Protein and wet gluten content detection method based on feature-level wave band fusion

Abstract

The invention discloses a protein and wet gluten content detection method based on feature-level wave band fusion, and belongs to the field of hyperspectral data processing. According to the method, a rapid prediction model of the content of the protein and the wet gluten in the wheat is constructed by utilizing a hyperspectral imaging technology in the spectral ranges of 400-1000 nm and 900-1700 nm in combination with different pretreatment and characteristic wavelength selection methods, and accurate prediction of the content of the protein and the wet gluten in the wheat is realized by performing characteristic-level wave band fusion. The method has good theoretical research significance and practical application value, and provides reference for estimation of the content of protein and wet gluten in other crops.

Description

Technical Field

[0001] The invention relates to a method for detecting protein and wet gluten content based on characteristic-level band fusion, and belongs to the field of hyperspectral data processing. Background Art

[0002] With the development of economy and the improvement of living standards, people's demand for food is no longer limited to satisfying hunger, but they are more concerned about their appearance quality, nutritional value and taste. Among them, protein and wet gluten content directly affect the nutritional value and quality of wheat, and are important indicators for evaluating wheat quality. Hyperspectral imaging technology has the advantages of being fast, objective, wide-range, and non-destructive. It can save a lot of chemical reagents without polluting the environment. In recent years, it has been widely used in non-destructive detection of protein and wet gluten content of wheat.

[0003] Among the current methods for detecting wheat protein and wet gluten content using hyperspectral imaging technology, Puneet Mishra et al. proposed a method for predicting the protein content in wheat grains using two different preprocessing fusion methods, sequential fusion and parallel fusion (see "Mishra, Puneet, and Santosh Lohumi." Improved prediction of protein content in wheat kernels with a fusion of scatter correction methods in NIR data modelling." Biosystems Engineering 203(2021):93-97."). Compared with the single scatter correction technology, the PLSR model established using sequential fusion and parallel fusion has prediction correlation coefficients of 0.89 and 0.90, respectively, and the prediction accuracy is improved. Zhang et al. proposed a method to predict the protein and wet gluten content in wheat flour by using hyperspectral and RGB sensors to obtain the original reflectance (OR), wavelet features (WF) and color index (CI) of wheat flour, and used the CARs-VIF algorithm to screen the features of OR and WF, and finally fused the three to improve the prediction accuracy (see "Zhang, Shaohua, et al." Estimation of wheat protein content and wet gluten content based on fusion of hyperspectral and RGB sensors using machine learning algorithms." Food Chemistry 448 (2024): 139103."). The results show that when OR+WF+CI is used in combination, the RFR model has the best prediction effect on the protein and wet gluten content of wheat flour, and the correlation coefficients of the prediction sets are 0.864 and 0.847, respectively.

[0004] Among the above methods, one was studied in the data processing stage, and the other was fused after screening different features in the feature screening stage. The method is not universal for different forms of wheat (including wheat grains and wheat flour), that is, a certain pretreatment method or a certain feature fusion method is not suitable for the prediction of protein or wet gluten content in different forms of wheat, and the combination of the pretreatment stage and the feature fusion stage is not considered. Therefore, a new method is needed to predict the protein or wet gluten content in wheat. Summary of the invention

[0005] In order to improve the prediction accuracy of protein or wet gluten content in wheat, the present invention provides a method for detecting protein and wet gluten content based on feature-level band fusion. Based on wheat in different forms, different methods are used to perform feature-level band fusion for the prediction of protein and wet gluten content to achieve high-precision prediction.

[0006] The first object of the present invention is to provide a protein content detection method based on feature-level band fusion, which is used to detect the protein content of wheat grains, comprising:

[0007] Step 1, obtaining hyperspectral image data of wheat grains in the 400-1000nm and 900-1700nm bands respectively;

[0008] Step 2: The hyperspectral image data in the 400-1000 nm band are preprocessed using 1st+MSC, and the hyperspectral image data in the 900-1700 nm band are preprocessed using 1st+SNV;

[0009] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0010] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0011] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat grains to be tested.

[0012] The second object of the present invention is to provide a protein content detection method based on feature-level band fusion, which is used to detect the protein content of wheat flour, comprising:

[0013] Step 1, obtaining hyperspectral image data of wheat flour in the 400-1000nm and 900-1700nm bands respectively;

[0014] Step 2: The hyperspectral image data in the 400-1000nm and 900-1700nm bands were preprocessed using 1st+SNV;

[0015] Step 3, using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using UVE+CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0016] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0017] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat flour to be tested.

[0018] The third object of the present invention is to provide a wet gluten content detection method based on feature-level band fusion, which is used to detect the wet gluten content of wheat grains, comprising:

[0019] Step 1, obtaining hyperspectral image data of wheat grains in the 400-1000nm and 900-1700nm bands respectively;

[0020] Step 2: Preprocess the hyperspectral image data in the 400-1000 nm band using 1st+SNV, and preprocess the hyperspectral image data in the 900-1700 nm band using 1st+MSC;

[0021] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm and 900-1700nm bands;

[0022] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0023] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat grains to be tested.

[0024] A fourth object of the present invention is to provide a wet gluten content detection method based on feature-level band fusion, which is used to detect the wet gluten content of wheat flour, comprising:

[0025] Step 1, obtaining hyperspectral image data of wheat flour in the 400-1000nm and 900-1700nm bands respectively;

[0026] Step 2: The hyperspectral image data in the 400-1000nm and 900-1700nm bands were preprocessed using 1st+SNV;

[0027] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0028] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0029] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat flour to be tested.

[0030] The beneficial effects of the present invention are:

[0031] The present invention takes wheat as the research object, collects wheat of different varieties in the same period, acquires spectral data of two forms of wheat grains and wheat flour, and determines their protein and wet gluten contents. Subsequently, a comprehensive analysis is conducted on the spectral characteristics of wheat and the quantitative relationship between its protein and wet gluten contents, and an estimation model for wheat protein and wet gluten contents based on the spectral ranges of 400-1000nm and 900-1700nm is established.

[0032] This application uses hyperspectral imaging technology in the 400-1000nm and 900-1700nm spectral ranges combined with different preprocessing and characteristic wavelength selection methods to construct a rapid prediction model for the protein and wet gluten content in wheat. Through further characteristic-level band fusion, the accurate prediction of wheat protein and wet gluten content is achieved, which has good theoretical research significance and practical application value. At the same time, it provides reference and reference for the estimation of protein and wet gluten content in other crops. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 It is the original spectrum of wheat grains based on the wavelength of 400-1000nm;

[0035] Figure 2 This is the original spectrum of wheat flour based on the wavelength of 400-1000nm;

[0036] Figure 3It is the original spectrum of wheat grains based on the wavelength of 900-1700nm;

[0037] Figure 4 It is the original spectrum of wheat flour based on the wavelength of 900-1700nm.

[0038] Figure 5 This is a trend chart of the number of sampling variables of CRAS at each wavelength of 900-1700nm during the prediction of wheat flour protein content;

[0039] Figure 6 This is the cross-validation root mean square error RMSECV graph of the number of sampling variables for each wavelength at 900-1700nm for the prediction process of wheat flour protein content;

[0040] Figure 7 This is a trend chart of the regression coefficient of CRAS for the number of sampling variables at each wavelength at 900-1700nm in the process of predicting the protein content of wheat flour.

[0041] Figure 8 It is the variation diagram of SPA root mean square error of wheat flour protein at wavelength of 900-1700nm;

[0042] Fig. 9 It is a graph of the selected characteristic wavelengths corresponding to the optimal pretreatment spectrum of SPA of wheat flour protein at a wavelength of 900-1700nm.

[0043] Fig.10 It is the characteristic wavelength diagram extracted by UVE algorithm of wheat flour protein at wavelength of 900-1700nm;

[0044] Fig.11 It is a graph of the selected characteristic wavelengths corresponding to the optimal pretreatment spectrum of the UVE of wheat flour protein at a wavelength of 900-1700nm. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0046] First, the sample source, spectral data acquisition method, protein content determination method, wet gluten content determination method and experimental instruments and equipment involved in the method of the present invention are described as follows:

[0047] (1) Sample preparation

[0048] ①Wheat grain preparation: Collect 160 kinds of wheat grains with a certain degree of commonality and store them in a cold storage at 4℃.

[0049] ② Preparation of wheat flour: According to NY / T 1094.1-2006 “Experimental Flour Milling of Wheat”, the wheat grains were tempered to make the moisture content of wheat distributed at about 14.5%, and then ground into powder using a Brabender mill, passed through a 100-mesh sieve, and stored for later use.

[0050] (2) Spectral data acquisition.

[0051] Turn on the hyperspectral instrument and preheat it for 30 minutes to allow the hyperspectral imaging system to reach a stable state, and perform system calibration; after debugging, measure the exposure time of 13ms, the platform movement speed of 6.93mm / s and the detection wavelength of 400-1000nm; measure the exposure time of 16ms, the platform movement speed of 1.95mm / s and the detection wavelength of 900-1700nm; then put the samples to be tested (including wheat flour and wheat grains) into the hyperspectral test platform, start the spectrometer Spectral Image software to scan, and obtain the hyperspectral images of wheat grains and wheat flour at 400-1000nm and 900-1700nm respectively.

[0052] (3) Protein content determination.

[0053] The Kjeldahl nitrogen determination method in GB / T 5009.5-2016 “Determination of protein in food” was used for determination.

[0054] ① Determination of protein content in wheat grains: Considering that the difference between the protein content in wheat grains and that in wheat flour is small, the protein content in wheat flour is subsequently used as the protein content in wheat grains.

[0055] ② Determination of protein content in wheat flour: Weigh 0.2g wheat flour, accurate to 0.001g, place it in a digestion tube, add 0.2g copper sulfate, 3g potassium sulfate and 10mL sulfuric acid to the digestion furnace for digestion. When the temperature of the digestion furnace reaches 420℃, continue digestion for 1h. When clear blue-green color appears, digestion is complete. Take it out and cool it, add 50mL water, and automatically add liquid, distill and titrate on the fully automatic Kjeldahl nitrogen analyzer to determine its nitrogen content.

[0056] Protein content of sample = total nitrogen content of flour × 5.7.

[0057] (4) Determination of wet gluten content.

[0058] The gluten content of wheat and wheat flour was determined by the instrumental method specified in GB / T 5506.2-2008.

[0059] ① Determination of wet gluten content in wheat grains: Considering that the difference between the wet gluten content in wheat grains and that in wheat flour is small, the wet gluten content in wheat flour is subsequently used as the wet gluten content in wheat grains.

[0060] ② Determination of wet gluten content in wheat flour: Weigh 10g of wheat flour, accurate to 0.01g, and put it all into the washing chamber of the gluten analyzer with a metal sieve with a bottom mesh size of 80μm. Use an adjustable pipette to add about 4.8mL of sodium chloride solution to the wheat flour, and gently shake the washing chamber to make the solution evenly distributed on the surface of the sample. The kneading time is 20s and the washing time is 5min. During the operation, 250ml~280ml of sodium chloride washing solution is usually required, and the flow rate is set to 50mL / min~56mL / min. After washing, use metal tweezers to take out the wet gluten from the washing chamber, gently press it on the sieve box of the centrifuge, start the centrifuge for 60s, remove the wet gluten with metal tweezers, and weigh it immediately (m1) with an accuracy of 0.01g.

[0061] The wet gluten content of the sample = m1 × 10%.

[0062] (5) Instruments used: push-broom hyperspectral imaging system, fully automatic Kjeldahl nitrogen analyzer, gluten washer, Brabender mill, Envi 5.3, The UnscramblerX 10.4 and Matlab 2023a software.

[0063] Embodiment 1

[0064] This embodiment provides a method for detecting protein and wet gluten content based on characteristic level band fusion, the method comprising:

[0065] Step 1, obtaining spectral data and corresponding protein and wet gluten content, including:

[0066] Step 1.1, prepare samples;

[0067] Among the 160 wheat varieties used in the research, the main ones are Zhumai No. 4, Zhongdai 100, Zhongxinmai 28, Zhongmai 998, Zhengmai 9023, Zhenmai 29, Yunhan 115, Yangmai 45, Yangjiangmai 1820, Yangfumai 21, Yanmai 996, Yannong 377, Xumai 919, Xindong 22, Weilong 169, Shannong 40, Ruihuamai 598, Qimin 8, Pinyu 8161, Nongmai 988, Ningzimai 188, Ningmai 40, Ninghongmai 199, Luyuan 502, Letu 808, Jinmai 95, Jimai 418, Jimai 55, Huaimai 44, Huamai 18, Daimai 366, Bainong 4199, etc.

[0068] According to NY / T 1094.1-2006 “Experimental Flour Milling of Wheat”, the wheat was tempered to a moisture content of about 14.5%. After being ground into powder using a Brabender mill, it was passed through a 100-mesh sieve to obtain 160 powder samples, which were placed in sealed bags, packaged, numbered, and stored for later use.

[0069] Step 1.2, collecting spectral data of each sample;

[0070] Hyperspectral images of each sample were collected at wavelengths of 400-1000nm and 900-1700nm.

[0071] Before the test, the hyperspectral instrument was turned on and preheated for 30 minutes to allow the hyperspectral imaging system to reach a stable state, and the system was calibrated; after debugging, the exposure time of 13ms, the platform movement speed of 6.93mm / s and the detection wavelength of 400-1000nm were measured; the exposure time of 16ms, the platform movement speed of 1.95mm / s and the detection wavelength of 900-1700nm were measured; then the sample to be tested was placed on the hyperspectral test platform, and the spectrometer Spectral Image software was turned on to scan, and hyperspectral images of wheat grains and wheat flour at 400-1000nm and 900-1700nm were obtained.

[0072] In order to reduce the influence of laboratory light source and dark noise, after taking the hyperspectral images, the collected original hyperspectral images of wheat and wheat flour were converted into corrected images through black and white correction.

[0073] like Figures 1 to 4 As shown, there are original spectra of wheat grains and wheat flour established at a wavelength of 400-1000nm and original spectra of wheat grains and wheat flour established at a wavelength of 900-1700nm.

[0074] Step 1.3, determining the protein and wet gluten content of the sample;

[0075] The protein content was determined by the Kjeldahl method in GB / T 5009.5-2016 “Determination of protein in foods”; the wet gluten content was determined by the instrumental method in GB / T 5506.2-2008 “Gluten content of wheat and wheat flour”.

[0076] Step 2, spectral data preprocessing;

[0077] According to the morphology of the samples to be tested and the detection objects, the optimal pretreatment method under different bands is determined in combination with the subsequent characteristic wavelength screening algorithm. The morphology of the samples to be tested includes wheat grains and wheat flour, and the detection objects include protein and wet gluten. The pretreatment methods include standard normal transformation (SNV), multivariate scattering correction (MSC), first-order derivative (1st), 1st+SNV and 1st+MSC.

[0078] The spectral data in different bands obtained in step 1 are preprocessed respectively using the determined optimal preprocessing method.

[0079] Step 3, screening the preprocessed spectral data for characteristic wavelengths;

[0080] According to the morphology of the samples to be tested and the detection object, corresponding algorithms are used to screen the characteristic wavelengths for the spectral data in different bands. The algorithms include competitive adaptive reweighting algorithm (CARS), continuous projection algorithm (SPA), uninformation variable elimination (UVE) and UVE+CARS algorithm to screen the characteristic wavelengths.

[0081] like Figures 5 to 7 As shown in the figure, in the process of predicting the protein content of wheat flour, CRAS screens the characteristic wavelengths at 900-1700nm. As the number of runs increases, the number of sampling variables gradually decreases. When the number of sampling runs reaches 22, the cross-validation root mean square error value RMSECV is the smallest at 0.2296, and the corresponding number of sampling variables is the number of selected characteristic wavelengths. Figure 8 As shown in the figure, in the process of predicting the protein content of wheat flour, SPA screens the characteristic wavelengths at 900-1700nm. As the number of variables increases, the root mean square error RMSE gradually decreases. When it reaches the minimum value of 0.41197 and tends to be stable, the corresponding number of variables is 35, which is the selected characteristic wavelength. The specific characteristic wavelength selected is Fig. 9 In the display. Fig.10 As shown in the figure, in the process of predicting the protein content of wheat flour, UVE screens the characteristic wavelengths at 900-1700nm. The blue curve on the left is the stable value of the spectral variable matrix of wheat flour, and the green curve on the right is the added spectral random noise matrix variable. The two black horizontal solid lines are the maximum thresholds of random noise. The variables between the two dotted lines are the eliminated spectral invalid information variables, and the variables outside the two dotted lines are the selected modeling characteristic wavelengths. Fig.11 shown.

[0082] Step 4: Link the selected spectral data in different bands to build the corresponding PLSR model.

[0083] The partial least squares PLSR model is a relatively mature model, and the structure of the model will not be introduced in this embodiment.

[0084] Step 5, using the PLSR model constructed in step 4 to predict the protein and wet gluten content in wheat grains or wheat flour.

[0085] In the above steps, the software Envi 5.3 was used to extract the spectral data of the region of interest (ROI) in the hyperspectral image, the raw spectral data was preprocessed in the software The UnscramblerX 10.4, and the PLSR model was established after the characteristic wavelengths were screened using the CARS, SPA, UVE and UVE+CARS methods in the software Matlab2023a, and the model after the fusion of the data level and the feature level was established.

[0086] Embodiment 2

[0087] This embodiment provides a protein content detection method based on feature-level band fusion, which is used to detect the protein content of wheat grains, including:

[0088] Step 1, obtaining hyperspectral image data of wheat grains in the 400-1000nm and 900-1700nm bands respectively;

[0089] 400-1000nm contains 558 bands in total, 900-1700nm contains 453 bands in total;

[0090] Step 2: The hyperspectral image data in the 400-1000 nm band are preprocessed using 1st+MSC, and the hyperspectral image data in the 900-1700 nm band are preprocessed using 1st+SNV;

[0091] The spectral features of the hyperspectral image data in the 400-1000nm band are first enhanced by the first-order derivative 1st. The small changes in the spectrum are highlighted by calculating the first-order derivative of the original spectral data, which can be achieved through the difference method. After enhancing the spectral features, the multivariate scattering correction MSC is used to eliminate the spectral errors caused by the surface scattered light of solid particles (wheat grains) and the change of optical path in the spectral data. This is achieved based on the first-order derivative of the original spectral data, and the scattering components of the data can be determined by principal component analysis PCA.

[0092] For the hyperspectral image data in the 900-1700nm band, the first-order derivative 1st is also used to enhance its spectral characteristics, but the standard normal transformation SNV is subsequently used to eliminate the influence of factors such as different particle sizes and uneven distribution of wheat grains on the spectrum. Specifically, the data is standardized by subtracting the mean of each spectrum and dividing it by the standard deviation.

[0093] Combined with the subsequent characteristic wavelength screening algorithm, in this step, the optimal preprocessing method is adopted for the hyperspectral image data of different bands to meet the necessary conditions for improving the accuracy of the final prediction results.

[0094] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0095] For hyperspectral image data in different bands, this step also uses the optimal algorithm to screen characteristic wavelengths. Specifically, for hyperspectral image data in the 400-1000nm band, the competitive adaptive reweighting algorithm CARS is used to screen characteristic wavelengths. CARS is based on the principle of survival of the fittest and optimizes the wavelength by evaluating the importance of each wavelength using the absolute regression coefficient. The optimal wavelength is determined by the minimum cross-validation RMSE value. For hyperspectral image data in the 900-1700nm band, the uninformation variable elimination method UVE is used to screen characteristic wavelengths. UVE uses the information statistics of noise-independent variables to select the characteristic variables of the spectrum itself, which can effectively filter out useful wavelength variables and eliminate non-contributory redundant variables. After adding noise, UVE judges the statistical distribution of the regression coefficient of the target matrix based on the independent variable matrix composed of spectral variables and noise. The statistical distribution of the regression coefficient is expressed as the ratio of the mean to the standard deviation. By determining the upper and lower limits and proposing variables within the corresponding range, the characteristic variables are finally determined.

[0096] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0097] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat grains to be tested.

[0098] Embodiment 3

[0099] This embodiment provides a protein content detection method based on feature-level band fusion, which is used to detect the protein content of wheat flour, including:

[0100] Step 1, obtaining hyperspectral image data of wheat flour in the 400-1000nm and 900-1700nm bands respectively;

[0101] Step 2: The hyperspectral image data in the 400-1000nm and 900-1700nm bands were preprocessed using 1st+SNV;

[0102] Combined with the subsequent characteristic wavelength screening algorithm, in this step, the first-order derivative 1st is used to enhance the spectral characteristics of the hyperspectral image data in the 400-1000nm and 900-1700nm bands, but the standard normal transformation SNV is subsequently used to eliminate the influence of factors such as different wheat flour particle size and uneven distribution on the spectrum. Specifically, the data is standardized by subtracting the mean of each spectrum and dividing it by the standard deviation.

[0103] Step 3, using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using UVE+CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0104] For the hyperspectral image data in the 900-1700nm band, the uninformation variable elimination method UVE is first used to screen the characteristic wavelengths, and the screened characteristic wavelengths are again screened using the competitive adaptive reweighting algorithm CARS. The results of the secondary characteristic wavelength screening are used as the characteristic wavelengths screened in the 900-1700nm band.

[0105] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0106] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat flour to be tested.

[0107] Embodiment 4

[0108] This embodiment provides a wet gluten content detection method based on feature-level band fusion, which is used to detect the wet gluten content of wheat grains, including:

[0109] Step 1, obtaining hyperspectral image data of wheat grains in the 400-1000nm and 900-1700nm bands respectively;

[0110] Step 2: Preprocess the hyperspectral image data in the 400-1000 nm band using 1st+SNV, and preprocess the hyperspectral image data in the 900-1700 nm band using 1st+MSC;

[0111] Considering the subsequent characteristic wavelength screening algorithm, for the prediction of the wet gluten content of wheat grains, this step first uses the first-order derivative 1st to enhance the spectral characteristics of the hyperspectral image data in the 400-1000nm band, and then uses the standard normal transformation SNV to eliminate the influence of factors such as different particle sizes and uneven distribution of wheat grains on the spectrum; for the hyperspectral image data in the 900-1700nm band, the first-order derivative 1st is first used to enhance its spectral characteristics, and after the spectral characteristics are enhanced, the multivariate scattering correction MSC is used to eliminate the spectral errors caused by the surface scattered light of solid particles (wheat grains) and the change of optical path in the spectral data;

[0112] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm and 900-1700nm bands;

[0113] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0114] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat grains to be tested.

[0115] Embodiment 5

[0116] This embodiment provides a wet gluten content detection method based on feature-level band fusion, which is used to detect the wet gluten content of wheat flour, including:

[0117] Step 1, obtaining hyperspectral image data of wheat flour in the 400-1000nm and 900-1700nm bands respectively;

[0118] Step 2: The hyperspectral image data in the 400-1000nm and 900-1700nm bands were preprocessed using 1st+SNV;

[0119] Step 3, using the competitive adaptive reweighting algorithm CARS to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and using the uninformation variable elimination method UVE to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band;

[0120] Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model;

[0121] Step 5, using the PLSR model constructed in step 4 to predict the protein content of the wheat flour to be tested.

[0122] To further illustrate the advantages of the protein and wet gluten content detection method based on feature-level band fusion provided by the present invention, the processing results of each step are given below. The performance of the constructed PLSR model is measured by the determination coefficient (R 2 C and R 2 P ), root mean square error (RMSEC and RMSEP), and mean absolute error (MAEC and MAEP) were used for evaluation.

[0123] Generally speaking, the R 2 C and R 2 P The higher the RMSEC and RMSEP as well as the MAEC and MAEP are, the better the robustness and predictive ability of the model are.

[0124] Specific:

[0125] For the prediction of wheat grain protein and wet gluten content, the modeling results of the PLSR model constructed based on the hyperspectral image data in the full band of 400-1000nm are shown in Table 1 below:

[0126] Table 1 Modeling results of wheat grains based on different pretreatments at 400-1000 nm

[0127]

[0128]

[0129] As shown in Table 1 above, by comparing the scheme proposed in this application with other existing preprocessing algorithms, it can be seen that for the prediction of wheat grain protein content, based on the hyperspectral image data in the 400-1000nm band, 1st+MSC is used for preprocessing, which can maximize the prediction accuracy of the model (as recorded in Example 2 of this application). For the prediction of wheat grain wet gluten content, based on the hyperspectral image data in the 400-1000nm band, 1st+SNV is used for preprocessing, which can maximize the prediction accuracy of the model (as recorded in Example 4 of this application).

[0130] For the prediction of wheat flour protein and wet gluten content, the modeling results of the PLSR model constructed based on the hyperspectral image data in the full band of 400-1000nm are shown in Table 2 below:

[0131] Table 2 Modeling results of wheat flour based on different pretreatments at 400-1000nm

[0132]

[0133] As shown in Table 2 above, by comparing the scheme proposed in this application with other existing preprocessing algorithms, it can be seen that for the prediction of wheat flour protein and wet gluten content, based on the hyperspectral image data in the 400-1000nm band, the use of 1st+SNV for preprocessing can maximize the prediction accuracy of the model (as recorded in Examples 3 and 5 of this application).

[0134] For the prediction of wheat grain protein and wet gluten content, the modeling results of the PLSR model constructed based on the hyperspectral image data in the full band of 900-1700nm are shown in Table 3 below:

[0135] Table 3 Modeling results of wheat grains based on different pretreatments at 900-1700 nm

[0136]

[0137]

[0138] As shown in Table 3 above, by comparing the scheme proposed in this application with other existing preprocessing algorithms, it can be seen that for the prediction of wheat grain protein content, based on the hyperspectral image data in the 900-1700nm band, 1st+SNV is used for preprocessing, which can maximize the prediction accuracy of the model (as recorded in Example 2 of this application). For the prediction of wheat grain wet gluten content, based on the hyperspectral image data in the 900-1700nm band, 1st+MSC is used for preprocessing, which can maximize the prediction accuracy of the model (as recorded in Example 4 of this application).

[0139] For the prediction of wheat flour protein and wet gluten content, the modeling results of the PLSR model constructed based on the hyperspectral image data in the full band of 900-1700nm are shown in Table 4 below:

[0140] Table 4 Modeling results of wheat flour based on different pretreatments at 900-1700 nm

[0141]

[0142] As shown in Table 4 above, by comparing the scheme proposed in this application with other existing preprocessing algorithms, it can be seen that for the prediction of wheat flour protein and wet gluten content, based on the hyperspectral image data in the 900-1700nm band, the use of 1st+SNV for preprocessing can maximize the prediction accuracy of the model (as recorded in Examples 3 and 5 of this application).

[0143] The effects of the screening algorithms of the present invention on the final prediction accuracy for different wheat morphologies, spectral data of different bands, and characteristic wavelengths determined by different prediction objects are further compared:

[0144] Based on the data preprocessed by the optimal preprocessing method adopted in each embodiment, the scheme proposed in this application is compared with other existing feature screening algorithms. The results of the PLSR model established by the characteristic wavelengths screened in the 400-1000nm band are shown in Table 5 below, and the results of the PLSR model established by the characteristic wavelengths screened in the 900-1700nm band are shown in Table 6 below

[0145] Table 5 Results of PLSR model established based on characteristic wavelengths screened in the 400-1000 nm band

[0146]

[0147] As shown in Table 5 above, in the 400-1000nm band, for the prediction of wheat grain protein and wet gluten content, the competitive adaptive reweighting algorithm CARS is used to screen characteristic wavelengths, which can maximize the prediction accuracy of the model (as recorded in Examples 2 and 4 of the present application); and for the prediction of wheat flour protein and wet gluten content, the uninformation variable elimination method UVE and the competitive adaptive reweighting algorithm CARS are used to screen characteristic waves, which can maximize the prediction accuracy of the model (as recorded in Examples 3 and 5 of the present application).

[0148] Table 6 Results of PLSR model established based on characteristic wavelengths at 900-1700 nm

[0149]

[0150]

[0151] As shown in Table 6 above, in the 900-1700nm band, for the prediction of the protein content of wheat grains, the characteristic wavelength screening using the uninformation variable elimination method UVE can maximize the prediction accuracy of the model (as recorded in Example 2 of the present application); for the prediction of the wet gluten content of wheat grains, the competitive adaptive reweighting algorithm CARS is used to screen the characteristic wavelengths, which can maximize the prediction accuracy of the model (as recorded in Example 4 of the present application); for the prediction of the protein content of wheat flour, it is necessary to first use the uninformation variable elimination method UVE to screen the characteristic wavelengths, and then use the competitive adaptive reweighting algorithm CARS to perform secondary characteristic wave screening to maximize the prediction accuracy of the model (as recorded in Example 3 of the present application); for the prediction of the wet gluten content of wheat flour, the characteristic wave screening using the uninformation variable elimination method UVE can maximize the prediction accuracy of the model (as recorded in Example 5 of the present application).

[0152] In order to further highlight the prediction accuracy of the method based on feature-level band fusion for wheat protein and wet gluten in the present application, the present application scheme is compared with the effect of data-level fusion alone, as shown in Tables 7 and 8 below. Table 7 is the modeling result of data-level fusion alone, that is, simply adding up the 400-1000nm and 900-1700nm band ranges for modeling, and Table 8 is the modeling result of feature-level band fusion processing using the present application scheme, that is, linking the characteristic wavelengths in the 400-1000nm and 900-1700nm bands after pretreatment and feature screening given in the above embodiment for modeling. This fusion method can adjust the number of features from different spectral data, especially when there is a large difference in the number of features between the two spectra, this fusion can avoid the situation where one spectrum dominates. At the same time, compared with the full-band data matrix, the feature fusion data matrix has less information overlap, lower covariance and higher correlation.

[0153] Table 7 Data-level fusion modeling effect

[0154]

[0155] Table 8. Modeling effects of feature-level band fusion

[0156]

[0157] In summary, the best prediction model for protein content using wheat grain spectral data is the 400-1000nm spectral data after 1st+MSC pretreatment and the 900-1700nm spectral data after 1st+SNV pretreatment. The feature-level fusion is performed after 32 and 65 characteristic wavelengths are screened out by CARS and UVE algorithms respectively. The prediction effect of the established PLSR model is R 2 P =0.946, RMSEP=0.363, MAEP=0.300.

[0158] The model with the best prediction effect on protein content using wheat flour spectral data is the PLSR model established after 1st+SNV preprocessing of 900-1700nm spectral data and 40 characteristic wavelengths screened by UVE+CARS algorithm. 2 P =0.970, RMSEP=0.251, MAEP=0.205.

[0159] The prediction of the wet gluten content using wheat grain spectral data showed that the best prediction model was the 400-1000nm spectral data after 1st+SNV pretreatment and the 900-1000nm spectral data after 1st+MSC pretreatment. The CARS algorithm was used to select 22 and 20 characteristic wavelengths respectively, and the prediction effect of the established PLSR model was R 2 P =0.899, RMSEP=1.967, MAEP=1.515.

[0160] Using wheat flour spectral data for prediction, the best model for predicting wet gluten content is the 400-1000nm spectral data after 1st+SNV preprocessing and the 900-1000nm spectral data after 1st+SNV preprocessing. The feature-level fusion is performed after 20 and 74 characteristic wavelengths are screened out by CARS and UVE algorithms respectively. The prediction effect of the established PLSR model is R 2 P =0.926, RMSEP=1.805, MAEP=1.421.

[0161] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0162] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting protein or wet gluten content based on characteristic level band fusion, characterized in that: The method comprises: Step 1, obtaining hyperspectral image data of the sample in the 400-1000nm and 900-1700nm bands respectively; Step 2: Based on the sample morphology, the detection object and the characteristic wavelength screening algorithm used later, the hyperspectral image data in different bands are preprocessed using corresponding preprocessing methods; Step 3: Based on the sample morphology and the detection object, the corresponding characteristic wavelength screening algorithm is used to perform feature screening for the hyperspectral image data in different bands; Step 4, linking the selected characteristic wavelengths in the 400-1000nm band and the characteristic wavelengths in the 900-1700nm band to construct a PLSR model; Step 5, using the PLSR model constructed in step 4 to predict the protein or wet gluten content in wheat grains or wheat flour.

2. The method according to claim 1, characterized in that When the sample morphology is wheat grain and the detection object is protein, in step 2, 1st+MSC is used to preprocess the hyperspectral image data in the 400-1000nm band, and 1st+SNV is used to preprocess the hyperspectral image data in the 900-1700nm band.

3. The method according to claim 2, characterized in that In the step 3, the competitive adaptive reweighting algorithm CARS is used to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and the uninformation variable elimination method UVE is used to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band.

4. The method according to claim 1, characterized in that: When the sample form is wheat flour and the detection object is protein, in step 2, the hyperspectral image data in the 400-1000nm and 900-1700nm bands are preprocessed using 1st+SNV.

5. The method according to claim 4, characterized in that In the step 3, the uninformation variable elimination method UVE is used to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and for the preprocessed hyperspectral image data in the 900-1700nm band, the uninformation variable elimination method UVE is first used to screen the characteristic wavelengths, and the screened characteristic wavelengths are again screened for the second time using the competitive adaptive reweighting algorithm CARS, and the results of the secondary characteristic wavelength screening are used as the characteristic wavelengths screened in the 900-1700nm band.

6. The method according to claim 1, characterized in that When the sample form is wheat grains and the detection object is wet gluten, in step 2, 1st+SNV is used to preprocess the hyperspectral image data in the 400-1000nm band, and 1st+MSC is used to preprocess the hyperspectral image data in the 900-1700nm band.

7. The method according to claim 6, characterized in that In the step 3, the competitive adaptive reweighting algorithm CARS is used to perform characteristic wavelength screening on the pre-processed hyperspectral image data in the 400-1000nm and 900-1700nm bands.

8. The method according to claim 1, characterized in that When the sample form is wheat flour and the detection object is wet gluten, in step 2, the hyperspectral image data in the 400-1000nm and 900-1700nm bands are preprocessed using 1st+SNV.

9. The method according to claim 8, characterized in that In the step 3, the competitive adaptive reweighting algorithm CARS is used to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 400-1000nm band, and the uninformation variable elimination method UVE is used to screen the characteristic wavelengths of the preprocessed hyperspectral image data in the 900-1700nm band.

10. The method according to claim 1, characterized in that When collecting hyperspectral image data in the 400-1000nm and 900-1700nm bands, after taking the hyperspectral images, the collected original hyperspectral images of wheat and wheat flour are converted into corrected images through black and white correction.

Images