Cold fresh beef online grading system based on computer vision

Abstract

The invention discloses a chilled beef online grading system based on computer vision, and relates to the technical field of food processing intellectualization, the system comprises an environment control unit (a camera obscura, an electronic scale, and a weighing precision + / -0.01 kg), an image acquisition unit (an LED light source, a camera with more than or equal to 20 million pixels, and an image processing unit), the frame rate is greater than or equal to 30fps), the image processing unit is used for high-grade part image processing and color analysis, the grading decision-making unit is used for weighting a decision-making tree model and grading according to different parts, and the label output unit is used for printing RFID labels within 300ms. According to the invention, multiple units and models are integrated, multi-dimensional detection of quality, pattern and meat color of chilled fresh beef is realized, the precision of an electronic scale is + / -0.01 kg, the image resolution is greater than or equal to 20 million pixels, an algorithm is combined to accurately quantify indexes, the accuracy rate of an XGBoost model is greater than or equal to 92%, different parts are differentially graded, the whole category is covered, the adaptability is improved by more than 30%, light interference is eliminated by a camera obscura nano-coating and a light source, and the detection precision is improved. The image processing ensures the data stability and the detection error rate lt; and 2%.

Description

Technical Field

[0001] This invention relates to the field of intelligent food processing technology, specifically to an online grading system for chilled beef based on computer vision. Background Technology

[0002] In the field of automated grading of chilled beef, existing technologies suffer from a significant deficiency in multi-dimensional detection capabilities. Traditional manual grading relies on visual inspection and experience, is significantly influenced by subjective factors, and struggles to meet the online detection efficiency requirements of slaughter lines at ≥30fps, making it difficult to implement the multi-index comprehensive grading required by the GB / T29392-2022 standard. Some automated systems employ single-dimensional grading (such as based solely on quality or meat color), failing to cover the differentiated grading needs of all types of beef, including premium cuts (sirloin / ribeye / chuck), tenderloin, and regular cuts. For example, the quantitative accuracy of marbling richness (ρ value) is insufficient, with detection error rates often exceeding 5%.

[0003] Meanwhile, existing testing environment control technologies have significant shortcomings: non-enclosed testing spaces are easily affected by ambient light interference, and fluctuations in the color temperature of conventional light sources (such as deviations from 5500K±200K) lead to deviations in flesh color detection; image acquisition resolution is generally lower than 20 million pixels, and when combined with traditional threshold segmentation algorithms (such as fixed thresholds), the accuracy of fat contour extraction is insufficient, making it difficult to accurately calculate the ρ value (e.g., unable to remove noisy contours with a perimeter <10 pixels). In addition, the existing grading system has weak coordination capabilities with the production line, lacking MES data traceability, process early warning (such as real-time response when the continuous C-grade yield exceeds 15%), and robotic arm sorting linkage functions. This results in grading data failing to feed back into production optimization, and overall production efficiency falling short of automation requirements by more than 40%.

[0004] In summary, existing technologies have significant technical bottlenecks in areas such as simultaneous detection of multi-dimensional indicators, stable control of ambient light, adaptive grading of all product categories, and intelligent linkage of production lines. There is an urgent need for an online grading system that integrates high-precision detection, adaptive algorithms, and intelligent control. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an online grading system for chilled beef based on computer vision. This invention integrates multiple units and models to achieve multi-dimensional detection of chilled beef quality, marbling, and color. The electronic scale accuracy is ±0.01kg, the image resolution is ≥20 million pixels, and the algorithm accurately quantifies the indicators. The XGBoost model has an accuracy of ≥92%, and it can perform differentiated grading for different parts, covering all categories and improving adaptability by more than 30%. The dark box nano-coating and light source eliminate light interference, and image processing ensures data stability, with a detection error rate of <2%.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In one aspect, an online grading system for chilled beef based on computer vision, the system comprising:

[0007] (1) Environmental control unit: The dark box (2) is integrated into the side of the conveyor belt (1) of the slaughtering production line. The inner wall is coated with a nano coating with an absorption rate of >99%. The bottom is equipped with an electronic scale (3) with a black platform. The weighing accuracy is ±0.01kg. It is used to obtain beef quality specification data M.

[0008] (2) Image acquisition unit: It consists of an LED light source (4) with a wavelength of 450–650nm and a digital camera (5) with a resolution of ≥20 million pixels. The light source color temperature is 5500±200K. The camera is connected to the image acquisition card (7) through the CameraLink interface. The frame rate is ≥30fps and the transmission delay is ≤50ms.

[0009] (3) Image processing unit, executes:

[0010] Processing of high-end areas: Cross-sectional images of the external spur, eyelids, and upper brain are processed sequentially.

[0011] ① Grayscale conversion and CLAHE contrast enhancement (ClipLimit=2.0, TileSize=8×8);

[0012] ②Otsu threshold segmentation to remove interference areas with backfat area >5%;

[0013] ③Canny operator contour extraction and 3×3 kernel morphological closing operation;

[0014] ④ Calculate the richness of marble patterns Where C r C represents the maximum fat contour circumference (mm). f This represents the total perimeter (mm) of the other contour points;

[0015] Color analysis: Convert the RGB values ​​of the muscle region to the HIS color space, and calculate:

[0016] ①Mean muscle color (R1, G1, B1);

[0017] ②Mean fat color (R2, G2, B2);

[0018] (4) Hierarchical decision-making unit, with built-in weighted decision tree model, executes:

[0019] High-grade parts: Simultaneously check the quality M, ρ value, flesh color (R1, G1, B1), and fat color (R2, G2, B2). If all four indicators are at the same level, the grade is determined; if any one of them is downgraded, the entire grade is downgraded.

[0020] Pork tenderloin: Graded by quality threshold: Grade S, M≥1.8kg; Grade A, 1.5kg≤M<1.8kg; Grade B, 1.3kg≤M<1.5kg; Grade C, M<1.3kg.

[0021] Regular cuts: graded by flesh color, those falling into the S / A / B range are considered superior, otherwise ordinary;

[0022] (5) Tag output unit (8): Prints an RFID tag containing the grade, location and time within 300ms after the decision is made.

[0023] Furthermore, the calculation of the ρ value includes:

[0024] (a) Divide the cross-sectional image into a 100×100 pixel grid and remove grids containing back fat;

[0025] (b) Fat contour extraction uses an adaptive threshold Thresh old = μ + 2σ, where μ is the mean gray value of the grid and σ is the standard deviation;

[0026] (c) Set the ρ-level threshold:

[0027] Grade S: ρ ≥ 2.5%; Grade A: 1.5% ≤ ρ < 2.5%;

[0028] Grade B: 0.5% ≤ ρ < 1.5% Grade C: ρ < 0.5%

[0029] Furthermore, the flesh color grading threshold is defined as:

[0030] Muscle color grade S / A / B: 178≤R1≤204, 10≤G1≤98, 15≤B1≤25;

[0031] Muscle color grade C: 150≤R1<178 or 204 <R1≤236,98<G1≤219,25<B1≤209;

[0032] Fat color grade S / A / B: 246≤R2≤254, 240≤G2≤251, 208≤B2≤244;

[0033] Fat color grade C: 218≤R2≤239, 198≤G2≤226, 104≤B2≤173; and the sampling points satisfy: per cm 2 The meat surface is sampled with ≥50 effective pixels, and the light source is a D65 standard light source.

[0034] Furthermore, the grading of high-end parts must simultaneously meet the following requirements:

[0035] Ribeye: Grade S requires M≥3.5kg and ρ≥2.5%, with flesh color grade S / A / B and fat color grade S / A / B;

[0036] For ribeye / croaker: Grade S requires M≥3.0kg and ρ≥2.5%, with meat color grade S / A / B and fat color grade S / A / B; if any indicator is downgraded, the lowest indicator shall be used for grading, and the measurement error of mass M shall be ≤±0.01kg.

[0037] Furthermore, the weighted decision tree model is constructed through the following steps:

[0038] (a) Constructing the sample dataset: Collect 2000 sets of beef cut sample data, each set containing:

[0039] Quality specification data M (unit: kg);

[0040] Cross-sectional digital image (resolution ≥ 20 million pixels);

[0041] Digital image of the epidermis (resolution ≥ 20 megapixels);

[0042] Manual grading labels (confirmed jointly by three certified graders in accordance with GB / T29392-2022 standard);

[0043] (b) Feature annotation:

[0044] The ρ value and color grade data were independently labeled by three graders holding national meat grading qualification certificates;

[0045] When the labeling results are inconsistent, a majority voting mechanism is used to determine the truth label;

[0046] (c) Feature extraction:

[0047] FeatureVector=[M,ρ,R1,G1,B1,R2,G2,B2];

[0048] in:

[0049] M represents the mass specification parameter (kg);

[0050] ρ represents the richness of marble patterns (%);

[0051] R1, G1, and B1 are the average values ​​of the RGB channels of muscle color;

[0052] R2, G2, and B2 are the average values ​​of the RGB channels for fat color;

[0053] (d) Model training:

[0054] ① The XGBoost algorithm is used as the basic architecture;

[0055] ②The key hyperparameters are set as follows:

[0056] Maximum tree depth: 7;

[0057] Learning rate: 0.1;

[0058] ③ Optimize the following auxiliary parameters using grid search:

[0059] Number of subtrees: range [50, 200];

[0060] Minimum number of leaf node samples: range [5, 20];

[0061] ④ Training termination condition: Five-fold cross-validation accuracy ≥ 92%.

[0062] Furthermore, the system communicates with the slaughter line's PLC control system to achieve:

[0063] Hierarchical data is uploaded to the MES system in real time to generate traceability QR codes;

[0064] When the C-grade product rate exceeds 15% for 10 consecutive batches, a process adjustment warning is triggered.

[0065] The robotic arm sorts and packs items by grade, with a sorting error rate of <0.5%.

[0066] On the other hand, an online grading method for chilled beef includes:

[0067] (1) Sorting by cut: The beef on the conveyor belt is sorted into high-grade (sirloin / ribeye / chuck), tenderloin, and regular (11 categories such as chili strips);

[0068] (2) Data Acquisition:

[0069] High-end section: Place the cross-section upwards on the dark box electronic scale to simultaneously acquire the M value and image;

[0070] Tenderloin / Regular cut: Harvested with the skin facing upwards;

[0071] (3) Feature extraction:

[0072] High-grade cuts: Calculate ρ value, mean RGB value of flesh color, and mean RGB value of fat color;

[0073] Standard areas: Only the RGB mean of flesh color is extracted;

[0074] (4) Hierarchical decision-making: Invoke the output level of the pre-trained weighted decision tree model;

[0075] (5) Tag binding: Print RFID tags within 300ms after the grading results are generated.

[0076] Furthermore, the calculation of the ρ value includes:

[0077] (a) Morphological closing operations are used to repair contour breakpoints, with a 3×3 circular kernel as the structural element;

[0078] (b) Remove noisy contours with a perimeter of less than 10 pixels.

[0079] Furthermore, analysis of flesh color and fat color needs to be performed:

[0080] Convert RGB values ​​to HIS color space: And calculate the mean values ​​of the H, I, and S channels.

[0081] Compared with existing technologies, this computer vision-based online grading system for chilled beef has the following advantages:

[0082] I. This invention integrates an environmental control unit, an image acquisition unit, and a weighted decision tree model to achieve multi-dimensional simultaneous detection of chilled beef quality specifications (M value), marbling richness (ρ value), and muscle and fat color. The electronic scale has a weighing accuracy of ±0.01kg, and the image acquisition resolution is ≥20 million pixels. Combined with image processing algorithms such as Otsu threshold segmentation and Canny operator contour extraction, it can accurately quantify beef grading indicators. The XGBoost weighted decision tree model, trained with 2000 sets of samples, achieves an accuracy of ≥92% through five-fold cross-validation, reducing subjective errors compared to traditional manual grading and ensuring the reliability of grading results and batch-to-batch consistency.

[0083] Second, this invention designs a differentiated grading strategy for different cuts of beef: high-grade cuts (sirloin, ribeye, and chuck) must simultaneously meet the four criteria of quality, ρ value, meat color, and fat color to be graded at the same level. If any one of these criteria is downgraded, the entire cut is downgraded. Tenderloin is graded according to a quality threshold, and regular cuts are graded according to meat color, achieving full category coverage from high-grade to regular cuts. This method is more than 30% more adaptable than a single-criteria grading method.

[0084] Third, this invention utilizes a darkroom structure with an inner wall coated with a nano-coating having an absorbance >99% in the environmental control unit, paired with a D65 standard light source (color temperature 5500±200K) to eliminate external light interference. In image processing, CLAHE contrast enhancement and adaptive thresholding (Thresh=μ+2σ) are used to extract fat contours, and a 3×3 kernel morphological closing operation is employed to repair contour breaks and eliminate noise with a perimeter <10 pixels. This design achieves a marble pattern detection error rate of <2%, and ensures that the sampling points for flesh and fat colors are accurate per cm. 2 ≥50 valid pixels ensure the stability of data collection and analysis, providing reliable data support for hierarchical decision-making.

[0085] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0086] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0087] Figure 1 A flowchart of a computer vision-based online grading system for chilled beef;

[0088] Figure 2 The flowchart shows a computer vision-based online grading method for chilled beef.

[0089] Figure 3 This is a schematic diagram of the device connections in an online grading system for chilled beef based on computer vision.

[0090] In the diagram: 1. Conveyor belt; 2. Dark box; 3. Electronic scale on a black platform; 4. LED light source; 5. Digital camera; 6. Computer; 7. Image acquisition card; 8. Tag output unit. Detailed Implementation

[0091] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0092] Example 1: An Example of an Online Grading System for High-Grade Beef Based on the XGBoost Model

[0093] 1.1 System Hardware Architecture Implementation Details

[0094] In this embodiment, the dark box 2 of the environmental control unit adopts a composite structure of aluminum alloy frame and light-absorbing coating. The inner layer is coated with nano-scale silicon carbide light-absorbing material, and the light absorption rate reaches 99.3% in the 450-650nm wavelength band as measured by a spectrometer. The bottom electronic scale 3 is an Ohaus AD4402 model, and the black stage is coated with polytetrafluoroethylene with a surface roughness Ra≤0.6μm. It communicates with the PLC in real time via Modbus RTU protocol and performs zero-point calibration every hour to ensure that the measurement error of the mass data M is controlled within ±0.008kg. Infrared beam sensors are installed on both sides of the dark box. After the beef is detected to be in place, the light-blocking curtain is automatically closed within 0.1 seconds. The light-blocking curtain is made of carbon fiber with a light transmittance of <0.01%.

[0095] The LED light source 4 of the image acquisition unit consists of a three-color LED matrix with 450nm blue light, 520nm green light, and 630nm red light. The color temperature is adjusted to 5500±180K via a constant voltage drive power supply, and the illumination uniformity is ≥95%. The digital camcorder 5 uses a Sony IMX411 sensor with a resolution of 6000×4000 pixels (24 million pixels), paired with a Computar M0814-MP2 lens, achieving a depth of field of 60mm-120mm at a distance of 280mm. The CameraLink interface uses Medium mode transmission. The image acquisition card 7 is a Euresys Picolo Plus2, with a measured frame rate of 32fps and a transmission latency of 45ms measured with an oscilloscope, meeting real-time processing requirements.

[0096] 1.2 Specific Implementation of Image Processing Algorithm

[0097] For cross-sectional images of the external spine, eyelids, and upper brain, the preprocessing workflow is as follows:

[0098] Grayscale conversion and CLAHE processing: The floating-point grayscale conversion formula Y = 0.299R + 0.587G + 0.114B is used to avoid the loss of precision in integer operations. During CLAHE processing, the image is divided into 8×8 pixel blocks, and ClipLimit is set to 1.8. By limiting the contrast enhancement, noise interference is reduced, and the entropy value of the processed image is increased by 1.9 times.

[0099] Accurate removal of back fat regions: Using Otsu thresholding combined with morphological opening operations (3×3 rhombus kernel), back fat regions are first identified by connected component labeling. When the area ratio exceeds 5%, a secondary segmentation based on the HSV color space (H∈[0, 15], S∈[0.2, 0.4], V∈[0.7, 1.0]) is adopted. The measured accuracy of back fat removal reaches 97.2%.

[0100] Contour extraction optimization: The high and low thresholds of the Canny operator were set to 80 and 20, respectively. Non-maximum suppression and dual threshold hysteresis tracking algorithms were adopted. Then, a 3×3 circular kernel was used to perform two morphological closing operations, which effectively repaired the broken parts of the fat contour. The contour integrity was measured to reach 98.5%.

[0101] The calculation of marble pattern richness ρ is improved as follows:

[0102] The cross-sectional image was divided into a grid of 120×120 pixels, and the backfat grid was identified using the α value (-10 to +10) of the Lab color space, achieving a rejection rate of 95.3%.

[0103] Fat contour extraction uses an adaptive threshold formula Threshold = μ + 1.9σ, and removes noisy contours with a perimeter < 11 pixels, reducing the false recognition rate by 2.3%.

[0104] A pixel-to-physical size mapping was established using a checkerboard calibration board (25mm×25mm), with a pixel equivalent of 0.045mm / pixel. C r and C f The perimeter calculation error is controlled within ±1.2%, and the ρ value grading threshold is strictly implemented according to the claims. When S-grade ρ≥2.5%, the corresponding actual fat area ratio is 3.3%-6.1%.

[0105] 1.3 Training and Execution of Hierarchical Decision-Making Model

[0106] The training process of the weighted decision tree model is as follows:

[0107] The sample dataset contains 2200 beef samples, including 600 samples of sirloin, 600 samples of ribeye, 600 samples of chuck, 300 samples of tenderloin, and 100 samples of other cuts. Each sample was independently labeled by three national first-level meat graders. When the labeling results were inconsistent, a majority vote plus expert arbitration mechanism was used (the arbitration experts have 15 years of grading experience). The labeling accuracy rate reached 99.3%.

[0108] The feature vector was kept as [M,ρ,R1,G1,B1,R2,G2,B2], and a new parameter for muscle brightness uniformity (calculated by local standard deviation) was added. However, feature importance analysis showed that its contribution rate was <2%, so it was not included in the model to maintain computational efficiency.

[0109] XGBoost model parameter optimization: The maximum tree depth was set to 7, the learning rate to 0.08, grid search was used to determine the number of subtrees to 140, the minimum number of leaf node samples to 8, and the accuracy of five-fold cross-validation reached 93.8%. L1 regularization coefficients of 0.02 and L2 regularization coefficients of 0.03 were added to reduce the risk of model overfitting.

[0110] Hierarchical decision-making execution logic:

[0111] For the outer rib to be classified as Grade S, the following conditions must be met: M≥3.5kg (actual measurement error ±0.007kg), ρ≥2.5%, and both muscle and fat color must be Grade S / A / B. A downgrade in any of these indicators results in an overall downgrade. Weight measurement employs a five-sample averaging method with a 30ms interval, effectively reducing vibration error.

[0112] For eyeballs / upper brain, the S-level requirement is M≥3.0kg, and other indicators are the same as above. When M is at the threshold boundary (e.g., 3.00kg), a tolerance range of ±0.01kg is enabled, that is, 2.99kg-3.01kg are all treated as S-level.

[0113] The tenderloin is graded strictly according to the quality threshold. Grade S has M≥1.8kg. When M=1.79kg, if ρ≥1.8%, it is exceptionally classified as Grade A, reflecting the flexible decision-making capability of the model.

[0114] 1.4 System Integration and Industrial Control

[0115] Communicating with Siemens S7-1200 PLC via the Profinet protocol to achieve:

[0116] The hierarchical data is uploaded to the MES system in real time. The data packet includes part code, level, quality and timestamp, with a transmission delay of ≤80ms and data integrity of 100%.

[0117] The traceability QR code uses the DataMatrix coding system, and the encoded content includes the slaughter date, the ear tag number of the cattle, and the grading station number. The QR code size is 15mm×15mm, and the decoding success rate is 99.8%.

[0118] When the C-grade product rate is >15% for 10 consecutive batches, the PLC sends a DI signal to the workshop control cabinet, triggering a buzzer alarm and illuminating a red warning light. At the same time, it pushes the warning information to the management system via OPCUA, with a response time of ≤20s.

[0119] The robotic arm sorting system uses the UR10e collaborative robot, and the control strategy is as follows:

[0120] Visual positioning is achieved using a binocular camera (1600×1200 resolution). A three-dimensional coordinate system is established based on Zhang's calibration method, with a positioning accuracy of ±0.3mm.

[0121] Adaptive gripping force adjustment: The pressure of the pneumatic gripper is automatically adjusted according to the weight M. The pressure is 0.4MPa when M<1kg and 0.6MPa when M≥1kg to avoid damaging the beef.

[0122] Double verification of sorting error: The sorting level is verified by a deep learning model (ResNet50) before picking and by a weighing sensor after placing the material in the bin. The actual sorting error rate is 0.25%, which is better than the target value in the claim.

[0123] Example 2: Example of an online grading method for chilled beef based on real-time vision

[0124] 2.1 Optimization of Part Sorting and Data Acquisition

[0125] In this embodiment, a linear laser scanning device (scanning frequency 150Hz) is deployed above the conveyor belt, which, in conjunction with a deep learning object detection model (Faster R-CNN), identifies beef cuts with an accuracy rate of 98.9%. High-grade cuts are grasped by a six-axis robotic arm and transferred to a darkroom electronic scale. The grasping trajectory utilizes cubic spline interpolation planning, improving motion smoothness by 30%. Tenderloin and regular cuts are processed via photoelectric sensors triggering skin image acquisition. The conveyor belt speed is controlled at 0.15m / s, and the camera exposure time is set to 1 / 1200s to ensure no motion blur in the images.

[0126] The dark box electronic scale 3 adopts a dual weighing sensor design with a weighing range of 0-50kg, an accuracy of ±0.005kg, and a sampling frequency of 100Hz. It uses a Kalman filter algorithm to fuse data and reduce interference from conveyor belt vibration. The LED light source 4 adds a diffuser plate, improving the uniformity of illumination to 97% and the color temperature stability to ±100K, ensuring consistent color acquisition.

[0127] 2.2 Feature Extraction and Real-time Computation

[0128] Image processing workflow for high-end areas:

[0129] Grayscale conversion and contrast enhancement: Grayscale conversion of the Y channel in the YUV color space is adopted, combined with adaptive CLAHE (window size 16×16, ClipLimit=2.2), which suppresses noise while preserving details. The average gradient magnitude of the processed image is increased by 2.1 times.

[0130] Intelligent filtering of back fat regions: First, suspected back fat regions are extracted through Otsu segmentation, and then semantic segmentation is performed using a pre-trained U-Net network. The back fat recognition accuracy reaches 98.1%, which is 1.4 percentage points higher than the traditional threshold method.

[0131] Accurate ρ value calculation: A 3×3 circular kernel is used for morphological closing operations to repair the contour, with 3 iterations. Noisy contours with a perimeter <13 pixels are then removed. A perspective transformation matrix is ​​established using a calibration board to correct image distortion, ensuring that the perimeter calculation error is ≤1.0%. A dynamic adjustment mechanism is added to the ρ-level threshold. When the ambient temperature is >15℃, the S-level threshold is temporarily adjusted to ρ≥2.3% to compensate for the effect of temperature on fat distribution.

[0132] Improved color analysis:

[0133] When converting RGB to HIS color space, for the special case of I=0 (R=G=B=0), set H=0 and S=0 to avoid calculation errors.

[0134] The grading thresholds for muscle and fat color are adjusted using a real-time temperature compensation formula:

[0135] Muscle color R1 threshold = standard value × (1 + 0.005 × (T - 25)), where T is the ambient temperature (°C).

[0136] Fat color R2 threshold = standard value × (1 - 0.003 × (T - 25))

[0137] When T = 10℃, the lower limit of muscle color R1 is automatically adjusted to 178 × (1 - 0.005 × 15) = 164.5, ensuring the consistency of grading standards under low temperature conditions.

[0138] 2.3 Hierarchical Decision Making and Label Output

[0139] The weighted decision tree model is deployed on an edge computing server (Intel i7-1185G7 + NVIDIA RTX A2000). Model quantization technology converts floating-point operations to 8-bit fixed-point operations, improving inference speed by 2.5 times and reducing single-image processing time to <30ms. Outlier detection is added to the decision logic: when ρ > 5.0% or M > 5.0kg, a manual review process is triggered to avoid misjudgments of extreme samples.

[0140] The RFID tag printing utilizes thermal transfer technology, employing a TSCTTP-345C printer with a 300dpi resolution. The tag material is a high-temperature resistant PET + coated paper composite layer, with a UV protective coating on the surface. The printed content includes grade, location, time, and temperature data (accuracy ±0.5℃), with a printing speed of 180mm / s. The time from decision completion to tag output is 250ms, meeting real-time requirements.

[0141] 2.4 System Integration and Quality Traceability

[0142] It communicates with the slaughter line PLC system (using Omron NJ series) via EtherCAT bus to achieve:

[0143] The hierarchical data is packaged and uploaded to the MES system in batches, with each batch containing 50 hierarchical records. The data packet verification uses the CRC-16 algorithm to ensure data integrity.

[0144] The traceability QR code contains a blockchain hash value and is linked to cattle breeding records, slaughter records, and processing procedures. The traceability information is uploaded to the blockchain in less than 1 second, meeting the requirements for food safety traceability.

[0145] The process adjustment early warning adds multi-dimensional judgment: when the C-grade product rate is >15% for 10 consecutive batches or the same part is classified as abnormal 5 times, a process adjustment work order is automatically generated and pushed to the equipment maintenance system. The work order includes information such as the abnormal part, classification index, and suggested adjustment parameters.

[0146] The robotic arm sorting system employs vision-force fusion control, monitoring contact force in real time during the grasping process. When the force exceeds a threshold (20N), an anti-collision mechanism is automatically triggered, stopping the operation and triggering an alarm. The sorting bins are equipped with RFID tag identification areas to ensure that the grade and category of each bin match the tag. Abnormalities are detected through audible and visual alarms. The actual measured mean time between failures (MTBF) of the sorting system reaches 1200 hours.

[0147]

[0148] Example 1 achieves highly stable grading results through precise hardware calibration and model parameter optimization, making it suitable for medium-sized slaughtering production lines. Example 2 introduces deep learning vision algorithms and real-time compensation mechanisms, improving grading accuracy and environmental adaptability, making it more suitable for large-scale intelligent production lines.

[0149] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A computer vision-based online grading system for chilled fresh beef, characterized in that, The system includes: (1) Environmental control unit: The dark box (2) is integrated into the side of the conveyor belt (1) of the slaughtering production line. The inner wall is coated with a nano coating with an absorption rate of >99%. The bottom is equipped with an electronic scale (3) with a black platform. The weighing accuracy is ±0.01kg. It is used to obtain beef quality specification data M. (2) Image acquisition unit: It consists of an LED light source (4) with a wavelength of 450–650nm and a digital camera (5) with a resolution of ≥20 million pixels. The light source color temperature is 5500±200K. The camera is connected to the image acquisition card (7) through the CameraLink interface. The frame rate is ≥30fps and the transmission delay is ≤50ms. (3) Image processing unit, executes: Processing of high-end areas: Cross-sectional images of the external spur, eyelids, and upper brain are processed sequentially. ① Grayscale conversion and CLAHE contrast enhancement (ClipLimit=2.0, TileSize=8×8); ②Otsu threshold segmentation to remove interference areas with backfat area >5%; ③Canny operator contour extraction and 3×3 kernel morphological closing operation; ④ Calculate the richness of marble patterns Where C r C represents the maximum fat contour circumference (mm). f This represents the total perimeter (mm) of the other contour points; Color analysis: Convert the RGB values ​​of the muscle region to the HIS color space, and calculate: ①Mean muscle color (R1, G1, B1); ②Mean fat color (R2, G2, B2); (4) Hierarchical decision-making unit, with built-in weighted decision tree model, executes: High-grade parts: Simultaneously check the quality M, ρ value, flesh color (R1, G1, B1), and fat color (R2, G2, B2). If all four indicators are at the same level, the grade is determined; if any one of them is downgraded, the entire grade is downgraded. Pork tenderloin: Graded by quality threshold: Grade S, M≥1.8kg; Grade A, 1.5kg≤M<1.8kg; Grade B, 1.3kg≤M<1.5kg; Grade C, M<1.3kg. Regular cuts: graded by flesh color, those falling into the S / A / B range are considered superior, otherwise ordinary; (5) Tag output unit (8): Prints an RFID tag containing the grade, location and time within 300ms after the decision is made.

2. The computer vision-based online grading system for chilled beef according to claim 1, characterized in that, The calculation of the ρ value includes: (a) Divide the cross-sectional image into a 100×100 pixel grid and remove grids containing back fat; (b) Fat contour extraction uses an adaptive threshold Thresh old = μ + 2σ, where μ is the mean gray value of the grid and σ is the standard deviation; (c) Set the ρ-level threshold: Grade S: ρ ≥ 2.5%; Grade A: 1.5% ≤ ρ < 2.5%; Grade B: 0.5% ≤ ρ < 1.5% Grade C: ρ < 0.5% 3. The computer vision-based online grading system for chilled beef according to claim 1, characterized in that, The flesh color grading threshold is defined as follows: Muscle color grade S / A / B: 178≤R1≤204, 10≤G1≤98, 15≤B1≤25; Muscle color grade C: 150≤R1<178 or 204 <R1≤236,98<G1≤219,25<B1≤209; Fat color grade S / A / B: 246≤R2≤254, 240≤G2≤251, 208≤B2≤244; Fat color grade C: 218≤R2≤239, 198≤G2≤226, 104≤B2≤173; and the sampling points meet the following requirements: ≥50 effective pixels are collected per cm2 of meat surface, and the light source is a D65 standard light source.

4. The computer vision-based online grading system for chilled beef according to claim 1, characterized in that, High-end parts must meet the following criteria for grading: Ribeye: Grade S requires M≥3.5kg and ρ≥2.5%, with flesh color grade S / A / B and fat color grade S / A / B; For ribeye / croaker: Grade S requires M≥3.0kg and ρ≥2.5%, with meat color grade S / A / B and fat color grade S / A / B; if any indicator is downgraded, the lowest indicator shall be used for grading, and the measurement error of mass M shall be ≤±0.01kg.

5. The computer vision-based online grading system for chilled beef according to claim 1, characterized in that, The weighted decision tree model is constructed through the following steps: (a) Constructing the sample dataset: Collect 2000 sets of beef cut sample data, each set containing: Quality specification data M (unit: kg); Cross-sectional digital image (resolution ≥ 20 million pixels); Digital image of the epidermis (resolution ≥ 20 megapixels); Manual grading labels (confirmed jointly by three certified graders in accordance with GB / T29392-2022 standard); (b) Feature annotation: The ρ value and color grade data were independently labeled by three graders holding national meat grading qualification certificates; When the labeling results are inconsistent, a majority voting mechanism is used to determine the truth label; (c) Feature extraction: FeatureVector=[M,ρ,R1,G1,B1,R2,G2,B2]; in: M represents the mass specification parameter (kg); ρ represents the richness of marble patterns (%); R1, G1, and B1 are the average values ​​of the RGB channels of muscle color; R2, G2, and B2 are the average values ​​of the RGB channels for fat color; (d) Model training: ① The XGBoost algorithm is used as the basic architecture; ②The key hyperparameters are set as follows: Maximum tree depth: 7; Learning rate: 0.1; ③ Optimize the following auxiliary parameters using grid search: Number of subtrees: range [50, 200]; Minimum number of leaf node samples: range [5, 20]; ④ Training termination condition: Five-fold cross-validation accuracy ≥ 92%.

6. The computer vision-based online grading system for chilled beef according to claim 1, characterized in that, The system communicates with the slaughter line's PLC control system to achieve: Hierarchical data is uploaded to the MES system in real time to generate traceability QR codes; When the C-grade product rate exceeds 15% for 10 consecutive batches, a process adjustment warning is triggered. The robotic arm sorts and packs items by grade, with a sorting error rate of <0.5%.

7. A computer vision-based online grading method for chilled beef, applicable to the computer vision-based online grading system for chilled beef described in claims 1-6, characterized in that, The method includes: (1) Sorting by cut: The beef on the conveyor belt is sorted into high-grade (sirloin / ribeye / chuck), tenderloin, and regular (11 categories such as chili strips); (2) Data Acquisition: High-end section: Place the cross-section upwards on the dark box electronic scale to simultaneously acquire the M value and image; Tenderloin / Regular cut: Harvested with the skin facing upwards; (3) Feature extraction: High-grade cuts: Calculate ρ value, mean RGB value of flesh color, and mean RGB value of fat color; Standard areas: Only the RGB mean of flesh color is extracted; (4) Hierarchical decision-making: Invoke the output level of the pre-trained weighted decision tree model; (5) Tag binding: Print RFID tags within 300ms after the grading results are generated.

8. The online grading method for chilled beef based on computer vision according to claim 7, characterized in that, The calculation of the ρ value includes: (a) Morphological closing operations are used to repair contour breakpoints, with a 3×3 circular kernel as the structural element; (b) Remove noisy contours with a perimeter of less than 10 pixels.

9. The online grading method for chilled beef based on computer vision according to claim 7, characterized in that, The analysis of flesh color and fat color needs to be performed as follows: Convert RGB values ​​to HIS color space: And calculate the mean values ​​of the H, I, and S channels.

Images