its dispersion image is used as the reference picture of calibration light source influence of fluctuations.

Spectral dispersion view data to above-mentioned collection is handled, and uses the Lorentz distribution function of revising (MLD) to calculate the profile parameters of each dispersion image

$R=a_1+\frac{a_2}{1+{\left(\frac{\left|z\right|}{a_3}\right)}^{a_4}}---\left(1\right)$

Wherein, R is the optical density of representing with the CCD grey; Z is the scattering distance; a ₁Be optical density asymmetry value, a ₂Be scattering profile peak value; a ₃It is the full scattering width (FWHM) of half peak-peak correspondence; a ₄Be the scattering slope around the FWHM.

After having gathered the spectral dispersion image, the MT hardness-testing device of use standard is measured the hardness number of each fruit, in the zone around each fruit equator identical, be that the probe of 11mm thrusts peeling fruit 9.0mm and measures its hardness number with diameter with the speed of fall of 2mm/s with images acquired.

Fruit hardness regressive prediction model comprises following 5 steps altogether:

1) by group I and every group of 450 apple sample of group II, all be divided into two independently sample sets separately, sample according to hardness is worth big or small descending sort.In per four apples first three is used for calibration, and the 4th is used for checking.This process finally makes from group I and group II and extracts two checking collection and two calibration collection respectively.

2) under different wavelength, adopt nonlinear regression analysis fitted figure astigmatism to penetrate profile to each sample, thus their MLD parameter in definite equation one.Can obtain a cover totally 32 MLD parameters (4 parameter *, 8 wavelength) from each sample.Further, with the MLD parameter of fruit sample divided by with Teflon normative reference corresponding parameters, obtain the influence that light source fluctuation brings the process of the test of dispersion image to proofread and correct from apple sample.

3) adopt multiple linear regression (MLR) calculating group I calibration collection and group II calibration to concentrate sample MT hardness and MLD hardness r value and calibration criterion poor (SEC) between the two, adopt the MLR method to set up the 8 wavelength forecast models of group I and group II respectively, then it is verified with their checking set pairs separately respectively.Checking is the result compare as obtaining the verification of model result with follow-up described new model update method.The fruit rigidity prediction model of group I can be represented with equation 2:

$F={\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0+c_1{a}_1+c_2{a}_2+\&CenterDot;\&CenterDot;\&CenterDot;+c_n{a}_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j---\left(2\right)$

Wherein, the MT prediction hardness of F for representing with N; Aj (j=1,2 ... n; N=32) be the MLD parameter; Subscript j represents each parameter; c _jFor each wavelength down with 4 MLD parameters in the regression coefficient that interrelates of each parameter; c ₀For being the intercept of unit with N.

4) use the sample correction chosen from group II calibration collection by the group I calibration collection forecast model of building, then verify this more new model with group II checking collection, update algorithm is derived by following process.

The regression coefficient c of forecast model (equation 1) will become c behind the adding new samples _j+ Δ c _j, can be expressed as:

$F+\mathrm{\&Delta;F}=({\mathrm{<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0+{\mathrm{\&Delta;c}}_0)+(c_1+{\mathrm{\&Delta;c}}_1)a_1+({\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">c</figure-callout>}}_2+{\mathrm{\&Delta;c}}_2)a_2+\&CenterDot;\&CenterDot;\&CenterDot;+\left(c_n{\mathrm{\&Delta;c}}_n\right)a_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j+\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&Delta;c}}_j{a}_j---\left(3\right)$

Wherein Δ F is the corrected value of prediction hardness, supposes that the MLD parameter of new samples and MT hardness are:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{2n}& f_2\end{array}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{m1}& a_{\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">m</figure-callout>}2}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)---\left(4\right)$

F wherein _iBe the new MT hardness that adds sample; (i=1,2 ... m) represent each sample; M is the new quantity that adds sample.The model that upgrades must guarantee newly to add the predicted value (F of sample _i+ Δ F _i) and reference value f _iBetween minimum difference, promptly utilize the least square of equation (5) to ask for the Δ c that upgrades forecast model _jValue.

$\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(F_i+{\mathrm{\&Delta;F}}_i-f_i)}^2---\left(5\right)$

According to least square method, can derive following equation to the single order differential of Δ cj by asking for:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{2n}& f_2\end{array}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{m1}& a_{\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="S2008100568531E00011.png"\; state="\{\{state\}\}">m</figure-callout>}2}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)\left(\begin{array}{c}{\mathrm{\&Delta;<figure-callout\; id="0"\; label="c"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">c</figure-callout>}}_0\\ {\mathrm{\&Delta;c}}_1\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ {\mathrm{\&Delta;c}}_n\end{array}\right)=\left(\begin{array}{c}{t}_0\\ t_1\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ t_n\end{array}\right)---\left(6\right)$

$s_{\mathrm{kj}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{a}_{\mathrm{ik}},t_k=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}(f_i-F_i)a_{\mathrm{ik}},a_{\mathrm{<figure-callout\; id="0"\; label="i"\; filenames="S2008100568531E00021.png"\; state="\{\{state\}\}">i</figure-callout>}0}=1,k=\mathrm{0,1},\&CenterDot;\&CenterDot;\&CenterDot;,n$

Regression coefficient c behind the adding new samples _jCorrection value delta c _jCan derive by equation 6.With Δ c _jInsert equation (3) and can derive the renewal forecast model.

5) select sample that following four kinds of methods are arranged from group II calibration collection: a) to select at random, b) only choose from ' soft ' fruit group; C) only choose from ' firmly ' fruit group; D) to choose from ' soft ' fruit group and ' firmly ' fruit group simultaneously. sample standard deviations all the group II are pressed descending sort according to the MT hardness number.' soft ' fruit sample since 1 by choosing to front sequence; ' firmly ' fruit sample from last sample by backward the order choose; To join respectively by the sample that above four kinds of systems of selection are chosen by group I calibration collection gained forecast model, to upgrade this forecast model.At last, verify more new model of institute with group II checking collection.

Above-mentioned steps is to enumerate application example illustrated.

Prediction in 1 group

Fig. 2 (a) shows, the r=0.89 between the MT hardness number of the calibration sample of group I and the estimation hardness number, SEC=5.74N.For the checking sample set, model prediction apple hardness value has r=0.89, SEP=6.04 N (Fig. 2 (b)).These results show that the MLR forecast model based on dispersion image MLD parameter can be used to predict apple sample then very easily.

Similarly, group II calibration collection is used to set up calibrating patterns, and group II checking collection is as verification model.The MT hardness number and under 8 wavelength and utilize the calibration set sample to return to draw the MLD parameter.Fig. 3 (a) and the result who (b) is respectively calibration and verifies.The result shows this model for calibration sample r=0.87 and SEC=5.91N, and for checking sample r=0.87, SEP=6.07N is a good forecast model.

Fig. 2 (a) seems similar with Fig. 3 (a), but, the forecast model of being set up by group I can not be directly used in the sample (Fig. 4) among the prediction group II, and this is because two sample sets have different fruit conditions, so this property difference of two groups will have a great difference.Like this, can obtain good predicting the outcome to the checking sample of this group based on the apple hardness forecast model that calibration samples in a certain group is set up, but not good at the estimated performance of another group.Therefore for guaranteeing that institute's established model to new group estimated performance, must be upgraded this model by the sample that adds in the new family.

2 predictions that straddle over year

Table 1 is depicted as in different ways and originally and with it joins the comparative result that group I calibration collection is used to predict fruit hardness from the sampling of group II calibration collection.The new samples number of adding is shown in table 1 first tabulation.

In the table 1 from per four row of secondary series to the 17 row respectively expression choose predicting the outcome of every kind of mode increasing sample, promptly picked at random, only choose from ' soft ' fruit group; Only choose from ' firmly ' fruit group; Choose from ' soft ' fruit group and ' firmly ' fruit group simultaneously.

Table 1 has provided at the correction of each sample system of selection and checking result.In the comparison sheet 1 the 2nd, 4,6,8,10th, 12,14 and related coefficient (r) value of th row, perhaps relatively the 3rd, 5,7,9, the standard deviation numerical value of 11,13,15 and 17 row can be seen, the mode of choosing sample from ' soft ' fruit group and ' firmly ' fruit group that replaces has best modelling effect.

Table 1 is concentrated from group II calibration and is selected sample adding group I to calibrate the comparison of four kinds of sample systems of selection of collection to prediction group II fruit hardness

Figure 5 shows that with adding the sample number purpose change the prediction r value of apple hardness corresponding to the change curve under four kinds of different sample selection modes.Along with the r value that increases of number of samples shows rising tendency, but selecting sample to be used to upgrade under this mode of calibration model at random, the rising tendency of curve is also unstable; Only from ' soft ' fruit group or the mode of only choosing sample from ' firmly ' fruit group has better and more stable model predict the outcome, along with adding the sample number purpose increase the r value stabilization, continue to increase.And replace under ' soft ' fruit group and ' firmly ' fruit group are chosen the mode of sample, can obtain best model prediction result.Along with adding the sample number purpose increase the r value stabilization and increase, r value arrives stationary value 0.86 when adding, sample number was 148, predicts standard deviation SEP=6.11N (table 1 and Fig. 5).To group II predict the outcome with at the r=0.87 that predicts the outcome of sample then, SEP=6.07N is close, the used forecast model of the latter is based upon on 338 calibration samples bases by group II group.

Use the related coefficient (r) of four kinds of different sample adding methods with the new samples adding group I fruit calibration set established model prediction group II of institute apple hardness

Result of study shows that if will obtain close predicting the outcome, this model update method is compared and relied on new group great amount of samples to rebulid the method for model, and required new samples is counted much less.Further, this research poises mode to four kinds of samples that upgrade forecast model and compares, result of study shows, these four kinds of modes have very big difference at apple hardness aspect predicting the outcome, in these four kinds of modes, the mode of choosing sample from ' soft ' fruit group and ' firmly ' fruit group that replaces has best model modification effect.Adopt the model update method of this paper, make and to organize the hardness of apple sample for group forecast model that I builds can be used for prediction group II, test result in 2005 is correlation coefficient r=0.86, the prediction standard deviation is 6.11N, and this result is close with predicting the outcome of the interior calibration samples institutes established model in a large number of dependence group II group.Therefore, the model update method of this paper can be used for to carrying out real-time estimate from the different harvest seasons with the hardness of the apple in different places together with multispectral scattering technology.

Claims (3)

1. apple hardness forecast model new method based on multiple optical spectrum dispersion image, it is characterized in that, adopt different homing methods to upgrade existing model, promptly developed a kind of by increasing the recurrence least square method that new samples upgrades existing forecast model to be used to predicting new one group of sample; And propose a kind ofly can obtain fruit Hardness Prediction result's sample increase method preferably, make model after the renewal can effectively predict the hardness of fruit, gather with the same time at last but have different two groups of different apple sample collection receiving the back storage time and test the Hardness Prediction effect of this model, and the fruit Hardness Prediction result of Forecasting Methodology compares with adopting in the past apple.

2. according to the described apple hardness forecast model update method of claim 1, it is characterized in that described apple hardness forecast model update method concrete steps are as follows based on multiple optical spectrum dispersion image:

(1) the apple sample material chooses

Before on-test, need choose suitable apple sample, and place it under the room temperature at least 15 hours, choose two groups of apple sample during test;

(2) data acquisition and processing

The apple sample of I of collection group respectively of multispectral image system that data acquisition is to use is pocket, be equipped with liquid crystal tunable filter (LCTF) and group II obtains its spectral dispersion image, obtains 8 width of cloth dispersion images of each apple sample under 8 wavelength;

Data processing uses the Lorentz distribution function of revising (MLD) to calculate the profile parameters of each dispersion image

$R=a_1+\frac{a_2}{1+{\left(\frac{\left|z\right|}{a_3}\right)}^{a_4}}---\left(1\right)$

In the following formula, R is the optical density of representing with the CCD grey; Z is the scattering distance; a ₁Be optical density asymmetry value, a ₂Be scattering profile peak value; a ₃It is the full scattering width (FWHM) of half peak-peak correspondence; a ₄Be the scattering slope around the FWHM;

After having gathered the spectral dispersion image, the MT hardness-testing device of use standard is measured the hardness number of each fruit, in the zone around each fruit equator identical, be that the probe of 11mm thrusts peeling fruit 9.0mm and measures its hardness number with diameter with the speed of fall of 2mm/s with images acquired;

(3) step of regressive prediction model, fruit hardness regressive prediction model comprise following 5 steps altogether:

1) by group I and every group of 450 apple sample of group II, all be divided into two independently sample sets separately, sample according to hardness is worth big or small descending sort, in per four apples first three is used for calibration, the 4th is used for checking, and this process finally makes from group I and group II and extracts two checking collection and two calibration collection respectively;

2) under different wavelength, adopt nonlinear regression analysis fitted figure astigmatism to penetrate profile, thereby their MLD parameter in definite equation one can obtain a cover totally 32 MLD parameter=4 parameter *, 8 wavelength from each sample to each sample; Further with the MLD parameter of fruit sample divided by with Teflon normative reference corresponding parameters, obtain the influence that light source fluctuation brings the process of the test of dispersion image to proofread and correct from apple sample;

3) adopt multiple linear regression (MLR) calculating group I calibration collection and group II calibration to concentrate sample MT hardness and MLD hardness r value and calibration criterion poor (SEC) between the two, adopt the MLR method to set up the 8 wavelength forecast models of group I and group II respectively, then it is verified with their checking set pairs separately respectively, checking is the result compare as obtaining the verification of model result with follow-up described new model update method, and the fruit rigidity prediction model of group I can be represented with equation (2):

$F=c_0+c_1{a}_1+c_2{a}_2+\&CenterDot;\&CenterDot;\&CenterDot;+c_n{a}_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j---\left(2\right)$

Wherein, the MT prediction hardness of F for representing with N; a _j(j=1,2 ... n; N=32) be the MLD parameter; Subscript j represents each parameter; c _jFor each wavelength down with 4 MLD parameters in the regression coefficient that interrelates of each parameter; c ₀For being the intercept of unit with N;

4) use the sample correction chosen from group II calibration collection by the group I calibration collection forecast model of building, then verify this more new model with group II checking collection, update algorithm is derived by following process,

The regression coefficient c of forecast model (equation 1) will become c behind the adding new samples _j+ Δ c _j, can be expressed as:

$F+\mathrm{\&Delta;F}=(c_0+{\mathrm{\&Delta;c}}_0)+(c_1+{\mathrm{\&Delta;c}}_1)a_1+(c_2+{\mathrm{\&Delta;c}}_2)a_2+\&CenterDot;\&CenterDot;\&CenterDot;+\left(c_n{\mathrm{\&Delta;c}}_n\right)a_n=\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{c}_j{a}_j+\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&Delta;c}}_j{a}_j---\left(3\right)$

Wherein Δ F is the corrected value of prediction hardness, supposes that the MLD parameter of new samples and MT hardness are:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{2n}& f_2\end{array}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{m1}& a_{m2}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)---\left(4\right)$

F wherein _iBe the new MT hardness that adds sample; (i=1,2 ... m) represent each sample; M is the new quantity that adds sample.The model that upgrades must guarantee newly to add the predicted value (f of sample _i+ Δ f _i) and reference value f _iBetween minimum difference, promptly utilize the least square of equation (5) to ask for the Δ c that upgrades forecast model _jValue,

$\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(F_i+{\mathrm{\&Delta;F}}_i-f_i)}^2---\left(5\right)$

According to least square method, can derive following equation to the single order differential of Δ cj by asking for:

$\left(\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{1n}& f_1\end{array}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{2n}& f_2\end{array}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{m1}& a_{m2}& \&CenterDot;\&CenterDot;\&CenterDot;& \begin{array}{cc}{a}_{\mathrm{mn}}& f_m\end{array}\end{array}\right)\left(\begin{array}{c}{\mathrm{\&Delta;c}}_0\\ {\mathrm{\&Delta;c}}_1\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ {\mathrm{\&Delta;c}}_n\end{array}\right)=\left(\begin{array}{c}{t}_0\\ t_1\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ t_n\end{array}\right)---\left(6\right)$

$s_{\mathrm{kj}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{a}_{\mathrm{ik}},t_k=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}(f_i-F_i)a_{\mathrm{ik}},a_{i0}=1,k=\mathrm{0,1},\&CenterDot;\&CenterDot;\&CenterDot;,n$

Regression coefficient c behind the adding new samples _jCorrection value delta c _jCan derive by equation 6.With Δ c _jInsert equation 3 and can derive the renewal forecast model;

5) select sample that following four kinds of methods are arranged from group II calibration collection: a) to select at random, b) only choose from ' soft ' fruit group; C) only choose from ' firmly ' fruit group; D) to choose from ' soft ' fruit group and ' firmly ' fruit group simultaneously. sample standard deviations all the group II are pressed descending sort according to the MT hardness number.' soft ' fruit sample since 1 by choosing to front sequence; ' firmly ' fruit sample from last sample by backward the order choose; To join respectively by the sample that above four kinds of systems of selection are chosen by group I calibration collection gained forecast model,, at last, verify more new model of institute with group II checking collection to upgrade this forecast model.

3. according to the described apple hardness forecast model update method of claim 1, it is characterized in that described 8 wavelength are 650,680,700,740,820,880,910, and 990nm based on multiple optical spectrum dispersion image

Images