Method for intelligent and accurate spot check of beef balls based on AI visual identification

Abstract

The invention discloses an intelligent and accurate beef ball spot check method based on AI visual identification, and relates to the technical field of intelligent food detection. Comprising the following steps: S1, constructing defect data: constructing a graph neural network through the microstructure and 3D surface topography of a beef ball sample, generating a virtual beef ball defect sample through multi-physics field modeling, and verifying the physical authenticity of the virtual beef ball defect sample through the graph neural network; s2, dual-channel confrontation training: identifying the beef ball model through a structure channel and a material channel, and determining a loss function of the beef ball model; and S3, quality detection: according to the composite loss function, setting a sampling inspection rate of the beef balls, and constructing a quality evaluation system. The geometric deformation of the beef balls is identified through the structure channel, and the fat distribution texture of the beef balls is analyzed through the material channel, so that the misjudgment rate can be synergistically reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent food detection, and specifically to a method for intelligent and precise sampling of beef balls based on AI visual recognition. Background Art

[0002] With the development of the modern food processing industry, consumers' demands for food safety and quality are increasing. Beef balls, a popular quick-frozen meat product, can face quality issues during production, such as mixed raw materials, uneven processing, and poor molding. Therefore, random quality inspections of beef balls are crucial for ensuring product quality.

[0003] Currently, beef meatball quality inspection primarily relies on manual visual inspection or traditional image processing for spot checks. Manual inspection relies on the experience and judgment of quality inspectors, which is not only inefficient and labor-intensive, but also highly subjective and susceptible to factors such as fatigue, leading to frequent missed and false positives. Furthermore, manual spot checks struggle to cover the entire batch, posing significant quality risks.

[0004] Chinese invention patent application publication number CN119643806A discloses a meat dish quality detection system, comprising a data acquisition module, a first determination module, a second determination module, a judgment module, a correction module, and an alarm transmission module. This invention effectively improves the quality control of meatball production lines through precise multi-parameter analysis and a step-by-step judgment mechanism. By real-time monitoring of the moisture content, sphericity, and shape of the meatballs, it accurately identifies potential quality issues such as cracks, excessive moisture, or excessive dryness. By dynamically adjusting the clamping pressure, the system can adapt to varying production conditions in real time, preventing the impact of excessive or insufficient clamping force on the shape and taste of the meatballs. The real-time feedback mechanism of the correction module ensures rapid adjustments to production process fluctuations, effectively reducing the defective rate during production and addressing the issues of low detection accuracy and slow response speed caused by reliance on static data.

[0005] Most existing production lines use AI vision models to monitor changes in the appearance and defect types of beef balls. However, in actual production, the proportion of qualified products exceeds 95%, and a single production line only sees 3-5 true defects per day, resulting in a scarcity of defect samples. Furthermore, defect labeling relies on professional quality inspectors, and manual labeling is subject to subjective bias. Consequently, within the same batch of monitored images, different annotators may have different judgments on "acceptable pore size." Therefore, when using AI vision models to monitor beef balls, switching to a new recipe will increase the model's error rate. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for intelligent and accurate sampling of beef balls based on AI visual recognition to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solutions: a method for intelligent and precise sampling of beef balls based on AI visual recognition, comprising: S1: Construct defect data: Build a graph neural network based on the micromorphology and 3D surface morphology of the beef ball sample. Simultaneously, generate a virtual beef ball defect sample through multi-physics modeling, and verify the physical authenticity of the virtual beef ball defect sample through the graph neural network. S2: Dual-channel adversarial training: Identify the beef ball model through the set structure channel and material channel, and determine the loss function of the beef ball model, including: S2.1: Constructing a network architecture: Identifying the geometric deformation of the beef ball model through the structure channel, and identifying the texture synthesis of the beef ball model through the material channel, thereby obtaining a geometric deformation field and a fat distribution map of the beef ball model; S2.2: Determine a loss function: Based on the geometric deformation field and the fat distribution map, obtain the adversarial loss of the structural channel, the adversarial loss of the material channel, the perceptual loss value, the density constraint term, the thermal conductivity constraint term, and the elasticity constraint term; S2.3: Determine a composite loss function: Determine a total adversarial loss based on the adversarial loss of the structural channel and the adversarial loss of the material channel, determine a physical constraint loss based on the density constraint term, the thermal conductivity constraint term, and the elasticity constraint term, and simultaneously obtain a composite loss function based on the total adversarial loss, the perceptual loss value, and the physical constraint loss using a weighted algorithm; S3: Quality inspection: According to the composite loss function, the sampling rate of beef balls is set and a quality assessment system is constructed.

[0008] Furthermore, multi-physics modeling was used to generate virtual beef meatball defect samples, including: S1.1: Thermal-mechanical coupling defect generation: Through the constructed material constitutive model, the elastic modulus change and anisotropic characteristics of the beef ball model during the cooking process are obtained, the real-time fracture stress of the beef ball model is determined, and the real-time fracture stress is compared with the critical fracture stress. Based on the comparison results, the crack state is determined, specifically: When the real-time fracture stress is greater than the critical fracture stress, the beef balls will crack; otherwise, the beef balls will not crack; S1.2: Fat flow modeling: Using the multiphase flow module and the set material properties, perform fat flow simulation on the beef ball model to determine the fat volume exudation rate of the beef ball model; S1.3: Obtaining defective samples: Based on the constructed material attribute library and the divided process areas, a defect simulation is performed on the beef ball model to obtain a beef ball model with defects.

[0009] Furthermore, a three-stage test was performed on the beef ball sample to obtain the displacement and strain of the beef ball sample, and the first and second parameters of the Mooney-Rivlin model were determined to construct a material constitutive model, specifically: ; in: is the strain energy density, is the first parameter of the Mooney-Rivlin model, is the second parameter of the Mooney-Rivlin model, is the first deformation invariant, is the second deformation invariant, is the volume ratio, is the compressibility parameter.

[0010] Furthermore, according to the process area and the production process flow of the beef balls, the spatial weight distribution of the beef balls is determined, and a 3D probability cloud map is obtained. The formula for obtaining the spatial weight distribution is specifically: ; in: For the The spatial weight coefficient of the region, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area, is the region index, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area.

[0011] Furthermore, we constructed a graph neural network based on the microscopic morphology and 3D surface morphology of the beef meatball samples, including: W1: Data acquisition: The beef meatball sample slices were examined using an optical microscope, a laser confocal microscope, and a near-infrared spectrometer to obtain a two-dimensional topography image, a three-dimensional surface elevation map, and a local near-infrared reflectance spectrum of the beef meatball sample; W2: Data processing: Convert the two-dimensional topography image into a grayscale image to construct a box counting double logarithmic curve, align the coordinates of the three-dimensional surface elevation map and the two-dimensional topography image, identify the marked foreign matter based on the Z value in the three-dimensional surface elevation map, and map the spectral data of the local near-infrared reflectance spectrum to the three-dimensional surface of the three-dimensional surface elevation map to construct a three-dimensional moisture distribution cloud map; W3: Feature extraction: Based on the box counting double logarithmic curve, the marked foreign matter and the three-dimensional moisture distribution cloud map, the time series graph network, the harmonic similarity k-nearest neighbor topology map and the voxel node map are obtained; W4: Construct a graph neural network: use the time series graph network, harmonic similarity k-nearest neighbor topology graph and voxel node graph as the top-level graph, middle-level graph and bottom-level graph of the graph neural network in turn, and connect the top-level graph, middle-level graph and bottom-level graph through edges.

[0012] Furthermore, based on the box counting double logarithmic curves corresponding to different defects, the morphological similarity between different defects is obtained, and a time series graph network is constructed; According to the coordinates of the marked foreign body, a Fourier descriptor vector is obtained through Fourier transform, and a normalized edge weight is determined, and a harmonic similarity k-nearest neighbor topological map is constructed based on the Fourier descriptor vector and the normalized edge weight; The three-dimensional moisture distribution cloud map is divided according to the set voxel grid size to construct a voxel node map.

[0013] Furthermore, the edge size between the bottom graph and the top graph of the graph neural network is specifically: ; in: is the moisture-thermodynamics comprehensive weight, is the moisture gradient vector, is the thermodynamic simulation similarity, is the maximum moisture gradient vector of the current batch.

[0014] Furthermore, according to the three-dimensional model of the beef ball model and the Marching Cubes algorithm, a binary mask of the beef ball model is obtained, and the binary mask is used as the input of the U-Net architecture model, and the geometric deformation field is obtained as the output; The geometric deformation field is used as the initial condition of the CycleGAN network model, and the muscle texture map of the beef ball model is used as the input of the CycleGAN network model, and the fat distribution map is obtained as the output.

[0015] Furthermore, we set a sampling rate for beef balls and built a quality assessment system, including: S3.1: Determine the sampling rate: Adjust the basic sampling rate according to the composite loss function, specifically: ; in: is the adjusted sampling rate, As the basic sampling rate, is the sampling rate adjustment coefficient, is the composite loss function, is the loss benchmark threshold, is the loss fluctuation range coefficient; S3.2: Constructing a quality assessment system: comparing the composite loss function with a preset judgment threshold, and determining the quality grade of the beef meatballs based on the comparison result.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention uses a structural channel to identify the geometric deformation of the beef balls and a material channel to analyze the fat distribution texture of the beef balls. This dual-channel adversarial training can collaboratively reduce the misjudgment rate. At the same time, based on physical constraints such as density, thermal conductivity, and elasticity, the generated defects are forced to conform to real physical laws, avoiding the misjudgment caused by traditional visual models that ignore physical laws, further reducing the misjudgment rate. Second, the present invention uses multi-physics field modeling to obtain high-fidelity virtual defect samples and verifies the physical authenticity of the generated samples through graph neural networks, thereby ensuring data reliability. At the same time, the generated virtual defect samples cover rare defects such as crack propagation and fat exudation that are difficult to obtain in actual production, thus expanding the training data set. Third: The present invention dynamically obtains the production line risk index through a composite loss function and adjusts the sampling rate in real time, so that the sampling rate of high-risk batches can be increased by 30%-50%, and the sampling rate of low-risk batches can be reduced to 5%, thereby achieving precise allocation of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the intelligent and precise sampling method for beef balls of the present invention; Figure 2 This is a comparison chart of the sampling inspection rates in the present invention; Figure 3 It is the probability cloud map of defect space distribution in the present invention; Figure 4 This is a comparison chart of the process effects in the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Most existing production lines use AI visual models to monitor the appearance changes and defect types of beef balls. However, in actual production, the proportion of qualified products exceeds 95%, and a single production line only has 3-5 real defects per day, which leads to a scarcity of defective samples. At the same time, defect labeling relies on professional quality inspectors, and manual labeling has subjective biases. As a result, in the same batch of monitoring images, different annotators have different judgments on "acceptable pore size". Therefore, in the process of monitoring beef balls through AI visual models, when the production line switches to beef balls with a new formula, the model's misjudgment rate will increase. This application generates virtual defect samples through multi-physics field modeling, and verifies the authenticity of the virtual defect samples through graph neural networks. At the same time, through a dual-channel adversarial training model containing structural channels and material channels, a composite loss function of the virtual defect samples is obtained. The composite loss function is used to dynamically adjust the sampling rate and establish a quality assessment system. This realizes intelligent detection of the entire process from data generation, model training to quality grading. Example 1

[0020] refer to Figure 1-Figure 4 This embodiment provides a method for intelligently and accurately sampling beef balls based on AI visual recognition. The method specifically includes the following steps: Step S1: Construct defect data. Defective beef ball samples are detected and identified using an optical microscope, a laser confocal microscope, and a near-infrared spectrometer. The microscopic morphology and 3D surface morphology of the defective beef ball samples are obtained. A graph neural network is constructed based on the obtained microscopic and 3D surface morphologies. Simultaneously, multi-physics modeling is used to generate virtual defective beef ball samples. The constructed graph neural network is then used to verify the physical authenticity of the virtual defective beef ball samples. The details are as follows: Step S1.1: Thermal-mechanical coupling defect generation. That is, the beef ball sample is subjected to a three-stage test using the TA.XT Plus texture analyzer, namely the compression stage, the holding stage, and the recovery stage. Specifically, the beef ball sample is compressed at a strain rate of 1 mm / s using a 5 mm diameter probe, that is, compressed to a 50% deformation, and maintained at this constant strain for 30 seconds. The beef ball sample is then rebounded at the same strain rate of 1 mm / s. Furthermore, during the three-stage test, the elastic modulus, relaxation modulus, and recovery rate of the beef ball sample are obtained.

[0021] Furthermore, after the three-stage test, a speckle pattern was sprayed onto the surface of the beef ball sample, and high-speed images and mechanical data were acquired to obtain the corresponding displacement and strain. Based on the acquired displacement and strain, the first and second parameters of the Mooney-Rivlin model were determined, specifically: ; in: is the total number of experimental data points, is the index of the experimental data point, For the The engineering stress of the experimental data point, is the first parameter of the Mooney-Rivlin model, is the second parameter of the Mooney-Rivlin model, is the elongation ratio.

[0022] Furthermore, based on the first and second parameters of the Mooney-Rivlin model, the Mooney-Rivlin hyperelastic constitutive model is constructed, specifically: ; in: is the strain energy density, is the first parameter of the Mooney-Rivlin model, is the second parameter of the Mooney-Rivlin model, is the first deformation invariant, is the second deformation invariant, is the volume ratio, is the compressibility parameter.

[0023] In this embodiment, geometric modeling is performed based on the diameter of the beef balls and the finite element model. Specifically, a 1 / 4 symmetrical model is established based on the size of the beef balls with a diameter of 30 mm, and preset micro-defects (such as 0.1 mm initial cracks) are randomly distributed on the established beef ball model. Furthermore, in the finite element model, the initial temperature (such as 25°C), the cooking temperature (such as a gradient temperature increase to 85°C), and the heat convection coefficient (such as 12 W / (m 2 .K) to simulate the cooking process of the established beef ball model and obtain the fracture stress in real time during the cooking process. Furthermore, the Mooney-Rivlin hyperelastic constitutive model was constructed to obtain the nonlinear deformation behavior of the beef ball model during the cooking process, namely the change in elastic modulus at different temperatures (for example, the elastic modulus at 20°C is 0.8MPa, which drops to 0.3MPa when cooked to 85°C) and the characterization of anisotropic properties (for example, the ratio of longitudinal to transverse deformation reaches 1.2:1).

[0024] Furthermore, based on the initial defect half-length, surface energy, and Young's modulus corresponding to the beef ball sample, the critical fracture stress of the beef ball sample was determined as follows: ; in: is the critical fracture stress, is Young's modulus, is the surface energy, is the half length of the initial defect.

[0025] Specifically, the fracture stress of the beef ball model obtained in real time during the cooking process is compared with the critical fracture stress, and the crack state is determined based on the comparison results. Specifically: When the real-time fracture stress is greater than the critical fracture stress, the initial defect begins to expand, and cracks appear in the beef balls. Conversely, when the real-time fracture stress is not greater than the critical fracture stress, the initial defect does not expand, and the beef balls do not crack.

[0026] Step S1.2: Fat flow modeling. This involves setting material properties using the ANSYS Fluent Multiphase Flow Module, including the material's phase (e.g., fat and minced meat), density, viscosity, and interfacial tension. This involves using the ANSYS Fluent Multiphase Flow Module to simulate fat flow in the beef meatball model. Furthermore, during the fat flow simulation, the corresponding fat volume exudation rate is obtained in real time, specifically: ; in: is the fat volume exudation rate, is the equivalent hydraulic radius of the crack channel, is the pressure difference across the crack, is the dynamic viscosity of fat, is the crack length.

[0027] Step S1.3: Obtain defect samples. This means building a material attribute library based on the metallic and non-metallic foreign matter found in the beef meatball production process. This is shown in Tables 1 and 2 below: Table 1: Metal foreign matter table

[0028] Table 2: Non-metallic foreign matter

[0029] Furthermore, according to the beef meatball production process, the process area is divided into the meat grinding area, the mixing area, the forming area and the conveying area. Specifically, according to the constructed material attribute library, the foreign matter type and weight coefficient of the process area are set, as shown in Table 3 below: Table 3: Regional area weight table

[0030] In this example, a defect simulation was performed on a beef ball model based on the regional area weight table to obtain a defective beef ball model. Specifically, the spatial weight distribution of the beef balls was determined based on the coordinates of the process area divisions and the process flow during the beef ball production process. Based on this spatial weight distribution, a 3D probability cloud map was generated. Simultaneously, a heat map of the foreign matter locations within the beef ball model was obtained based on the 3D probability cloud map.

[0031] In this embodiment, the formula for obtaining the spatial weight distribution of beef balls is specifically: ; in: For the The spatial weight coefficient of the region, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area, is the region index, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area.

[0032] Step S2: Dual-channel adversarial training. This involves identifying the beef ball model through the set structure channel and material channel, and obtaining the corresponding loss function. The details are as follows: Step S2.1: Build a network architecture. This involves identifying the geometric deformation of the beef ball model through the structure channel and identifying the texture synthesis of the beef ball through the material channel.

[0033] Specifically, a slice of a beef meatball sample was scanned using micro-CT to obtain a 3D model. Using the Marching Cubes algorithm, the isosurface was extracted and projected to generate a corresponding binary mask. This binary mask was then used as input to a U-Net architecture model, which output the corresponding geometric deformation field.

[0034] Furthermore, the obtained geometric deformation field is used as the initial condition of the CycleGAN network model, and the muscle texture map of the beef ball model is used as the input of the CycleGAN network model, and the corresponding fat distribution map is obtained as the output.

[0035] Step S2.2: Determine the loss function. That is, based on the geometric deformation field and the original CT image obtained in step S2.1, determine the adversarial loss of the corresponding structural channel, specifically: ; in: is the adversarial loss of the structural channel, is the structure discriminator network, is the real CT scan data, is the real displacement field, To generate the defect CT image, is the predicted deformation field, is the expectation operator.

[0036] Furthermore, based on the fat distribution map and the real micrograph obtained in step S2.1, the adversarial loss of the corresponding material channel is determined, specifically: ; in: is the adversarial loss of the material channel, Index for experts, is the scoring function of the k-th expert, For the generated material texture image, This is the real fat distribution map.

[0037] In this embodiment, the corresponding perceptual loss value is obtained based on the feature vector of the beef meatball in the material texture image and the feature vector in the real microscopic image, specifically: ; in: is the perceptual loss value, is the feature map height, is the feature map width, For location Feature vectors in the material texture image, For location The eigenvector in the real microscopic image, is the Gram matrix weight, To generate the Gram matrix of the image, is the Gram matrix of the real image, is the Frobenius norm.

[0038] In this embodiment, based on the virtual CT image corresponding to the beef ball model, the virtual CT image is sampled by partitioning the sample points according to the number of sampling points. Simultaneously, the corresponding temperature rise rate is obtained using a transient heat conduction model. Specifically, the corresponding density constraint term is determined based on the CT Hounsfield units and the true density value at each sampling point. Simultaneously, the corresponding thermal conductivity constraint term is determined based on the obtained temperature rise rate. The corresponding elasticity constraint term is determined based on the stiffness tensor and strain tensor of each beef ball sample.

[0039] Furthermore, in this embodiment, the formulas for obtaining the density constraint term, the thermal conductivity constraint term, and the elasticity constraint term are specifically as follows: ; in: is the density constraint term, is the number of sampling points, is the CT Hounsfield unit of the pth sampling point, is the true density value of the pth sampling point, is the thermal conductivity constraint term, is the temperature change rate, To generate the sample thermal conductivity, is the temperature field Laplace operator, is the detection area, is the observation time, is the elastic constraint term, To generate the sample stiffness tensor, is the strain tensor, is the true Cauchy stress.

[0040] Step S2.3: Determine the composite loss function. Specifically, the total adversarial loss is determined based on the adversarial losses of the structural channel and the material channel obtained in step S2.2. Simultaneously, the physical constraint loss is determined based on the density constraint, thermal conductivity constraint, and elasticity constraint terms obtained in step S2.2.

[0041] Furthermore, through a weighted algorithm, the total adversarial loss, physical constraint loss and perceptual loss values ​​are combined to obtain the corresponding composite loss function.

[0042] Step S3: Quality inspection. That is, according to the composite loss function obtained in step S2.3, the sampling rate of beef balls is set, and the corresponding quality assessment system is constructed. The details are as follows: Step S3.1: Determine the sampling rate. That is, according to the composite loss function obtained in step S2.3, adjust the set basic sampling rate to determine the adjusted sampling rate, specifically: ; in: is the adjusted sampling rate, As the basic sampling rate, is the sampling rate adjustment coefficient, is the composite loss function, is the loss benchmark threshold, is the loss fluctuation range coefficient.

[0043] Step S3.2: Constructing a quality assessment system. That is, based on the composite loss function obtained in step S2.3, the composite loss function is compared with a preset judgment threshold, and the quality grade of the beef meatballs is determined based on the comparison result.

[0044] Specifically, when the composite loss function is less than 0.5, the corresponding beef meatball quality grade is excellent. When the composite loss function is between 0.5 and 1.2, the corresponding beef meatball quality grade is good. When the composite loss function is less than 1.2-2, the corresponding beef meatball quality grade is qualified. When the composite loss function is greater than 2, the corresponding beef meatball quality grade is unqualified.

[0045] refer to Figure 2 , Figure 2 is a comparison chart of sampling rates in this embodiment, Figure 2 As can be seen, for the high-risk batch (Batch C), the sampling rate was increased from the baseline value of 20% to 48%, significantly increasing the intensity of sampling to capture more defects. For the low-risk batch (Batch A), the sampling rate was reduced from 20% to 8%, reducing the number of sampling inspections by 60%, thus avoiding wasting resources on low-risk batches. For the medium-risk batch (Batch B), the sampling rate was slightly adjusted to 22%, close to the baseline value.

[0046] refer to Figure 3 , Figure 3 is the probability cloud diagram of defect space distribution in this embodiment, Figure 3It can be seen that the minced meat area has the highest defect density, with the main defect being metal foreign matter. The steaming area has the second highest defect density, with the main defect being surface cracks. The forming area has the lowest defect density, with the main defect being shape distortion.

[0047] refer to Figure 4 , Figure 4 is a comparison chart of the process effects in this embodiment, Figure 4 It can be seen that increasing the magnetic separation frequency in the mincing zone reduced metal foreign matter defects by 62%. Implementing a temperature gradient in the cooking zone reduced surface crack defects by 45%. Adjusting the mold pressure in the forming zone reduced shape distortion defects by 38%. Example 2

[0048] This embodiment provides a method for intelligent and precise sampling of beef balls based on AI visual recognition. Its specific implementation method is the same as that of Example 1, except that defective beef ball samples are detected and identified using an optical microscope, a laser confocal microscope, and a near-infrared spectrometer to obtain the microscopic morphology and 3D surface topography of the defective beef ball samples. A graph neural network is constructed based on the obtained microscopic and 3D surface topography. The present invention is illustrated below with reference to the specific implementation methods of this embodiment.

[0049] In this example, a graph neural network was constructed based on the microscopic morphology and 3D surface morphology of defective beef meatball samples. The details are as follows: Step W1: Data Collection. Defective beef ball samples were taken from the production line and quickly frozen in liquid nitrogen to stabilize their microstructure. Simultaneously, the frozen meatball samples were sliced ​​into 5μm slices using a cryostat to preserve typical defect areas, such as crack tips and embedded foreign matter.

[0050] Furthermore, the cut beef ball sample slices were tested using an optical microscope, a laser confocal microscope and a near-infrared spectrometer to obtain a two-dimensional morphology map, a three-dimensional surface elevation map and a local near-infrared reflectance spectrum corresponding to each beef ball sample.

[0051] Step W2: Data processing. The two-dimensional topography image obtained in step W1 is converted into an eight-bit grayscale image. The box size sequence is set to 1 pixel, 2 pixels, 4 pixels, 8 pixels, 16 pixels, 32 pixels, and 64 pixels. The box counting method double logarithmic curve is constructed by obtaining the number of covered boxes at each scale in the grayscale image.

[0052] Furthermore, the coordinates of the three-dimensional surface elevation map obtained in step W1 are aligned with the two-dimensional topography map obtained in step W2. At the same time, the Z value in the aligned three-dimensional surface elevation map is compared with a preset Z threshold (which is set based on specific data and is not specifically described in this embodiment). Based on the comparison result, the foreign matter markers in the three-dimensional surface elevation map are determined. Specifically, When the actual Z value is greater than the preset Z threshold, the object corresponding to the actual Z value is a marked foreign object. Conversely, when the actual Z value is not greater than the preset Z threshold, the object corresponding to the actual Z value is not a marked foreign object.

[0053] In this embodiment, the X and Y values ​​corresponding to the marked foreign matter are obtained based on the determined Z value corresponding to the marked foreign matter, and the corresponding Fourier descriptor is determined through Fourier transform.

[0054] Furthermore, in the local near-infrared reflectance spectrum obtained in step W1, the 1730 cm -1 The characteristic peaks of the fat content were identified and correlated with the histogram overflow areas in the 2D topography to construct a fat content-reflectance calibration curve. The spectral data from the local near-infrared reflectance spectrum was also mapped to the 3D surface of the 3D surface elevation map and annotated with color scales.

[0055] In this example, the bound water state is determined by the reflectivity of the fat content-reflectivity calibration curve, namely, the reflectivity at 1450 nm. The free water content is determined by mapping the spectral data to the absorption peak intensity in the three-dimensional surface elevation map, namely, the near-infrared absorption peak intensity at 1200 nm. Specifically, a corresponding three-dimensional water distribution cloud map is constructed based on the bound water state and free water content.

[0056] Step W3: Feature extraction. That is, based on the fractal dimension of the box counting double logarithmic curve obtained in step W2, the corresponding fractal feature vector is obtained and used as the node feature of the graph neural network. Specifically, ; in: is the fractal eigenvector, is the fractal dimension of the current crack, is the rate of change of fractal dimension over time, is the largest fractal dimension in history, is the historical minimum fractal dimension.

[0057] In this embodiment, the morphological similarity between different defects is obtained based on the box counting double logarithmic curves corresponding to different defects, and the corresponding timing diagram network is constructed based on the obtained morphological similarity. Specifically, the formula for obtaining the morphological similarity is: ; in: is the shape similarity, is the base of natural logarithm, is the fractal dimension of the i-th crack node, is the fractal dimension of the j-th crack node, is the Gaussian kernel width.

[0058] Furthermore, the Fourier descriptor vector obtained in step W2 is normalized, and the normalized Fourier descriptor vector is used as the node feature of the graph neural network. Specifically, the corresponding normalized edge weights are determined based on the Fourier descriptor vectors corresponding to the different labeled foreign objects. And based on the obtained normalized edge weights, the corresponding harmonic similarity k-nearest neighbor topology graph is constructed. Specifically, the formula for obtaining the normalized edge weights is as follows: ; in: is the normalized edge weight between the i-th foreign node and the j-th foreign node, is the Fourier descriptor vector of the i-th foreign node, is the Fourier descriptor vector of the j-th foreign node, is the Euclidean distance.

[0059] Furthermore, the three-dimensional water distribution cloud map obtained in step W2 is divided into 1mm 3 A voxel grid is constructed, and each voxel contains the water content, gradient direction, and Raman spectral fingerprint.

[0060] Step W4: Construct a graph neural network. The voxel node graph divided in step W3 is used as the bottom layer of the graph neural network, the harmonic similarity k-nearest neighbor topology graph divided in step W3 is used as the middle layer of the graph neural network, and the time series graph network divided in step W3 is used as the top layer of the graph neural network.

[0061] Furthermore, the bottom graph, middle graph, and top graph of the graph neural network are connected in sequence, and the bottom graph and top graph of the graph neural network are connected according to the set edge size to construct the corresponding graph neural network. In this embodiment, the edge size between the bottom graph and the top graph in the graph neural network is specifically: ; in: is the moisture-thermodynamics comprehensive weight, is the moisture gradient vector, is the thermodynamic simulation similarity, is the maximum moisture gradient vector of the current batch.

[0062] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is limited by the accompanying embodiments and their equivalents.

Claims

1. A method for intelligent and precise sampling of beef balls based on AI visual recognition, characterized in that: Includes: S1: Construct defect data: Build a graph neural network based on the micromorphology and 3D surface morphology of the beef ball sample. Simultaneously, generate a virtual beef ball defect sample through multi-physics modeling, and verify the physical authenticity of the virtual beef ball defect sample through the graph neural network. S2: Dual-channel adversarial training: Identify the beef ball model through the set structure channel and material channel, and determine the loss function of the beef ball model, including: S2.1: Constructing a network architecture: Identifying the geometric deformation of the beef ball model through the structure channel, and identifying the texture synthesis of the beef ball model through the material channel, thereby obtaining a geometric deformation field and a fat distribution map of the beef ball model; S2.2: Determine a loss function: Based on the geometric deformation field and the fat distribution map, obtain the adversarial loss of the structural channel, the adversarial loss of the material channel, the perceptual loss value, the density constraint term, the thermal conductivity constraint term, and the elasticity constraint term; S2.3: Determine a composite loss function: Determine a total adversarial loss based on the adversarial loss of the structural channel and the adversarial loss of the material channel, determine a physical constraint loss based on the density constraint term, the thermal conductivity constraint term, and the elasticity constraint term, and simultaneously obtain a composite loss function based on the total adversarial loss, the perceptual loss value, and the physical constraint loss using a weighted algorithm; S3: Quality inspection: According to the composite loss function, the sampling rate of beef balls is set and a quality assessment system is constructed.

2. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 1 is characterized in that: Through multi-physics modeling, virtual beef ball defect samples are generated, including: S1.1: Thermal-mechanical coupling defect generation: Through the constructed material constitutive model, the elastic modulus change and anisotropic characteristics of the beef ball model during the cooking process are obtained, the real-time fracture stress of the beef ball model is determined, and the real-time fracture stress is compared with the critical fracture stress. Based on the comparison results, the crack state is determined, specifically: When the real-time fracture stress is greater than the critical fracture stress, the beef balls will crack; otherwise, the beef balls will not crack; S1.2: Fat flow modeling: Using the multiphase flow module and the set material properties, perform fat flow simulation on the beef ball model to determine the fat volume exudation rate of the beef ball model; S1.3: Obtaining defective samples: Based on the constructed material attribute library and the divided process areas, a defect simulation is performed on the beef ball model to obtain a beef ball model with defects.

3. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 2 is characterized in that: A three-stage test was performed on the beef ball sample to obtain the displacement and strain of the beef ball sample, and the first and second parameters of the Mooney-Rivlin model were determined to construct the material constitutive model, specifically: ； in: is the strain energy density, is the first parameter of the Mooney-Rivlin model, is the second parameter of the Mooney-Rivlin model, is the first deformation invariant, is the second deformation invariant, is the volume ratio, is the compressibility parameter.

4. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 2 is characterized in that: According to the process area and the production process flow of beef balls, the spatial weight distribution of beef balls is determined, and a 3D probability cloud map is obtained. The formula for obtaining the spatial weight distribution is specifically: ； in: For the The spatial weight coefficient of the region, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area, is the region index, For the The equivalent effective area of ​​the region, For beef balls The residence time in the area.

5. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 1 is characterized in that: Through the microscopic morphology and 3D surface morphology of beef meatball samples, a graph neural network was constructed, including: W1: Data acquisition: The beef meatball sample slices were examined using an optical microscope, a laser confocal microscope, and a near-infrared spectrometer to obtain a two-dimensional topography image, a three-dimensional surface elevation map, and a local near-infrared reflectance spectrum of the beef meatball sample; W2: Data processing: Convert the two-dimensional topography image into a grayscale image to construct a box counting double logarithmic curve, align the coordinates of the three-dimensional surface elevation map and the two-dimensional topography image, identify the marked foreign matter based on the Z value in the three-dimensional surface elevation map, and map the spectral data of the local near-infrared reflectance spectrum to the three-dimensional surface of the three-dimensional surface elevation map to construct a three-dimensional moisture distribution cloud map; W3: Feature extraction: Based on the box counting double logarithmic curve, the marked foreign matter and the three-dimensional moisture distribution cloud map, the time series graph network, the harmonic similarity k-nearest neighbor topology map and the voxel node map are obtained; W4: Construct a graph neural network: use the time series graph network, harmonic similarity k-nearest neighbor topology graph and voxel node graph as the top-level graph, middle-level graph and bottom-level graph of the graph neural network in turn, and connect the top-level graph, middle-level graph and bottom-level graph through edges.

6. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 5 is characterized in that: Based on the box counting double logarithmic curves corresponding to different defects, the morphological similarity between different defects is obtained, and a timing graph network is constructed; According to the coordinates of the marked foreign body, a Fourier descriptor vector is obtained through Fourier transform, and a normalized edge weight is determined, and a harmonic similarity k-nearest neighbor topological map is constructed based on the Fourier descriptor vector and the normalized edge weight; The three-dimensional moisture distribution cloud map is divided according to the set voxel grid size to construct a voxel node map.

7. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 5 is characterized in that: The edge size between the bottom graph and the top graph of the graph neural network is specifically: ； in: is the moisture-thermodynamics comprehensive weight, is the moisture gradient vector, is the thermodynamic simulation similarity, is the maximum moisture gradient vector of the current batch.

8. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 1 is characterized in that: According to the three-dimensional model of the beef ball model and the Marching Cubes algorithm, a binary mask of the beef ball model is obtained, and the binary mask is used as an input of a U-Net architecture model to obtain a geometric deformation field as an output; The geometric deformation field is used as the initial condition of the CycleGAN network model, and the muscle texture map of the beef ball model is used as the input of the CycleGAN network model to obtain the fat distribution map as the output.

9. The method for intelligent and precise sampling of beef balls based on AI visual recognition according to claim 1 is characterized in that: Set the sampling rate for beef balls and build a quality assessment system, including: S3.1: Determine the sampling rate: Adjust the basic sampling rate according to the composite loss function, specifically: ； in: is the adjusted sampling rate, As the basic sampling rate, is the sampling rate adjustment coefficient, is the composite loss function, is the loss benchmark threshold, is the loss fluctuation range coefficient; S3.2: Constructing a quality assessment system: comparing the composite loss function with a preset judgment threshold, and determining the quality grade of the beef meatballs based on the comparison result.

Images