CN107271375B - A kind of high spectral image detecting method of quality of mutton index - Google Patents
A kind of high spectral image detecting method of quality of mutton index Download PDFInfo
- Publication number
- CN107271375B CN107271375B CN201710598787.XA CN201710598787A CN107271375B CN 107271375 B CN107271375 B CN 107271375B CN 201710598787 A CN201710598787 A CN 201710598787A CN 107271375 B CN107271375 B CN 107271375B
- Authority
- CN
- China
- Prior art keywords
- spectrum
- sample
- quality
- index
- mutton
- 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.)
- Active
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
Abstract
A kind of high spectral image detecting method of quality of mutton index, more particularly to a kind of method for making full use of high spectrum image information preferably to obtain a plurality of modeling spectrum and then improvement quality of mutton Indexs measure effect, including first establishing the prediction model of the mutton high spectrum image index of quality and prediction model being recycled to carry out quality of mutton Indexs measure.The present invention can exponentially expand modeling sample amount, overcome because of the difficult problem for causing modeling sample amount limited of sample chemical value measurement by making full use of high spectrum image information to the preferably a plurality of establishment of spectrum prediction model of each sample；The modeling effect that model built can be improved keeps its calibration set and verifying collection precision higher；Characteristic wavelength extraction is carried out on the basis of preferred a plurality of spectrum, it is more preferable to model effect for result stable convergence；The index of quality is detected using preferred prediction model, detection performance can be improved；To improving, mutton and other meat products intellectualized detections are horizontal, ensure that meat products Quality Safety is meaningful.
Description
Technical field
The invention belongs to meat products high spectrum image technical field of nondestructive testing more particularly to one kind to make full use of EO-1 hyperion
Image information preferably obtains a plurality of modeling spectrum and then improves the method for quality of mutton Indexs measure effect, and this method can be used for changing
The qualitative and quantitative analysis modeling effect and testing result of kind livestock products high spectrum image.
Background technique
High spectrum image detection technique can obtain the two-dimensional image information and sample of beef and mutton sample wavelength points simultaneously
Spectral information at each point may be implemented beef and mutton sample inside and outside quality while detect, and have non-destructive, quick, nothing
It the features such as pollution, collection of illustrative plates, is rapidly developed and is widely applied in agricultural and animal products context of detection in recent years.Utilize EO-1 hyperion
The qualitatively and quantitatively detection that image carries out quality of mutton index needs to establish its prediction model first, and the foundation of prediction model needs standard
The Testing index physics and chemistry information of sample is really obtained, but since physics and chemistry value measurement general operation is comparatively laborious, time and effort consuming, causes to use
It is limited in the sample size of detection, and sample size is expanded by the physical and chemical value that test obtains a large amount of samples and is extremely not easy.In order to
The deficiency of Small Sample Database analysis is solved, existing research person is built using the method for Bootstrap sampling (bootstrapping) to expand
Apperance notebook data, but when sample size is seldom not enough to provide overall distribution information, it is easy to produce covering and conflict.Cattle and sheep meat sample
Product high-spectrum seems the three-dimensional data cube being formed by stacking by the image of hundreds of continuous wave bands by subband order, information content
It is abundant, generally comprise 200,000 or more pixels.In existing document report, high spectrum image is carried out to the beef and mutton index of quality
Modeling first has to obtain its area-of-interest, and the rectangle region of 50x50 or 100x100 size is usually chosen in meat central area
Domain or the pure muscle region of meat, then to one averaged spectrum of this region of interesting extraction for modeling and detecting, due to sample
Product amount is limited to cause model inspection precision lower.From line scanning high spectrum image acquisition process discovery, with line scanning into
Row, the acquisition of high spectrum image is the continuous process of dynamic, causes the image obtained on perpendicular to line scanning direction
Pixel has a certain difference, and in addition the content distribution in space of meat samples index of quality itself is also variant, because
This can attempt to improve modeling and detection effect using two species diversity.
Summary of the invention
The purpose of the present invention is to provide a kind of high spectral image detecting methods of quality of mutton index, by meat samples height
The spatial information and index of quality distribution character of spectrum picture, which combine, obtains preferred spectrum to improve modeling effect
With testing result, it is intended to solve in current high spectrum image modeling process sample size is limited, model prediction is indifferent, detection
The problems such as characteristic wavelength that effect is poor, extracts is unstable.
The technical solution adopted by the invention is as follows:
A kind of high spectral image detecting method of quality of mutton index first establishes prediction model, recycles model inspection sheep
Meat index, it is characterised in that: first establish the prediction model of the mutton high spectrum image index of quality, recycle prediction model into
Row quality of mutton Indexs measure；
Specific step is as follows for the prediction model for establishing the mutton high spectrum image index of quality:
Step 1: sample preparation, to the meat sample removal surrounding fascia and connective tissue after government official, sample is using slice or is made
Two ways is mixed after meat gruel prepares sample；Using slice whens being grade of freshness, Volatile Base Nitrogen etc. for Testing index
Mode sample preparation, for meat detection of adulterations, mixing is placed in sample preparation in surface plate after meat gruel is made in meat；It need to be examined when sample preparation
Consider the representativeness of Testing index distribution；
Step 2: the sample of preparation is placed on phase shift platform, is obtained by the movement of translation stage and line scan camera
Meat samples high spectrum image three-dimensional data block；
Step 3: meat samples index of quality physics and chemistry value is obtained, including grade of freshness, TVB-N content or is mixed
False amount ratio；
Step 4: the area-of-interest of meat samples high spectrum image relevant to Testing index is obtained；
Step 5: according to Testing index sample area-of-interest distribution character to the interested of sample high spectrum image
Each pixel in region is reconstructed, and sequentially extracts discrepant a plurality of spectrum；
Step 6: the rejecting of exceptional spectrum is carried out to extracted meat samples spectroscopic data；
Step 7: by remaining spectroscopic data after rejecting abnormalities spectrum and grade of freshness, Volatile Base Nitrogen, adulterated amount
The Testing index such as ratio correspond, and collect according to the ratio cut partition calibration set and verifying of 3:1；
Step 8: for different Testing index, using calibration set spectroscopic data preferred pretreatment method, and using identical
Preprocess method to verifying collection spectrum handle, analyzed using pretreated data；
Step 9: for different spectrum extract item number established respectively according to Testing index offset minimum binary quantitative model or
Offset minimum binary qualitative discrimination model, and verified using verifying collection sample, item number is extracted according to the preferred spectrum of modelling effect,
Determine preferred all band prediction model, it is feasible that modeling effect meets the requirements then representative model；Otherwise, step 5 is repeated extremely
Nine, until meeting the requirements；
Step 10: be utilized respectively competitive adaptive weight weighting algorithm, the Stepwise Regression Algorithm and successive projection algorithm and its
Combination is to quality of mutton index extraction characteristic wavelength；
Step 11: using the characteristic wavelength extracted as input variable, each index of quality is established respectively and is based on characteristic wave
Long preferred prediction model, and the preferred prediction model of all band obtained with step 9 is compared, and determines final prediction mould
Type；
It is described using prediction model to carry out quality of mutton Indexs measure specific step is as follows:
Step A: meat samples to be detected are prepared；
Step B: the high spectrum image three-dimensional data block of meat samples to be detected is obtained；
Step C: meat samples high spectrum image area-of-interest relevant to Testing index is obtained；
Step D: extracting an averaged spectrum for detecting to each pixel of sample high spectrum image area-of-interest, if
Visual analyzing is carried out, then extracts the spectral information of each pixel for detecting；
Spectral information: being substituted into built index of quality prediction model by step E, and the index of quality for obtaining mutton to be detected is pre-
Measured value, until the detection of all mutton test samples finishes.
In above-mentioned steps one, about 4 cm of cm × 1.5 of cm × 4 of beef and mutton sample size prepared using slicing mode is used
It is initially formed the meat gruel of about 2 ~ 3 millimeters of partial sizes in the sample of adulterated content detection, prepares according to different proportion and is laid in after mixing
In surface plate, surface plate general requirements diameter used is less than 10 cm, about 50 cm of floor space2, according to above-mentioned sample making course, obtain
Sample shape feature have rectangle or a nearly rectangle, circle or subcircular, rectangle or nearly rectangle are general slicing mode sample preparation, round
Or subcircular is the sample preparation of meat gruel blending manner.
In above-mentioned steps three, the quality quantitative target of meat samples is the adulterated amount ratio of Volatile Base Nitrogen and adulterated meat
Example, or total number of bacteria, tenderness, pH value etc.；Quality qualitative index is the freshness obtained according to freshness overall merit
Rating, or adulterated classification, tenderness grade etc..
In above-mentioned steps four, the acquisition of area-of-interest is carried out in two steps；On the one hand, it is transported with wave band subtraction, binaryzation
Calculation obtains the binary image of sample, then carries out exposure mask to high spectrum image, removes the background and shadow part of high spectrum image
Point, on the other hand, with wave band add operation and binaryzation, the fat of mask process removal sample, bright spot position, acquisition can
Reaction detection index property, the sample area-of-interest with different, generally comprises 12 ~ 150,000 pixels.
In above-mentioned steps five, Testing index is distributed with approaches uniformity distribution and dissipates distribution around from center, such as works as sample
When product are the adulterated sample mixed, each location detection index value difference is not distributed for approaches uniformity significantly；When meat evaluation refers to
When being designated as Volatile Base Nitrogen, since the decay process of sample gradually develops generally since surrounding to center, Testing index
In dissipating distribution around from center；To each pixel in region of interest domain space with perpendicular or parallel with line scanning direction
Direction by scanning sequencing be reassembled as a column, all pixels point is respectively extracted according to the spectrum item number that need to be extracted,
A starting cut-off rule position can also be selected, to all pixels point according to the spectrum that need to be extracted by selecting sample space central point
Item number is respectively extracted by 360 ° of regions；The spectrum item number of extraction is related with the pixel number that every spectrum is included, and is total
The ratio of pixel number and spectrum item number, in order to keep the spectrum extracted representative, it is desirable that the pixel that every spectrum is included
Point is greater than 2500, can extract the combination spectrums of the different spectrum item numbers and a plurality of spectrum that meet above-mentioned pixel requirement into
Row analysis as a further preferred embodiment of the present invention, is extracted 8,4,2,1 spectrum and is compared for ease of operation.
In above-mentioned steps six, since the factors such as instrument, environment, thickness of sample influence, can exist in modeling process some different
Ordinary light spectrum, to influence model accuracy, need to use Q residual error boundary combination Hotelling T2Boundary rejects these exceptional spectrums.
In above-mentioned steps eight, in the case of each sample preferably extracts 8 spectrum, freshness is classified, each sample collection is preferred
The preprocess method of spectrum is the method that first derivative, 15 point S-G smooth, variable standardization, centralization processing combine；For
Volatile Base Nitrogen index prediction, each sample integrate the preprocess method of preferred spectrum be second dervative, 23 point S-G smoothly, center
Change handles the method combined；It is flat for 17 point S-G for the preprocess method of the preferred spectrum of adulterated amount scale prediction each sample collection
The preprocess method that sliding, centralization combines.
In above-mentioned steps nine, all band grade of freshness is established respectively using a plurality of spectrum that each sample collection of division extracts
The PLS quantitative model of PLS-DA qualitative model or Volatile Base Nitrogen, mutton adulteration extracts item according to the preferred spectrum of modelling effect
Several and preferred prediction model, the root-mean-square error and phase relation that PLS quantitative model collects according to calibration set, validation-cross collection, verifying
Number synthesis evaluates preferred spectrum and its prediction model, and PLS-DA qualitative model is quasi- according to calibration set, validation-cross collection and verifying collection
The really preferred spectrum of rate overall merit and its prediction model.
In above-mentioned steps ten, as a further preferred embodiment of the present invention, characteristic wave that grade of freshness is extracted with
11, respectively 530,558,634,668,753,765,776,785,833,859,957 nm；To the spy of TVB-N index extraction
Sign wavelength has 12, respectively 527,571,605,637,649,753,770,799,809,838,848,969 nm；To adulterated
The characteristic wave that amount ratio is extracted is with 8, respectively 472,552,573,767,772,945,963,975 nm.
In above-mentioned steps 11, the preferred prediction model of each index of quality is established as branch using the characteristic wavelength information of extraction
Vector machine SVM model is held, evaluation index is identical as step 9.
Advantage is the present invention compared with prior art:
The first, it is built by making full use of high spectrum image information that can exponentially expand a plurality of spectrum of each sample extraction
Mould sample size overcomes because of the difficult problem for causing modeling sample amount limited of sample chemical value measurement.
The second, the modeling effect of model built can be improved using preferred spectrum.It is preferred that extracting a plurality of spectrum modeling phase
It is representative more stronger than extracting an averaged spectrum modeling, and the calibration set of model and verifying collection precision are higher.
Third extracts the modeling effect that a plurality of spectrum modeling not only improves all-wave segment model, and in preferred a plurality of spectrum
On the basis of carry out characteristic wavelength extraction, result is also stable convergence, and be better than single spectrum modeling effect, can be subsequent
Multispectral detecting instrument exploitation lay the foundation.
4th, the index of quality is detected using finally determining prediction model, detection accuracy can be improved, improve detection effect
Fruit.
5th, the present invention can provide theory support to improve mutton and other meat products intellectualized detections levels and technology
And technical support, for ensureing meat products Quality Safety, maintenance consumer health has direct realistic meaning.
Detailed description of the invention
Fig. 1 be in the present invention each pixel of sample area-of-interest by perpendicular to line scanning direction recombination divide equally extract spectrum
Schematic diagram；
Fig. 2 be in the present invention each pixel of sample area-of-interest by be parallel to line scanning direction recombination divide equally extract spectrum
Schematic diagram；
Fig. 3 be in the present invention each pixel of sample area-of-interest by from center to surrounding divide equally extract spectrum signal
Figure；
Fig. 4 is the meat samples primary light spectrogram under the preferred spectrum extraction item number that the embodiment of the present invention 1 is related to；
Fig. 5 is the exceptional spectrum rejecting figure that the embodiment of the present invention 1 is related to；
Fig. 6 is the characteristic variable distribution map under the preferred spectrum extraction item number that the embodiment of the present invention 1 is related to；
Fig. 7 is the final optimization model grade of freshness testing result figure that the embodiment of the present invention 1 is related to；
Fig. 8 is the meat samples primary light spectrogram under the preferred spectrum extraction item number that the embodiment of the present invention 2 is related to；
Fig. 9 is the characteristic variable distribution map under the preferred spectrum extraction item number that the embodiment of the present invention 2 is related to；
Figure 10 is the final optimization model TVB-N Indexs measure result figure that the embodiment of the present invention 2 is related to；
Figure 11 is the adulterated meat samples primary light spectrogram under the preferred spectrum extraction item number that the embodiment of the present invention 3 is related to；
Figure 12 is the characteristic variable distribution map under the preferred spectrum extraction item number that the embodiment of the present invention 3 is related to；
Figure 13 is the adulterated amount ratio testing result figure of final optimization model fox meat that the embodiment of the present invention 3 is related to.
Specific embodiment
In order to which the purpose of the present invention, advantage is more clearly understood, the content of present invention is made combined with specific embodiments below
It is further described.
A kind of embodiment 1: high spectral image detecting method of mutton grade of freshness
The part a: the high spectrum image prediction model of mutton grade of freshness is established
A1, sample preparation
Experimental material is the outer fillet of sheep and goat carcass, is purchased from the Shihezi market of farm produce.Meat sample is gone after muscle to be sliced to obtain
The fritter of 60 about 4 cm × 1.5 of cm × 4 cm, weighs about 25 g, with being deposited in 4 DEG C of insulating boxs after hermetic bag bale number
1 ~ 14 day, freshness distribution was in fresh, secondary fresh, corrupt distribution, representative.
A2, line scanning high spectrum image acquisition
Line scans high spectrum image acquisition system by imaging spectrometer (ImSpector V10E, Finland), line array CCD
Camera (hamamastsu), 150W optical fiber halogen lamp white light source (SCHOTT DCR III, China), electronic control translation stage and meter
The composition such as calculation machine.
30 minutes opening Hyperspectral imagers are fully warmed-up it in advance before data acquisition.Acquire meat samples EO-1 hyperion
Parameter when data are as follows: 10 ms of camera exposure time, 1.25 mm/s of electronic control translation stage movement speed, spectra collection range be 400 ~
1000 nm, 0.63 nm of spectral resolution include 948 spectroscopic data points altogether, carry out spectrum school after having acquired high spectrum image
Just.
A3, grade of freshness divide
Reference GB/T 5009.44 in 2003 measures Volatile Base Nitrogen TVB-N after having acquired meat samples high-spectral data
Physics and chemistry value measures total number of bacteria TVC referring to GB 4789.2 in 2016.Mutton freshness is determined according to TVB-N, TVC overall merit
Grade, wherein TVB-N < 15 mg/100g, TVC < 5.7logCFU is fresh, the mg/100g of 15 < TVB-N < 25,5.7 < TVC <
For 6.7logCFU to be secondary fresh, TVB-N > 25 mg/100g, TVC > 6.7logCFU are corruption.
The acquisition of a4, sample high spectrum image area-of-interest
In the present invention, discovery spectrum it is larger in 400 ~ 460 section nm noises, therefore select the wave spectrum of 460 ~ 1000 nm as
Subsequent analysis.In addition sample gray value difference at wavelength 544.15 nm and 818.98 nm is larger, and background, dash area are grey
Angle value difference is smaller, so that background, shade is tended to be completely black using wave band subtraction.Then gone by binary conversion treatment and mask method
Except the sample image after background, shade.Again using wave band add operation, mask method removal fat and bright spot, pure muscle portion is obtained
It is allocated as region of interesting extraction spectroscopic data.
The extraction of a5, a plurality of spectrum of sample area-of-interest
Due to observing the muscle image information of sample, there is some difference, sweeps to the pure muscle region of each sample according to line
The vertical direction retouched forms a line and recombinates after extracting each pixel spectroscopic data, extracts 1,2,4,8 spectrum respectively.Data mention
Process is taken to see Fig. 1.Wherein total pixel number is N, n N/1, N/2, N/4, N/8, i.e., every N/1, N/2, N/4, N/8 pixel
Spectrum is averaged can extract 1,2,4,8 spectrum respectively.If total pixel number is not the integral multiple for extracting spectrum item number, right
N is rounded.The high spectrum image original spectrum curve that spectrum extracts each sample when item number is 8 is as shown in Figure 4.
A6, exceptional spectrum are rejected
Exceptional spectrum can change the overall distribution characteristic of sample set, and then influence model prediction accuracy.In test, using Q
Residual error boundary combination Hotelling T2Boundary rejecting abnormalities spectrum.The sample sets for extracting 8,4,2 spectrum eliminate 13,12,3 respectively
Spectrum.The abnormal data elimination for wherein extracting 8 spectrum is as shown in Figure 5.
A7, different spectrum extract the modeling sample collection statistics and analysis of item number
To 60 mutton high spectrum image samples for establishing prediction model, freshness is divided into three according to TVB-N value
Class: fresh 16, fresh 24 secondary, corruption 20.It uses every the model split of one-out-three as 45 correcting samples, 15 verifyings
Sample.Table 1 is the modeling sample collection statistical result under different spectrum extraction item number after rejecting abnormalities spectrum.
The different spectrum of table 1 extract the modeling sample collection statistical result of item number
When each mutton high spectrum image extracts 8 spectrum, Calibration is tested by 350 spectral compositions of 45 samples
Card collects 117 spectral compositions by 15 samples.As can be seen that extracting modeling sample after a plurality of spectrum compared to 1 spectrum is extracted
About corresponding multiple has also been enlarged in collection quantity.
A8, modeling sample light harvesting Spectrum data processing
The present embodiment determines optimal preprocess method by establishing 10 folding cross validation of PLS-DA models coupling.Wherein extract
The preferred pretreatment method of 8 spectrum is the side that first derivative, 15 point S-G smooth, variable standardization, centralization processing combine
Method, the preferred pretreatment method for extracting 1 spectrum is that first derivative, 11 point S-G smooth, variable standardization, centralization handle phase
In conjunction with method.
A9, preferably spectrum extract the determination of item number and the foundation of the preferred prediction model of all band
PLS-DA freshness classification results such as 2 institute of table that different spectrum item numbers are established is extracted to each sample high spectrum image
Show.As can be seen that the model Grading accuracy rate that each sample extracts the foundation of 1 averaged spectrum is minimum, calibration set, validation-cross
Collection, verifying collection accuracy rate are respectively 84.44%, 81.25%, 73.33%.Extract the PLS-DA modelling effect of 8 establishment of spectrum most
Excellent, calibration set, validation-cross collection, verifying collection differentiate that accuracy rate is respectively 94.29%, 91.51%, 93.16%.It is flat compared to extracting 1
For equal spectrum, the PLS-DA modelling effect for extracting 2,4 establishment of spectrum is improved.Item number is extracted according to each spectrum
Model result it can be concluded that extract 8 establishment of spectrum PLS-DA model be optimization model, improve average light well
Compose the mutton freshness classification results of modeling.
The different spectrum of table 2 extract the PLS-DA mutton freshness classification results of item number
A10, characteristic wavelength extract and final preferably prediction model determines
8 spectrum peace that each sample high spectrum image is extracted respectively using the competitive adaptive CARS algorithm of weighting again
Equal spectrum extracts characteristic wave bands, extracts to obtain 11 characteristic wavelengths to 8 spectrum of each sample: 530,558,634,668,
753,765,776,785,833,859,957 nm, variable distribution are as shown in Figure 6.Sample average spectrum is extracted to obtain 8
Characteristic wavelength are as follows: 672,688,778,788,837,880,927,992 nm.
Table 3 is the mutton freshness established using the characteristic wavelength information of 8 spectrum preferably extracted and averaged spectrum
SVM classification results, wherein parameter C is penalty factor, and g is RBF nuclear parameter.As can be seen that extracting the CARS- of 8 establishment of spectrum
For PSO-SVM model calibration set accuracy rate up to 96.57%, validation-cross collection and verifying collection accuracy rate are respectively 94.58%, 95.73%.
Be only 84.92% using the CARS-PSO-SVM model validation-cross collection accuracy rate of sample average establishment of spectrum, modelling effect compared with
Difference.Experiments have shown that being extracted using preferred 8 spectrum, each sample sets nicety of grading of Feature information modeling is higher, and stability is more
It is good.
The different spectrum of table 3 extract the mutton freshness SVM model classification results of item number
Compare each sample high spectrum image and extracts all band PLS-DA prediction model of 8 establishment of spectrum and based on CARS spy
Sign wavelength extracts the SVM model established, and SVM model freshness Grading accuracy rate is better than PLS-DA model, for finally preferred pre-
It surveys model and is used for subsequent detection.
The part b: the detection of mutton grade of freshness is carried out using prediction model
The detection process that sample freshness is classified is carried out according to step A ~ E sequence using final preferred prediction model,
Sample to be tested is prepared first, then obtain sample high spectrum image area-of-interest and extracts averaged spectrum, substitutes into final prediction
Model carries out freshness classification.Preparation, the high spectrum image acquisition, the division of grade of freshness, area-of-interest of sample to be tested
Acquisition referring to the present embodiment a1 ~ a4 operating process carry out.Meanwhile for the detection effect of comparison model, sample is being acquired
The grade of freshness information that sample to be tested is also obtained after high spectrum image, the inspection based on sample freshness classification results to model
Effect is surveyed to be evaluated.
Experiment is prepared for 22 samples for model inspection, including 7 fresh samples, 7 fresh samples and 8 corruption altogether
Sample, freshness testing result are as shown in table 4.Utilize the SVM under the built characteristic wavelength of preferred spectrum and averaged spectrum of extraction
Model carries out greenness determination to 22 test samples, and overall accuracy is respectively 95.45% and 81.82%, extracts 8 spectrum and builds
Vertical final optimization model only has 1 detection mistake, and detection effect comparison is as shown in Figure 7.The result shows that utilizing preferred 8 light
The mutton freshness hierarchy model stability for composing foundation is more preferable, and detection accuracy is higher, effectively improves the classification of mutton freshness
Detection effect.
Table 4 utilizes the freshness testing result of SVM prediction model
Freshness classification test shows to extract a plurality of spectrum compared to sample average spectrum is extracted and expand modeling sample amount,
Final optimization model based on preferably 8 establishment of spectrum is greatly better than the detection effect of 1 averaged spectrum modeling, preferably a plurality of
Quick, the accurate detection to mutton grade of freshness may be implemented in spectrum modeling.
A kind of embodiment 2: high spectral image detecting method of mutton Volatile Base Nitrogen
The part c: the high spectrum image prediction model of mutton Volatile Base Nitrogen is established
C1, sample preparation
Meat samples needed for testing are derived from the ridge position of 12 fresh Suffolks butchered, and are purchased from Shihezi City In Xinjiang
The market of farm produce.Meat is transported to livestock products laboratory using antistaling box, removes the fat and connective tissue of sheep ridge meat, is cut into about 40
The sample of the mm × 20 of mm × 40 mm amounts to 57 samples.Sample packaging label is placed in 4 DEG C of constant temperature refrigerator place 2 ~
14 d。
C2, line scanning high spectrum image acquisition
It mainly includes imaging spectrometer (ImSpector V10E, Finland), CMOS phase that line, which scans Hyperspectral imager,
Machine (MV-1024E, China), light source (3900, Illumination science and technology), sample transporting equipment (23000 Y of DP), outside
The composition such as shading black box.Spectra collection range is 400 ~ 1000 nm.
Bloom spectrometer is first opened before acquisition to be preheated, and is taken out after sample stands 30 minutes at room temperature and is acquired sample again
High spectrum image information.System setting before data acquire: 10 ms of time for exposure, 0.85 mm/s of objective table movement speed, object distance
For 38 cm.And it needs to carry out black and white correction before acquiring high spectrum image.
C3, the measurement of Volatile Base Nitrogen physics and chemistry value
TVB-N physics and chemistry value measurement experiment is measured according to GB/T 5009.44 in 2003 using semi-micro nitrogen method.Each sample
Measurement twice, is averaged the TVB-N value as the sample, and ensure error sub-average 10%.
The acquisition of c4, sample high spectrum image area-of-interest
It is identical as described in a4, pure muscle parts are extracted as sample high spectrum image area-of-interest.Firstly, with wave band
Subtraction, binaryzation and background and dash area in masking method removal high spectrum image；Secondly, with wave band addition
Operation and masking method removal fat, bright spot etc..
The extraction of c5, a plurality of spectrum of sample area-of-interest
The meat samples high spectrum image of acquisition have in 400 ~ 460 nm wave-length coverages lower spectral response value and
Noise, therefore the spectral information for choosing 460 ~ 1000 nm is used for modeling analysis.To the pure muscle region of each sample in the present embodiment
It is in line after extracting each pixel spectroscopic data according to the parallel direction of line scanning, is extracted after being reassembled as 1,2,4,8 rows respectively
Each row averaged spectrum.Data extraction procedure is as shown in Figure 2.Wherein total pixel number be N, n N/1, N/2, N/4, N/8, i.e., often
N/1, N/2, N/4, N/8 pixel spectrum are averaged can extract 1,2,4,8 spectrum respectively.If total pixel number does not mention
The integral multiple for taking spectrum item number, then be rounded n.Fig. 8 is the original spectrum curve that spectrum extracts that item number is 8.
C6, exceptional spectrum are rejected
Using Q residual error boundary combination Hotelling T2Boundary rejecting abnormalities spectrum.After exceptional spectrum is rejected, 8 light are extracted
In spectrum, modeling sample collection includes 331 spectrum of 42 samples, and verifying collection is 118 spectrum of 15 samples.
C7, different spectrum extract the modeling sample collection statistics and analysis of item number
Table 5 is each sample collection TVB-N measured value statistical result that different spectrum extract item number, including average value, standard deviation
And maximum value and minimum value.The TVB-N content range of all samples is 10.43 ~ 40.90 mg/100g.
5 mutton TVB-N assay result of table
C8, modeling sample light harvesting Spectrum data processing
It is relatively ground in the present embodiment using the PLS quantitative model that pretreatment combined method establishes mutton Testing index TVB-N
Study carefully, in conjunction with staying a cross verification to determine optimal preprocess method.Wherein the preferred pretreatment method of 1 averaged spectrum of extraction is
The method that second dervative, 13 point S-G are smooth, centralization processing combines, the preferred pretreatment method for extracting 8 spectrum is second order
The method that derivative, 23 point S-G are smooth, centralization processing combines.
C9, preferably spectrum extract the determination of item number and the foundation of the preferred prediction model of all band
PLS model TVB-N prediction result such as 6 institute of table that different spectrum item numbers are established is extracted to each sample high spectrum image
Show.The results show that extracting the modelling effect of 1 averaged spectrum of each sample most using in the PLS model of all band establishment of spectrum
Difference, RMSEC, RMSECV and RMSEP are respectively 4.52,4.71 and 4.87, RC、RCVAnd RPRespectively 0.84,0.82 and 0.79.It mentions
Take the PLS modelling effect of 8 establishment of spectrum optimal, RMSEC, RMSECV and RMSEP are respectively 2.83,3.37 and 3.24, RC、RCV
And RPRespectively 0.96,0.92 and 0.92.The modeling of sample single averaged spectrum is extracted in comparison, and 2,4 spectrum modelings of extraction are pre-
Survey effect is more excellent, and validation-cross effect is also improved.According to each spectrum extract item number PLS prediction model result it can be concluded that
It is 8 that preferred spectrum, which extracts item number, and the optimization model of foundation is stable and validation-cross effect is preferable, is better than extracting sample single light
Spectrum.
The different spectrum of table 6 extract the PLS model TVB-N index modeling result of item number
C10, characteristic wavelength extract and final preferably prediction model determines
The single averaged spectrum that each sample high spectrum image is extracted respectively in conjunction with CARS algorithm and successive Regression SR algorithm
It is preferred that 8 spectrum extracted carry out spectral signature Variable Selection, 8 spectrum preferably extracted are chosen to obtain 12 variables:
527,571,605,637,649,753,770,799,809,838,848 and 969 nm, variable distribution are as shown in Figure 9.To sample
Averaged spectrum is extracted to obtain 15 variables: 508,516,533,539,545,556,572,578,600,650,753,760,835,
911、983 nm。
Table 7 is 8 spectrum preferably extracted using each sample high spectrum image and the characteristic wave that averaged spectrum is screened
The SVM model prediction result for the mutton TVB-N index that long message is established.As can be seen from Table 7,8 establishment of spectrum of extraction
CARS-SVM model inspection effect is substantially better than 1 averaged spectrum modeling, calibration set RCIt is respectively 0.96 and 2.60 with RMSEC.
The R of validation-cross collectionCVIt is respectively 0.93 and 3.27 with RMSECV, verifies the R of collectionPIt is respectively 0.93 and 3.04 with RMSEP.Using
The CARS-SVM model R of sample average establishment of spectrumCV、RP、RVOnly 0.80 or so, it is equal to the RMSE of index of quality TVB-N prediction
In 4.5mg/100mg or more, prediction effect is poor.
The different spectrum of table 7 extract the SVM model TVB-N modeling result of item number
Compare each sample high spectrum image and extracts all band PLS prediction model of 8 establishment of spectrum and based on CARS feature
Wavelength extracts the SVM model established, and SVM model TVB-N index prediction effect is better than all band PLS model, for finally preferred
Prediction model is used for subsequent detection.
The part d: the detection of mutton Volatile Base Nitrogen is carried out using prediction model
It is carried out using quantitative analysis process of the final preferred prediction model to TVB-N index according to step A ~ E sequence, it is first
Sample to be tested is first prepared, sample high spectrum image area-of-interest is then obtained and extracts averaged spectrum, is substituted into final preferably pre-
It surveys model and obtains TVB-N detected value.Meanwhile for the detection effect of comparison model, having acquired, sample high spectrum image is laggard
It has gone corresponding experiment, has obtained the TVB-N measured value of meat samples to be measured, by comparing the correlation of measured value and detected value,
The detection effect of model is evaluated.It is the preparation of sample to be tested, high spectrum image acquisition, the measurement of TVB-N index, interested
The acquisition in region is carried out referring to the present embodiment c1 ~ c4 operating process.
Experiment is prepared for 14 samples for model inspection altogether, and table 8 is the TVB-N Indexs measure result of sample to be tested.Benefit
Sample TVB-N value is detected with the SVM prediction model for extracting 1 averaged spectrum and preferably 8 spectrum respectively, root mean square misses
Poor RMSEV is respectively 4.75 and 3.33, coefficient RVRespectively 0.80 and 0.92.Figure 10 is the TVB-N index of sample to be tested
Detected value and measured value distribution map.The result shows that using preferred 8 establishment of spectrum preferred SVM prediction model stability more
Good, predictive ability is stronger, effectively improves the quantitative detection result of mutton TVB-N index.
Table 8 utilizes the TVB-N value testing result of SVM prediction model
A kind of embodiment 3: the high spectral image detecting method of the adulterated amount ratio of fox meat in mutton
The part e: the high spectrum image prediction model of the adulterated amount ratio of fox meat in mutton is established
E1, sample preparation
Experimental material for mutton adulteration detection includes mutton and fox meat.Wherein mutton is derived from the back leg position of sheep,
Fox meat is derived from the fox meat sample product of 3 freezings.Meat removes fat after being transported to laboratory and connective tissue, stripping and slicing simultaneously sufficiently twist
It is broken into meat gruel, is 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50% totally 10 gradients according to practical adulterated amount ratio
It is 20g that fox meat, which is mixed, in mixing sample preparation, each sample quality in surface plate, and 80 adulterated meat samples are made altogether.It will be real
Test sample be packed into vacuum bag hermetic package label after it is to be measured in being stored in 0 DEG C of refrigerator-freezer.Table 9 is the comparison knot of the adulterated amount of fox meat
Fruit.
The adulterated amount comparing result of 9 fox meat of table
E2, line scanning high spectrum image acquisition
Hyperspectral imager mainly includes imaging spectrometer (ImSpector V10E-QE, Finland), line array CCD phase
Machine (hamamastsu), light source (150W optical fiber halogen lamp, China), pulse conveying device (Zolix, SC300-1A, Beijing),
Camera bellows etc..Spectra collection range is 408 ~ 1013 nm, 0.61 nm of spectral resolution.
Before high spectrum image acquisition, opens light source and camera preheats half an hour, sample is taken out before sweeping sample and places 5 points
Clock, so that the physicochemical properties such as color are restored.High spectrum image acquisition system parameter setting: the time for exposure is set as 16 ms, object
Away from for 33.5 cm, 1.35 mm/s of Image Acquisition speed, high spectrum image acquisition is carried out again after carrying out black and white correction.
The acquisition of e3, sample high spectrum image area-of-interest
It is identical as described in a4, firstly, in wave band subtraction, binaryzation and masking method removal high spectrum image
Background and dash area；Secondly, with wave band add operation and masking method removal bright spot etc..Remainder is as sample height
Spectrum picture area-of-interest.
The extraction of e4, a plurality of spectrum of sample area-of-interest
The spectral information that 460 ~ 1013 nm of adulterated meat samples are chosen in the present embodiment is tested and analyzed.It extracts more
When spectrum, to each adulterated meat samples using sample center as the center of circle, pixel is pressed in the fan-shaped region for dividing same size
It is respectively extracted according to the spectrum item number that need to be extracted, 1,2,4,8 spectrum is extracted respectively to each adulterated sample, data were extracted
Journey is as shown in figure 3, Figure 11 is the primary light spectrogram for extracting 8 spectrum.
E5, exceptional spectrum are rejected
Using Q residual error boundary combination Hotelling T2Boundary rejecting abnormalities spectrum.Distinguish 1,2,4,8 spectroscopic data is extracted
Eliminate 2,5,8,13 exceptional spectrums.
E6, different spectrum extract the modeling sample collection statistics and analysis of item number
80 adulterated each gradients of meat samples of preparation include 8 samples, and 5 are randomly selected from each gradient as school
Positive collection sample, remaining 3 as verifying collection sample.Table 10 is the sample set stroke of different spectrum extraction item numbers after rejecting abnormalities spectrum
Divide result.
The different spectrum of table 10 extract the adulterated meat samples collection statistical result of item number
E7, the processing of adulterated sample set spectroscopic data
The present embodiment examines the adulterated mutton of incorporation different proportion fox meat by establishing PLS quantitative detection model
It surveys.In modeling process, using the preferred preprocessing procedures of a validation-cross are stayed, for extracting the preferred pre- of 1 averaged spectrum
The processing method preprocess method that smooth, centralization combines for 13 point S-G, the preferred pretreatment method for extracting 8 spectrum are
The preprocess method that 17 point S-G are smooth, centralization combines.
E8, preferably spectrum extract the determination of item number and the foundation of the preferred prediction model of all band
The adulterated amount ratio of PLS model for extracting 1,2,4,8 establishment of spectrum to adulterated meat samples high spectrum image is built
Mould result is as shown in table 11.Each extraction item number modeling achieves good modeling result, and each evaluation it can be seen from table
Index has some difference, and modeling sample collection and verification sample collection related coefficient are up to 0.9 or more.Extract 8 spectrum modeling effects
It is optimal, RC、RCVAnd RPRespectively 0.99,0.97 and 0.96.RMSEC, RMSECV and RMSEP are respectively 1.17,2.27 and
2.83,1,2,4 spectrum is extracted in comparison, and modeling effect has obtained certain improvement.Therefore, each sample extracts 8 establishment of spectrum
PLS model is preferred prediction model under all band.
The different spectrum of table 11 extract the adulterated amount ratio modeling result of PLS model of item number
E9, characteristic wavelength extract and final preferably prediction model determines
Characteristic wavelength is extracted to 8 spectrum preferably extracted in conjunction with CARS algorithm and SR algorithm and obtains 8 variables: 472,
552,573,767,772,945,963,975 nm, variable distribution as shown in figure 12, extract sample average spectrum to obtain 11
Variable are as follows: 472,560,579,592,646,764,769,834,953,972,991 nm.
Table 12 is the adulterated amount ratio SVM model prediction result of fox meat established according to characteristic wavelength information.As can be seen that
Preferred 8 spectrum modeling results are superior to 1 averaged spectrum modeling, calibration set RCIt is respectively 0.99 and 1.92 with RMSEC, hands over
The mutually R of verifying collectionCVIt is respectively 0.97 and 2.07 with RMSECV, verifies the R of collectionPIt is respectively 0.97 and 2.36 with RMSEP.
The different spectrum of table 12 extract the adulterated amount scale prediction result of SVM model of item number
Compare each sample high spectrum image and extracts all band PLS prediction model of 8 establishment of spectrum and based on CARS-SR spy
It levies wavelength and extracts the SVM model established, SVM modelling effect is better than PLS-DA model, after being used for for final preferred prediction model
Continuous detection.
The part f: the adulterated amount ratio of fox meat in mutton is carried out using prediction model and is detected
Using final preferably prediction model to the detection process of the adulterated amount ratio of fox meat in adulterated meat samples according to step
Rapid A ~ E sequence carries out, and prepares adulterated sample to be measured first, then obtains sample high spectrum image area-of-interest and extracts average
Spectrum substitutes into final preferably prediction model and obtains adulterated amount ratio detected value.By comparing practical adulterated amount ratio and detected value
Correlation, the detection effect of model is evaluated.Preparation, high spectrum image acquisition and the area-of-interest of adulterated sample
Acquisition referring to the present embodiment e1 ~ e3 operating process carry out.
Experiment is prepared for 30 adulterated samples for model inspection altogether.After extracting sample average spectrum, preferred 8 are utilized
Spectrum and extraction the built SVM prediction model of averaged spectrum detect the adulterated amount ratio of sample fox meat, and test sample is square
Root error RMSEV is respectively 2.83 and 3.98, coefficient RVRespectively 0.95 and 0.91.Model inspection result such as table 13 and figure
Shown in 13.The result shows that it is higher using final preferred SVM model inspection precision, improve the detection effect of mutton adulteration.
Table 13 utilizes the adulterated amount ratio testing result of SVM prediction model
The present invention by three embodiments elaborates that mutton high spectrum image information is made full use of to improve from different perspectives
The detection effect of quality of mutton index, the result detected from model foundation and application model can be seen that either for
The qualitative discrimination of mutton grade of freshness, or for the quantitative detection of mutton Volatile Base Nitrogen index, also or for mixing
The adulterated amount ratio quantitative detection of false mutton extracts a plurality of spectrum and exponentially expands modeling sample amount, and passes through feature extraction
There has also been further increasing, the effect that finally built prediction model carries out quality of mutton Indexs measure obtains prediction model effect afterwards
Preferable improvement is arrived.Testing result show the present invention by making full use of high spectrum image information preferably to extract a plurality of spectrum, into
And prediction model is established, realize the accurate detection of quality of mutton index.
Based on hyper-spectral image technique to other index of quality of mutton, and the index of quality of other meats is carried out
When non-destructive testing, it can refer to proposed detection method and testing process operated.
The foregoing is merely three embodiments of the invention, are not intended to limit the invention, all in spirit of the invention
Within principle, any modification, equivalent replacement, improvement on the basis of the technology of the present invention essence should be included in this
Within the scope of invention.
Claims (10)
1. a kind of high spectral image detecting method of quality of mutton index first establishes prediction model, model inspection mutton is recycled
The index of quality, it is characterised in that: first establish the prediction model of the mutton high spectrum image index of quality, prediction model is recycled to carry out
Quality of mutton Indexs measure；
Specific step is as follows for the prediction model for establishing the mutton high spectrum image index of quality:
Step 1: sample preparation, to the meat sample removal surrounding fascia and connective tissue after government official, sample is using slice or meat gruel is made
Two ways preparation is mixed afterwards；Slicing mode sample preparation is used when being grade of freshness, Volatile Base Nitrogen for the index of quality, it is right
In meat detection of adulterations, mixing is placed in sample preparation in surface plate after meat gruel is made in meat；The index of quality need to be considered when sample preparation
The representativeness of distribution；
Step 2: the sample of preparation is placed on phase shift platform, mutton is obtained by the movement of translation stage and line scan camera
Sample high spectrum image three-dimensional data block；
Step 3: obtaining meat samples index of quality physics and chemistry value, including grade of freshness, TVB-N content or adulterated amount
Ratio；
Step 4: obtaining the area-of-interest of meat samples high spectrum image relevant to the index of quality；
Step 5: according to the index of quality sample area-of-interest distribution character to the area-of-interest of sample high spectrum image
Each pixel be reconstructed, sequentially extract discrepant a plurality of spectrum；
Step 6: carrying out the rejecting of exceptional spectrum to extracted meat samples spectroscopic data；
Step 7: by remaining spectroscopic data after rejecting abnormalities spectrum and grade of freshness, Volatile Base Nitrogen, adulterated amount ratio
It corresponds, collects according to the ratio cut partition calibration set and verifying of 3:1；
Step 8: the different index of quality is directed to, using calibration set spectroscopic data preferred pretreatment method, and using identical pre-
Processing method handles verifying collection spectrum, is analyzed using pretreated data；
Step 9: extract item number for different spectrum establishes according to the index of quality offset minimum binary quantitative model or partially most respectively
Small two multiply qualitative discrimination model, and are verified using verifying collection sample, extract item number according to the preferred spectrum of modelling effect, determine
Preferred all band prediction model, it is feasible that modeling effect meets the requirements then representative model；Otherwise, step 5 is repeated to nine, until
It meets the requirements；
Step 10: being utilized respectively competitive adaptive weight weighting algorithm, the Stepwise Regression Algorithm and successive projection algorithm and combinations thereof
Mode is to quality of mutton index extraction characteristic wavelength；
Step 11: establishing each index of quality respectively based on characteristic wavelength using the characteristic wavelength extracted as input variable
It is preferred that prediction model, and the preferred prediction model of all band obtained with step 9 is compared, and determines final prediction model；
It is described using prediction model to carry out quality of mutton Indexs measure specific step is as follows:
Step A, meat samples to be detected are prepared；
Step B, the high spectrum image three-dimensional data block of meat samples to be detected is obtained；
Step C, acquisition meat samples high spectrum image area-of-interest relevant to the index of quality；
Step D, an averaged spectrum is extracted for detecting to each pixel of sample high spectrum image area-of-interest, if carrying out
Visual analyzing then extracts the spectral information of each pixel for detecting；
Step E, spectral information is substituted into built index of quality prediction model, obtains the index of quality prediction of mutton to be detected
Value, until the detection of all mutton test samples finishes.
2. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 1, the meat samples size for using slicing mode to prepare is used for adulterated content detection for the cm of cm × 1.5 of 4 cm × 4
Sample be initially formed the meat gruels of 2 ~ 3 millimeters of partial sizes, prepare according to different proportion and be laid in surface plate after mixing, surface used
Ware format diameter is less than 10 cm, floor space 50cm2, according to above-mentioned sample making course, obtained sample shape feature have rectangle or
Nearly rectangle, circle or subcircular, rectangle or nearly rectangle are slicing mode sample preparation, and round or subcircular is meat gruel blending manner system
Sample.
3. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 3, the quality quantitative targets of meat samples is Volatile Base Nitrogen, the adulterated amount ratio of adulterated meat, total number of bacteria, tender
Degree, pH value；Quality qualitative index is the grade of freshness obtained according to freshness overall merit, adulterated classification, tenderness grade.
4. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 4, the acquisition of area-of-interest is carried out in two steps；On the one hand, sample is obtained with wave band subtraction, binaryzation operation
Then binary image carries out exposure mask to high spectrum image, removes the background and dash area of high spectrum image, on the other hand,
With wave band add operation and binaryzation, the fat of mask process removal sample, bright spot position, acquisition can react the index of quality
Property, the sample area-of-interest with different, the pixel for including are 12 ~ 150,000.
5. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 5, the index of quality is distributed with approaches uniformity distribution and dissipates distribution around from center, when sample is the adulterated of mixing
When sample, each position quality index values difference is not distributed for approaches uniformity significantly；When meat evaluation index is volatility alkali
When nitrogen, since the decay process of sample is gradually to develop since surrounding to center, the index of quality is in send out around from center
Dissipate distribution.
6. a kind of high spectral image detecting method of quality of mutton index according to claim 1 or 4, it is characterised in that:
In the step 5, to each pixel in region of interest domain space with the direction perpendicular or parallel with line scanning direction by sweeping
It retouches sequencing and is reassembled as a column, all pixels point is respectively extracted according to the spectrum item number that need to be extracted, or by selected
Sample space central point selects a starting cut-off rule position, to all pixels point according to the spectrum item number that need to be extracted by 360 ° of areas
It is respectively extracted in domain；The spectrum item number of extraction is related with the pixel number that every spectrum is included, and is total pixel number and light
The ratio of item number is composed, in order to keep the spectrum extracted representative, it is desirable that the pixel that every spectrum is included is greater than 2500,
Meet the different spectrum item numbers of spectrum of the pixel greater than 2500 with extraction and the combination spectrum of a plurality of spectrum is analyzed；
For ease of operation, 8,4,2,1 spectrum are extracted to be compared.
7. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 6, since noise of instrument, experimental situation, thickness of sample factor influence, there can be exceptional spectrum in modeling process, thus
Model accuracy is influenced, Q residual error boundary combination Hotelling T need to be used2Boundary rejecting abnormalities spectrum.
8. a kind of high spectral image detecting method of quality of mutton index according to claim 1 or 3, it is characterised in that:
In the step 8, in the case of each sample preferably extracts 8 spectrum, for freshness be classified, the preferred spectrum of each sample collection it is pre-
Processing method is the method that first derivative, 15 point S-G smooth, variable standardization, centralization processing combine；For volatile salts
The prediction of base nitrogen index, it is that second dervative, 23 point S-G are smooth, centralization processing phase that each sample, which integrates the preprocess method of preferred spectrum,
In conjunction with method；Preprocess method smooth, centralization for 17 point S-G for the preferred spectrum of adulterated amount scale prediction each sample collection
The method combined.
9. a kind of high spectral image detecting method of quality of mutton index according to claim 1, it is characterised in that: described
In step 10, the characteristic wave that grade of freshness is extracted with 11, respectively 530,558,634,668,753,765,776,
785,833,859,957 nm；To the characteristic wave of TVB-N index extraction with 12, respectively 527,571,605,637,649,
753,770,799,809,838,848,969 nm；The characteristic wave that adulterated amount ratio is extracted with 8, respectively 472,552,
573、767、772、945、963、975 nm。
10. according to claim 1 or a kind of high spectral image detecting method of quality of mutton index, feature described in 3,9 exist
In: in the step 9, all band grade of freshness PLS- is established respectively using a plurality of spectrum that each sample collection of division extracts
The PLS quantitative model of DA qualitative model or Volatile Base Nitrogen, mutton adulteration, according to the preferred spectrum of modelling effect extract item number and
Prediction model；PLS quantitative model is commented according to calibration set, validation-cross collection, the root-mean-square error of verifying collection and related coefficient synthesis
The preferred spectrum of valence and its prediction model, PLS-DA qualitative model are comprehensive according to calibration set, validation-cross collection and verifying collection accuracy rate
Evaluate preferred spectrum and its prediction model；In the step 11, each index of quality is established using the characteristic wavelength information of extraction
Preferred prediction model be support vector machines model, evaluation index is identical as step 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710598787.XA CN107271375B (en) | 2017-07-21 | 2017-07-21 | A kind of high spectral image detecting method of quality of mutton index |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710598787.XA CN107271375B (en) | 2017-07-21 | 2017-07-21 | A kind of high spectral image detecting method of quality of mutton index |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107271375A CN107271375A (en) | 2017-10-20 |
| CN107271375B true CN107271375B (en) | 2019-10-01 |
Family
ID=60078887
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710598787.XA Active CN107271375B (en) | 2017-07-21 | 2017-07-21 | A kind of high spectral image detecting method of quality of mutton index |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107271375B (en) |
Families Citing this family (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108444798A (en) * | 2018-01-29 | 2018-08-24 | 华中农业大学 | A kind of beef adulteration detection method based on Biospeckles and the moment of inertia spectrum analysis |
| CN108593565A (en) * | 2018-03-19 | 2018-09-28 | 河南科技学院 | The method of on-line quick detection chicken pseudomonad content |
| CN110931087B (en) * | 2018-09-03 | 2023-06-16 | 上海烟草集团有限责任公司 | Method for monitoring conventional chemical components of tobacco product |
| CN110779467B (en) * | 2019-10-24 | 2020-10-27 | 中国科学技术大学 | Shadow phase error compensation method and device |
| CN110675400B (en) * | 2019-11-15 | 2023-08-01 | 石河子大学 | Mutton quality index rapid intelligent detection method based on mobile phone APP |
| CN110987824A (en) * | 2019-12-04 | 2020-04-10 | 中国农业科学院兰州畜牧与兽药研究所 | Beef sample analysis method |
| CN111445469A (en) * | 2020-04-15 | 2020-07-24 | 天津商业大学 | Hyperspectrum-based apple damage parameter lossless prediction method after impact |
| CN112285031A (en) * | 2020-09-30 | 2021-01-29 | 江苏大学 | Method for rapidly representing umami intensity of salmon based on hyperspectral imaging technology |
| CN112697718A (en) * | 2020-09-30 | 2021-04-23 | 江苏大学 | Method for quickly detecting freshness of cooked beefsteak based on hyperspectral imaging technology and wavelength selection algorithm tandem strategy |
| CN112415050B (en) * | 2020-11-13 | 2024-01-23 | 石河子大学 | Mutton mixing supposition discriminating method based on temperature distribution difference |
| CN113406013A (en) * | 2021-04-13 | 2021-09-17 | 江苏师范大学 | Meat product quality detection device and method based on hyperspectral and near-infrared fusion |
| CN113109272B (en) * | 2021-04-20 | 2023-01-24 | 石河子大学 | Drip irrigation cotton nitrogen nutrition deficiency diagnosis method and system |
| CN113567359B (en) * | 2021-08-10 | 2022-05-20 | 江苏大学 | Raw cut meat and high meat-imitation identification method thereof based on component linear array gradient characteristics |
| CN113533581B (en) * | 2021-08-11 | 2022-08-02 | 贵州中烟工业有限责任公司 | Method for analyzing substance quality based on information entropy and computer-readable storage medium |
| CN113607660B (en) * | 2021-08-27 | 2023-11-28 | 石河子大学 | Adulterated mutton qualitative detection method based on visual imaging in time sequence temperature change process |
| CN114295611B (en) * | 2021-08-27 | 2023-10-03 | 石河子大学 | Mutton mixing assumption detection method based on visual imaging in time sequence temperature change process |
| CN113959961B (en) * | 2021-12-22 | 2022-03-08 | 广东省农业科学院动物科学研究所 | Hyperspectral image-based tannin additive anti-counterfeiting detection method and system |
| CN114371144A (en) * | 2021-12-31 | 2022-04-19 | 中国农业科学院草原研究所 | Establishment method and application of database for detecting mutton biogenic amine content based on near infrared spectrum imaging technology |
| CN116992259B (en) * | 2023-09-28 | 2023-12-08 | 奥谱天成（厦门）光电有限公司 | Spectral feature extraction method and detection method suitable for biological quality detection |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102507459A (en) * | 2011-11-23 | 2012-06-20 | 中国农业大学 | Method and system for quick lossless evaluation on freshness of fresh beef |
| CN102564964A (en) * | 2011-12-29 | 2012-07-11 | 南京林业大学 | Spectral image-based meat quality visual non-contact detection method |
| CN103257118A (en) * | 2013-04-22 | 2013-08-21 | 华南理工大学 | Fish tenderness hyperspectral detection method based on characteristic wave band |
| CN103439271A (en) * | 2013-08-29 | 2013-12-11 | 华南理工大学 | Visual detecting method for pork maturity condition |
| CN104655586A (en) * | 2015-02-28 | 2015-05-27 | 华南理工大学 | Hyperspectral-data-fusio-based fast non-contact fish fat oxidation monitoring method |
| CN105548029A (en) * | 2015-12-14 | 2016-05-04 | 北京农业质量标准与检测技术研究中心 | Meat product freshness detection method based on spectral imaging technology |
| CN106596416A (en) * | 2016-11-25 | 2017-04-26 | 华中农业大学 | Chilled fresh meat quality non-destructive testing method based on hyperspectral imaging technology |
-
2017
- 2017-07-21 CN CN201710598787.XA patent/CN107271375B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102507459A (en) * | 2011-11-23 | 2012-06-20 | 中国农业大学 | Method and system for quick lossless evaluation on freshness of fresh beef |
| CN102564964A (en) * | 2011-12-29 | 2012-07-11 | 南京林业大学 | Spectral image-based meat quality visual non-contact detection method |
| CN103257118A (en) * | 2013-04-22 | 2013-08-21 | 华南理工大学 | Fish tenderness hyperspectral detection method based on characteristic wave band |
| CN103439271A (en) * | 2013-08-29 | 2013-12-11 | 华南理工大学 | Visual detecting method for pork maturity condition |
| CN104655586A (en) * | 2015-02-28 | 2015-05-27 | 华南理工大学 | Hyperspectral-data-fusio-based fast non-contact fish fat oxidation monitoring method |
| CN105548029A (en) * | 2015-12-14 | 2016-05-04 | 北京农业质量标准与检测技术研究中心 | Meat product freshness detection method based on spectral imaging technology |
| CN106596416A (en) * | 2016-11-25 | 2017-04-26 | 华中农业大学 | Chilled fresh meat quality non-destructive testing method based on hyperspectral imaging technology |
Non-Patent Citations (5)
| Title |
|---|
| Detection of internal qualities of hami melons using hyperspectral imaging technology based on variable selection algorithms;Jingtao Sun et al.;《Journal of Food Process Engineering》;20161231;第1-10页 * |
| 基于GA和CARS的真空包装冷却羊肉细菌菌落总数高光谱检测;段宏伟 等;《光谱学与光谱分析》;20170310;第37卷(第3期);第847-852页 * |
| 基于高光谱图像和偏最小二乘的羊肉pH值特征波段筛选研究;朱荣光 等;《光谱学与光谱分析》;20160930;第36卷(第9期);第2925-2929页 * |
| 感兴趣区域和预处理方法对羊肉高光谱图像提取光谱的影响;黄旭阳 等;《新疆农垦科技》;20151231;第36-39页 * |
| 羊肉挥发性盐基氮的高光谱图像快速检测研究;朱荣光 等;《光谱学与光谱分析》;20160331;第36卷(第3期);第806-810页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107271375A (en) | 2017-10-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107271375B (en) | A kind of high spectral image detecting method of quality of mutton index | |
| Che et al. | Pixel based bruise region extraction of apple using Vis-NIR hyperspectral imaging | |
| Barbin et al. | Non-destructive determination of chemical composition in intact and minced pork using near-infrared hyperspectral imaging | |
| Barbin et al. | Predicting quality and sensory attributes of pork using near-infrared hyperspectral imaging | |
| ElMasry et al. | Image analysis operations applied to hyperspectral images for non-invasive sensing of food quality–a comprehensive review | |
| Kamruzzaman et al. | Non-destructive prediction and visualization of chemical composition in lamb meat using NIR hyperspectral imaging and multivariate regression | |
| Kamruzzaman et al. | Fast detection and visualization of minced lamb meat adulteration using NIR hyperspectral imaging and multivariate image analysis | |
| Briottet et al. | Military applications of hyperspectral imagery | |
| Wu et al. | Potential of hyperspectral imaging and multivariate analysis for rapid and non-invasive detection of gelatin adulteration in prawn | |
| Wu et al. | Novel non-invasive distribution measurement of texture profile analysis (TPA) in salmon fillet by using visible and near infrared hyperspectral imaging | |
| Hyyppä et al. | Accuracy comparison of various remote sensing data sources in the retrieval of forest stand attributes | |
| Park et al. | AE—Automation and emerging technologies: Co-occurrence matrix texture features of multi-spectral images on poultry carcasses | |
| Huang et al. | Assessment of intramuscular fat content of pork using NIR hyperspectral images of rib end | |
| Weyermann et al. | Correction of reflectance anisotropy effects of vegetation on airborne spectroscopy data and derived products | |
| CN109100323B (en) | Nondestructive quantitative evaluation method for transmission spectrum of apple water core disease | |
| Grotti et al. | An intensity, image-based method to estimate gap fraction, canopy openness and effective leaf area index from phase-shift terrestrial laser scanning | |
| Scheffler et al. | Spectral harmonization and red edge prediction of Landsat-8 to Sentinel-2 using land cover optimized multivariate regressors | |
| Calvini et al. | Data dimensionality reduction and data fusion for fast characterization of green coffee samples using hyperspectral sensors | |
| Huang et al. | Rapid and non-invasive quantification of intramuscular fat content of intact pork cuts | |
| Blanchette et al. | Predicting wood fiber attributes using local-scale metrics from terrestrial LiDAR data: A case study of Newfoundland conifer species | |
| Stocco et al. | Accuracy and biases in predicting the chemical and physical traits of many types of cheeses using different visible and near-infrared spectroscopic techniques and spectrum intervals | |
| CN108956604A (en) | A method of Eriocheir sinensis quality is identified based on hyper-spectral image technique | |
| Naganathan et al. | Hyperspectral imaging of ribeye muscle on hanging beef carcasses for tenderness assessment | |
| Li et al. | On the utilization of novel spectral laser scanning for three-dimensional classification of vegetation elements | |
| Li et al. | Seasonal change of leaf and woody area profiles in a midlatitude deciduous forest canopy from classified dual-wavelength terrestrial lidar point clouds |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |