Intelligent discrimination method and system for frozen damage degree of meat based on multi-modal information

Abstract

The application discloses a kind of intelligent discrimination method and system of frozen damage degree of meat based on multi-modal information, belong to food nondestructive testing and intelligent processing technical field.Aiming at the problem that existing frozen damage degree discrimination method is strong, cannot be monitored in real time, multi-source data is isolated, the three-dimensional MRI image of meat, microstructure image and color difference, weight loss rate, shear force quality index after thawing are synchronously collected, and frozen damage feature database of multi-source fusion is constructed and damage grading is carried out;Further, the frozen damage grade is intelligently discriminated using support vector machine, and the thawing quality is discriminated using long short-term memory network, to realize the quantitative discrimination of quality parameters from multi-modal data input.The system has the characteristics of nondestructive, real-time, multi-dimensional and discriminable, and is suitable for quality control of meat low-temperature storage whole chain.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing and intelligent processing technology for food. More specifically, this invention relates to a method and system for intelligently determining the degree of freezing damage in meat based on multimodal information. Background Technology

[0002] During the low-temperature storage and cold chain distribution of meat, temperature fluctuations can easily cause the meat to freeze, especially under various low-temperature environments such as ice-temperature storage, ultra-ice-temperature preservation, non-crystallization treatment, and frozen storage. Once the temperature drops below the freezing point, ice crystals easily form inside the meat. The growth of ice crystals can cause irreversible physical damage to muscle fiber structure and cell membrane integrity, leading to quality degradation problems such as juice loss, deterioration in texture, and color deterioration after thawing. Currently, monitoring the degree of this type of freezing damage mainly relies on traditional destructive methods, such as thawing loss rate determination, colorimeter measurement, texture analysis, and observation of ice crystal morphology and distribution by combining tissue sections (HE staining) with optical microscopy or scanning electron microscopy. Although these methods can reflect the freezing damage to some extent, they all require destructive treatment of the samples, making it impossible to achieve continuous, real-time monitoring of the same sample, and even more difficult to apply to online quality control. Furthermore, existing technologies often analyze macroscopic quality indicators or microscopic structural features in isolation, lacking systematic integration and correlation analysis of multi-source data (such as temperature history, image features, and physicochemical indicators). This results in limited accuracy in judging the degree of freezing damage and an inability to achieve cross-scale discrimination from structural damage to quality evolution. Therefore, developing a technology capable of real-time, non-destructive, and multi-dimensional monitoring and intelligent identification of the degree of freezing damage in cold-stored meat has become an urgent need in the field of cold chain quality control. Summary of the Invention

[0003] The object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages described below.

[0004] To achieve these objectives and other advantages according to the present invention, a method for intelligently determining the degree of freezing damage in meat based on multimodal information is provided, comprising: Real-time three-dimensional MRI images, microstructure images, temperature dynamics during the cooling process, and quality index data after thawing of low-temperature stored meat training samples were collected. The quality indexes include color difference, weight loss rate, and shear force. Preprocessing and feature extraction are performed on real-time 3D MRI images and microscopic structure images of the training samples to obtain the feature vectors of the corresponding images; The feature vectors of real-time three-dimensional MRI images of the training samples are used as input, and the frozen injury level divided by the quality index data of the training samples is used as the classification label to train a frozen injury level discrimination model. The frozen injury level discrimination model determines the frozen injury level based on the feature vectors of real-time three-dimensional MRI images. The thawing quality discrimination model is trained by using the feature vectors and temperature dynamic features of real-time three-dimensional MRI images of training samples as the main input, the feature vectors of microscopic structure images of training samples as auxiliary supervision information, and the quality index data of training samples as the output label. By introducing a microscopic feature-assisted loss function for joint optimization, a thawing quality discrimination model is obtained. The thawing quality discrimination model judges the quality index value after thawing based on the feature vectors and temperature dynamic features of real-time three-dimensional MRI images. The trained freezing damage level discrimination model, thawing quality discrimination model, and their corresponding mapping relationship between freezing damage level and quality index data range are stored in the freezing damage database. Real-time three-dimensional MRI images of meat samples stored at low temperatures and temperature dynamics during the cooling process were collected. The real-time three-dimensional MRI images of the samples were preprocessed and feature extracted to obtain the feature vector of the real-time three-dimensional MRI images of the samples. The feature vector of the real-time three-dimensional MRI image of the sample to be tested is input into the freezing damage level discrimination model to obtain the freezing damage level of the sample to be tested; the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested are input into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; according to the output freezing damage level, the corresponding quality index data range is queried from the freezing damage database, and the discriminated quality index predicted value is compared and verified with the queried quality index data range to generate the final thawing quality discrimination result.

[0005] Preferably, the steps of preprocessing and feature extraction of real-time 3D MRI images specifically include: The real-time three-dimensional MRI images were sequentially subjected to Gaussian filtering and non-local mean denoising processing, wherein the kernel size of the Gaussian filter was 3×3×3 and the standard deviation was 0.5-1.0; The Otsu thresholding algorithm is used to extract binary images of ice crystal regions from the denoised image. The texture features of the ice crystal region are extracted using a three-dimensional local binary mode algorithm, and the local gradient histogram of each voxel is calculated to form an initial feature vector. Principal component analysis is used to reduce the dimensionality of the initial feature vector. The number of principal components is dynamically determined based on the cumulative variance contribution rate reaching 90%-95%, resulting in an MRI image feature vector containing 12-18 principal components.

[0006] Preferably, after performing Gaussian filtering and nonlocal mean denoising on the real-time 3D MRI image, the process further includes an enhancement step using Savitzky-Golay smoothing filtering: The three-dimensional MRI image was divided into two-dimensional slice sequences in the XY plane according to the slice thickness direction, and each slice was subjected to Savitzky-Golay filtering in the three-dimensional spatial domain. The filter window size is set to 5×5×3, the polynomial order is 2-3, and the local pixel polynomial surface is fitted by the least squares method to calculate the smoothing value of the center pixel. To address the artifacts that are prone to occur in the ice crystal boundary region, an adaptive weight coefficient adjustment strategy is adopted: when the local pixel gradient value is greater than the threshold, the polynomial fitting weight is shifted to the boundary pixel by 15%-25% to preserve the details of the ice crystal edge. Anisotropic diffusion enhancement was performed on the filtered 3D image, with the diffusion coefficient set to 0.2-0.5 and the number of iterations to 3-5 times, to further eliminate noise while maintaining the clarity of tissue interfaces; The filtering effect is evaluated in real time using the structural similarity index. When the SSIM value is below 0.92, the window size is automatically adjusted to 7×7×3 and the above filtering process is repeated.

[0007] Preferably, the frozen injury level discrimination model is a support vector machine classification model, which takes the MRI image feature vector as input and outputs the corresponding frozen injury level. The support vector machine adopts a Gaussian radial basis kernel function and optimizes the penalty parameter C and kernel function parameter δ through grid search.

[0008] Preferably, the preprocessing steps for the microscopic structure image specifically include: The microscopic structure image was divided into 50×50 pixel blocks with a 30% overlap, and median filtering was applied to each block to remove salt-and-pepper noise. To assess the spectral characteristics of ice crystal micrographs, reflectance spectral data within the 450-700 nm wavelength range were selected. Using the average spectrum of the image patch as a reference spectrum, a multivariate scattering correction algorithm was employed to calculate the correction coefficient matrix. The correction formula is as follows:

[0009] in X The original spectral matrix, The average spectrum of the image patch. b To correct the slope, The average value is the reference spectrum. Wavelet transform is introduced to perform multi-scale decomposition on the corrected image patch. In the high-frequency subband, the residual noise of scattering is eliminated by soft thresholding, while the microstructure contour is preserved in the low-frequency subband. An adaptive weight model based on local variance is constructed. When the gray-level variance of an image patch is greater than 1.5 times the global variance, the weight of the MSC correction coefficient is increased by 20%-30% to enhance the background suppression effect in the ice crystal edge region. The anisotropic diffusion coefficient of the corrected image is calculated using structural tensor analysis, and edge-preserving filtering is performed iteratively 2-4 times.

[0010] Preferably, the step of preprocessing the microscopic structure image before feature extraction includes: Gaussian pyramid downsampling was used to generate a three-layer scale space for the microstructure image. Local binary pattern and local gradient histogram features were extracted from each layer and combined to form a multi-scale feature matrix. Dynamic PCA dimensionality reduction of the multi-scale feature matrix: Set the initial number of principal components to 25-30, calculate the Pearson correlation coefficient between each principal component and the quality index, and retain the principal components with an absolute value of correlation coefficient greater than 0.6; The final number of principal components is dynamically determined by the cumulative variance contribution rate. When the cumulative variance contribution rate reaches 85%-90%, dimensionality reduction is stopped, and the principal component feature vectors of the 8-12 dimensional microstructure are obtained as the auxiliary supervision information.

[0011] Preferably, the temperature dynamic characteristics during the cooling process include a temperature change rate sequence and a signal intensity attenuation rate sequence, and their acquisition and calculation methods are as follows: Temperature data were collected during the cooling process of low-temperature stored meat at time intervals of 0-10s. The temperature-time curve was fitted by cubic spline interpolation, and the temperature change rate at adjacent time points was calculated to obtain the temperature change rate sequence. Simultaneously extract the MRI signal intensity values ​​corresponding to each time point, and calculate the signal intensity attenuation rate γ per unit time using the signal intensity S0 at the initial temperature as the benchmark: γ=[(S0-S t ) / S0]×(1 / t), where S t Let t be the signal strength at time t, where t is the sampling time point, to obtain the signal strength attenuation rate sequence.

[0012] Preferably, the thawing quality discrimination model is a temporal discrimination model that includes a long short-term memory network. The model input layer receives feature vectors and temperature dynamic features from real-time three-dimensional MRI images. The hidden layer is configured with two 128-dimensional LSTM units and a regularization layer with a dropout rate of 0.3. The output layer uses a fully connected layer to discriminate the quality index values ​​after thawing.

[0013] Preferably, the thawing quality discrimination model is optimized using a joint loss function during training. This joint loss function is composed of a weighted main loss function and an auxiliary loss function: the main loss function is the mean squared error between the quality index discrimination value and the true value; the auxiliary loss function is the mean squared error between the hidden state of the last time step of the long short-term memory network and the feature vector of the microstructure image. By constraining the temporal features extracted by the LSTM to approximate the microstructure features through the auxiliary loss function, the model learns the mapping relationship from MRI image features and temperature dynamic features to microstructure features, thereby improving the discrimination accuracy of the quality index.

[0014] This invention also provides an intelligent system for judging the degree of freezing damage in meat based on multimodal information, comprising: The data acquisition module acquires real-time three-dimensional MRI images of low-temperature stored meat training samples and / or test samples, microstructure images of training samples, temperature dynamic characteristics of training samples and / or test samples during the cooling process, and quality index data of training samples after thawing. The quality indexes include color difference, weight loss rate, and shear force. The feature extraction module preprocesses and extracts features from real-time 3D MRI images of training samples and / or test samples, as well as microscopic images of training samples, to obtain feature vectors of the corresponding images. The frozen injury level discrimination model training module takes the feature vector of the real-time three-dimensional MRI image of the training sample as input and the frozen injury level divided by the quality index data of the training sample as the classification label to train the frozen injury level discrimination model. The frozen injury level discrimination model discriminates the frozen injury level based on the feature vector of the real-time three-dimensional MRI image. The thawing quality discrimination model training module uses the feature vectors and temperature dynamic features of real-time three-dimensional MRI images of training samples as the main input, the feature vectors of microscopic structure images of training samples as auxiliary supervision information, and the quality index data of training samples as the output label. By introducing a microscopic feature auxiliary loss function for joint optimization, the thawing quality discrimination model is trained. The thawing quality discrimination model judges the quality index value after thawing based on the feature vectors and temperature dynamic features of real-time three-dimensional MRI images. The database module stores the trained freezing damage level discrimination model, thawing quality discrimination model, and their corresponding mapping relationship between freezing damage level and quality index data range in the freezing damage database. The thawing quality discrimination module inputs the feature vector of the real-time three-dimensional MRI image of the sample to be tested into the freeze damage level discrimination model to obtain the freeze damage level of the sample; it also inputs the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; based on the output freeze damage level, it queries the freeze damage database for the corresponding quality index data range, compares and verifies the discriminated quality index predicted value with the queried quality index data range, and generates the final thawing quality discrimination result.

[0015] This invention offers at least the following advantages: By acquiring real-time three-dimensional MRI images, microscopic images, and quality index data of cryogenically stored meat and constructing a database, this invention establishes a comprehensive non-destructive monitoring system covering the entire process from image acquisition to freezing damage grading. Compared to traditional destructive testing, MRI imaging and microscopic image analysis do not require sample destruction and can directly obtain the dynamic three-dimensional structure and microscopic damage characteristics of ice crystal growth during the cooling process of cryogenically stored meat. Combined with macroscopic quality indicators such as color difference, weight loss rate, and shear force, it achieves multi-dimensional quantitative analysis of the degree of freezing damage, solving the problem of real-time non-destructive monitoring of freezing damage in cold chain transportation. Specifically, based on ice crystal texture feature extraction and PCA dimensionality reduction from MRI images, it can accurately capture the structural damage differences corresponding to different freezing damage levels. The dynamically updated database and intelligent classification model can adapt to changes in freezing damage characteristics of different varieties of cryogenically stored meat, keeping the error in freezing damage level discrimination and quality discrimination within an acceptable range for the industry, providing a quantitative basis for meat sorting and sales strategy formulation in cold chain logistics. Furthermore, this technical solution upgrades the monitoring of freezing damage in cryogenically stored meat from a single physicochemical indicator detection to a cross-scale correlation analysis of "structure-function-quality" through multimodal data acquisition (3D MRI features, microstructural features, and quality indicators) and spatiotemporal correlation analysis. By analyzing the rate of temperature change and the attenuation rate of MRI signals, the influence mechanism of different cooling gradients on ice crystal growth and muscle tissue damage can be revealed, providing scientific support for optimizing temperature control strategies in cold chain transportation, and ultimately achieving precise control of the quality of cryogenically stored meat throughout the entire chain from production to distribution.

[0016] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent method and system for judging the degree of frozen damage of meat based on multimodal information as described in this invention; Figure 2The images are real-time three-dimensional MRI images of the S001 sample described in this invention at 0, 20, 40, 60, 80, and 100 minutes (the time increments from left to right). Figure 3 The images are real-time microstructure images of the S001 sample described in this invention at 0, 20, 40, 60, 80, and 100 min (the time increases sequentially from left to right). Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0019] like Figure 1 As shown, this invention provides a method and system for intelligently determining the degree of freezing damage in meat based on multimodal information, comprising: S1. Collect real-time three-dimensional MRI images, microstructure images, temperature dynamic characteristics during the cooling process, and quality index data after thawing of low-temperature stored meat training samples. The quality indexes include color difference, weight loss rate, and shear force. The term "low-temperature stored meat" refers to meat stored below freezing point. It can be frozen meat, or unfrozen or slightly frozen meat. As long as it is stored below freezing point, the issue of freezing damage will arise.

[0020] Several types of low-temperature stored meat were selected as training samples, such as fresh beef longissimus dorsi muscle, with a size of 2×2×2cm. 3 The initial temperature was 4℃. The cryogenically stored meat was placed on the sample stage of an MRI scanner (located in the central region of the magnetic field). The temperature of the meat was lowered from 4℃ to -16.5℃ in an environment of -16 to -18℃. Simultaneously, a three-dimensional scan was performed using a spin echo (SE) sequence to acquire dynamic three-dimensional image data of ice crystal formation and growth during the cooling process. The scanning parameters were set as follows: repetition time (TR) 2000-3000ms, echo time (TE) 10-30ms, slice thickness 1-3mm, matrix size 256×256×64, number of excitations (NEX) 2-4, and images were acquired every 1 minute. This scanning method can capture the critical temperature change stage near the freezing point (-1.5℃), at which point ice crystal growth causes the most significant damage to muscle tissue, providing a dynamic temporal basis for analyzing the degree of freezing damage. The MRI scanner can be a commercially available medical or industrial superconducting magnetic resonance imaging device, with scanning parameters controlled and image data acquired through accompanying software.

[0021] The temperature dynamics during the cooling process include the temperature change rate sequence and the signal intensity attenuation rate sequence.

[0022] The temperature change rate sequence was obtained by the following method: temperature data of meat stored at low temperature were collected at time intervals of 0-10s during the cooling process, and the temperature-time curve was fitted by cubic spline interpolation to calculate the temperature change rate at adjacent time points, thus obtaining the temperature change rate sequence. The signal intensity attenuation rate sequence was obtained using the following method: MRI signal intensity values ​​at each time point were extracted synchronously, and the signal intensity S0 at the initial temperature was used as a reference to calculate the signal intensity attenuation rate γ per unit time: γ=[(S0-S t ) / S0]×(1 / t), where S t Let t be the signal strength at time t, where t is the sampling time point, to obtain the signal strength attenuation rate sequence.

[0023] Align the two sequences on the time axis to form the input features of the thawing quality discrimination model.

[0024] Temperature data was collected at 0-10 s intervals, and curves were fitted using cubic spline interpolation. This method accurately captures the dynamic temperature changes during the cooling process of meat stored at low temperatures. The calculated rate range of 0-1℃ / min reflects the influence of different cooling gradients on ice crystal growth. For example, rapid cooling (close to 0.5℃ / min) easily forms small ice crystals, while slow cooling (close to 0.01℃ / min) may lead to the formation of large ice crystals, providing a thermodynamic basis for analyzing the degree of freezing damage. The simultaneously calculated MRI signal intensity attenuation rate (0.02-0.15% / s) quantifies the influence of water freezing and ice crystal growth on proton signals. This attenuation rate is positively correlated with the ice crystal integral number, which can indirectly reflect the degree of physical damage to muscle tissue.

[0025] For each training sample, after cooling to the target temperature, its microstructure image is acquired. This image can be obtained using an optical microscope equipped with a reflectance spectral acquisition module for wavelengths of 450-700 nm, at a magnification of 200× and a resolution of 2048×2048 pixels, by photographing sample slices. A standard biological microscope with imaging capabilities can be selected as the optical microscope.

[0026] The samples were then thawed, and the quality indicators were measured after thawing. These indicators included color difference (ΔE), weight loss (%), and shear force (N). Color difference was measured using a portable colorimeter (such as a Minolta CR-400) to measure the L, a, and b* values ​​and calculate ΔE. Weight loss was calculated by weighing the samples before and after thawing using a precision balance. Shear force was measured using a texture analyzer (such as a TA.XT Plus).

[0027] This step involves multi-dimensional data collection to obtain the microstructural characteristics and macroscopic quality parameters of meat stored at low temperatures, providing basic data for database construction.

[0028] S2. Preprocess and extract features from the real-time 3D MRI images and microscopic structure images of the training samples to obtain the feature vectors of the corresponding images.

[0029] S2.1 Real-time 3D MRI image preprocessing and feature extraction; Gaussian filtering (kernel size 3×3×3, standard deviation 0.5-1.0) and nonlocal mean (NLM) denoising were sequentially applied to the acquired real-time 3D MRI images to effectively suppress random noise and speckle artifacts. Simultaneously, the Otsu thresholding algorithm was used to accurately extract binary images of the ice crystal region, avoiding blurring of the boundaries between muscle tissue and ice crystals. This preprocessing improves the clarity of the ice crystal structure, making subsequent texture feature extraction more accurate, such as preserving key morphological features like ice crystal edges and size distribution. The 3D Local Binary Pattern (3D-LBP) algorithm is used to extract texture features from the ice crystal region and calculate the local gradient histogram (3D-GLH) of each voxel. This can quantify the spatial distribution complexity of ice crystals, edge gradient changes, and other microstructural parameters. Specifically, the preprocessed real-time 3D MRI images are divided into temperature-image data groups according to the time series. Each group contains 3D images corresponding to every 0.5℃ as the temperature drops from 4℃ to 16.5℃. 10-15 feature slices are extracted along the Z-axis (slice thickness direction) for each group of images. Each slice is divided into 8×8×8 voxel blocks. The mean gray level, variance, and 3D-LBP texture features of each voxel block are calculated to form the initial feature vector.

[0030] Supervised feature selection and dimensionality reduction of the initial feature vector are performed through principal component analysis, as follows: The initial number of principal components is set to 25-30. The Pearson correlation coefficient between each principal component and the quality indicators (color difference, weight loss rate, shear force) is calculated. Principal components with an absolute correlation coefficient greater than 0.6 are retained. The number of principal components is dynamically determined based on the cumulative variance contribution rate reaching 90%-95%, resulting in an MRI image feature vector containing 12-18 principal components. This reduces data redundancy while retaining more than 90% of the key information, making the feature vector more focused on core parameters related to freezing damage (such as ice crystal size and distribution density), providing low-dimensional and efficient input data for subsequent classification models. As a preferred embodiment, after Gaussian filtering and nonlocal mean denoising of the real-time 3D MRI image, the process further includes an enhancement step using Savitzky-Golay smoothing filtering: The three-dimensional MRI image was divided into two-dimensional slice sequences in the XY plane according to the slice thickness direction, and each slice was subjected to Savitzky-Golay filtering in the three-dimensional spatial domain. The filter window size is set to 5×5×3, the polynomial order is 2-3, and the local pixel polynomial surface is fitted by the least squares method to calculate the smoothing value of the center pixel. To address the artifacts that are prone to occur in the ice crystal boundary region, an adaptive weight coefficient adjustment strategy is adopted: when the local pixel gradient value is greater than the threshold (0.05-0.15), the polynomial fitting weight is shifted to the boundary pixel by 15%-25% to preserve the details of the ice crystal edge. Anisotropic diffusion enhancement was performed on the filtered 3D image, with the diffusion coefficient set to 0.2-0.5 and the number of iterations to 3-5 times, to further eliminate noise while maintaining the clarity of tissue interfaces; The filtering effect is evaluated in real time using the structural similarity index. When the SSIM value is below 0.92, the window size is automatically adjusted to 7×7×3 and the above filtering process is repeated.

[0031] Here, the 3D MRI image is divided into 2D slices according to slice thickness and then filtered in the 3D spatial domain. Combined with a 5×5×3 window and least-squares fitting of a 3rd-order polynomial, this approach can suppress image noise while preserving gradient changes between pixels, avoiding the blurring of ice crystal edges caused by traditional filtering. An adaptive weight adjustment strategy for ice crystal boundary artifacts, triggered by a gradient threshold to shift the weights of boundary pixels, effectively preserves the details of the ice crystal contour, making the morphological features of ice crystals in the microstructure more clearly discernible.

[0032] Anisotropic diffusion enhancement, using a diffusion coefficient of 0.2-0.5 and 3-5 iterations, can directionally eliminate random noise in the image while maintaining the clarity of the muscle tissue and ice crystal interface, avoiding interference from residual noise in subsequent feature extraction. A dynamic evaluation mechanism based on the Structural Similarity Index (SSIM) automatically expands the window to 7×7×3 when the filtering effect is below an SSIM value of 0.92. This adaptively optimizes filtering parameters under different noise levels, ensuring a balance between detail preservation and noise suppression in the MRI image, providing a high-quality image data foundation for subsequent frozen lesion severity assessment.

[0033] S2.2 Microscopic structure image preprocessing and feature extraction; The microscopic structure images undergo preprocessing to eliminate background interference, including multivariate scattering correction and wavelet denoising. The specific process is as follows: The microscopic structure image was divided into 50×50 pixel blocks with a 30% overlap. Median filtering (kernel size 3×3) was applied to each block to remove salt-and-pepper noise. To assess the spectral characteristics of ice crystal micrographs, reflectance spectral data within the 450-700 nm wavelength range were selected. Using the average spectrum of the image patch as a reference spectrum, a multivariate scattering correction algorithm was employed to calculate the correction coefficient matrix. The correction formula is as follows:

[0034] in X The original spectral matrix, The average spectrum of the image patch. b To correct the slope, The average value is the reference spectrum. Wavelet transform (using Daubechies-4 wavelet basis) is introduced to perform multi-scale decomposition on the corrected image patch. In the high-frequency subband, the residual noise of scattering is eliminated by soft thresholding (the threshold is 2.5 times the standard deviation of noise), while the microstructure contour is preserved in the low-frequency subband. An adaptive weight model based on local variance is constructed. When the gray-level variance of an image patch is greater than 1.5 times the global variance, the weight of the MSC correction coefficient is increased by 20%-30% to enhance the background suppression effect in the ice crystal edge region. The anisotropic diffusion coefficient (range 0.1-0.8) of the corrected image is calculated by structural tensor analysis. Edge-preserving filtering is performed iteratively 2-4 times, ultimately improving the gray-scale uniformity of the background area by no less than 40% and preserving the edge sharpness of the ice crystal structure by no less than 85%.

[0035] Here, the microscopic image is divided into 50×50 pixel image blocks with a 30% overlap. Combined with a 3×3 kernel median filter, salt-and-pepper noise is effectively removed while preserving the detailed features of the ice crystal edges. For the reflectance spectral characteristics of ice crystals in the 450-700nm wavelength range, a multivariate scattering correction algorithm is used with the average spectrum of the image blocks as a reference. This significantly reduces the interference of background scattered light, making the spectral differences between ice crystals and muscle tissue more prominent.

[0036] A Daubechies-4 wavelet basis is introduced for multi-scale decomposition, combined with a soft thresholding method with a noise standard deviation of 2.5 times to denoise the high-frequency subband. This eliminates residual scattering noise while preserving the contour information of the microstructure in the low-frequency subband. Based on an adaptive weighting model of local variance, when the gray-level variance of the image patch is greater than 1.5 times the global variance, the weight of the MSC correction coefficient is increased by 20%-30%, which can specifically enhance the background suppression effect in the ice crystal edge region and improve the contrast between the ice crystal boundary and the background.

[0037] By calculating the anisotropic diffusion coefficient in the range of 0.1-0.8 using structural tensor analysis and performing 2-4 iterations of edge-preserving filtering, the gray-scale uniformity of the background area can be significantly improved, while effectively preserving the edge clarity of the ice crystal structure. This provides high-quality image data for subsequent quantitative analysis of the degree of microstructural damage and the determination of the level of freezing damage, ensuring a more accurate correspondence between microscopic features and the level of freezing damage.

[0038] The steps for feature extraction after preprocessing the microscopic structure image include: Gaussian pyramid downsampling was used to generate a three-layer scale space for the microstructure image. Local binary pattern and local gradient histogram features were extracted from each layer and combined to form a multi-scale feature matrix. Supervised feature selection and dimensionality reduction of the multi-scale feature matrix: Set the initial number of principal components to 25-30, calculate the Pearson correlation coefficient between each principal component and the quality index, and retain the principal components with an absolute value of correlation coefficient greater than 0.6; The final number of principal components is dynamically determined by the cumulative variance contribution rate. When the cumulative variance contribution rate reaches 85%-90%, dimensionality reduction is stopped, and the principal component feature vectors of the 8-12 dimensional microstructure are obtained as the auxiliary supervision information.

[0039] As a preferred approach, during the supervised feature selection and dimensionality reduction process for the 3D MRI feature vectors and microstructure feature matrices, a temperature weighting factor can be introduced to correct the PCA loading matrix. The formula for calculating the weighting factor is as follows:

[0040] in T The current temperature. T f This is the freezing point temperature (-1.5℃). λ Using values ​​of 0.8-1.2 increases the feature weights corresponding to temperatures near the freezing point by 15%-25%.

[0041] Here, the preprocessed 3D MRI images are segmented according to temperature sequence and voxel block features (grayscale mean, variance, and 3D-LBP texture) are extracted. Combined with Gaussian pyramid multi-scale feature extraction (LBP and GLH) of microscopic images, it is possible to comprehensively capture the changes in ice crystal structure and tissue damage features of low-temperature stored meat during the cooling process from a spatiotemporal perspective, providing multi-dimensional feature data including temperature dynamic response for subsequent dimensionality reduction.

[0042] For the multi-scale feature matrix, preliminary PCA dimensionality reduction is first performed, yielding 25-30 initial principal components. To further filter out the features most relevant to the freeze damage quality indicators, this invention introduces a supervised feature selection step: calculating the Pearson correlation coefficient between each principal component and the quality indicators (color difference, weight loss rate, shear force). Only principal components with an absolute correlation coefficient greater than 0.6 with at least one quality indicator are retained, thus constructing a feature subset highly correlated with the quality indicators. Finally, this feature subset undergoes final PCA dimensionality reduction, stopping when the cumulative variance contribution rate reaches 85%-90%, resulting in an 8-12 dimensional microstructure feature vector. This ensures that the extracted features not only summarize image information but also directly serve the final quality prediction task.

[0043] Introducing a temperature weighting factor (λ ranging from 0.8 to 1.2) to correct the PCA load matrix increases the feature weights corresponding to temperatures near the freezing point (-1.5℃) by 15%-25%, thereby strengthening the feature representation of key stages of ice crystal formation. Since temperature changes near the freezing point of cold-stored meat significantly affect ice crystal growth rate and microstructural damage, this weighting adjustment highlights the characteristic differences at this stage and improves the sensitivity of freezing damage level determination.

[0044] The number of principal components is dynamically determined by the cumulative variance contribution rate. Dimensionality reduction is stopped when the cumulative variance contribution rate of the 3D MRI image features reaches 90%-95% and that of the microstructure features reaches 85%-90%. In practice, 12-18 dimensional 3D MRI principal component feature vectors and 8-12 dimensional microstructure principal component feature vectors are typically obtained, avoiding the curse of dimensionality while ensuring no loss of key information. The lower-dimensional feature vectors after dimensionality reduction reduce the computational complexity of subsequent models and improve the training efficiency and generalization ability of frozen lesion level classification models (such as SVM) through feature redundancy removal, laying a data foundation for intelligent identification of frozen lesion severity.

[0045] S3. Using the feature vector of the real-time three-dimensional MRI image of the training sample as input, and the frozen injury level divided by the quality index data of the training sample as the classification label, a frozen injury level discrimination model is trained. The frozen injury level discrimination model is based on the feature vector of the real-time three-dimensional MRI image to discriminate the frozen injury level. The frozen injury level discrimination model can be a support vector machine classification model, which takes the MRI image feature vector as input and outputs the corresponding frozen injury level. The support vector machine adopts a Gaussian radial basis kernel function and optimizes the penalty parameter C and kernel function parameter δ through grid search.

[0046] S4. Using the feature vector and temperature dynamic features of the real-time three-dimensional MRI images of the training samples as the main input, the feature vector of the microstructure images of the training samples as auxiliary supervision information, and the quality index data of the training samples as the output label, the thawing quality discrimination model is trained by introducing a microscopic feature auxiliary loss function for joint optimization. The thawing quality discrimination model is based on the feature vector and temperature dynamic features of the real-time three-dimensional MRI images to discriminate the quality index value after thawing. The thawing quality discrimination model can adopt a temporal discrimination model that includes a long short-term memory network. The input layer of the model receives the feature vector and temperature dynamic features of real-time three-dimensional MRI images. The hidden layer is set with two 128-dimensional LSTM units and a regularization layer with a dropout rate of 0.3. The output layer uses a fully connected layer to discriminate the quality index value after thawing.

[0047] During model training, a joint loss function is used for optimization. This joint loss function is composed of a weighted main loss function and an auxiliary loss function: the main loss function is the mean squared error between the quality indicator's discriminant value and the true value; the auxiliary loss function is the mean squared error between the hidden state of the last time step of the Long Short-Term Memory network and the feature vector of the microscopic structure image. By constraining the temporal features extracted by the LSTM to approximate the microscopic structure features, the model learns the mapping relationship from MRI image features and temperature dynamics features to microscopic structure features, thereby improving the accuracy of quality indicator discrimination. Thus, in the prediction stage, no microscopic image input is required; the thawing quality can be accurately determined based on MRI features and temperature dynamics features. The weighting coefficients are determined through cross-validation, for example, a main loss weight of 0.7 and an auxiliary loss weight of 0.3.

[0048] This time-series discriminant model utilizes a two-layer, 128-dimensional LSTM unit to capture the temporal dependencies between temperature, signal attenuation, and structural features during the cooling process of meat stored at low temperatures. Combined with a discard rate of 0.3 to suppress overfitting, it can effectively learn the dynamic evolution of quality indicators (color difference, weight loss rate, and shear force) during freezing damage. For example, the model can identify the effect of sudden temperature changes near the freezing point on muscle protein denaturation, and then discriminate the trend of shear force changes after thawing, keeping the deviation between the discriminant value and the actual thawing quality within the range of color difference ±0.3, weight loss rate ±0.5%, and shear force ±3.0N.

[0049] S5. Store the trained freezing damage level discrimination model, thawing quality discrimination model and their corresponding mapping relationship between freezing damage level and quality index data range (refer to Table 6) in the freezing damage database. The database can also store the original data, feature vectors, model parameters, etc. of the training samples, and supports dynamic updates.

[0050] For example, spatiotemporal alignment of three types of data (MRI feature vector, microscopic feature vector, and quality index) from the same sample at the same temperature and time point; Based on a pre-defined freezing damage grading standard, a freezing damage level label is assigned to each sample or each temperature time point. The grading criteria include comprehensive indicators such as color difference ΔE value, weight loss rate, and shear force. Structured storage is performed using a relational database (such as MySQL) or a time-series database. The database must contain at least the following tables: Sample Information Table: Stores basic sample information; MRI Feature Table: Stores the dimensionality-reduced MRI feature vectors and their acquisition temperature and timestamp; Microscopic feature table: stores the dimensionality-reduced microscopic feature vectors and their corresponding field of view and temperature; Quality Indicators Table: Labels for stored color difference, weight loss rate, shear force measurements, and calculated freeze damage level.

[0051] Frozen Damage Registration Standard Table: Stores the quality index range and typical image feature descriptions corresponding to each level of frozen damage.

[0052] Create a related index for the above table to enable quick retrieval of all related data by sample number, temperature point, or freezing damage level.

[0053] The database can reserve a data interface to support the import of raw data, preprocessed features and model discrimination results of new detection samples into the database; When the accumulated new sample data reaches a certain scale (such as an increase of 10% in data volume), the model retraining process can be triggered. The frozen damage level discrimination model (SVM) and the thawing quality discrimination model (LSTM) can be retrained using the updated database, and the classification boundary and feature weights can be optimized. Establish data version management, record logs for each data update and model iteration, and ensure the traceability of the database.

[0054] Thus, a database of freezing damage in cryogenically stored meat, containing multimodal raw data, processed features, hierarchical labels, and intelligent models, has been completed. This database not only serves the comparison and discrimination of subsequent samples, but its dynamic update mechanism also ensures that the system can continuously learn and optimize.

[0055] S6. Collect real-time three-dimensional MRI images of the low-temperature stored meat sample and the temperature dynamic characteristics during the cooling process. Preprocess and extract features from the real-time three-dimensional MRI images of the sample to obtain the feature vector of the real-time three-dimensional MRI image of the sample. The real-time three-dimensional MRI image acquisition of the test sample was performed using the same method as the training sample (same as step S1).

[0056] For the sample to be tested, its temperature dynamic characteristics during the cooling process can be obtained in two ways: 1) Online monitoring: The sample to be tested is placed in a cold chain environment with real-time temperature monitoring, and its complete temperature process from the initial temperature to the target temperature is recorded from scratch.

[0057] 2) Offline traceability and simulation: For samples that have undergone a cooling process, their historical temperature curves can be traced through embedded temperature recording tags (RFID temperature tags); if traceability is not possible, a standard cooling curve of the same type and specification as the sample is used as the model input for quality prediction.

[0058] The real-time three-dimensional MRI images of the test sample are preprocessed and feature extracted using the same method as in the training phase (same as step S2.1) to obtain the MRI image feature vector of the test sample.

[0059] S7. Input the feature vector of the real-time three-dimensional MRI image of the sample to be tested into the freezing damage level discrimination model to obtain the freezing damage level of the sample; input the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; according to the output freezing damage level, query the corresponding quality index data range from the freezing damage database, compare and verify the discriminated quality index prediction value with the queried quality index data range, and generate the final thawing quality discrimination result. This method can assess the reliability of the discrimination result through historical data distribution and provide a quantitative basis for freezing damage risk warning in cold chain logistics, helping enterprises to formulate graded processing strategies for low-temperature stored meat with different freezing damage levels (such as prioritizing the sale of meat with high freezing damage levels) and reduce economic losses caused by quality deterioration.

[0060] In practical applications, we conducted an experiment using beef as an example. The following section uses 103 beef samples as an example to explain in detail the entire process from data input to model output.

[0061] 1. Basic information of the beef sample: Fresh beef sirloin, longissimus dorsi muscle, measuring 2×2×2 cm. 3 The initial temperature was 4℃.

[0062] 2. Raw data collection 1) Three-dimensional MRI images Acquisition conditions included: ambient temperature -16 to -18℃; sample temperature range: 4℃ → -16.5℃; scanning interval: images were acquired every 1 minute. Scanning parameters include: sequence: spin echo (SE); TR / TE: 2500ms / 20ms; slice thickness: 2mm; matrix: 256×256×64; number of excitations: 3. like Figure 2 The image shows real-time three-dimensional MRI images of sample S001 at (0 min, 4℃), (20 min, -1℃), (40 min, -1.1℃), (60 min, -1.2℃), (80 min, -1.5℃), and (100 min, -3.9℃). Due to the large number of sample images, they are not listed one by one.

[0063] 2) Microscopic structural images The sampling conditions included: ambient temperature -16 to -18℃; sample temperature range: 4℃ → -16.5℃. Microscope: Equipped with a 450-700nm reflectance spectroscopy module Resolution: 2048×2048 pixels Magnification: 200× like Figure 3 The image shows real-time microstructure images of sample S001 at (0 min, 4℃), (20 min, -1℃), (40 min, -1.1℃), (60 min, -1.2℃), (80 min, -1.5℃), and (100 min, -3.9℃). Due to the large number of sample images, they are not listed one by one.

[0064] 3) Quality indicator data Color difference (∆E): The original values ​​of the three channels L, a, b* are measured using a portable colorimeter (such as Minolta CR-400). The difference between the current time / temperature and the initial sample is then calculated. Finally, ∆E = √[(ΔL)² + (Δa)² + (Δb*)²] is calculated. Weight loss rate (%): Using a precision balance (accuracy 0.01g), the initial weight and the weight after thawing of the sample were measured, and the weight loss rate (%) was calculated as follows: (initial weight - thawed weight) / initial weight × 100%; Shear force (N): A texture analyzer (such as TA.XT Plus) was used at a test speed of 1 mm / s and a probe diameter of 1.27 mm. The output was a force-displacement curve, and the peak shear force (N) was read. Table 1 below shows the quality index data of sample S001 at (0 min, 4℃), (20 min, -1℃), (40 min, -1.1℃), (60 min, -1.2℃), (80 min, -1.5℃), and (100 min, -3.9℃). Due to the large number of sample data, they are not listed one by one.

[0065] Table 1 4) Temperature-time curve data Data collection frequency: Record once every 2 seconds; Sensor: Thermocouple (accuracy ±0.1℃); Table 2 below shows some temperature-time data for sample S001. Due to the large amount of sample data, it will not be listed one by one.

[0066] Table 2

[0067] 3. Feature Extraction 1) MRI feature vectors Preprocessing: Savitzky-Golay filtering + anisotropic diffusion; Segmentation: Otsu thresholding method for extracting ice crystal regions; Features: 3D-LBP, local gradient histogram, grayscale statistics; Dimensionality reduction: PCA to 18 dimensions; Table 3 below shows some of the MRI feature vectors of sample S001.

[0068] Table 3

[0069] 2) Microscopic feature vectors Preprocessing: Multivariate scattering correction + wavelet denoising; Features: Local Binary Pattern (LBP), Gradient Histogram (GLH); Dimensionality reduction: PCA to 12 dimensions; Table 4 below shows some of the microscopic feature vectors of sample S001.

[0070] Table 4

[0071] 3) Temperature dynamic characteristics Includes: temperature change rate sequence (°C / min), signal attenuation rate sequence (% / s); 4. Database construction (SQLite / MySQL table structure) This includes: a table of standards for the grading of frozen skin injuries (see Table 5), a sample information table, an MRI feature table, and a quality index table.

[0072] Table 5

[0073] It should be noted that freezing damage is a complex and integrated physical process, and changes in a single indicator (such as color difference, weight loss rate, or shear force) may not be synchronous. To assign a unique and representative freezing damage level label to each sample, this invention employs the following comprehensive weighted scoring method: Index normalization: The measured color difference ΔE, weight loss rate (%), and shear force (N) are normalized to 0-100 points according to their respective proportions within their respective ranges; Weighting: Weights are assigned based on the degree of impact of indicators on actual edible quality and commercial value. For example, weight loss rate (juice loss) is directly related to economic benefits and is assigned a weight of 0.4; shear force (tenderness) is a core taste indicator and is assigned a weight of 0.4; color difference (appearance) is assigned a weight of 0.2. Calculate the comprehensive damage index: Comprehensive damage index = weight loss rate score × 0.4 + shear force score × 0.4 + color difference score × 0.2; Mapping level: The level is determined based on the comprehensive damage index. For example: 0-20 is considered fresh meat, 21-40 is considered level I, 41-60 is considered level II, and so on.

[0074] According to the above rules, the data of sample S001 in Table 1 were recalculated and classified.

[0075] 5. Model training and validation data 1) Dataset partitioning (103 samples) Training set: 70 samples (S001-S070); Validation set: 30 samples (S071-S100); Test set: 3 samples (S101-S103); 2) Model output: The results of the freezing damage level determination are shown in Table 6 below: Table 6

[0076] The results of the thawing quality assessment are shown in Table 7 below: Table 7

[0077] It should be noted that the number of training samples used in this embodiment is only for principle verification, aiming to demonstrate the complete process and feasibility of the method. Those skilled in the art will understand that in practical engineering applications, the model can be retrained or subjected to transfer learning by collecting more representative samples to further improve the discrimination accuracy.

[0078] In the task of determining the level of frozen damage, the Support Vector Machine (SVM) model performed stably, with an accuracy of 97.8% on the training set and 96.4% on the validation set. In the test set, sample S103 had a relatively unique ice crystal distribution morphology (Level II damage but ice crystal size close to Level I), leading to a misclassification as Level I and a lower confidence level (79.4%). The other two samples were correctly identified, and the overall classification accuracy basically met the initial validation requirements. Further improvements in model robustness can be made by increasing sample diversity.

[0079] In terms of quantitative judgment of thawing quality, the regression model combining LSTM and microscopic feature-assisted supervision showed good prediction accuracy: color difference ΔE: mean absolute error 0.27, root mean square error 0.28; weight loss rate (%): mean absolute error 0.27%, root mean square error 0.28%; shear force (N): mean absolute error 2.33N, root mean square error 2.35N. All error values ​​were controlled within the design requirements (color difference ±0.3, weight loss rate ±0.5%, shear force ±3.0N), confirming that the model can accurately map from MRI time-series features and temperature dynamic features to the final quality parameters. Table 7 shows the detailed prediction results for three test samples, and their error levels are all within the acceptable range.

[0080] In terms of practicality, the model completes the entire process from feature extraction to freezing damage identification and quality assessment for a single sample, with an average latency of approximately 3.2 seconds. This speed meets the needs of most offline or near real-time monitoring scenarios. Overall, the model not only demonstrates high classification and discrimination accuracy but also good operational efficiency, providing a feasible technical solution for non-destructive monitoring of freezing damage and intelligent quality assessment during the low-temperature storage of meat.

[0081] Based on the same inventive concept, the present invention also provides an intelligent system for judging the degree of frozen damage of meat based on multimodal information. The system may be a personal computer, a server, or other system that implements the aforementioned intelligent system for judging the degree of frozen damage of meat based on multimodal information.

[0082] The intelligent meat freezing damage assessment system based on multimodal information includes: The data acquisition module acquires real-time three-dimensional MRI images of low-temperature stored meat training samples and / or test samples, microstructure images of training samples, temperature dynamic characteristics of training samples and / or test samples during the cooling process, and quality index data of training samples after thawing. The quality indexes include color difference, weight loss rate, and shear force. The feature extraction module preprocesses and extracts features from real-time 3D MRI images of training samples and / or test samples, as well as microscopic images of training samples, to obtain feature vectors of the corresponding images. The frozen injury level discrimination model training module takes the feature vector of the real-time three-dimensional MRI image of the training sample as input and the frozen injury level divided by the quality index data of the training sample as the classification label to train the frozen injury level discrimination model. The frozen injury level discrimination model discriminates the frozen injury level based on the feature vector of the real-time three-dimensional MRI image. The thawing quality discrimination model training module uses the feature vectors and temperature dynamic features of real-time three-dimensional MRI images of training samples as the main input, the feature vectors of microscopic structure images of training samples as auxiliary supervision information, and the quality index data of training samples as the output label. By introducing a microscopic feature auxiliary loss function for joint optimization, the thawing quality discrimination model is trained. The thawing quality discrimination model judges the quality index value after thawing based on the feature vectors and temperature dynamic features of real-time three-dimensional MRI images. The database module stores the trained freezing damage level discrimination model, thawing quality discrimination model, and their corresponding mapping relationship between freezing damage level and quality index data range in the freezing damage database. The thawing quality discrimination module inputs the feature vector of the real-time three-dimensional MRI image of the sample to be tested into the freeze damage level discrimination model to obtain the freeze damage level of the sample; it also inputs the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; based on the output freeze damage level, it queries the freeze damage database for the corresponding quality index data range, compares and verifies the discriminated quality index predicted value with the queried quality index data range, and generates the final thawing quality discrimination result.

[0083] All relevant content of each step involved in the aforementioned embodiment of the intelligent method for judging the degree of frozen damage to meat based on multimodal information can be referred to in the functional description of the corresponding functional module of the judgment system in the embodiment of this application, and will not be repeated here.

[0084] The division of units in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this invention can be integrated into a single processor, exist as separate physical units, or two or more units can be integrated into a single module. The integrated units described above can be implemented in hardware or as software functional units.

[0085] The present invention also provides an electronic device comprising: at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the aforementioned intelligent method for determining the degree of meat freezing damage based on multimodal information. This electronic device can be any terminal device including mobile phones, laptops, desktop computers, tablets, PDAs (Personal Digital Assistants), POS (Point of Sales) terminals, and in-vehicle computers.

[0086] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described intelligent method for judging the degree of frozen damage to meat based on multimodal information.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the present invention, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, portable hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0088] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for intelligently determining the degree of freezing damage in meat based on multimodal information, characterized in that, include: Real-time three-dimensional MRI images, microstructure images, temperature dynamics during the cooling process, and quality index data after thawing of low-temperature stored meat training samples were collected. The quality indexes include color difference, weight loss rate, and shear force. Preprocessing and feature extraction are performed on real-time 3D MRI images and microscopic structure images of the training samples to obtain the feature vectors of the corresponding images; The feature vectors of real-time three-dimensional MRI images of the training samples are used as input, and the frozen injury level divided by the quality index data of the training samples is used as the classification label to train a frozen injury level discrimination model. The frozen injury level discrimination model determines the frozen injury level based on the feature vectors of real-time three-dimensional MRI images. The thawing quality discrimination model is trained by using the feature vectors and temperature dynamic features of real-time three-dimensional MRI images of training samples as the main input, the feature vectors of microscopic structure images of training samples as auxiliary supervision information, and the quality index data of training samples as the output label. By introducing a microscopic feature-assisted loss function for joint optimization, a thawing quality discrimination model is obtained. The thawing quality discrimination model judges the quality index value after thawing based on the feature vectors and temperature dynamic features of real-time three-dimensional MRI images. The trained freezing damage level discrimination model, thawing quality discrimination model, and their corresponding mapping relationship between freezing damage level and quality index data range are stored in the freezing damage database. Real-time three-dimensional MRI images of meat samples stored at low temperatures and temperature dynamics during the cooling process were collected. The real-time three-dimensional MRI images of the samples were preprocessed and feature extracted to obtain the feature vector of the real-time three-dimensional MRI images of the samples. The feature vector of the real-time three-dimensional MRI image of the sample to be tested is input into the freezing damage level discrimination model to obtain the freezing damage level of the sample to be tested; the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested are input into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; according to the output freezing damage level, the corresponding quality index data range is queried from the freezing damage database, and the discriminated quality index predicted value is compared and verified with the queried quality index data range to generate the final thawing quality discrimination result.

2. The intelligent method for determining the degree of meat freezing damage based on multimodal information as described in claim 1, characterized in that, The specific steps for preprocessing and feature extraction of real-time 3D MRI images include: The real-time three-dimensional MRI images were sequentially subjected to Gaussian filtering and non-local mean denoising processing, wherein the kernel size of the Gaussian filter was 3×3×3 and the standard deviation was 0.5-1.0; The Otsu thresholding algorithm is used to extract binary images of ice crystal regions from the denoised image. The texture features of the ice crystal region are extracted using a three-dimensional local binary mode algorithm, and the local gradient histogram of each voxel is calculated to form an initial feature vector. Principal component analysis is used to reduce the dimensionality of the initial feature vector. The number of principal components is dynamically determined based on the cumulative variance contribution rate reaching 90%-95%, resulting in an MRI image feature vector containing 12-18 principal components.

3. The intelligent method for judging the degree of meat freezing damage based on multimodal information as described in claim 2, characterized in that, After performing Gaussian filtering and nonlocal mean denoising on the real-time 3D MRI image, the process further includes an enhancement step using Savitzky-Golay smoothing filtering: The three-dimensional MRI image was divided into two-dimensional slice sequences in the XY plane according to the slice thickness direction, and each slice was subjected to Savitzky-Golay filtering in the three-dimensional spatial domain. The filter window size is set to 5×5×3, the polynomial order is 2-3, and the local pixel polynomial surface is fitted by the least squares method to calculate the smoothing value of the center pixel. To address the artifacts that are prone to occur in the ice crystal boundary region, an adaptive weight coefficient adjustment strategy is adopted: when the local pixel gradient value is greater than the threshold, the polynomial fitting weight is shifted to the boundary pixel by 15%-25% to preserve the details of the ice crystal edge. Anisotropic diffusion enhancement was performed on the filtered 3D image, with the diffusion coefficient set to 0.2-0.5 and the number of iterations to 3-5 times, to further eliminate noise while maintaining the clarity of tissue interfaces; The filtering effect is evaluated in real time using the structural similarity index. When the SSIM value is below 0.92, the window size is automatically adjusted to 7×7×3 and the above filtering process is repeated.

4. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 1, characterized in that, The frozen injury level discrimination model is a support vector machine classification model. It takes the MRI image feature vector as input and outputs the corresponding frozen injury level. The support vector machine adopts a Gaussian radial basis kernel function and optimizes the penalty parameter C and kernel function parameter δ through grid search.

5. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 1, characterized in that, The specific steps for preprocessing microscopic structure images include: The microscopic structure image was divided into 50×50 pixel blocks with a 30% overlap, and median filtering was applied to each block to remove salt-and-pepper noise. To assess the spectral characteristics of ice crystal micrographs, reflectance spectral data within the 450-700 nm wavelength range were selected. Using the average spectrum of the image patch as a reference spectrum, a multivariate scattering correction algorithm was employed to calculate the correction coefficient matrix. The correction formula is as follows: in X The original spectral matrix, The average spectrum of the image patch. b To correct the slope, The average value is the reference spectrum. Wavelet transform is introduced to perform multi-scale decomposition on the corrected image patch. In the high-frequency subband, the residual noise of scattering is eliminated by soft thresholding, while the microstructure contour is preserved in the low-frequency subband. An adaptive weight model based on local variance is constructed. When the gray-level variance of an image patch is greater than 1.5 times the global variance, the weight of the MSC correction coefficient is increased by 20%-30% to enhance the background suppression effect in the ice crystal edge region. The anisotropic diffusion coefficient of the corrected image is calculated using structural tensor analysis, and edge-preserving filtering is performed iteratively 2-4 times.

6. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 5, characterized in that, The steps for feature extraction after preprocessing the microscopic structure image include: Gaussian pyramid downsampling was used to generate a three-layer scale space for the microstructure image. Local binary pattern and local gradient histogram features were extracted from each layer and combined to form a multi-scale feature matrix. Dynamic PCA dimensionality reduction of the multi-scale feature matrix: Set the initial number of principal components to 25-30, calculate the Pearson correlation coefficient between each principal component and the quality index, and retain the principal components with an absolute value of correlation coefficient greater than 0.6; The final number of principal components is dynamically determined by the cumulative variance contribution rate. When the cumulative variance contribution rate reaches 85%-90%, dimensionality reduction is stopped, and the principal component feature vectors of the 8-12 dimensional microstructure are obtained as the auxiliary supervision information.

7. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 1, characterized in that, The temperature dynamics during the cooling process include a temperature change rate sequence and a signal intensity attenuation rate sequence, which are acquired and calculated as follows: Temperature data were collected during the cooling process of low-temperature stored meat at time intervals of 0-10 seconds. Temperature-time curves were fitted using cubic spline interpolation, and the rate of temperature change at adjacent time points was calculated to obtain a temperature change rate sequence. Simultaneously extract the MRI signal intensity values ​​corresponding to each time point, and calculate the signal intensity attenuation rate γ per unit time using the signal intensity S0 at the initial temperature as the benchmark: γ=[(S0-S t ) / S0]×(1 / t), where S t Let t be the signal strength at time t, where t is the sampling time point, to obtain the signal strength attenuation rate sequence.

8. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 1, characterized in that, The thawing quality discrimination model is a temporal discrimination model that includes a long short-term memory network. The model input layer receives feature vectors and temperature dynamic features from real-time three-dimensional MRI images. The hidden layer is set with two 128-dimensional LSTM units and a regularization layer with a dropout rate of 0.

3. The output layer uses a fully connected layer to discriminate the quality index values ​​after thawing.

9. The intelligent method for determining the degree of frozen damage to meat based on multimodal information as described in claim 8, characterized in that, During the training of the thawing quality discrimination model, a joint loss function is used for optimization. The joint loss function is composed of a weighted main loss function and an auxiliary loss function: the main loss function is the mean squared error between the quality index discrimination value and the true value; the auxiliary loss function is the mean squared error between the hidden state of the last time step of the long short-term memory network and the feature vector of the microstructure image. By constraining the temporal features extracted by LSTM to approximate the microstructure features through the auxiliary loss function, the model learns the mapping relationship from MRI image features and temperature dynamic features to microstructure features, thereby improving the discrimination accuracy of the quality index.

10. A smart system for judging the degree of frozen damage to meat based on multimodal information, characterized in that, include: The data acquisition module acquires real-time three-dimensional MRI images of low-temperature stored meat training samples and / or test samples, microstructure images of training samples, temperature dynamic characteristics of training samples and / or test samples during the cooling process, and quality index data of training samples after thawing. The quality indexes include color difference, weight loss rate, and shear force. The feature extraction module preprocesses and extracts features from real-time 3D MRI images of training samples and / or test samples, as well as microscopic images of training samples, to obtain feature vectors of the corresponding images. The frozen injury level discrimination model training module takes the feature vector of the real-time three-dimensional MRI image of the training sample as input and the frozen injury level divided by the quality index data of the training sample as the classification label to train the frozen injury level discrimination model. The frozen injury level discrimination model discriminates the frozen injury level based on the feature vector of the real-time three-dimensional MRI image. The thawing quality discrimination model training module uses the feature vectors and temperature dynamic features of real-time three-dimensional MRI images of training samples as the main input, the feature vectors of microscopic structure images of training samples as auxiliary supervision information, and the quality index data of training samples as the output label. By introducing a microscopic feature auxiliary loss function for joint optimization, the thawing quality discrimination model is trained. The thawing quality discrimination model judges the quality index value after thawing based on the feature vectors and temperature dynamic features of real-time three-dimensional MRI images. The database module stores the trained freezing damage level discrimination model, thawing quality discrimination model, and their corresponding mapping relationship between freezing damage level and quality index data range in the freezing damage database. The thawing quality discrimination module inputs the feature vector of the real-time three-dimensional MRI image of the sample to be tested into the freeze damage level discrimination model to obtain the freeze damage level of the sample; it also inputs the feature vector and temperature dynamic features of the real-time three-dimensional MRI image of the sample to be tested into the thawing quality discrimination model to obtain the predicted quality index value of the sample after thawing; based on the output freeze damage level, it queries the freeze damage database for the corresponding quality index data range, compares and verifies the discriminated quality index predicted value with the queried quality index data range, and generates the final thawing quality discrimination result.

Images