Mutton quality index intelligent hyperspectral image detection method

Abstract

The invention discloses a mutton quality index intelligent hyperspectral image detection method, and belongs to the technical field of hyperspectral image detection.The mutton quality index intelligent hyperspectral image detection method comprises the steps that the surface of a mutton sample is scanned by rotating a polarized light source, and a muscle fiber four-direction light field matrix is generated. A three-dimensional muscle fiber topological graph is constructed, Gaussian curved surface fitting and iterative optimization of an energy function are combined, segmentation precision and robustness of geometric modeling are remarkably improved, reliability and multi-scale adaptability of muscle fiber spatial distribution analysis are improved, and by fusing the sarcomedo contraction state, intermuscular fat marks and metabolic activity, the segmentation accuracy is improved. And a tenderness-freshness comprehensive decision is generated through dynamic weight distribution and a nonlinear scoring formula, so that multi-dimensional data complementation is realized, and the comprehensiveness and automatic decision-making capability of quality grading are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of hyperspectral image detection, and particularly relates to a sheep meat quality index intelligent hyperspectral image detection method. BACKGROUND

[0002] Sheep meat quality detection usually relies on artificial sensory evaluation, chemical analysis or physical detection methods, and has problems of low efficiency, strong subjectivity, destructive detection and the like. In recent years, hyperspectral imaging technology has been widely applied in food detection due to its non-contact, rich information, multi-band response and other advantages. However, the existing method still has significant defects in muscle fiber structure modeling, metabolic activity evaluation, multi-index fusion decision and frozen damage analysis, and it is difficult to meet the demand of modern sheep meat industry chain for intelligent and standardized quality detection.

[0003] However, the existing sheep meat quality index hyperspectral image detection still has certain defects. The existing method only uses single feature or fixed threshold to segment muscle fiber bundles and perimysium connective tissue, lacks utilization of muscle fiber anisotropy optical characteristics, causes fuzzy segmentation boundary and high misjudgment rate, the model relies on single spectral feature, a quantitative mapping relationship between optical characteristics and metabolic parameters is not established, ice crystal distribution is observed by naked eye or detected by X-ray, quantitative correlation analysis of dielectric loss and muscle fiber membrane damage is lacked, and multi-angle scanning and frequency domain response data are not combined, resulting in strong subjectivity of damage evaluation. Therefore, the sheep meat quality index intelligent hyperspectral image detection method is proposed. SUMMARY

[0004] The purpose of the present application is to provide a sheep meat quality index intelligent hyperspectral image detection method to solve the problems in the background.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a sheep meat quality index intelligent hyperspectral image detection method, comprising the following steps:

[0006] S1, scanning the surface of a sheep meat sample by rotating a polarized light source to generate a muscle fiber four-directional light field matrix;

[0007] S2, decomposing muscle fiber bundles and perimysium connective tissue of the sheep meat based on the muscle fiber four-directional light field matrix to construct a three-dimensional muscle fiber topology graph;

[0008] S3, locating a mitochondria-rich area of the sheep meat according to the three-dimensional muscle fiber topology graph to establish a mapping model of ATP activity and characteristic waveband quantum efficiency;

[0009] S4, fusing muscle segment contraction state, intermuscular fat lines and metabolic activity of the sheep meat to perform multi-index decision of tenderness and freshness;

[0010] S5, starting terahertz scanning according to the decision result to quantify dielectric loss of muscle fiber membrane caused by ice crystals of the sheep meat.

[0011] S6, integrating the lamb optical characteristics and verification data to generate a blockchain traceable hash value.

[0012] Preferably, S1 is achieved by rotating the polarized light halogen lamp through a three-dimensional electric translation table, carrying a pressure sensor, real-time monitoring the flatness of the lamb sample surface and automatically leveling, driving the polarizer by a motor, and synchronously triggering the hyperspectral camera to collect data.

[0013] Specifically, the lamb sample is placed on a three-dimensional electric translation table, and a topographic map is generated by scanning the surface with a laser range finder. The scanning path is planned according to the topographic map to cover the entire surface area of the sample. The polarizer is rotated to the horizontal direction (0°), vertical direction (90°), reverse horizontal direction (180°), and reverse vertical direction (270°) in turn. After each rotation, the light source is kept stable.

[0014] It should be understood that the light field data of the sample surface is synchronously collected by the hyperspectral camera, and four sets of spectral data are recorded for each pixel point. The four sets of polarized direction light field data are stored as independent matrices respectively. The spatial coordinate systems of the four sets of data are aligned by image registration technology. The local missing areas caused by the difference in polarized direction are interpolated and completed. The four sets of matrices are stacked according to the direction dimension to generate the final muscle fiber four-direction light field matrix M 4D

[0015] It should be understood that its dimension is HxWx4, H is the height, W is the width, and 4 is the four polarization directions. The four sets of matrices represent the light field intensity data of the lamb sample surface collected under the four polarization directions of the rotating polarized light source. Each matrix records the reflectivity, scattering characteristics, and other optical information of the sample under the corresponding polarization direction. The light field intensity difference of each pixel point under four-direction polarization is used to represent the anisotropy of muscle fibers. According to the typical arrangement direction of muscle fibers, the four-direction data are weighted and fused to strengthen the optical contrast of muscle fibers.

[0016] Preferably, S2 includes the steps of decomposing the lamb muscle fiber bundle and perimysium connective tissue, extracting the main direction of the muscle fiber by direction gradient, dividing the pixel points into muscle fiber bundle and perimysium connective tissue regions according to the polarization response difference and main direction information, and realizing the formula:

[0017]

[0018] In the formula, E sep represents the separation energy function, which is used to quantify the separation effect of muscle fiber bundle and perimysium connective tissue, N represents the total number of image pixels, ΔI i represents the light field intensity difference of the i-th pixel point in the horizontal and vertical polarization directions, and ΔI i represents the​

[0019] Myofibril bundles have high reflectivity in longitudinal polarization and low reflectivity in transverse polarization, so ΔI i is large, while the perimysium connective tissue has no obvious directionality, and ΔI i is small.

[0020] G i represents the directional gradient of the i-th pixel point, the myofibril bundles are arranged in the longitudinal direction, and the gradient amplitude is large; while the perimysium connective tissue is distributed randomly, and the gradient amplitude is small.

[0021] s i represents the segmentation label of the i-th pixel point, δ(s i ) represents that if s i is an effective segmentation result (i.e., s i ∈{0, 1}), then δ(s i ) = 1, otherwise δ(s i ) = 0, and α and β represent weight coefficients, respectively controlling the contribution of the polarization response difference and the directional gradient.

[0022] s i is adjusted through iterative optimization to minimize E sep , so as to realize the separation of the myofibril bundles and the perimysium connective tissue.

[0023] Preferably, the step of constructing a three-dimensional muscle fiber topology map S2 comprises: extracting all pixel points marked as myofibril bundles as two-dimensional coordinates (x i , y j ) according to the segmentation result s j , combining the information of the hyperspectral camera to obtain three-dimensional coordinates z j , and generating a local surface of a Gaussian distribution for each center point (x j , y j , z j ) of the myofibril bundles to reflect the thickness distribution of the myofibril bundles.

[0024] Specifically, the implementation formula of the three-dimensional muscle fiber topology map is:

[0025]

[0026] In the formula, S(x, y, z) represents the three-dimensional muscle fiber topology map, M represents the three-dimensional coordinates of the center points of the myofibril bundles after segmentation, K j represents the curvature weight of the j-th center point, K j represents the curvature weight of the j-th center point, K j = α·ΔI j + β·G j , σ represents the standard deviation of the Gaussian kernel, controlling the smoothness of the surface, and θ(z-zj ) represents a step function, exp represents a natural exponential function.

[0027] Preferably, the S3, the step of locating the mutton mitochondrial enrichment area in the three-dimensional muscle fiber topology map comprises: extracting a high-density area of muscle fiber bundles from the three-dimensional topology map, performing local spectral analysis on the high-density area, extracting a waveband feature related to mitochondria, marking out a mitochondrial enrichment area in combination of the high-density muscle fiber area and the local spectral feature, and outputting a candidate area set R mito ={r1, r2,..., r K} in three-dimensional space.

[0028] Extracting hyperspectral data from the mitochondrial enrichment area R mito , screening a waveband related to ATP activity, determining an optimal feature waveband set Λ={λ1, λ2,..., λ L} through partial least squares regression, calculating quantum efficiency Q l for each feature waveband λ l ∈Λ, and realizing the formula as follows:

[0029]

[0030] In the formula, w r represents the weight of the mitochondrial enrichment area r, I emit,l,t represents the emission light intensity of the mitochondrial enrichment area r at the waveband λ l , I absorb,l,t represents the absorption light intensity of the mitochondrial enrichment area r at the waveband λ l , and ηl represents the photon conversion efficiency of the waveband λ l .

[0031] Preferably, the S3, the step of establishing a mapping model of ATP activity and feature waveband quantum efficiency comprises: based on the quantum efficiency Q l and the ATP activity A, establishing a mapping model, and realizing the formula as follows:

[0032]

[0033] In the formula, V mito represents the total volume of mitochondrial enrichment, the volume V j of each mitochondrial enrichment area r mito ∈R mito , a geometric volume in three-dimensional space, w l and β represent weight coefficients, and ε represents a model error term, the samples are divided into a training set and a test set, and the model accuracy is evaluated through mean square error.

[0034] Preferably, the S4, the step of multi-index decision-making comprises:

[0035] Sarcemere contraction state: Extract the contraction degree of the sarcomere based on the three-dimensional muscle fiber topology map, and analyze the Z-line spacing of the sarcomere through microscopic imaging;

[0036] Intermuscular fat lines: Extract the distribution pattern of intermuscular fat from hyperspectral data, and extract the fat area by image segmentation, and calculate the area ratio and texture uniformity;

[0037] Metabolic activity evaluation: Combine the mapping model with the volume of mitochondrial enrichment area to quantify the energy metabolism level;

[0038] Specifically, the multi-source features are converted into quantifiable indicators of tenderness and freshness, and the comprehensive score of tenderness and freshness is realized by the formula:

[0039]

[0040] In the formula, S represents the comprehensive score of mutton quality, C myo represents the contraction degree of the sarcomere, C max represents the maximum sarcomere contraction threshold, A ref represents the reference ATP activity, F fat represents the intermuscular fat line uniformity index, and w1, w2, and w3 represent weight coefficients.

[0041] It should be understood that the greater C myo , the lower the tenderness, and therefore represents the tenderness contribution, and the ratio of ATP activity A to the reference ATP activity A ref reflects the freshness, and the higher the intermuscular fat line uniformity index F fat , the better the tenderness.

[0042] Through multi-index comprehensive scoring, the tenderness and freshness grade of mutton are output, and multi-index decision is executed.

[0043] Preferably, according to the decision result, if the sample belongs to frozen meat or chilled meat, and the tenderness or freshness score is lower than the threshold, the terahertz scanning process is triggered, and the terahertz scanning device is automatically started to send the sample into the test platform.

[0044] The frequency range and resolution of the terahertz scanner are set to cover the dielectric response characteristics of ice crystals and muscle fiber membranes, the mutton sample is placed in a low-temperature constant-temperature box to keep the surface ice crystals stable, and the S916 test fixture is used to fix the sample to ensure the consistency of the sample thickness and position during scanning.

[0045] Specifically, start the terahertz time-domain spectroscopy system, transmit short pulse terahertz waves to penetrate the sample, record the time delay and amplitude attenuation of the transmitted waves, scan the sample from different directions to cover the omnidirectional response of the muscle fiber bundle, capture the difference in ice crystal distribution, perform Fourier transform on the transmitted wave data, extract the frequency domain response, calculate the dielectric constant and loss tangent, and the larger the ice crystal volume, the higher the dielectric loss, and the damage degree and change of the muscle fiber membrane are positively correlated;

[0046] The spatial distribution of ice crystals is reconstructed by scanning images, the ice crystal dielectric loss data is associated with indicators such as muscle contraction state and ATP activity, and the ice crystal damage level is divided according to a threshold value.

[0047] Preferably, the S6 obtains the tenderness and freshness decision results output by the S4 and the terahertz scanning data of the S5, introduces the laboratory test results, and performs correlation analysis on the optical feature data; the consistency of the optical feature and the traditional indicators is verified through the Pearson correlation coefficient; the optical feature of the current sample is compared with a historical database to identify abnormal patterns;

[0048] Specifically, the optical feature, laboratory data and terahertz scanning results are integrated into a unified data package, a unique hash value is generated for the integrated data package through the Keccak-256 algorithm, only the hash value and key metadata are chained, the complete optical feature data and verification report are saved through a distributed storage system, the hash value points to the off-chain data address, and a smart contract rule is written to define the data chaining condition, and the contract automatically records operation logs.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] 1. The present application combines the directional gradient and the polarization response to segment the muscle fiber bundle and the perimysium connective tissue, and constructs a three-dimensional muscle fiber topology graph, which significantly improves the segmentation accuracy and the robustness of geometric modeling, and improves the reliability and multi-scale adaptability of muscle fiber spatial distribution analysis;

[0051] 2. The present application locates the mitochondria-rich area through the three-dimensional topology graph, establishes a mapping model of ATP activity and characteristic waveband quantum efficiency, optimizes the weight parameters through the partial least squares regression, quantifies the correlation between metabolic activity and optical feature, and improves the prediction accuracy of energy metabolism evaluation;

[0052] 3. The present application fuses the muscle contraction state, the intermuscular fat line and the metabolic activity, generates a tenderness and freshness comprehensive decision through dynamic weight distribution and a nonlinear scoring formula, realizes multi-dimensional data complementation, and improves the comprehensiveness and automatic decision-making ability of quality grading;

[0053] 4、The application quantifies the relationship between ice crystal dielectric loss and muscle fiber membrane damage by triggering terahertz scanning according to the decision result, improves the reliability of frozen damage detection through multi-angle scanning and frequency domain response analysis, and combines ice crystal damage grade division. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The running process of the intelligent hyperspectral image detection method for the sheep meat quality index Figure One ;

[0055] Figure 2 The running process of the intelligent hyperspectral image detection method for the sheep meat quality index Figure Two ;

[0056] Figure 3 The running process of the intelligent hyperspectral image detection method for the sheep meat quality index Figure Three ;

[0057] Figure 4 The running process of the intelligent hyperspectral image detection method for the sheep meat quality index Figure Four . DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0059] EMBODIMENT

[0060] Please refer to Figures 1-4 , the application provides a technical solution: comprising the following steps:

[0061] S1, scanning the surface of the sheep meat sample by rotating the polarized light source to generate a muscle fiber four-way light field matrix;

[0062] S2, decomposing the muscle fiber bundle and perimysium connective tissue of the sheep meat based on the muscle fiber four-way light field matrix to construct a three-dimensional muscle fiber topology graph;

[0063] S3, positioning the mitochondria-rich area of the sheep meat according to the three-dimensional muscle fiber topology graph to establish a mapping model of ATP activity and characteristic waveband quantum efficiency;

[0064] S4, fusing the muscle segment contraction state, intermuscular fat lines and metabolic activity of the sheep meat to perform multi-index decision of tenderness and freshness;

[0065] S5, start terahertz scanning according to the decision result, quantify the muscle fiber membrane dielectric loss caused by ice crystals in mutton;

[0066] S6, integrate the mutton optical characteristics and verification data to generate a blockchain traceable hash value.

[0067] Preferably, S1, through the light halogen lamp of the rotatable vibration piece, through the three-dimensional electric translation table, carries the pressure sensor, monitors the flatness of the mutton sample surface in real time and automatically adjusts the level, drives the polarizer through the motor, and synchronously triggers the hyperspectral camera to collect data;

[0068] In this embodiment, the mutton sample is placed on the three-dimensional electric translation table, the surface is scanned by the laser range finder to generate a topographic map, the scanning path is planned according to the topographic map, the complete surface area of the sample is covered, and the polarizer is rotated to the horizontal direction (0°), the vertical direction (90°), the reverse horizontal direction (180°) and the reverse vertical direction (270°) in turn. After each rotation, the light source stability is maintained;

[0069] It should be understood that the light field data of the sample surface is synchronously collected by the hyperspectral camera, four sets of spectral data are recorded for each pixel point, the four sets of polarized direction light field data are respectively stored as independent matrices, the spatial coordinate systems of the four sets of data are aligned through image registration technology, the local missing area caused by the difference in polarized direction is interpolated and completed, and the four sets of matrices are stacked according to the direction dimension to generate the final muscle fiber four-way light field matrix M 4D ;

[0070] In this embodiment, the dimension is HxWx4, H is the height, W is the width, and 4 is the four polarization directions. Four sets of matrices respectively represent the light field intensity data of the mutton sample surface collected under the four polarization directions of the rotating polarized light source. Each matrix records the reflectivity, scattering characteristics and other optical information of the sample under the corresponding polarization direction. The light field intensity difference of each pixel point under four-way polarization is used to represent the anisotropy of muscle fibers. According to the typical arrangement direction of muscle fibers, the four-way data is weighted and fused to strengthen the optical contrast of muscle fibers.

[0071] Preferably, S2, the step of decomposing mutton myofibril bundles and perimysium connective tissue includes extracting the main direction of muscle fibers through direction gradient, dividing the pixel points into myofibril bundle area and perimysium connective tissue area according to polarization response difference and main direction information, and realizing the formula:

[0072]

[0073] In the formula, E sep represents the separation energy function, which is used to quantify the separation effect of myofibril bundles and perimysium connective tissue, N represents the total number of image pixels, ΔI iΔI i represents the difference of light field intensity between horizontal and vertical polarization directions of the i-th pixel point, ΔI i = |I0°,i - I90°,i| 90 °,i|;

[0074] Myofibril bundles have high reflectivity in longitudinal polarization and low reflectivity in transverse polarization, so ΔI i is large, while the perimysium connective tissue has no obvious directionality, and ΔI i is small;

[0075] G i represents the direction gradient of the i-th pixel point. Myofibril bundles are arranged along the longitudinal direction, and the gradient amplitude is large. The perimysium connective tissue is distributed randomly, and the gradient amplitude is small;

[0076] s i represents the segmentation label of the i-th pixel point, δ(s i ) represents that if s i is an effective segmentation effect (i.e., s i ∈{0,1}), then δ(s i ) = 1, otherwise δ(s i ) = 0, and α and β represent weight coefficients, respectively controlling the contribution of polarization response difference and direction gradient;

[0077] s i is adjusted by iterative optimization to minimize E sep , so as to realize the separation of myofibril bundles and perimysium connective tissue.

[0078] Preferably, the S2, constructing a three-dimensional muscle fiber topology map step comprises: according to the segmentation result s i , all pixel points marked as myofibril bundles are extracted as two-dimensional coordinates (x j , y j ), combined with the information of the hyperspectral camera, three-dimensional coordinates z j are obtained, and a local surface of a Gaussian distribution is generated for each myofibril bundle center point (x j , y j , z j ) to reflect the thickness distribution of the myofibril bundle;

[0079] Specifically, the three-dimensional muscle fiber topology map is realized by the formula:

[0080]

[0081] In the formula, S(x, y, z) represents the three-dimensional muscle fiber topology map, M represents the three-dimensional coordinates of the center points of the segmented myofibril bundles, K j represents the curvature weight of the j-th center point, and K j is represented as Kj = a * Delta I j + b * G j , sigma denotes the standard deviation of the Gaussian kernel, controlling the smoothness of the surface, theta (z-z j ) denotes the step function, and exp denotes the natural exponential function.

[0082] Preferably, the S3, the step of locating the mutton mitochondrial enrichment area in the three-dimensional muscle fiber topology map comprises: extracting a high-density area of muscle fiber bundles from the three-dimensional topology map, performing local spectral analysis on the high-density area, extracting a wave band feature related to mitochondria, marking out a mitochondrial enrichment area in combination with the high-density muscle fiber area and the local spectral feature, and outputting a candidate area set R mito = {r1, r2,..., r K} in three-dimensional space.

[0083] In this embodiment, hyperspectral data is extracted from the mitochondrial enrichment area R mito , wave bands related to ATP activity are screened, an optimal feature wave band set Lambda = {lambda1, lambda2,..., lambda L} is determined through partial least squares regression, and a quantum efficiency Q l of each feature wave band lambda l e Lambda is calculated, and the implementation formula is:

[0084]

[0085] In the formula, w r denotes the weight of the mitochondrial enrichment area r, I emit,l,t denotes the emission light intensity of the mitochondrial enrichment area r at the wave band lambda l , I absorb,l,t denotes the absorption light intensity of the mitochondrial enrichment area r at the wave band lambda l , and eta l denotes the photon conversion efficiency of the wave band lambda l .

[0086] In this embodiment, the S3, the step of establishing a mapping model of ATP activity and feature wave band quantum efficiency comprises: based on the quantum efficiency Q l and the ATP activity A, a mapping model is established, and the implementation formula is:

[0087]

[0088] In the formula, V mito denotes the total volume of the mitochondrial enrichment, the volume V j of each mitochondrial enrichment area r mito e R mito , is the geometric volume in three-dimensional space, and w l, β represents a weight coefficient, and ε represents a model error term, the samples are divided into a training set and a test set, and the model accuracy is evaluated by mean square error.

[0089] In this embodiment, the S4, the multi-index decision step includes:

[0090] Muscle contraction state: extract the contraction degree of muscle segment based on three-dimensional muscle fiber topology, and analyze the Z-line spacing of muscle segment by microscopic imaging;

[0091] Intermuscular fat lines: extract the distribution pattern of intermuscular fat from hyperspectral data, extract the fat area by image segmentation, and calculate the area ratio and texture uniformity;

[0092] Metabolic activity evaluation: combine the mapping model with the volume of mitochondrial enrichment area to quantify the energy metabolism level;

[0093] Specifically, the multi-source features are converted into quantifiable indexes of tenderness and freshness, and the comprehensive score of tenderness and freshness is realized by the formula:

[0094]

[0095] In the formula, S represents the comprehensive score of mutton quality, C myo represents the muscle contraction degree, C max represents the maximum muscle contraction threshold, A ref represents the reference ATP activity, F fat represents the intermuscular fat line uniformity index, and w1, w2 and w3 represent weight coefficients.

[0096] It should be understood that the greater C myo , the lower the tenderness, and therefore represents the tenderness contribution, the ratio of ATP activity A to the reference ATP activity A ref reflects the freshness, and the higher the intermuscular fat line uniformity index F fat , the better the tenderness;

[0097] Through multi-index comprehensive score, the tenderness and freshness grade of mutton are output, and multi-index decision is executed.

[0098] Preferably, the S5, according to the decision result, if the sample belongs to frozen meat or chilled meat, and the tenderness or freshness score is lower than the threshold, the terahertz scanning process is triggered, and the terahertz scanning device is automatically started to send the sample into the test platform.

[0099] In this embodiment, the frequency range and resolution of the terahertz scanner are set to cover the dielectric response characteristics of ice crystals and muscle fiber membranes, the mutton sample is placed in a low-temperature incubator to keep the surface ice crystals stable, the S916 test fixture is used to fix the sample to ensure the consistency of the sample thickness and position during scanning.

[0100] Specifically, start the terahertz time-domain spectroscopy system, emit short pulse terahertz waves to penetrate the sample, record the time delay and amplitude attenuation of the transmitted wave, scan the sample from different directions to cover the omnidirectional response of the muscle fiber bundle, capture the difference in ice crystal distribution, perform Fourier transform on the transmitted wave data, extract the frequency domain response, calculate the dielectric constant and loss tangent, and the larger the ice crystal volume, the higher the dielectric loss, and the damage degree of the muscle fiber membrane is positively correlated with the change;

[0101] Reconstruct the spatial distribution of ice crystals by scanning images, correlate the ice crystal dielectric loss data with indicators such as muscle contraction state and ATP activity, and divide the ice crystal damage level according to the threshold value.

[0102] In this embodiment, the S6 obtains the tenderness and freshness decision results output by the S4 and the terahertz scanning data of the S5, introduces the laboratory test results, and performs correlation analysis on the optical feature data, verifies the consistency of the optical features and traditional indicators through the Pearson correlation coefficient, compares the optical features of the current sample with the historical database, and identifies abnormal patterns;

[0103] Specifically, the optical features, laboratory data, and terahertz scanning results are integrated into a unified data package, a unique hash value is generated for the integrated data package through the Keccak-256 algorithm, only the hash value and key metadata are chained, the complete optical feature data and verification report are saved through a distributed storage system, the hash value points to the off-chain data address, a smart contract rule is written, the data chaining condition is defined, and the contract automatically records operation logs.

[0104] Working principle: Through the cooperation of the halogen lamp with the rotatable vibration piece and the three-dimensional electric translation stage, automatic leveling and multi-angle scanning of the sheep meat sample surface are realized. After the laser range finder generates a topographic map, the scanning path is planned to cover the entire sample area. The polarizer is rotated to the horizontal direction, vertical direction, reverse horizontal direction, and reverse vertical direction in turn, and the hyperspectral camera is triggered to collect light field data at the same time. Four sets of data are aligned in space coordinates through image registration, and the missing areas are completed through interpolation. Finally, the four-way light field matrix is superimposed, and the anisotropic optical contrast of muscle fibers is enhanced through weighted fusion;

[0105] The main direction of muscle fibers is extracted by directional gradient, and the muscle fiber bundle and perimysium area are divided by combining the polarization response difference (longitudinal polarization reflectivity is high, and transverse polarization reflectivity is low). The center points of the segmented muscle fiber bundles are fitted by Gaussian surface, and a three-dimensional topological graph is constructed by combining the depth information of the hyperspectral camera. The smoothness is controlled by the standard deviation of the Gaussian kernel. The high-density muscle fiber region is extracted from the three-dimensional topological graph, and the mitochondria-related waveband features are screened by combining local spectral analysis. The optimal feature waveband is determined by partial least squares regression, and the quantum efficiency is calculated. A mapping model is established based on the quantum efficiency and ATP activity. The model accuracy is verified by the training set and the test set. The contraction degree is quantified by analyzing the sarcomere Z-line spacing through microscopic imaging. The area ratio and texture uniformity of intermuscular fat are extracted by hyperspectral segmentation. Combined with the mapping model, the multi-source features are fused by weighted fusion. The tenderness and freshness comprehensive score formula is normalized and weighted to output the grading result. When the tenderness or freshness score is lower than the threshold value, the terahertz scanning process is triggered. The sample is placed in a low-temperature incubator and fixed by an S916 clamp. Short-pulse terahertz waves are emitted to penetrate the sample. The time delay and amplitude attenuation of the transmitted wave are recorded. The frequency domain response is extracted by Fourier transform. The dielectric constant and loss tangent are calculated, and the ice crystal spatial distribution is reconstructed. Combined with the sarcomere contraction state and ATP activity, the damage level is divided. The tenderness, freshness decision result, terahertz scanning data and laboratory test result of S4 are integrated into a data package. A unique hash value is generated by Keccak-256 algorithm. Only the hash value and key metadata are chained. The complete data is saved by distributed storage. The hash value points to the off-chain address. The smart contract defines the on-chain condition and records the operation log.

[0106] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

[0107] The above describes the present application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In summary, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which should belong to the protection scope of the present application.

Claims

1. An intelligent hyperspectral image detection method for mutton quality indicators, characterized in that, The method comprises the following steps: S1, scanning the surface of the mutton sample by rotating a polarized light source to generate a muscle fiber four-directional light field matrix; S2, decomposing mutton myofibril bundles and perimysium connective tissue based on the muscle fiber four-directional light field matrix to construct a three-dimensional muscle fiber topology map; S3, locating the mutton mitochondria-rich area according to the three-dimensional muscle fiber topology map to establish a mapping model of ATP activity and characteristic waveband quantum efficiency; S4, fusing the mutton sarcomere contraction state, intermuscular fat lines and metabolic activity to perform multi-index decision of tenderness and freshness; S5, starting terahertz scanning according to the decision result to quantify the muscle fiber membrane dielectric loss caused by ice crystals in mutton; S6, integrating the mutton optical characteristics and verification data to generate a block chain traceable hash value.

2. The lamb quality index intelligent hyperspectral image detection method according to claim 1, characterized in that: The S1, through the light ray halogen lamp of rotatable vane, motor drives the polarizer, and triggers the hyperspectral camera to collect data synchronously, places the mutton sample on the three-dimensional electric translation table, generates the topographic map through the laser range finder scanning surface, plans the scanning path according to the topographic map, covers the complete surface area of the sample, rotates the polarizer to the horizontal direction, the vertical direction, the reverse horizontal and the reverse vertical four directions in turn, after each rotation, keeps the light source stability, synchronously collects the light field data of the sample surface through the hyperspectral camera, records four groups of spectral data for each pixel point, stores the light field data of four groups of polarization directions as independent matrix respectively, aligns the spatial coordinate system of four groups of data through the image registration technology, interpolates and completes the local missing area caused by the polarization direction difference, superimposes four groups of matrix according to the direction dimension, generates the final muscle fiber four-way light field matrix M 4D .

3. The lamb quality index intelligent hyperspectral image detection method according to claim 1, characterized in that: The S2 decomposing mutton myofibril bundles and perimysium connective tissue step comprises extracting the main direction of the muscle fiber by direction gradient, dividing the pixel points into myofibril bundle areas and perimysium connective tissue areas according to the polarization response difference and the main direction information, and realizing the formula: In the formula, E sep represents the separation energy function, N represents the total number of image pixels, ΔI i represents the light field intensity difference of the i-th pixel point in the horizontal and vertical polarization directions, ΔI i represents ΔI i = |I0°,i-I 90 °,i|, G i represents the direction gradient of the i-th pixel point, s i represents the segmentation label of the i-th pixel point, δ(s i ) represents that if s i is an effective segmentation effect, then δ(s i )=1, otherwise δ(s i )=0, and α and β represent weight coefficients.

4. The lamb quality index intelligent hyperspectral image detection method according to claim 3, characterized in that: S2, the step of constructing a three-dimensional muscle fiber topology includes: according to the segmentation result s i In the method, all pixel points marked as myofibril bundles are extracted as two-dimensional coordinates (x j ,y j ), combined with the information of the hyperspectral camera, three-dimensional coordinates z j are obtained, and a local surface of a Gaussian distribution is generated for each myofibril bundle center point (x j ,y j , z j ), the thickness distribution of the myofibril bundle is reflected, and the three-dimensional muscle fiber topology is realized according to the formula: In the formula, S(x, y, z) represents a three-dimensional muscle fiber topology map, M represents three-dimensional coordinates of center points of muscle fiber bundles after segmentation, K j represents a curvature weight of the jth center point, σ represents a standard deviation of a Gaussian kernel, θ(z-z j ) represents a step function, and exp represents a natural exponential function.

5. The lamb quality index intelligent hyperspectral image detection method according to claim 1, characterized in that: The S3, the three-dimensional muscle fiber topology map locates the mutton mitochondria enrichment area step includes: extracting the high density area of muscle fiber bundle from the three-dimensional topology map, performing local spectral analysis on the high density area, extracting the wave band characteristics related to mitochondria, marking out the mitochondria enrichment area combined with the high density muscle fiber area and the local spectral characteristics, and outputting as a candidate area set R in three-dimensional space mito ={r1, r2,..., r K} Extracting hyperspectral data from the mitochondria enrichment area R mito , screening the wave bands related to ATP activity, determining the optimal feature wave band set Lambda = {lambda1, lambda2,..., lambda L} through partial least squares regression, each feature wave band being lambda l e Lambda, calculating the quantum efficiency Q l , and realizing the formula as In the formula, w r represents the weight of the mitochondria-rich region r, I emit,l,t represents the emission light intensity of the mitochondria-rich region r, I l represents the absorption light intensity of the mitochondria-rich region r, and ηl represents the photon conversion efficiency of the wave band λ absorb,l,t . l l In the formula, w r represents the weight of the mitochondria-rich region r, I emit,l,t represents the emission light intensity of the mitochondria-rich region r, I l represents the absorption light intensity of the mitochondria-rich region r, and ηl represents the photon conversion efficiency of the wave band λ absorb,l,t . l l In the formula, w r represents the weight of the mitochondria-rich region r 6. The lamb quality index intelligent hyperspectral image detection method according to claim 1, characterized in that: The S3, the mapping model step of establishing the ATP activity and the characteristic waveband quantum efficiency includes: based on the quantum efficiency Q l and the ATP activity A, the mapping model is established, and the formula is: In the formula, V mito represents the total volume of mitochondria-rich cells, w l , β represents the weight coefficient, and ε represents the model error term. The samples are divided into a training set and a test set, and the model accuracy is evaluated by the mean square error.

7. The lamb quality index intelligent hyperspectral image detection method according to claim 1, characterized in that: The S4 multi-index decision step comprises: Sarcomere contraction state: extracting the contraction degree of the sarcomere based on the three-dimensional muscle fiber topology map, and analyzing the Z-line spacing of the sarcomere by microscopic imaging; Intermuscular fat lines: extracting the distribution pattern of intermuscular fat from the hyperspectral data, and extracting the fat area by image segmentation; Metabolic activity evaluation: combining the mapping model and the volume of the mitochondria-rich area to quantify the energy metabolism level; Converting the multi-source features into quantifiable indexes of tenderness and freshness, performing comprehensive scoring of the tenderness and freshness, and outputting the tenderness and freshness grades of the mutton through multi-index comprehensive scoring.

Images