CN110243765A - The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model - Google Patents
The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model Download PDFInfo
- Publication number
- CN110243765A CN110243765A CN201910590833.0A CN201910590833A CN110243765A CN 110243765 A CN110243765 A CN 110243765A CN 201910590833 A CN201910590833 A CN 201910590833A CN 110243765 A CN110243765 A CN 110243765A
- Authority
- CN
- China
- Prior art keywords
- fruit
- layer
- double
- model
- photon
- 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
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/02—Food
- G01N33/025—Fruits or vegetables
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N2201/00—Features of devices classified in G01N21/00
- G01N2201/12—Circuits of general importance; Signal processing
- G01N2201/129—Using chemometrical methods
- G01N2201/1296—Using chemometrical methods using neural networks
Abstract
The fruit EO-1 hyperion quality detecting method for the photon transmission simulation based on fruit double-layer plate model that the invention discloses a kind of, comprising the following steps: building fruit double-layer plate model；Photon transmission simulation, which is carried out, based on monte carlo method obtains light brightness distribution figure；Light brightness distribution figure is input in the pre- convolutional neural networks mixed up and carries out convolution and deconvolution；Analysis convolution and deconvolution visit distance as a result, obtaining source；Distance is visited using the source, and Quality Detection is carried out to fruit.Double-layer plate model can be substituted for surface model, reduce the complexity of surface model in this way, while guarantee accuracy.Compared to apple surface model, simulated time can be reduced by simplified double-layer plate model, while is applicable in this method (the stone fruit of the pulp such as pears thickness) with more universality.
Description
Technical field
The present invention relates to detection field, specifically a kind of photon transmission simulation based on based on fruit double-layer plate model
Fruit EO-1 hyperion quality detecting method.
Background technique
Portable spectrum detector because its is small in size, portability, it is easy to operate the features such as be widely applied.At present for
The research of fruit portable detector is existing very much, but the less light source for fruit portable detector and distance of popping one's head in
Research, Tsinghua University's Ding Haishu seminar benefit wear light being averaged in multi-layer biological body tissue model using monte carlo method
Saturating depth, photon average flight distance and spatial sensitivity profile are emulated and have been analyzed, the results showed that light source and light detection
The distance between device is an important factor for influencing detection sensitivity, to propose for specific organization position, there is optimal light
Source-detector distance, i.e. source are visited apart from this concept.
At present there are also for source visit distance research, Guo Zhiming using hyperspectral technique carry out attributional analysis when,
Square region and corresponding border circular areas provided with the several different horizontal pixel sizes in 50,100,150,200 and 250.It is right
The soluble solid Quantitative Analysis Model that Partial Least Squares establishes apple is respectively adopted in these different zones.Demonstrate EO-1 hyperion
The selection of detection zone shapes and sizes will affect testing result in image, finally show that area-of-interest transverse direction MAXIMUM SELECTION is
The model result obtained when 150 pixel is best, while the optimum modeling and prediction result that find rounded interested area are better than side
The result of shape area-of-interest.Zhao Fan extracts region to the magic heart precision of fruit pol detection to understand spectrum in high spectrum image
Influence, be extracted the square area of 10*10,20*20,30*30 (pixel * pixel) respectively using " Hua You " Kiwi berry as object
Pol fitting is carried out, the increase that discovery spectrum extracts area is able to ascend the related coefficient of pol fitting, illustrates in high-spectrum
Suitable spectrum is selected to extract the precision of prediction that region helps to improve model as in.
These researchs for visiting distance to source are carried out on the basis of measured data.It is studied using measured data
There is two o'clock disadvantage, first is that since research in this respect can not obtain a large amount of enough data, and influenced by kind batch, it obtains
Result out has significant limitations.Second is that in the case where data are less, measured data there is also inevitable error,
More cause the not accurate enough of result.Therefore it proposes to replace measured data using based on the data of the generation of monte carlo method
Method.
CN109856064A proposes a kind of apple EO-1 hyperion quality detecting method based on photon transmission simulation, it includes
Following steps: (1) apple model is constructed based on Ellipsoidal Surface equation；(2) point light source is directed at apple equator, distance selection is visited in source
1.5mm~10.15mm carries out the simulation of Monte Carlo photon transmission based on apple model；(2.1) photon initializes；(2.2) it calculates
The photon direction of motion and step-length, while crossing the border and judging with out-of-bounds；(2.21a) crosses the border judgement, crosses the border, carry out step
(2.21b) otherwise carries out step (2.22a)；(2.21b) out-of-bounds judgement, out-of-bounds then carry out out-of-bounds processing and go to step (2.5),
Otherwise step (2.22b) is carried out；(2.22a) photonic absorption and scattering calculate, and are transferred to step (2.3)；The refraction of (2.22b) photon,
Reflection calculates, and is transferred to step (2.3)；(2.3) judge whether photon weight is too small, is, carry out step (2.4), otherwise return to step
Suddenly (2.2)；(2.4) judge whether photon lives or dies, be, carry out step (2.5), otherwise return step (2.2)；(2.5) judgement be
No is the last one photon, is to terminate, otherwise return step (2.1)；(3) a large amount of photons and different optical parameters is taken to combine weight
Multiple step (2) obtain the noise-free light intensity map of apple model.
The program constructs apple model using Ellipsoidal Surface equation, and apple surface model is in the simulation to practical apple tissue
Reduction degree is higher, and the data simulated are more accurate, but this model is excessively complicated in operation, this to exchange essence for complexity
The method of exactness practical degree in the life class research having is not high.In addition, universality is very just to a kind of fruit tectonic model
It is low.
Summary of the invention
The present invention is directed to the problem of background technique, proposes a kind of photon biography based on fruit double-layer plate model
The fruit EO-1 hyperion quality detecting method of defeated simulation.
Technical solution:
A kind of fruit EO-1 hyperion quality detecting method of the photon transmission simulation based on fruit double-layer plate model, it includes
Following steps:
(1) fruit double-layer plate model is constructed；
(2) photon transmission simulation is carried out based on monte carlo method and obtains light brightness distribution figure；
(3) light brightness distribution figure is input in the pre- convolutional neural networks mixed up and carries out convolution and deconvolution；
(4) analysis convolution and deconvolution visit distance as a result, obtaining source；
(5) distance being visited using the source, Quality Detection is carried out to fruit.
Preferably, in step (2), photon transmission simulation is carried out to double-layer plate model based on monte carlo method, specifically
Step is:
(2.1) photon initializes；
(2.2) the photon direction of motion and step-length are calculated, while carrying out cross the border judgement (2.3)；
(2.3) it crosses the border judgement, crosses the border, carry out step (2.4b), otherwise carry out step (2.4a)；
(2.4a) photonic absorption and scattering calculate, and are transferred to step (2.5)；
The refraction of (2.4b) photon, reflection calculate, and are transferred to step (2.5)；
(2.5) judge whether photon weight is too small, is, carry out step (2.6), otherwise return step (2.2)；
(2.6) judge whether photon lives or dies, be, carry out step (2.7), otherwise return step (2.2)；
(2.7) judge whether it is the last one photon, be, terminate, otherwise return step (2.1)；
(2.8) it takes a large amount of photons and different optical parameters to combine and repeats step (2), obtain the nothing of fruit double-layer plate model
Noise light intensity map.
Preferably, in step (1), double-layer plate model is two parallel plates, indicates pericarp, two plates between two plates
Between distance, that is, peel thickness be denoted as d1, pulp thickness is denoted as d2=∞ mm；First plate, that is, pericarp absorption coefficient is denoted as μa1,
The scattering coefficient of pericarp is denoted as μs1；Second plate, that is, pulp absorption coefficient is denoted as μa2, the scattering coefficient of pulp is denoted as μs2。
Preferably, in step (3), obtained light brightness distribution figure is input to trained neural network, specifically:
I pericarp absorption coefficient μa1It falls into 5 types, pulp absorption coefficient μa2It is divided into 4 classes, pericarp scattering coefficient μs1It falls into 5 types,
Pulp scattering coefficient μs2It falls into 5 types, median is respectively taken to be combined 500 groups of optical parameter combinations of acquisition；
Ii carries out Monte Carlo simulation on double-layer plate model, obtains 8000 surface surfaces of intensity distribution, extracts wherein
6000 are used as training set, and 2000 are used as test subset that convolutional neural networks is used to be trained.
Preferably, in step (4), convolutional neural networks are finely adjusted using back-propagation algorithm, training 1000 times, directly
To network convergence；500 are randomly selected in all data, carries out the operation of convolution sum deconvolution, after saving deconvolution calculating
Result；Adjacent picture is subtracted each other；Point-by-point statistics absolute difference is greater than 0.001 number, final to determine that distance is visited in source.
Specifically, convolutional neural networks structure includes:
First layer input layer carries out convolution to image using the convolution of 30 2*2 sizes, and convolution kernel moving step length is 1, volume
Product mode uses valid, initial learning rate 0.07, batch size 50, total the number of iterations 2000 times；
Second layer pond layer is averaged pond using 2*2；
Third layer uses the convolution kernel of 20 3*3 sizes, and convolution kernel moving step length is 1；
The 4th layer of convolution kernel using 20 3*3 sizes, convolution kernel moving step length are 1, are operated without pondization；
Layer 5 carries out convolution to input data using 15 2*2 convolution kernels；
The output of layer 6 network is pulled into one-dimensional vector by full articulamentum, and output data size is 60*1；
Layer 7 is network output layer.
Preferably, the fruit is the stone fruit of pulp thickness.
Preferably, the fruit is apple or pears.
Preferably, when the fruit is apple, in step (2) i, pericarp absorption coefficient μa1It takes: 0.7,1.55,1.85,
2.25,4.5, pulp absorption coefficient μa2: 0.5,1.15,1.45,5, pericarp scattering coefficient μs1: 30,67.5,82.5,100,190,
Pulp scattering coefficient μs2: 12,25.5,29,34,56, unit: mm-1。
Beneficial effects of the present invention
It compares and is analyzed by EO-1 hyperion measured data, this method is on the basis of simulation model by simulation number
According to being analyzed, the various experimental errors in measurement process are eliminated, accuracy is higher.
The similarity of double-layer plate model plot of light intensity and apple surface model plot of light intensity similarity in effective range is high,
Comparison range is bigger, and double-layer plate model can be substituted for surface model, reduces the complexity of surface model in this way, protects simultaneously
Demonstrate,prove accuracy.
Compared to apple surface model, simulated time can be reduced by simplified double-layer plate model, while also making
This method (is applicable in the stone fruit of the pulp such as pears thickness) with more universality.
Detailed description of the invention
Fig. 1 is double-layer plate model schematic
Fig. 2 is double-layer plate model Monte Carlo simulation figure
Fig. 3 is the plot of light intensity of double-layer plate model Monte Carlo simulation
Fig. 4 is bright figure sectional drawing in Fig. 3
Fig. 5 is convolutional network structural schematic diagram
Fig. 6 a is the Controlling UEP of 10 pixel sections in Fig. 4
Fig. 6 b is the Controlling UEP of 20 pixel sections in Fig. 4
Fig. 6 c is the Controlling UEP of 30 pixel sections in Fig. 4
Fig. 6 d is the Controlling UEP of 40 pixel sections in Fig. 4
Fig. 7 is Pearson's charts for finned heat
Pearson's coefficient curve that Fig. 8 is Ua1 when being 5.5
Fig. 9 is the feature difference and location diagram of double-layer plate model deconvolution result
Figure 10 is that distance results schematic diagram is visited in three kinds of method sources in embodiment
Figure 11 is the apple illustraton of model of surface equation
Figure 12 is the ray trace figure of Figure 11 model
Figure 13 is the feature difference and location diagram of surface model deconvolution result
Figure 14 is that interested position selects schematic diagram
Specific embodiment
Below with reference to embodiment, the invention will be further described, and but the scope of the present invention is not limited thereto:
Embodiment part is in a manner of curve model, double-layer plate model, three kinds of actual measurement verifying, to technical scheme
Effect compares verifying, and the fruit selects apple.
1, the photon transmission simulation based on monte carlo method of two kinds of different models
1.1 curve model
Curve model joins existing Ellipsoidal Surface using the Apple formative method deformed based on ellipsoid, this method
Number equation adds interference function, the deformation to surface equation is realized, to achieve the effect that be close with practical apple shape.Now will
Its modeling process is summarized as follows:
1) formal parameter a=r is setx, b=ry, c=rz, wherein a, b, c are that ellipsoid apple model is three corresponding respectively
The axial length later period generally realizes the control to apple model size by changing the value of a, b, c.
2) its upper and lower ends has different degrees of recess for apple, thus need to Ellipsoidal Surface along u parameter to
Interior of articles carries out negative exponent interference.
According to the feature of the recess of upper end, interference function is added for it:
g1(u, w)=- p1*exp (- 2u) (1-1)
According to the recess feature of lower end, interference function is added for it:
g2(u, w)=- p2*exp (2u) (1-2)
In conjunction with the interference function of upper and lower ends, that is, obtain the resultant interference function of apple
G (u, w)=g1(u,w)+g2(u, w)=- p1*exp (- 2u)-p2*exp (2u) (1-3)
In 1-3 formula, p1, p2 are the parameter for controlling Ellipsoidal Surface upper and lower ends sinking degree.Thus apple outer layer shape is obtained
The parametric equation of state are as follows:
xQ(u, w)=a cosu cos w+g (u, w) a (u, w)
yQ(u, w)=b cosu sin w+g (u, w) b (u, w)
zp(u, w)=c sin u+g (u, w) c (u, w)
(-π/2≤u≤π/2,0≤w≤2π)
The apple model such as Figure 11 constructed selects an apple model and one group of optical parameter (pericarp absorption at random
Coefficient μa1For 0.70mm-1, pulp absorption coefficient μa2For 0.50mm-1, pericarp scattering coefficient μs1For 30.00mm-1, pulp scattering system
Number μs2For 12.00mm-1), set number of photons as 100,000, the motion profile after apple is entered with monte carlo method simulated photons,
It is as shown in figure 12 that ray trace figure can be obtained.
The motion profile of most of photon is uniformly scattered around as the center of circle using incidence point and concentrates on incidence
Around point.Most of photon trajectories concentrate under apple surface layer, and closer to apple internal fruit stone, photon trajectories are fewer, only
Extremely least a portion of photon energy reaches fruit stone, this is consistent substantially with reality.It can quickly be obtained not by Monte Carlo simulation
With the noise-free light intensity map under optics parameter combination, compensate for that actual measurement noise is big, acquisition data cover face is small, consumption
The shortcomings that taking manpower and material resources.
1.2 double flat Slabs
Though apple surface model can effectively simulated photons transmit, this model is excessively complicated, and only for apple
A kind of this fruit of fruit, the scope of application is too narrow, does not have universality.Apple tissue multilayered model Monte Carlo simulation figure has following spy
Point: firstly, photon trajectories concentrate on around incidence point, range very little, meanwhile, most of photon motion range is only in Apples
Only only a few photon can reach apple core after skin and pulp this two layers of photon enter apple epidermis.It is right for these features
Apple surface model is simplified, and double-layer plate model is constructed, and is intended to carry out the simulation of photon transmission on the basis of this model.
Fig. 1 is simplified double-layer plate model, and d1 represents peel thickness, since photon seldom reaches fruit stone, therefore double-layer plate second
The thickness approximation of layer regards infinitely great as, thus achievees the purpose that model simplification.
It equally selects at random and one group of optical parameter (pericarp absorption coefficient μa1For 0.70mm-1, pulp absorption coefficient μa2For
0.50mm-1, pericarp scattering coefficient μs1For 30.00mm-1, pulp scattering coefficient μs2For 12.00mm-1), d1=0.4mm is taken, in order to
Relatively clear displaying ray trace, we, which emit 20 photons and are simulated with monte carlo method, can obtain photon
Motion profile figure is as shown in Figure 2.
2, the bright distribution map of two kinds of different models
2.1 apple surface models
Since apple model is surface model, the bright figure on obtained evolution surface is also curved surface, and analysis is got up more
Complexity, therefore obtained bright figure is mapped, it is projected on the image of 50 pixel *, 50 pixel according to a certain percentage, wherein often
A pixel represents 1mm, completes available plot of light intensity after projection.
2.2 double-layer plate models
Double-layer plate model is also mapped to the image of 50 pixel *, 50 pixel in the same scale, and Fig. 3 is the light that simulation obtains
Strong distribution map example.
3, the source based on convolutional neural networks is visited distance and is calculated
The 3.1 convolutional network structures based on apple surface model
When irradiating apple using point light source, due to there is a large amount of mirror-reflection at recent photograph exit point, the quality of apple is visited
Survey does not have major significance, and apart from radiation source remote position since brightness is lower, signal that detector receives compared with
It is weak, also do not have detection value.It is thus necessary to determine that a suitable distance enables probe to obtain in this position sensing
Best Effect on Detecting.This range is referred to as source and visits position.CN109856064A passes through EO-1 hyperion measurement research apple product
Distance is visited in source when quality detection, and light source incidence position is selected in equator by her, acquires the point light source irradiation image of 200 samples,
It is mapped to the surface of intensity distribution of 50*50 pixel.The pol moisture data of these samples is acquired simultaneously.It adjusts the distance central point
Pol and moisture apart from different pixel and sample do correlation analysis.Fit the luminous intensity and apple product of different pixels
The degree of correlation of matter.Thinking that optimum detection regional location is by comprehensive analysis is the center of circle, 3~13 pixels apart from spot center point
Point is in the annular region of radius.It is distance center point 2.7mm~11.7mm through conversion.Convolutional Neural net is used in the present embodiment
It visits the source spy distance that distance is obtained with measured data and compares in the source that network analyzes.
In order to compare with measured data, we quote optical parameter value obtained in actual measurement, take pericarp transmission coefficient
Ua1:0.7,1.55,1.85,2.25,4.5 (mm-1), pulp transmission coefficient ua2:0.5,1.15,1.45,5 (mm-1), pericarp dissipate
Penetrate coefficient us1:30,67.5,82.5,100,190 (mm-1), pulp scattering be us2:12,25.5,29,34,56 (mm-1), preceding
Monte Carlo simulation is carried out on the apple simulation model of text building, shares 5*4*5*5=500 optical parameter combination, 120 apples
Fruit model obtains 60000 emulation apple surface surfaces of intensity distribution, randomly selects the analogous diagram conduct of 100 models therein
Training set, the analogous diagram of 20 models share 50000, test set has 10000 pictures as test set in training set.
Training set is inputted default network to be trained, adjusts every layer of convolution kernel number, the classification accuracy of comparing cell.After training
Network degree appropriate is increased to the inverting accuracy rate of apple optical parameter.Finally obtained network structure is as shown in Figure 5:
First layer is input layer, carries out convolution to image using the convolution of 30 2*2 sizes, and convolution kernel moving step length is 1,
Convolution mode uses valid, initial learning rate 0.07, batch size 50, total the number of iterations 2000 times.Second layer pond layer, is adopted
It is averaged pond with 2*2.Third layer uses the convolution kernel of 20 3*3 sizes, and convolution kernel moving step length is 1.4th layer uses 20
The convolution kernel of 3*3 size, convolution kernel moving step length are 1, are operated without pondization.Layer 5 is using 15 2*2 convolution kernels to defeated
Enter data and carries out convolution.The output of layer 6 network is pulled into one-dimensional vector by full articulamentum, and output data size is 60*1.Finally
One layer is network output layer.
The 3.2 convolutional network structures based on apple double-layer plate model
Double-layer plate model equally using optical parameter value obtained in actual measurement, takes pericarp transmission coefficient ua1:0.7,
1.55,1.85,2.25,4.5(mm-1), pulp transmission coefficient ua2:0.5,1.15,1.45,5 (mm-1), pericarp scattering coefficient
us1:30,67.5,82.5,100,190(mm-1), pulp scattering be us2:12,25.5,29,34,56 (mm-1), above
To double-layer plate model on carry out Monte Carlo simulation, shared 5*4*5*5=500 optical parameter is combined, is obtained
8000 surface surfaces of intensity distribution, extraction are wherein used as training set for 6000, and 2000 as test subset.Training set is inputted
Default network, is trained using network structure identical with Fig. 5.
3.3 based on deconvolution source visit position determine
Convolutional neural networks are finely adjusted using back-propagation algorithm, training 1000 times, until network convergence.Using anti-
Convolutional Neural can carry out shielding to picture the recovery of unrelated label noise.Think that the biggish position of difference is just between reconstructed picture
It is feature position.500 are randomly selected in all data, carries out the operation of convolution sum deconvolution, save deconvolution meter
Result after calculation.Adjacent picture is subtracted each other.Think that position feature of the absolute difference less than 0.001 is not much different.Point-by-point system
Count the number that absolute difference is greater than 0.001.It is final to determine that distance is visited in source.
4, interpretation of result
A more regular apple model is chosen, using the T method of inspection to apple model and double-layer plate model at 500
Bright figure curved surface under parameter combination carries out correlation analysis, and brightness figure is carried out section by us herein, forms 50 and cuts
Piece, difference label 1 to 50, as shown in Figure 4.Mutual corresponding slice carries out the similarity analysis of curve.It has counted without significant
The section of sex differernce accounts for the percentage of the number of slices of selection.
The comparative analysis of plot of light intensity under 4.1 two kinds of models
4.1.1 bright figure Controlling UEP
Due in brightness figure light intensity concentrate on distance center it is relatively close in, the farther away slice practical study of distance center is worth
Not high, in order to avoid these slices have an impact experimental result, we, which divide, has chosen most intermediate 10,20,30,40 pixels
Section carried out four times analysis, as a result as shown in Figure 6.Abscissa is 500 different parameter combinations in figure, and ordinate is not
There is the slice of significant difference to account for the percentage of the number of slices of selection.It is assumed herein that ratio is in 0.8 this optical parameter indicated above
Plot of light intensity under combination is similar.It can be seen from the figure that the similarity of plot of light intensity is got over when the section range of selection is bigger
It is low, it was demonstrated that distance center point is remoter, and escape intensity randomness is bigger, and reference value is smaller.So alternative in analysis model
When property, these parts should be excluded.As can be seen that the lower pattern number of similar proportion is in have well-behaved distribution in figure.These moulds
Model extracts, it is found that they belong to the combination of Ua2=5.5, therefore the combination of independent analysis Ua2=5.5.Two kinds extracted
Bright figure of the model in Ua2=5.5, it can be seen that light distribution is extremely concentrated when Ua2 is 5.5 from the surface of intensity distribution,
Substantially concentrate within 5 pixels of distance center, which results in except this section randomness it is very strong, similarity is very low.
4.1.2 Pearson's coefficient analysis
Pearson's coefficient can indicate the similarity of two curves, and Pearson's coefficient value is in -1 to 1 time, closer to 1 liang
Curve is more similar.Pearson's coefficient that department pattern finds out 50 sections of model is selected at random, as a result as shown in Figure 7.
Figure is that Ua1 is all totally 25 combinations that 0.7, Ua2 is 1.55, and abscissa is section number, and ordinate is cut thus
Pearson's coefficient under face.Other model results are similar with this figure, seldom show herein, and figure is it is found that under optical parameter combination
It is substantially higher than 0.8 in section Pearson's coefficient that pixel is 15 to 35, distance center point is remoter, and Pearson's coefficient is lower.By
After analysis above, independent analysis Ua2 be 5.5 parameter combination.Selected part plot of light intensity draws to obtain Fig. 8.Light in this section
It learns Pearson's coefficient under parameter within 10 pixel of distance and is substantially higher than 0.8, according to the particularity of plot of light intensity, in efficient intensity
In range, the similarity of two kinds of models is still very high.
4.1.3 conclusion
By the analysis of both the above method it can be concluded that double-layer plate model plot of light intensity and apple surface model light
The similarity of strong figure similarity in effective range is high, and comparison range is bigger, when the region of selection tends to be invalid, two moulds
The plot of light intensity similarity of type can also reduce.Therefore in effective range, double-layer plate model can be substituted for surface model, in this way
Reduce the complexity of surface model, while the reduction of accuracy and little.
4.2 optical parameter inversion results
(25:50,25:50) of the bright figure that apple model Imitating obtains is input to trained network.As a result such as
Table 4-1:
It follows that the optical parameter inversion result of apple pulp pericarp is respectively pericarp transmission coefficient ua1 from table:
51.92%, pulp transmission coefficient ua2:94.78%, pericarp scattering coefficient us1:67%, pulp scattering coefficient us2:89.49%.
The optical parameter efficiency of inverse process of pulp is far better than pericarp, this is because in simulation process, the peel thickness that we are arranged is
0.02mm, than relatively thin, photon travel in pericarp is shorter, and travel is longer in pulp, the pulp information of carrying compared with
More, pericarp information is relatively fewer.In compared with apple quality information association, the optical parameter and the correlation of quality of pulp
More preferably, therefore it may be concluded that the inversion result to pulp has more researching value.
(25:50,25:50) of the bright figure that double-layer plate model Imitating obtains is input to trained network.Knot
Fruit such as table 4-2:
Based on double-layer plate model to the inversion result of apple optical parameter are as follows: pericarp transmission coefficient ua1:84.16%, fruit
Meat transmission coefficient ua2:95.90%, pericarp scattering coefficient us1:84.22%, pulp scattering coefficient us2:90.74%, plate mould
The inversion result far better than apple surface model of type, especially compared to apple model to the parametric inversion poor effect of pericarp,
Double-layer plate model has also reached 80 or more percent to the efficiency of inverse process of pericarp.And more to pulp optical coefficient inversion result
It is good.This is because the complexity of double-layer plate model is lower, it is brighter that photon escapes tissue optical parameter feature entrained by tissue
It is aobvious.
4.3 deconvolution results and source are visited distance and are determined
4.3.1 distance analysis is visited in apple surface model source
Determine that distance is visited in source based on deconvolution on trained convolutional neural networks.
It can carry out shielding the recovery of unrelated label noise to picture using deconvolution nerve.Think poor between reconstructed picture
Not biggish position is exactly feature position.When carry out the operation of convolution sum deconvolution, adjacent picture is carried out
Subtract each other.Think that position feature of the absolute difference less than 0.001 is not much different.Point-by-point statistics absolute difference is greater than 0.001 number,
As a result such as Figure 13.From statistical result it can be seen that the absolute difference in 2-10 location of pixels is more greater than 0.001 quantity, explanation
This Partial Feature is obvious, and wherein 4-10 pixel partial amt is most.Wherein 0 represent the position where light source, 4-10 picture
The feature of vegetarian refreshments becomes apparent.
The result of deconvolution visits distance with the source that Actual measurement goes out and matches.
4.3.2 double-layer plate model source visits distance analysis
With same method, adjacent picture is subtracted each other.Think position feature difference of the absolute difference less than 0.001
Less.Point-by-point statistics absolute difference is greater than 0.001 number, as a result such as Fig. 9.From statistical result it can be seen that in 5mm~10mm
The absolute difference of position is more greater than 0.001 quantity, illustrates that the result of the obvious deconvolution of this Partial Feature and actual measurement are counted
The source of calculating visits distance and matches, while distance range is visited in the source that actual measurement obtains and is reduced.This is also demonstrated using volume
Product neural network carries out the validity that the calculating of distance is visited in source.
4.3.3 Comparative result
Distance is visited between 2-11 pixel based on the source that apple surface model obtains from known to the analysis of front, and based on double
Distance is visited between 5-10 pixel in the source that layer model obtains.And in the analysis of the point light source EO-1 hyperion actual measurement experiment based on apple
Distance is visited between 3-13 pixel in the source reached, and distance range now is visited in the source that three kinds of methods obtain and is depicted as shown in Figure 10.
In the analysis of the point light source EO-1 hyperion actual measurement experiment based on apple, the point light source irradiation of 200 samples is acquired
Image chooses 15 interested points as shown, the pol moisture with sample does correlation analysis on spectrogram.
Pol, moisture related coefficient result R be as follows:
The root-mean-square error RMSEP of pol moisture forecast set is as follows:
Actual measurement, AUS model, DLF model these three methods are labeled as 1,2,3, the SDD value obtained according to these three methods
Range, calculate R and RMSEP value average value.Obtain following table (such as in measured data SDD be 3-13 pixel, Wo Menji
0.663) average value for calculating the pol R value of position 3-13 obtains)
As can be seen from the above table, the sugar R value average value of DLF model (i.e. No. 3 models) highest in three models, and it is sugared
Divide RMSEP value average value minimum in three models, this shows in pol, and the SDD scoped features that No. 3 models obtain are best,
At the same time, No. 2 model performances are medium, and No. 1 model is worst.And the moisture R value average value of actual measurement (i.e. No. 1 model) is three
Highest in a model, moisture RMSEP value average value is minimum in three models, this shows on moisture, what No. 3 models obtained
SDD scoped features are best, and at the same time, No. 3 models show the golden mean of the Confucian school, and No. 2 model performances are worst.But it is poor between these three models
Away from being not obvious.It can be seen that being emulated using double-layer plate model, and next pregnant true with this according to characteristic value using CNN deconvolution
The effect for determining source spy distance is best.
Present applicant proposes a kind of analysis methods based on convolutional neural networks.The surface plot of light intensity that two kinds of models are generated
It is inputted as data, the convolutional neural networks of one energy inverting optical parameter of training pass through and adjust ginseng, normalization, Optical Parametric array
The methods of conjunction makes it reach acceptable degree to the inversion accuracy of optical parameter combination.The side of deconvolution is used on this basis
Method is trained again.Feature position is found by comparing the difference between reconstructed picture and input picture, so that it is determined that source
Visit distance.Showing that distance radius is visited in the source of apple quality detection based on apple surface model is 2 to 11 pixels, changes actual range into
For 1.8mm-9.9mm.And it is 5mm-10mm that distance is visited in the source of the book based on double-layer plate model, the old actual range that converts is
Also the conclusion range of actual measurement is reduced while the conclusion that 4.5mm-9.0mm is studied before demonstrating.It verifies simultaneously
Using the feasibility of flat plate model substitution apple surface model, and using convolutional neural networks the source of progress visit analysis can
Row.It compares and is analyzed by EO-1 hyperion measured data, this method is to pass through analogue data on the basis of simulation model
It is analyzed, eliminates the various experimental errors in measurement process, accuracy is higher.Solves sample in actual measurement experiment simultaneously
With the insufficient problem of data.Simulated time can be reduced by simplified double-layer plate model, while also making this method more
With universality, distance acquisition and the EO-1 hyperion quality inspection of fruit are visited in the source that can solve the stone fruit of any pulp thickness
It surveys.
Specific embodiment described herein is only to illustrate to spirit of that invention.The neck of technology belonging to the present invention
The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (9)
1. a kind of fruit EO-1 hyperion quality detecting method of the photon transmission simulation based on fruit double-layer plate model, feature exist
In it the following steps are included:
(1) fruit double-layer plate model is constructed；
(2) photon transmission simulation is carried out based on monte carlo method and obtains light brightness distribution figure；
(3) light brightness distribution figure is input in the pre- convolutional neural networks mixed up and carries out convolution and deconvolution；
(4) analysis convolution and deconvolution visit distance as a result, obtaining source；
(5) distance being visited using the source, Quality Detection is carried out to fruit.
2. according to the method described in claim 1, it is characterized in that being based on monte carlo method to double-layer plate in step (2)
Model carries out photon transmission simulation, comprises the concrete steps that:
(2.1) photon initializes；
(2.2) the photon direction of motion and step-length are calculated, while carrying out cross the border judgement (2.3)；
(2.3) it crosses the border judgement, crosses the border, carry out step (2.4b), otherwise carry out step (2.4a)；
(2.4a) photonic absorption and scattering calculate, and are transferred to step (2.5)；
The refraction of (2.4b) photon, reflection calculate, and are transferred to step (2.5)；
(2.5) judge whether photon weight is too small, is, carry out step (2.6), otherwise return step (2.2)；
(2.6) judge whether photon lives or dies, be, carry out step (2.7), otherwise return step (2.2)；
(2.7) judge whether it is the last one photon, be, terminate, otherwise return step (2.1)；
(2.8) it takes a large amount of photons and different optical parameters to combine and repeats step (2), obtain the noiseless of fruit double-layer plate model
Light brightness distribution figure.
3. according to the method described in claim 1, double-layer plate model is two parallel plates it is characterized in that in step (1),
Pericarp is indicated between two plates, distance, that is, peel thickness is denoted as d between two plates1, pulp thickness is denoted as d2=∞ mm；First
Plate, that is, pericarp absorption coefficient is denoted as μa1, the scattering coefficient of pericarp is denoted as μs1；Second plate, that is, pulp absorption coefficient is denoted as
μa2, the scattering coefficient of pulp is denoted as μs2。
4. according to the method described in claim 3, it is characterized in that obtained light brightness distribution figure is input in step (3)
Trained neural network, specifically:
I pericarp absorption coefficient μa1It falls into 5 types, pulp absorption coefficient μa2It is divided into 4 classes, pericarp scattering coefficient μs1It falls into 5 types, pulp
Scattering coefficient μs2It falls into 5 types, median is respectively taken to be combined 500 groups of optical parameter combinations of acquisition；
Ii carries out Monte Carlo simulation on double-layer plate model, obtains 8000 surface surfaces of intensity distribution, extracts wherein 6000
Zhang Zuowei training set, 2000 are used as test subset that convolutional neural networks is used to be trained.
5. according to the method described in claim 4, it is characterized in that in step (4), using back-propagation algorithm to convolutional Neural
Network is finely adjusted, training 1000 times, until network convergence；500 are randomly selected in all data, carry out convolution sum warp
Long-pending operation saves the result after deconvolution calculates；Adjacent picture is subtracted each other；Point-by-point statistics absolute difference is greater than 0.001
Number, it is final to determine that distance is visited in source.
6. according to the method described in claim 4, it is characterized in that convolutional neural networks structure includes:
First layer input layer carries out convolution to image using the convolution of 30 2*2 sizes, and convolution kernel moving step length is 1, convolution side
Formula uses valid, initial learning rate 0.07, batch size 50, total the number of iterations 2000 times；
Second layer pond layer is averaged pond using 2*2；
Third layer uses the convolution kernel of 20 3*3 sizes, and convolution kernel moving step length is 1；
The 4th layer of convolution kernel using 20 3*3 sizes, convolution kernel moving step length are 1, are operated without pondization；
Layer 5 carries out convolution to input data using 15 2*2 convolution kernels；
The output of layer 6 network is pulled into one-dimensional vector by full articulamentum, and output data size is 60*1；
Layer 7 is network output layer.
7. according to the method described in claim 1, it is characterized in that the fruit is the stone fruit of pulp thickness.
8. according to the method described in claim 7, it is characterized in that the fruit is apple or pears.
9. according to the method described in claim 4, it is characterized in that in step (2) i, pericarp absorbs when the fruit is apple
Coefficient μa1It takes: 0.7,1.55,1.85,2.25,4.5, pulp absorption coefficient μa2: 0.5,1.15,1.45,5, pericarp scattering coefficient
μs1: 30,67.5,82.5,100,190, pulp scattering coefficient μs2: 12,25.5,29,34,56, unit: mm-1。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910590833.0A CN110243765A (en) | 2019-07-02 | 2019-07-02 | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910590833.0A CN110243765A (en) | 2019-07-02 | 2019-07-02 | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110243765A true CN110243765A (en) | 2019-09-17 |
Family
ID=67890624
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910590833.0A Pending CN110243765A (en) | 2019-07-02 | 2019-07-02 | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110243765A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160478A (en) * | 2019-12-31 | 2020-05-15 | 北京理工大学重庆创新中心 | Hyperspectral target significance detection method based on deep learning |
| CN111175248A (en) * | 2020-01-22 | 2020-05-19 | 中国农业科学院农产品加工研究所 | Intelligent meat quality online detection method and detection system |
| CN116592774A (en) * | 2023-07-18 | 2023-08-15 | 成都洋湃科技有限公司 | Pipe wall dirt detection method and device, storage medium and electronic equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008103486A1 (en) * | 2007-02-23 | 2008-08-28 | Duke University | Scaling method for fast monte carlo simulation of diffuse reflectance spectra |
| CN106770047A (en) * | 2017-01-09 | 2017-05-31 | 浙江大学 | The inversion method and measurement apparatus of a kind of rice leaf tissue optical property parameter |
| CN107862326A (en) * | 2017-10-30 | 2018-03-30 | 昆明理工大学 | A kind of transparent apple recognition methods based on full convolutional neural networks |
| CN109856064A (en) * | 2019-01-28 | 2019-06-07 | 南京农业大学 | Apple EO-1 hyperion quality detecting method based on photon transmission simulation |
| CN109924949A (en) * | 2019-03-05 | 2019-06-25 | 北京工业大学 | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks |
-
2019
- 2019-07-02 CN CN201910590833.0A patent/CN110243765A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008103486A1 (en) * | 2007-02-23 | 2008-08-28 | Duke University | Scaling method for fast monte carlo simulation of diffuse reflectance spectra |
| CN106770047A (en) * | 2017-01-09 | 2017-05-31 | 浙江大学 | The inversion method and measurement apparatus of a kind of rice leaf tissue optical property parameter |
| CN107862326A (en) * | 2017-10-30 | 2018-03-30 | 昆明理工大学 | A kind of transparent apple recognition methods based on full convolutional neural networks |
| CN109856064A (en) * | 2019-01-28 | 2019-06-07 | 南京农业大学 | Apple EO-1 hyperion quality detecting method based on photon transmission simulation |
| CN109924949A (en) * | 2019-03-05 | 2019-06-25 | 北京工业大学 | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| MENGYUN ZHANG 等: "Optical properties of blueberry flesh and skin and Monte Carlo multi-layered simulation of light interaction with fruit tissues", 《POSTHARVEST BIOLOGY AND TECHNOLOGY》 * |
| 石舒宁: "水果组织的光学描述与光传输规律研究", 《中国优秀硕士学位论文全文数据库 农业科技辑》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160478A (en) * | 2019-12-31 | 2020-05-15 | 北京理工大学重庆创新中心 | Hyperspectral target significance detection method based on deep learning |
| CN111160478B (en) * | 2019-12-31 | 2022-07-26 | 北京理工大学重庆创新中心 | Hyperspectral target significance detection method based on deep learning |
| CN111175248A (en) * | 2020-01-22 | 2020-05-19 | 中国农业科学院农产品加工研究所 | Intelligent meat quality online detection method and detection system |
| CN116592774A (en) * | 2023-07-18 | 2023-08-15 | 成都洋湃科技有限公司 | Pipe wall dirt detection method and device, storage medium and electronic equipment |
| CN116592774B (en) * | 2023-07-18 | 2023-09-19 | 成都洋湃科技有限公司 | Pipe wall dirt detection method and device, storage medium and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110243765A (en) | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model | |
| Baranyai et al. | Analysis of laser light propagation in kiwifruit using backscattering imaging and Monte Carlo simulation | |
| CN108257132A (en) | A kind of method of the CT image quality measures based on machine learning | |
| WO2017167030A1 (en) | Modelling method for human skin spectrum and mathematical modelling method for plurality of skin parameters with high fitting degree | |
| CN106340000A (en) | Bone age assessment method | |
| CN108460391A (en) | Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network | |
| CN107392977A (en) | Single-view Cherenkov lights tomography rebuilding method | |
| Hu et al. | Monte Carlo: A flexible and accurate technique for modeling light transport in food and agricultural products | |
| CN110010229A (en) | A kind of ultrasonic image analogy method based on generation confrontation network | |
| Shao et al. | Detection and analysis of sweet potato defects based on hyperspectral imaging technology | |
| CN109856064B (en) | Apple EO-1 hyperion quality detecting method based on photon transmission simulation | |
| CN114494217A (en) | Method and device for detecting artificial tissues and organoids | |
| Mei et al. | An overview on optical non-destructive detection of bruises in fruit: Technology, method, application, challenge and trend | |
| Simpson et al. | Impact of quantization algorithm and number of gray level intensities on variability and repeatability of low field strength magnetic resonance image-based radiomics texture features | |
| Zou et al. | Ultrasound-enhanced Unet model for quantitative photoacoustic tomography of ovarian lesions | |
| Li et al. | Establishment of a multi-position general model for evaluation of watercore and soluble solid content in ‘Fuji’apples using on-line full-transmittance visible and near infrared spectroscopy | |
| Wu et al. | Diffuse optical Imaging using decomposition methods | |
| Potlov et al. | Localization of inhomogeneities in diffuse optical tomography based on late arriving photons | |
| Liu et al. | Improving heterogeneous classification accuracy based on the MDFAT and the combination feature information of multi-spectral transmission images | |
| Wang et al. | Pixel-wise body composition prediction with a multi-task conditional generative adversarial network | |
| CN107280644A (en) | A kind of knee osteoarthritis course of disease detection method based near infrared light | |
| Ceyhan et al. | Statistical analysis of cortical morphometrics using pooled distances based on labeled cortical distance maps | |
| Sun et al. | Investigations of optical geometry and sample positioning in NIRS transmittance for detecting vascular browning in apples | |
| Manojlović et al. | Rapid extraction of skin physiological parameters from hyperspectral images using machine learning | |
| CN113129286B (en) | Method, device, equipment and storage medium for detecting osteoporosis |
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: 20190917 |