CN117092041A

CN117092041A - Rapid detection method for muscle quality of living carp based on hyperspectral imaging technology

Info

Publication number: CN117092041A
Application number: CN202311061490.1A
Authority: CN
Inventors: 曹逸铭; 李炯棠; 张研; 王�琦; 赵然; 侯明喜; 于双婷; 陈颖杰; 王开阔
Original assignee: Chinese Academy Of Fishery Science
Current assignee: Chinese Academy Of Fishery Science
Priority date: 2023-08-22
Filing date: 2023-08-22
Publication date: 2023-11-21

Abstract

The application discloses a living carp muscle quality rapid detection method based on hyperspectral imaging technology, which comprises the steps of obtaining hyperspectral data and texture indexes of a carp sample; extracting spectral information of a region of interest, calculating to generate a spectral information mean value of the region of interest, carrying out variable screening extraction according to the spectral information mean value to generate a characteristic variable, and generating a sample set according to the characteristic variable and a texture index; constructing a meat quality index prediction model, and optimizing and screening the meat quality index prediction model according to a sample set to generate an optimal prediction model; predicting pixel spectrum values of the actually measured carp sample through an optimal prediction model to generate texture results of different muscle areas of the actually measured carp sample; through the technical scheme, the method and the device can be used for efficiently, quickly, conveniently and quickly detecting the quality of the carp muscles, and are high in reliability.

Description

Rapid detection method for muscle quality of living carp based on hyperspectral imaging technology

Technical Field

The application relates to the technical field of carp meat quality detection, in particular to a rapid carp muscle quality detection method based on hyperspectral imaging technology.

Background

The carp is one of important freshwater aquaculture fishes in the world, and has high nutritive value and great economic benefit. And with the rapid development of the aquaculture genetic breeding industry, the consumption of high-quality fish and other aquatic products is continuously increased, and consumers and the whole industry have higher requirements on the quality requirements and the identification of aquatic animals. Texture is one of the key qualities of fish meat, and relates to the overall quality of fish and related products and consumer acceptance. Traditional texture evaluation methods include subjective evaluation and objective evaluation. Subjective evaluation relies on professional training and experience-rich professional reviews by touch or the like. Subjective evaluation is subject to larger subjective influence factors and larger errors by panelists. The objective evaluation is to detect the texture index of the fish meat by using an instrument to carry out strict standardized control through mechanical measurement. However, the traditional objective evaluation method still has the bottleneck problems of long detection period, high cost, lethal sampling and the like. Therefore, development of a rapid, nondestructive and efficient carp texture detection technology is needed, the speed and the accuracy of identifying phenotypic characters are improved, rapid and accurate screening of genetic breeding and excellent characters of genetic improvement is realized, and the breeding efficiency is improved. Therefore, the development and construction of the rapid, efficient and nondestructive living fish quality detection method based on the hyperspectral imaging technology has a very wide application prospect in the field of aquaculture genetic breeding.

The hyperspectral imaging technology is a mature rapid nondestructive testing method and has the advantages of no sample damage, no pretreatment, no sample pollution, short testing time, low cost and the like. The method is mainly characterized by comprising the two aspects of high spectral resolution and spectral imaging, and combining the image and spectral technology, so that the method can obtain the 'space' information and the 'spectrum' information of a sample at the same time. When the hyperspectral imager is used for detecting the sample, light emitted by the light source irradiates the surface of the sample, one part of the light signals are absorbed by the sample, the other part of the light signals enter an optical element (PGP unit) through the objective lens and the incident slit, and the PGP unit separates the received light signals into different wavelengths after interference of the ambient light is filtered. The separated dispersed optical signals are mapped onto a CCD image sensor to form a two-dimensional image matrix. The distribution of physical and morphological characteristics (such as color, size, shape, texture, etc.) and some inherent chemical and molecular information (such as water, fat, protein, etc.) inside the sample is obtained by matrix analysis of the two-dimensional images. Therefore, the spectrum data and image data analysis advantages of the hyperspectral imaging technology are utilized, and the quality index optimal prediction model is associated by combining a machine learning data processing method, so that the rapid nondestructive testing of the muscle quality of the living carp is realized. However, in current studies to identify meat quality based on hyperspectral imaging techniques, spectral information is typically obtained from the surface of the meat slices of the organism, still being a lethal sample. At present, a rapid detection method for detecting meat quality by utilizing spectral information obtained by scanning the surface of a living fish by using a hyperspectral imaging system does not exist.

Disclosure of Invention

In order to solve the problems of long period, high cost, lethal sampling and the like in the prior art, the application provides a carp muscle quality rapid detection method based on a hyperspectral imaging technology and a machine learning method, and the method has the advantages of high efficiency, high speed, convenience, high speed and high reliability for rapid detection of the carp muscle quality.

In order to better realize the technical purposes, the application provides the following technical scheme: a living body carp muscle quality rapid detection method based on hyperspectral imaging technology comprises the following steps:

obtaining hyperspectral data of a carp sample and texture indexes of different muscle parts;

extracting spectral information of an interest region in hyperspectral data, and calculating to generate a spectral information mean value of the interest region, wherein the interest region corresponds to different parts of the carp sample;

performing variable screening extraction according to the spectrum information mean value to generate a characteristic variable, and generating a sample set according to the characteristic variable and the texture index;

constructing a meat quality index prediction model, optimizing and predicting the meat quality index prediction model through a sample set, and screening the meat quality index prediction model according to a prediction result to generate an optimal prediction model;

and obtaining an actually measured carp sample, obtaining pixel spectrum values in a hyperspectral image of the actually measured carp sample according to the characteristic variables, and predicting the pixel spectrum values through an optimal prediction model to generate texture results of different muscle areas of the actually measured carp sample.

Optionally, collecting hyperspectral data of the carp sample by a hyperspectral imager, wherein the translation speed of a sample stage of the hyperspectral imager is 90mm/s; the travel is 180mm; the camera exposure time is set to 150ms; spectral resolution was 2.5nm; the image resolution is 2558×960pixel.

Optionally, texture indexes of different muscle parts of the carp sample are obtained through a texture analyzer, wherein the texture indexes comprise adhesiveness, elasticity, cohesiveness, resilience, hardness, brittleness, viscosity and chewiness.

Optionally, after the hyperspectral data is acquired, the original spectrum of the hyperspectral data is preprocessed through convolution smoothing, standard canonical transformation and multiplicative scattering correction.

Optionally, variable screening and extracting are carried out on the spectrum information mean value by a regression coefficient variable screening method.

Optionally, the meat quality index prediction model comprises a least square regression, an interval partial least square method, a joint interval partial least square method, a reverse interval partial least square method, a least square support vector machine and a reverse propagation artificial neural network prediction model.

Optionally, the characteristic variable and the texture index are processed through a spectrum-physicochemical value symbiotic distance method to generate a sample set, wherein the sample set comprises a correction set and a verification set.

Optionally, calculating a corresponding prediction index according to the prediction result, and screening a meat quality index prediction model according to the prediction index; the prediction index comprises correction set correlation coefficient, prediction set correlation coefficient, correction set root mean square error and prediction set root mean square error.

Optionally, the optimal prediction model includes: partial least squares regression, least squares support vector machine and back propagation artificial neural network.

Optionally, the generating the texture result of the different muscle areas of the measured carp sample further comprises: and visually displaying the texture results of different muscle areas of the actually measured carp sample.

Compared with the prior art, the application has the advantages that:

(1) Compared with the existing detection method of the carp muscle texture indexes, the rapid nondestructive detection of the texture indexes of different muscle parts of the living carp is realized. Based on hyperspectral imaging technology, an attempt is made for the first time to detect the quality of fish meat at different muscle parts by acquiring hyperspectral information on the surface of a living carp. The living carp muscle quality model is constructed by utilizing the hyperspectral imaging technology, lethal sampling is not performed when the target population of the breeding fish is screened, and the living carp is reserved to be used for subsequent breeding, so that the effect of nondestructive detection is achieved.

(2) The application establishes a living body carp meat quality rapid detection method based on hyperspectral imaging construction, combines a hyperspectral imaging technology with a machine learning method, and realizes the visual distribution of texture indexes of different muscle parts of the carp through an optimal prediction model, so that the texture index prediction is more rapid, efficient and visual. Meanwhile, according to the application, the carp muscle texture indexes are rapidly detected through screening the characteristic wave bands, so that the detection workload is reduced, the data processing time is shortened, the detection efficiency is improved, and theoretical basis and technical support are provided for achieving rapid nondestructive detection of the carp muscle texture indexes.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a method for rapidly detecting the muscle quality of a living carp based on hyperspectral imaging technology;

FIG. 2 is a schematic diagram of a hyperspectral imager used in the method for rapidly detecting the muscle quality of living carp based on hyperspectral imaging technology;

FIG. 3 is a schematic diagram of a three-dimensional data block acquired based on a hyperspectral imager in accordance with the present application;

FIG. 4 is a graph showing the index value of the muscle texture of carp according to the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

The technical scheme of the application is as follows:

1) Preparing a plurality of living carp samples and numbering;

2) Carrying out hyperspectral image acquisition on living carp by using a hyperspectral imaging system to obtain a hyperspectral image of a living carp sample;

3) Extracting a spectrum value mean value of an interested region in the living body carp hyperspectral image acquired in the step 2) by using ENVI 5.3 software to serve as original spectrum data;

4) Preprocessing the spectrum value of the original spectrum in the hyperspectral image extracted in the step 3) by adopting different preprocessing methods, and removing redundant information and noise to obtain preprocessed spectrum data;

5) Measuring the texture indexes of different muscle parts of which the hyperspectral image data of the living carp are acquired in the step 2) by using a texture analyzer;

6) The regression coefficient variable screening method is used for recombining the extracted characteristic variable and the corresponding texture index into a new sample for the characteristic variable in the spectrum data after pretreatment in the step 4);

7) Dividing the new sample extracted by the characteristic variable in the step 6) into a correction set and a prediction set;

8) And 7) establishing a living carp muscle texture index prediction model by using the sample set divided in the step 7).

9) And 3) utilizing the prediction model established in the step 8) to realize the prediction and the distribution visualization of the texture indexes of different muscle parts of the living carp.

In the step 1), one side of a living carp sample is descaled, hyperspectral detection is performed, after hyperspectral detection is completed, the muscle tissue of the dorsal, pectoral, abdominal and gluteal muscles of the carp is cut by a blade, and a tissue block with the size of about 2cm multiplied by 1.5cm is cut for detecting the actual measurement value of the texture index by a texture analyzer.

In the step 2), hyperspectral information of the carp sample is collected by using a hyperspectral imager, and the specific method for collecting hyperspectral images of the living carp sample comprises the following steps: and (3) placing a tail of living carp to be tested on the mobile platform to align with the camera. And placing a tail of living carp to be tested on the mobile platform to align with the camera. When the platform moves, the spectrum information of one line of the spatial position of the living carp in the whole spectrum area is obtained through the hyperspectral imager, then the living carp is driven to move through the platform, and the spectrum information of the living carp in other positions is obtained until the complete spectrum information of the sample is obtained. The translation speed of the sample stage in the spectrum imaging system is 90mm/s; the travel is 180mm; the camera exposure time is set to 150ms; the spectral resolution was 2.5nm. The image resolution is 2558×960pixel. Since the spectral scan range is 300 effective wavelengths in the 400-1000nm band, the size of the resulting three-dimensional data block is 2558×960×300.

In step 3) above, the spectral values of the region of interest of the sample are extracted using ENVI 5.3 software. The size of the ROI is selected according to different muscle parts of the carp and is 200 pixels multiplied by 200 pixels. For each wavelength, the average spectrum of the ROI is calculated by averaging the spectra of all pixels. Averaging reflectance values of all pixels for each wavelength variable to obtain an average value representing each sample; the ROI and different parts of each sample correspond to a new spectrum value as the original spectrum data of the sample.

In the step 4), in order to eliminate the influence of the factors such as equipment, samples and environment in the process of acquiring the original spectrum data, the original spectrum is analyzed by using Matlab 2021a software through convolution Smoothing (SG), standard regular transformation (Standard normal variable, SNV) and multiplicative scattering correction (Multiplicative scattering correction, MSC) preprocessing methods.

In the step 5), the texture index of the carp muscle is measured, and the texture value of the muscle sample is obtained by a texture analyzer, and the specific detection method is to use the texture analyzer to measure eight texture indexes of the muscle, including adhesiveness, elasticity, cohesiveness, resilience, hardness, brittleness, viscosity and masticability, by matching with a TA/36 cylindrical probe. The measurement speed was 2 mm/sec, the trigger force was 5N, and the compression set of the sample was 60%. Thereby obtaining the actually measured values of the texture indexes of different muscle parts.

In the step 6), extracting characteristic variables in the spectral data of the preprocessed region of interest by using a variable screening method; the method selects a regression coefficient (Regression coefficient) RC variable screening method, establishes a quantitative model of a texture index by taking the screened spectrum value as a new variable, and compares the prediction effect of the model.

In the above step 7), the sample set is divided using a spectrum-physicochemical value co-occurrence distance method (Sample setpartitioning based on joint x-y distances, SPXY). The SPXY method is a new method based on the Kenneard-Stone method, and can be used for considering X in a variable matrix and an index Y to be measured.

In the step 8), a hyperspectral imaging system is used for collecting spectral image data results of living carp samples, carp muscle texture data measured by a texture analyzer are used as actual measurement values, the spectral information obtained through pretreatment and variable screening is used as variables, PLSR, iPLS, siPLS, biPLS, LS-SVM and BP-ANN prediction models are built, fitting optimization and judgment of the prediction results are carried out on the prediction models according to a sample set, and an optimal prediction model is screened according to the prediction results.

In the step 9), the hyperspectral result of the carp is predicted by using PLSR, LS-SVM and BP-ANN methods as optimal prediction models; and screening the hyperspectral results of the carps by using an RC method, taking the screened spectral values as variables, taking different texture indexes as dependent variables, and establishing a quantitative model of the texture index values. After the optimal quantitative model is determined, extracting spectral values corresponding to all pixel points in the hyperspectral image of the living carp sample, and substituting the spectral values into the established optimal quantitative model for prediction. And finally, obtaining distribution diagrams of different muscle texture indexes of the living carp on a plane according to the space coordinate information of each pixel and the corresponding index content of each pixel.

The application will now be further illustrated by means of a specific implementation in conjunction with the accompanying drawings without limiting the application.

Example 1:

1. sample preparation

As shown in fig. 1, several tail living carps are prepared, weighing about 600g, and one side of the scale is removed for subsequent hyperspectral imaging analysis; according to different muscle areas (dorsal, pectoral, abdominal and gluteal), the carp after scanning by the hyperspectral imager is cut into small muscle blocks with the length of about 2cm, the width of about 2cm and the height of 1.5cm by using a sterilized scalpel, and the muscle blocks are used for subsequent texture value measurement, and the carp sample is sterilized after treatment.

2. Acquiring hyperspectral data using hyperspectral imaging systems

The hyperspectral imaging camera (FigSpec hyperspectral camera FS-13, color spectrum science and technology (Zhejiang) limited company, china) in the hyperspectral imaging system used by the application mainly comprises the following components: hyperspectral imagerFS-13, figSpec), spectral range 400-1000nm; precision displacement platform device, six of 50W fiber halogen lamps, reduce ambient light and influence camera bellows (50 cm x 40cm x 30 cm), configuration image acquisition card's LENOVO computer. The spectrometer is used as the most core structure of the whole hyperspectral imaging system, and can acquire the spectral information of each point on the sample surface in the wave band of 400-1000nm during testing. The main component schematic diagram of the system hardware part is shown in fig. 2. Before the spectral image of the sample is acquired, the hyperspectral image acquisition system needs to be preheated for half an hour to be adjusted to the optimal working state. After the instrument is preheated, placing the carp with one side of scales removed on a displacement platform, and setting system parameters: the translation speed of the sample table is 90mm/s; the travel is 180mm; the camera exposure time is set to 150ms; the image resolution is 2558×960pixel. Because the spectral scan range is 300 effective wavelengths in the 400-1000nm band, the size of the resulting three-dimensional data block is 2558×960×300, as shown in fig. 3, where λ _n For the nth effective wavelength, x, y are the row and column directions of the image, respectively.

Hyperspectral imaging systems are prone to adverse factors such as uneven illumination and dark current in the process of spectrum image acquisition, so that acquired spectrum image information is subjected to large noise. The interference of noise on spectrum information can be effectively eliminated by adopting a black-and-white correction method. The black and white correction formula is:

wherein the corrected reflectance hyperspectral image I is represented in units of relative reflectance (%); d represents an original hyperspectral image; i ₀ For dark images (0% reflectivity), W is a white reference image (100% reflectivity).

3. Spectral data preprocessing

In order to eliminate the influence of equipment, samples, environment and other factors on the hyperspectral imaging system in the process of acquiring the original spectrum data, the original spectrum is analyzed by using Matlab 2021a software through convolution Smoothing (SG), standard canonical transformation (Standard normal variable, SNV) and multiplicative scattering correction (Multiplicative scattering correction, MSC) preprocessing methods. According to the application, compared with standard regular transformation and multiplicative scattering correction, the method has the advantages that the determination coefficient of the data set is increased after the collected original spectrum is subjected to convolution smoothing pretreatment, and the spectrum accuracy is improved.

Spectral values of the region of interest of the sample were extracted using ENVI 5.3 software. The size of the region of interest (ROI) is selected according to the different muscle parts of the carp to be 200 pixels by 200 pixels. For each wavelength, the average spectrum of the ROI is calculated by averaging the spectra of all pixels. The spectral reflectance values for all pixels of each wavelength variation are averaged to obtain an average value representing a different ROI for each sample.

4. Obtaining hyperspectral characteristic values of samples

Different nutritional ingredients have different spectral characteristics on the spectrogram. Wherein, the size of the wave peak in the spectrum data also reflects the content of the corresponding nutrient components. Meanwhile, the change of the nutritional ingredients in the carp muscle can also be reflected in the aspects of physical properties and the like, such as the texture indexes of the carp muscle, such as cohesiveness, chewiness, brittleness, hardness and the like. Therefore, a link can be established between the spectral image of the living carp and the texture index of the muscle thereof. According to the application, through analyzing the spectral image and spectral information of the skin of the living carp, the connection between the texture and spectrum of the muscle of the living carp is established, and a theoretical basis is provided for the establishment of a muscle texture model of the living carp.

The carp at different muscle parts images the same spectrum curve change trend in a spectrum band of 400-1000nm, but the spectrum intensities of different parts have differences, and the differences are related to the types and the characteristics of organic chemical components of the carp at different parts of the carp muscle. Mainly related to the frequency multiplication and the vibration of the combined frequency of chemical bonds of substances such as water, proteins, fat and the like in the muscle, such as O-H, C-H, C-O, N-H, S-H and the like.

Extracting characteristic variables in the pretreated spectrum data by using a variable screening method; the method selects a regression coefficient (Regression coefficient) RC variable screening method, establishes a quantitative model of a texture index by taking the screened spectrum value as a new variable, and compares the prediction effect of the model.

5. Obtaining texture index data

The texture value of the muscle sample is obtained through a texture instrument, and the specific detection method comprises the following steps: we measured eight texture indexes of the muscle, including cohesiveness, elasticity, cohesiveness, recovery, hardness, crispness, tackiness, and chewiness, using a texture analyzer (TA. XTC-18, baosheng, shanghai, china) and a TA/36 cylindrical probe. The measurement speed was 2 mm/sec, the trigger force was 5N, and the compression set of the sample was 60%. Thereby obtaining the actually measured values of the texture indexes of different muscle parts.

6. Constructing a predictive model

In order to make the obtained spectral characteristics correspond to the measured quality indexes one by one, quality indexes of different muscle parts of the carp are rapidly detected by using Partial Least Squares Regression (PLSR), interval Partial Least Squares (iPLS), joint interval partial least squares (SiPLS), reverse interval partial least squares (BiPLS), least square support vector machine (LS-SVM) and reverse propagation artificial neural network (BP-ANN) quantitative models respectively.

PLSR projects the predicted variables (data matrix Y) and the observable variables (data matrix X) into the new feature space to build a linear regression model. PLSR breaks down the independent variables X and Y into several X fractions (T), and a PLSR model is constructed with the following equation:

Y＝XB+E＝XW*C+E＝TC+E

W*＝W(P′W) ^-1

where B is the PLSR coefficient, E is the Y residual matrix, T is the fractional matrix of X, W is the PLS weight, P' and C are the loadings of X and Y, respectively, and W is the regression coefficient matrix. The set of data projected from the spectral data is referred to as the orthogonality factor of the "latent variables". The optimal number of orthogonality factors depends on the prediction error, typically by using the lowest value of the sum of squares of Prediction Residual Errors (PRESS). Cross-validation to minimize errors between predicted and observed response values is detailed below:

wherein Minimized is Minimized, PRESS _j (h) To predict the sum of squares of residual errors for sample h, n is the number of samples, i is the sample number, Y _ij In order to be able to predict the value,for the observed value, j is the tag number and p is the number of tags.

in the iPLS algorithm, the full spectral region is divided into smaller equidistant subintervals, and a PLS regression model is generated on a per subinterval basis. The best interval and principal component scores are selected according to the principle of lowest RMSEC value.

The sips algorithm is a modification of the original iPLS, the whole spectral region is divided equally into several sub-intervals, and then a PLS model is built by all possible combinations of two, three and four sub-intervals, each combination being considered as a subset, and a PLSR model is generated for each combination. The combination of lowest RMSEC values was selected.

Bipsl can improve the predictive performance of the model by eliminating high noise spectral regions. The whole spectrum area is divided into N subintervals with equal width, PLS regression is carried out, each interval is omitted in sequence, the worst RMSEC value is obtained in modeling, and the subintervals are deleted until the lowest RMSEC value is obtained.

The LS-SVM first projects the input spectral information into a high-dimensional feature space through a nonlinear mapping and then creates a linear model in the same feature space. A non-linear function that reduces the complexity of the training process is used with Radial Basis Functions (RBFs). Regularization parameters gamma (gamma) and kernel parameters (sigma) that can reduce complexity ² ) Representing the width of the RBF kernel is an important parameter in the LS-SVM model. And a good fitting effect is achieved by adjusting the two parameters.

The BP-ANN uses an error-reversal propagation algorithm to train a multi-layer positive feedback neural network. The multi-layer sensor consists of an input layer, a hidden layer and an output layer. Parameters (e.g., learning rate, decay rate, momentum) are correctly adjusted to reduce the chance of under-fitting or over-fitting. The update amount of the weight is called a learning rate, the momentum is responsible for accelerating the learning speed, and the attenuation rate is responsible for preventing the weight from growing too much.

The inventors used PLSR, iPLS, siPLS, biPLS, LS-SVM and BP-ANN models, respectively, to establish a link between spectral information of the epidermis of a living carp and texture indexes of different muscle regions. The predictive power of the model uses correction set correlation coefficients (correlation coefficient ofcalibration, r _C ) Prediction set correlation coefficients (correlation coefficient ofprediction, r _P ) The correction set root mean square error (RootMean Square Error of Calibration, RMSEC), the predicted root mean square error (Root Mean Square Error of Prediction, RMSEP). r is (r) _C And r _P The closer to 1 the values of (c) and the smaller RMSEC and RMSEP represent the better predictive performance of the model.

In order to improve the prediction accuracy of the model and the prediction efficiency of the model, the application adopts an RC method to screen the spectrum variable. The predictive model results for the different muscle region texture indexes based on the optimal wavelength for RC screening are shown in tables 1-4. Dorsum muscle and pectoral muscle60-114 characteristic wavelengths are selected from the abdominal muscles and gluteus muscles, the selected characteristic wavelengths and the samples are recombined into a new variable matrix, and a texture index prediction model of different muscle areas of the carp is established. As shown in Table 1, BP-ANN and PLSR predictive models established by RC-screened variables in the dorsal muscle were superior in predicting adhesiveness, chewiness, cohesiveness, hardness and tackiness, and r _P For 0.9191 to 0.9847, rmsep is 0.1070 to 0.3165. As shown in Table 2, BP-ANN, LS-SVM and PLSR predictive models established by RC-screened variables were better in predicting adhesiveness, chewiness, stickiness, cohesiveness and hardness in pectoral muscle, r _P 0.9033 to 0.9574, rmsep 0.2285 to 0.3930. As shown in Table 3, BP-ANN and PLSR predictive models established by RC-screened variables in abdominal muscles were superior in the predictive effect on masticatory, gummy, hardness and tackiness, r _P 0.9070 to 0.9776, rmsep 0.1649 to 0.3601. As shown in Table 4, BP-ANN predictive model established by RC-screened variables in gluteus muscle gave better results in masticatory, gummy and sticky predictions, r _P 0.9304 to 0.9768 and rmsep 0.1804 to 0.2861. Therefore, PLSR, LS-SVM and BP-ANN models are selected to construct a prediction model of the texture indexes of different muscle areas of the living carp. Table 1 shows the result of predicting the index of the dorsum muscle texture of the carp based on the reflectivity value of the optimal wavelength screening; table 2 shows the result of predicting the texture index of the pectoral muscle of the carp based on the reflectivity value of the optimal wavelength screening; table 3 shows the result of predicting the Cyprinus Carpio abdominal muscle texture index based on the reflectance value of the optimal wavelength screening; table 4 shows the results of predicting the gluteus texture index of the carp based on the reflectance values of the optimal wavelength screen.

TABLE 1

Index (I)	Model	r _C	RMSEC	r _P	RMSEP
						Tackiness of the adhesive	BP-ANN	0.9912	0.1361	0.9847	0.1070
Masticatory properties	BP-ANN	0.9755	0.2450	0.9469	0.2164
						Cohesion of	PLSR	0.9714	0.2535	0.9367	0.2836
Hardness of	PLSR	0.9586	0.2972	0.9298	0.3165
						Viscosity of the adhesive	BP-ANN	0.9432	0.3653	0.9191	0.3004
Brittleness of the product	PLSR	0.9112	0.4423	0.8804	0.3848
						Elasticity of	BP-ANN	0.8380	0.5171	0.7194	0.7554
Recovery of	BP-ANN	0.8346	0.5676	0.6493	0.7074

TABLE 2

Index (I)	Model	r _C	RMSEC	r _P	RMSEP
						Tackiness of the adhesive	BP-ANN	0.9883	0.1605	0.9574	0.2439
Masticatory properties	BP-ANN	0.9862	0.1771	0.9552	0.2285
						Viscosity of the adhesive	LS-SVM	0.9424	0.3643	0.9370	0.2499
Cohesion of	PLSR	0.9556	0.3081	0.9056	0.3765
						Hardness of	BP-ANN	0.9379	0.3561	0.9033	0.3930
Brittleness of the product	PLSR	0.8835	0.5213	0.8846	0.3753
						Elasticity of	LS-SVM	0.8866	0.4490	0.7376	0.6971
Recovery of	LS-SVM	0.7095	0.6983	0.5964	0.7960

TABLE 3 Table 3

Index (I)	Model	r _C	RMSEC	r _P	RMSEP
						Masticatory properties	BP-ANN	0.9858	0.1779	0.9776	0.1649
Tackiness of the adhesive	PLSR	0.9740	0.2418	0.9517	0.2360
						Hardness of	PLSR	0.9592	0.2860	0.9392	0.3158
Cohesion of	PLSR	0.9631	0.2859	0.9070	0.3601
						Viscosity of the adhesive	LS-SVM	0.9564	0.3245	0.8623	0.3358
Brittleness of the product	BP-ANN	0.9309	0.4130	0.8617	0.3860
						Elasticity of	PLSR	0.8954	0.4320	0.8322	0.5928
Recovery of	LS-SVM	0.9367	0.3525	0.5609	0.7414

TABLE 4 Table 4

Index (I)	Model	r _C	RMSEC	r _P	RMSEP
						Masticatory properties	BP-ANN	0.9768	0.2174	0.9768	0.1804
Tackiness of the adhesive	BP-ANN	0.9614	0.2804	0.9339	0.2856
						Viscosity of the adhesive	BP-ANN	0.9421	0.3614	0.9304	0.2861
Cohesion of	LS-SVM	0.9826	0.1891	0.8726	0.3938
						Hardness of	PLSR	0.9092	0.4186	0.8486	0.4933
Brittleness of the product	PLSR	0.8883	0.4723	0.7613	0.4921
						Elasticity of	PLSR	0.7668	0.6437	0.6396	0.7508
Recovery of	PLSR	0.7066	0.7431	0.6316	0.7230

7. Texture prediction

According to the sample preparation, sample spectrum information acquisition, data preprocessing and hyperspectral characteristic value acquisition, the hyperspectral characteristic values of different muscle areas of the living carp to be detected are substituted into the established PLSR, LS-SVM and BP-ANN prediction models, and the texture results of the different muscle areas of the living carp to be detected are obtained.

The PLSR, LS-SVM and BP-ANN prediction models constructed in four different muscle areas of the living carp have good prediction effects, the correlation coefficient of the prediction set is higher than 0.9, and the method is suitable for rapid nondestructive detection of the muscle texture indexes of the living carp and has practical application significance for the texture prediction of the living carp.

8. Distribution visualization of texture index

And extracting spectral values corresponding to pixel points in the living carp skin hyperspectral image, substituting the selected spectral values into an optimal model, predicting the texture index value on each pixel point in the living carp skin hyperspectral image, and finally reconstructing according to coordinate information of the pixels and the corresponding texture values through hyperspectral imaging technology information fusion to obtain a distribution diagram of the texture index of the living carp on a plane, as shown in fig. 4. In the texture index value distribution map, the color is from dark to light. The darker the color the lower the texture value, and the lighter the color the higher the texture value.

The present application is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present application are intended to be included in the scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology is characterized by comprising the following steps of:

extracting spectral information of an interest region in hyperspectral data, and calculating to generate a spectral information mean value of the interest region as an original spectrum, wherein the interest region corresponds to different parts of the carp sample;

performing variable screening extraction according to the original spectrum to generate a characteristic variable, and generating a sample set according to the characteristic variable and a texture index;

2. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

collecting hyperspectral data of a carp sample through a hyperspectral imager, wherein the translation speed of a sample stage of the hyperspectral imager is 90mm/s; the travel is 180mm; the camera exposure time is set to 150ms; spectral resolution was 2.5nm; the image resolution is 2558×960pixel.

3. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

texture indexes of different muscle parts of the carp sample are obtained through a texture instrument, wherein the texture indexes comprise adhesiveness, elasticity, cohesiveness, resilience, hardness, brittleness, viscosity and chewiness.

4. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

after the original spectrum is obtained, the original spectrum is preprocessed through convolution smoothing, standard regular transformation and multiplication scattering correction.

5. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

and carrying out variable screening extraction on the original spectrum by a regression coefficient variable screening method.

6. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

the meat quality index prediction model comprises a partial least square regression, an interval partial least square method, a joint interval partial least square method, a reverse interval partial least square method, a least square support vector machine and a reverse propagation artificial neural network prediction model.

7. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

and processing the characteristic variables and the texture indexes by a spectrum-physicochemical value symbiotic distance method to generate a sample set, wherein the sample set comprises a correction set and a verification set.

8. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

calculating a corresponding prediction index according to the prediction result, and screening a meat quality index prediction model according to the prediction index; the prediction index comprises correction set correlation coefficient, prediction set correlation coefficient, correction set root mean square error and prediction set root mean square error.

9. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

the best prediction model comprises: partial least squares regression, least squares support vector machine and back propagation artificial neural network.

10. The method for rapidly detecting the muscle quality of the living carp based on the hyperspectral imaging technology, which is characterized by comprising the following steps of:

the method further comprises the following steps of: and visually displaying the texture results of different muscle areas of the actually measured carp sample.