Macrobrachium rosenbergii quality grading system based on machine vision

Abstract

The invention provides a macrobrachium rosenbergii quality grading system based on machine vision. The macrobrachium rosenbergii quality grading system based on machine vision comprises a prawn body placing platform used for placing macrobrachium rosenbergii to be tested; the three-dimensional acquisition module comprises a structured light projection device, a binocular stereo camera and a flight time depth sensor; and the three-dimensional reconstruction module is used for constructing the data of the three-dimensional acquisition module into a complete three-dimensional point cloud model through a data fusion algorithm and a point cloud registration algorithm. According to the macrobrachium rosenbergii quality grading system based on machine vision, the overall scheme has robustness and expandability, the system is suitable for various large-scale farms and grading scenes, and the intelligent level of macrobrachium rosenbergii quality control is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aquaculture, in particular to a Macrobrachium rosenbergii quality grading system based on machine vision. BACKGROUND

[0002] Macrobrachium rosenbergii is widely favored by the market for its delicious meat and rich protein. Due to its strong adaptability and low breeding difficulty, Macrobrachium rosenbergii has formed a large-scale breeding in China. With the improvement of breeding technology, the Macrobrachium rosenbergii industry has gradually entered the quality and efficiency stage, and more large-sized and high-quality Macrobrachium rosenbergii products have been produced, greatly improving its economic benefits. With the increasing concern of consumers for food safety and quality, the demand for quality grading of Macrobrachium rosenbergii is increasing. The traditional quality grading method mainly relies on manual means, and the meat quality, appearance, size and other aspects of the shrimp body are judged manually to grade. However, manual grading not only has a large workload, but also has low efficiency and is easily affected by human factors, resulting in unstable grading results.

[0003] At present, although there are some automatic grading technologies in the market, most of the systems still have obvious defects. First, the existing automatic grading system has weak judgment ability for the living state, and cannot effectively handle the grading demand of living Macrobrachium rosenbergii. Secondly, many existing machine vision systems cannot achieve high-precision grading standards in the processing process, and may cause inconsistent grading results due to inaccurate image acquisition or changes in environmental factors. In addition, the existing technology often fails to provide a comprehensive and standardized grading model, and cannot finely grade Macrobrachium rosenbergii of different specifications and different qualities according to market demand. In summary, the existing technology still has deficiencies in grading accuracy, speed and environmental adaptability, and innovative intelligent systems are needed to improve the grading efficiency and accuracy. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a Macrobrachium rosenbergii quality grading system based on machine vision, which solves the problems of lack of unified and objective grading standards for Macrobrachium rosenbergii and inability to non-destructively and quickly grade.

[0005] To achieve the above purpose, the present application is implemented by the following technical scheme: a Macrobrachium rosenbergii quality grading system based on machine vision, comprising:

[0006] a shrimp body placing platform for placing Macrobrachium rosenbergii to be tested;

[0007] a three-dimensional acquisition module, the three-dimensional acquisition module comprising a structured light projection device, a binocular stereo camera and a time-of-flight depth sensor;

[0008] a three-dimensional reconstruction module, configured to construct the three-dimensional acquisition module data into a complete three-dimensional point cloud model through a data fusion algorithm and a point cloud registration algorithm;

[0009] a quality mapping and error self-correction module, configured to output a shrimp body quality prediction value and a prediction uncertainty through an end-to-end deep regression network, and dynamically perform error self-correction according to the prediction uncertainty;

[0010] a grade determination module, configured to objectively grade the Macrobrachium rosenbergii according to the shrimp body quality prediction value and a pre-set weight threshold, and output grade information.

[0011] Preferably, the shrimp body placement platform has a bottom integrated position detection sensor, which is configured to confirm whether the shrimp body is located in the best field of view of the three-dimensional acquisition module.

[0012] Preferably, the structured light projection device comprises a dual-wavelength laser structured light module, which projects a coded light pattern on the shrimp body, the wavelength of the coded light pattern being switched between 660 nm and 850 nm, and the projection resolution being ≤0.1 mm.

[0013] Preferably, the binocular stereo camera and the time-of-flight depth sensor work simultaneously and synchronously with an error Δt≤1 ms, the binocular camera has a baseline distance B∈[100 mm, 200 mm], a focal length f≥8 mm, and a resolution ≥1920×1080, and the time-of-flight depth sensor has a ranging range of 0.2 m–2 m and a depth resolution ≤1 mm.

[0014] Preferably, the three-dimensional reconstruction module comprises the following three steps:

[0015] 5.1 preliminary registration: rigid registration is performed using the output data of the binocular stereo camera and the time-of-flight depth sensor;

[0016] 5.2 weighted fusion: each point cloud source data is weighted and fused based on color consistency and surface normal similarity;

[0017] 5.3 surface reconstruction: a Poisson reconstruction method is used to generate a smooth three-dimensional model, and the reconstruction error of the smooth three-dimensional model is <0.5 mm.

[0018] Preferably, the end-to-end deep regression network adopts a multilayer perceptron structure, and the model formula is:

[0019] L=α·MSE(y p ,y t )+β·Var(σ p ),

[0020] where L denotes the total loss of the network, y p denotes the predicted body weight of the shrimp by the network, y t denotes the actual weight measured by a precision electronic balance, denotes the mean square error of all training samples, σ p denotes the uncertainty estimation of each predicted value output by the network, denotes the sample variance of the uncertainty estimation, and α and β are weight coefficients for balancing the quality error term and the uncertainty constraint term, which are set by cross-validation.

[0021] Preferably, the error self-correction includes adjustment of structure light projection parameters, adjustment of binocular stereo camera acquisition parameters, and adjustment of time-of-flight depth sensor parameters.

[0022] Preferably, the grade determination module comprises:

[0023] an integrity analysis submodule: based on the shape features and edge integrity of the shrimp body, an integrity score I is calculated, ranging from 0 to 1;

[0024] a color and luster analysis submodule: a color histogram in HSV space is extracted, and a color and luster score C is calculated, ranging from 0 to 1;

[0025] a quality score submodule: the predicted quality value is mapped to a normalized score M, ranging from 0 to 1;

[0026] The grade determination module determination model formula is:

[0027] S = ω1I + ω2C + ω3M

[0028] where ω1 is the integrity score weight coefficient, ω2 is the color and luster score weight coefficient, and ω3 is the normalized score weight coefficient, and ω1 + ω2 + ω3 = 1, and S is the final comprehensive score.

[0029] Preferably, the grade information comprises:

[0030] superior grade: predicted quality > 100g, comprehensive score S ≥ 0.90;

[0031] premium grade: predicted quality 50g < M ≤ 100g, comprehensive score 0.80 ≤ S < 0.90;

[0032] good grade: predicted quality 20g < M ≤ 50g, comprehensive score 0.70 ≤ S < 0.80;

[0033] qualified grade: predicted quality ≤ 20g, comprehensive score 0.60 ≤ S < 0.70;

[0034] unqualified: comprehensive score S < 0.60.

[0035] The application provides a machine vision-based macrobrachium rosenbergii quality grading system.

[0036] The machine vision-based macrobrachium rosenbergii quality grading system realizes accurate three-dimensional volume reconstruction and online quality estimation of the macrobrachium rosenbergii by fusing structured light, binocular stereo vision and time-of-flight depth sensing technology, and overcomes the deficiency that the existing two-dimensional projection method is susceptible to posture and occlusion; the designed end-to-end depth regression network can not only give a high-precision shrimp body quality prediction value, but also output a prediction uncertainty, and based on this, dynamically corrects the acquisition parameters to ensure the measurement consistency and stability in long-time continuous operation. The weighted fusion of multi-source data and Poisson surface reconstruction ensure that the reconstruction error is less than 0.5 mm, significantly improves the grading accuracy, and makes the system have a unified and objective grading standard, and can realize high-precision grading under the conditions of nondestructive and fast (capacity not less than 600 per hour).

[0037] In addition, the system relies on a programmable rotating table and a position detection sensor to realize automatic positioning and multi-view acquisition of the shrimp body, and eliminates the tediousness of manual resetting and calibration; the three-dimensional comprehensive scoring model of integrity, color and quality makes the grading results more comprehensive, transparent and traceable; the overall scheme has robustness and scalability, and is suitable for various large-scale breeding farms and grading scenes, greatly improving the intelligent level of the quality control of the macrobrachium rosenbergii. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is a flowchart for realizing the application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0040] Embodiment one

[0041] As shown in Figure 1 The application provides a machine vision-based macrobrachium rosenbergii quality grading system, which comprises a shrimp body placing platform for placing the macrobrachium rosenbergii to be measured. The position detection sensor is integrated at the bottom of the shrimp body placing platform support plate, and is used to confirm whether the shrimp body is located within the best field of view range of the three-dimensional acquisition module.

[0042] The three-dimensional acquisition module comprises a structured light projection device, a binocular stereo camera and a time-of-flight depth sensor.

[0043] The structured light projection device comprises a dual-wavelength laser structured light module, which projects a coded light pattern on the shrimp body, the wavelength of the coded light pattern is switched between 660 nm and 850 nm, and the projection resolution is ≤0.1 mm.

[0044] The binocular stereo camera and the time-of-flight depth sensor work simultaneously and the synchronization error Δt≤1 ms, the binocular camera baseline distance B∈[100 mm, 200 mm], the focal length f≥8 mm, the resolution ≥1920×1080, the time-of-flight depth sensor ranging range 0.2 m-2 m, the depth resolution ≤1 mm.

[0045] The specific implementation is as follows:

[0046] 1. The structured light projection device

[0047] The dual-wavelength laser structured light module is built-in 660 nm and 850 nm laser diodes, and each output power is 50 mW.

[0048] The coded light pattern adopts 16×16 black and white checkerboard alternately projection, the switching frequency is 50 Hz, the wavelength switching time sequence is controlled by FPGA, and the switching delay is ≤2 μs.

[0049] The projection resolution is ≤0.1 mm, the projection area size is 200 mm×200 mm, the working distance is fixed at 300 mm, and a single shrimp body with a length of ≤180 mm can be covered.

[0050] In the indoor light environment, the 640 lx ambient light is effectively suppressed by a narrow-band filter (center wavelength 660 nm / 850 nm, bandwidth 10 nm), and the projection signal-to-noise ratio is guaranteed to be >35 dB.

[0051] 2. Binocular stereo camera

[0052] Two Basler acA1920-150um industrial cameras are used, the pixel resolution is 1920×1080, and the pixel size is 3.45 μm.

[0053] The lens focal length f is selected to be 12 mm (F1.8), and the field of view can reach 150 mm×85 mm.

[0054] The baseline distance B is set to 120 mm, and it is verified by experiments that the depth resolution of 0.25 mm can be realized under the working distance of 300 mm.

[0055] The camera frame rate is 30 fps, the single frame exposure time is 20 ms, the synchronization trigger signal is output by the NI-6602 data acquisition card, and the trigger jitter Δt<1 ms.

[0056] Camera gain is fixed at 1.0, white balance is off, gray noise is less than or equal to 0.4 DN, and dynamic range is 70 dB.

[0057] 3. Time-of-flight depth sensor

[0058] Select Infineon Real3 TM IMX556 ToF sensor, ranging from 0.2 m to 2 m, depth resolution less than or equal to 1 mm.

[0059] Emission wavelength 940 nm, emission power 20 mW, integration time adjustable 10 ms-33 ms.

[0060] It has an ambient light compensation function and can work stably under 100 lx-10000 lx illumination.

[0061] ToF and binocular camera share the same trigger source, synchronization accuracy Δt<1 ms, depth data is output in real time through USB3.0, and average delay is less than 5 ms.

[0062] 4. Module integration and workflow

[0063] The structured light projector, binocular camera and ToF sensor are placed in a dark box, the optical centers are parallel to the placement platform plane, and the working distance is set to 300 mm.

[0064] Single scanning process:

[0065] FPGA sends TTL trigger signal, structured light module starts chessboard projection.

[0066] Binocular camera and ToF collect one frame of image at the same time, and original point cloud and depth map are acquired synchronously.

[0067] FPGA switches to another wavelength coding pattern at a frequency of 50 Hz, and completes two waveband alternating projection collection.

[0068] Each shrimp body collects 8 frames, and the total time consumption is approximately 320 ms.

[0069] After the collection is completed, all original data is transmitted to the three-dimensional reconstruction module through Ethernet for point cloud fusion and registration.

[0070] Three-dimensional reconstruction module, used for constructing complete three-dimensional point cloud model by three-dimensional acquisition module data through data fusion algorithm and point cloud registration algorithm.

[0071] The three-dimensional reconstruction module runs including the following three steps:

[0072] 5.1 Preliminary registration: rigid registration is performed using binocular stereo camera and time-of-flight depth sensor output data.

[0073] 5.2 Weighted Fusion: Weighted fusion of cloud source data for each point based on color consistency and surface normal similarity.

[0074] 5.3 Surface Reconstruction: A smooth 3D model is generated using the Poisson reconstruction method, with a reconstruction error of <0.5mm.

[0075] The specific implementation method is as follows:

[0076] 5.1 Preliminary Registration

[0077] Point cloud preprocessing:

[0078] The binocular stereo camera and the ToF sensor each output approximately 200,000 points.

[0079] By using a 1.0mm voxel grid for downsampling, the two sets of point clouds were compressed to approximately 55,000 points each, improving the speed of subsequent calculations.

[0080] Rigid registration:

[0081] The registration algorithm uses the Iterative Closest Point (ICP) library function provided by Open3D.

[0082] The maximum number of iterations is set to 50.

[0083] The sensor has been calibrated for external parameters during installation, with an initial rotational error of less than 0.5° and a translational error of no more than 1mm.

[0084] Effects and time consumption:

[0085] The average residual between the two sets of point clouds after registration was 0.85 mm.

[0086] Single-frame registration takes approximately 120ms, with a registration success rate maintained above 99.5%.

[0087] 5.2 Weighted Fusion

[0088] Feature extraction:

[0089] Calculate the surface normal for each point using a radius of 5mm as the neighborhood.

[0090] Read the RGB information saved in the binocular point cloud and convert it to HSV space for use as color features.

[0091] Weighting principles:

[0092] When the hue difference between two source point clouds at the same physical location is less than 30 degrees and the included normal angle is less than 10 degrees, they are assigned a high weight.

[0093] The greater the color difference or normal difference, the lower the weight; if the color difference exceeds 30 degrees or the angle between the normals exceeds 30 degrees, the weight of the source point is reduced to zero.

[0094] Fusion results:

[0095] The final merged point cloud has a scale of approximately 120,000 points.

[0096] The point cloud color statistical difference is less than 5%, and the average angle of the normal is 7–8 degrees.

[0097] The fusion process takes approximately 100ms.

[0098] 5.3 Surface Reconstruction

[0099] Poisson reconstruction parameters:

[0100] The octree depth is set to 9.

[0101] Each leaf node retains at least 4 sampling points.

[0102] Memory usage is approximately 280MB, and rebuild time is 250–260ms.

[0103] Reconstruction results and error assessment:

[0104] The output mesh contains approximately 200,000 triangular faces with an average side length of 1.2 mm.

[0105] Randomly select 10,000 vertices from the mesh model and calculate the distance between them and the corresponding points in the fused point cloud:

[0106] The average error is 0.42 mm.

[0107] The maximum error is 0.87mm, which is far below the 0.5mm error limit requirement.

[0108] Overall performance:

[0109] The total time from initial data acquisition to completing a smooth 3D mesh was approximately 470ms.

[0110] It can support 7,000–7,500 frames of 3D reconstruction per hour, providing high-precision input data for subsequent quality prediction and classification.

[0111] The quality mapping and error self-correction module is used to process the 3D point cloud model through an end-to-end depth regression network and output the predicted value of shrimp body quality and the prediction uncertainty. The quality mapping and error self-correction module dynamically performs error self-correction based on the prediction uncertainty.

[0112] The end-to-end deep regression network uses a multilayer perceptron structure, and the model formula is as follows:

[0113] L=α·MSE(y p ,y t )+β·Var(σ p ),

[0114] Where L represents the total network loss, yp y represents the shrimp body mass predicted by the network. t This indicates the actual mass measured using a precision electronic balance. σ represents the mean squared error of all training samples. p This represents the uncertainty estimate for each predicted value in the network output. The sample variance represents the uncertainty estimate. α and β are weighting coefficients used to balance the quality error term and the uncertainty constraint term, and are set through cross-validation.

[0115] Error self-correction includes adjusting the structured light projection parameters, the binocular stereo camera acquisition parameters, and the time-of-flight depth sensor parameters.

[0116] The specific implementation method is as follows:

[0117] 1. Hardware integration and calibration

[0118] 1.1 Sensor Layout

[0119] The structured light projector, binocular stereo camera, and time-of-flight depth sensor are fixed in a triangular arrangement on the top plate of the dark box, with the optical center 300mm from the shrimp placement surface. The alignment error of the three channels, as measured by a micrometer, does not exceed 0.3mm.

[0120] 1.2 Core Components

[0121] Laser structured light: dual wavelengths 660nm / 850nm, 50mW each.

[0122] Binocular camera: 1920×1080 resolution, 120mm baseline, 12mm focal length.

[0123] ToF: Operating wavelength 940nm, ranging 0.2–2m, depth resolution 1mm.

[0124] 1.3 External parameters at the factory

[0125] The three sensors are aligned in one go using a calibration board, with an initial rotation error of less than 0.5° and a translation error of no more than 1mm; the external parameters are written into the EEPROM, and subsequent adjustments only require incremental fine-tuning.

[0126] 2. Data Acquisition and 3D Reconstruction

[0127] 2.1 Single-frame scanning

[0128] Each shrimp was captured in 8 frames (two wavelengths alternating, 4 frames per wavelength), with a total time of approximately 320ms.

[0129] 2.2 Point Cloud Processing

[0130] Each of the two original data points was approximately 200,000 points, which were reduced to 55,000 points after 1mm voxel downsampling.

[0131] ICP rigid registration 120ms, mean square residual 0.85mm.

[0132] After weighted fusion based on color and normal consistency, a high-density cloud of 120,000 points is obtained.

[0133] Poisson surface reconstruction depth 9, time 250ms; mesh of approximately 200,000 triangular faces, average side length 1.2mm.

[0134] The measured average error of the reconstruction accuracy is 0.42 mm, with a maximum of 0.87 mm, which meets the target of <0.5 mm.

[0135] 3. Weight prediction and error self-correction

[0136] 3.1 Standard Network

[0137] The default operating system is a 4-layer MLP (512→256→128→64 nodes). Training was performed on 5000 high-precision samples over 50 epochs, with a single RTX 4090 graphics card taking 7 hours. The validation set had an absolute error of 0.85g and a mean uncertainty of 0.70g.

[0138] 3.2 Scenario-based model

[0139] The high-speed production line version is compressed to 3 layers (256→128→64), with an error of 2.3g.

[0140] The outdoor adaptive version is expanded to 5 layers and incorporates light intensity characteristics, with an error of 1.8g.

[0141] The low-cost cooperative version has only 2 layers (128→64), an error of 3.7g, and can run smoothly on a GTX1650.

[0142] 3.3 Self-calibration logic

[0143] Uncertainty of single-unit prediction monitoring: laboratory threshold 0.80g, high-speed production line 1.50g, outdoor adaptive 1.0g (dynamically adjusted according to light intensity).

[0144] Adjust the laser power, then the ToF emission or integration time, and finally the camera exposure or gain.

[0145] The correction process takes 80–150 ms to complete and can typically reduce uncertainty by 25–35%.

[0146] 4. Objective grading

[0147] 4.1 Three scores

[0148] Integrity is based on the continuity of severed limbs, shell cracks, and trunk; the fewer the defects, the higher the score.

[0149] Color is matched with the standard color chart using the HSV histogram; the smaller the color deviation, the higher the score.

[0150] The mass fraction is linearly mapped to four levels from 0 to 120g (≤20g, 21–50g, 51–100g, >100g).

[0151] 4.2 Weights and Levels

[0152] Fixed weights are 0.35 (completeness), 0.25 (color), and 0.40 (quality); a comprehensive score ≥ 0.90 is judged as excellent, 0.80–0.89 as good, 0.70–0.79 as good, 0.60–0.69 as acceptable, and a score below 0.60 or any single item below 0.50 is judged as unacceptable.

[0153] 4.3 Singleton Output Example

[0154] A shrimp has a body integrity score of 0.92, a color score of 0.88, a mass fraction of 0.85, a comprehensive score of 0.88, and a weight of 83g—the system immediately returns a grade of "Superior (B)".

[0155] 5. Operating data under five typical working conditions

[0156] Laboratory high precision: average 0.8s / piece, error ±1g, daily processing capacity 4000 pieces.

[0157] High-speed production line: average 0.6s / piece, error ±3g, processing capacity 1100–1200 pieces per hour.

[0158] Outdoor adaptive: average 0.9s / piece, error ±2g, processing 650 pieces per hour.

[0159] Maritime cold chain inspection: 12,000 items were scanned in 3 hours after packing, with an error of 1.9g, and a PDF report was automatically generated before unloading.

[0160] Low-cost cooperative: average 0.9s / unit, error ±4g, compatible with GTX1650 and low-end hardware with single / dual-lens + structured light.

[0161] 6. Operation and Upgrade

[0162] Online OTA: The terminal automatically receives the new weighting file from headquarters every month, and the upgrade takes 30 seconds.

[0163] Log retention: Inference logs and hardware parameter snapshots are compressed daily and retained for 180 days.

[0164] Optical cleaning: Wipe the windows with alcohol swabs every shift; if the structured light power decreases by 15%, the laser component should be replaced.

[0165] Safety protection: If the casing is opened, the internal temperature exceeds 75°C, or the laser drive exceeds the current limit, the system will immediately cut off the power and record the error code.

[0166] The grading module is used to objectively grade giant freshwater prawns based on predicted shrimp body weight and pre-set weight thresholds, and output grading information.

[0167] The rating module includes:

[0168] Integrity Analysis Submodule: Based on the shrimp body shape features and edge integrity scores, calculate the integrity score I, ranging from [0,1].

[0169] Color Analysis Submodule: Extracts the color histogram in HSV space and calculates the color score C, ranging from [0,1].

[0170] Quality score submodule: maps predicted quality values ​​to normalized scores M, ranging from [0,1].

[0171] The formula for the grade determination module is as follows:

[0172] S=ω1I+ω2C+ω3M

[0173] Where: ω1 is the integrity score weight coefficient, ω2 is the color score weight coefficient, ω3 is the normalized score weight coefficient, and ω1+ω2+ω3=1, and S is the final comprehensive score.

[0174] The rating information includes;

[0175] Excellent grade: Predicted weight > 100g, overall score S ≥ 0.90.

[0176] Excellent grade: Predicted weight 50g < ≤ 100g, overall score 0.80 ≤ S < 0.90.

[0177] Good grade: Predicted weight 20g < M ≤ 50g, comprehensive score 0.70 ≤ S < 0.80.

[0178] Qualified grade: Predicted weight ≤20g, comprehensive score 0.60≤S<0.70.

[0179] Unsatisfactory: Overall score S < 0.60.

[0180] The specific implementation method is as follows:

[0181] 1. Sub-processes and scoring ranges

[0182] 1.1 Integrity Analysis

[0183] The number of severed limbs, missing segments, and shell cracks was statistically analyzed, and the trunk continuity rate was calculated.

[0184] The score ranges from 0.00 to 1.00; in the experimental data, the score is usually ≥0.90 when there is no trunk defect and no limb amputation; between 0.70 and 0.89 when there are minor missing segments or small cracks; and below 0.70 when there are obvious defects.

[0185] 1.2 Color Analysis

[0186] Extract the HSV color histogram and compare its hue, saturation, and brightness with a standard color chart.

[0187] The score range is 0.00–1.00; when the hue deviation is ≤5° and the saturation difference is ≤3%, the score is ≥0.85; when the deviation is moderate, the score is 0.60–0.84; when the hue or brightness is severely abnormal, the score is below 0.60.

[0188] 1.3 Mass fraction

[0189] Predicted weight 0–120g mapped to 0.00–1.00:

[0190] Less than 20g: approximately 0.25; 21–50g: approximately 0.55; 51–100g: approximately 0.85; more than 100g: 1.00. The overall score is the weighted average of the above three factors.

[0191] 2. Level Thresholds and Judgment Logic

[0192] Premium grade: Weight > 100g, and overall score not lower than 0.90.

[0193] Superior grade: Weight between 50–100g, overall score 0.80–0.89.

[0194] Good grade: Weight between 20–50g, overall score 0.70–0.79.

[0195] Qualified grade: weight ≤20g, overall score 0.60–0.69.

[0196] Unacceptable: Overall score below 0.60, or any single item score below 0.50 (e.g., severe damage or abnormal color).

[0197] 3. Single Sample Demonstration

[0198] A shrimp was evaluated through a sub-process:

[0199] Integrity score 0.92 (truncusa intact, no missing segments), color score 0.88 (hue deviation 4°), mass score 0.85 (predicted weight 83g).

[0200] The weighted overall score is 0.88. The weight falls within the 50–100g range and the overall score is between 0.80 and 0.89, therefore the system outputs "Excellent".

[0201] The results are then generated, including ID, predicted weight, three scores, overall score, and final grade, and sent to the PLC and database via Ethernet.

[0202] 4. Batch operation performance

[0203] In indoor high-density breeding batch (1000 birds) testing, the system's average judgment time per bird was 0.75 seconds; the grading ratios were: excellent 8.1%, superior 39.5%, good 31.2%, qualified 18.6%, and unqualified 2.6%.

[0204] The batch of 800 pieces in outdoor natural light was slightly affected by light fluctuations, with an average judgment time of 0.92 seconds. 34.8% were rated as excellent and 37.0% as good.

[0205] In the high-speed production line mode (1500 units), the single-unit judgment is completed in 0.61 seconds, with a superior grade ratio of 41.1%, and the overall error is still controlled within ±3g.

[0206] 5. Reliability and Traceability Measures

[0207] Once the system detects that the integrity or color score is below 0.50, it immediately affixes a red label and sends an alarm to the host computer. Every 50 animals processed, a CSV report (including individual scores and grades) is automatically compiled, and the 3D model index and scoring data are written to the database and retained for at least one year. Through this process, the grading module completes objective and traceable grading within 0.6–1 seconds, meeting the real-time quality control requirements of large-scale breeding and processing.

[0208] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based quality grading system for giant freshwater prawns, characterized in that, include: The shrimp placement platform is used to place the giant freshwater prawns to be tested. A three-dimensional acquisition module, which includes a structured light projection device, a binocular stereo camera, and a time-of-flight depth sensor; The 3D reconstruction module is used to construct a complete 3D point cloud model from the data acquired by the 3D acquisition module through data fusion algorithm and point cloud registration algorithm. The quality mapping and error self-correction module is used to pass the three-dimensional point cloud model through an end-to-end depth regression network and output the shrimp body quality prediction value and prediction uncertainty. The quality mapping and error self-correction module dynamically performs error self-correction based on the prediction uncertainty. The grading module is used to objectively grade the giant freshwater prawns based on the predicted shrimp body weight and a pre-set weight threshold, and output the grading information.

2. The machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The bottom of the shrimp placement platform tray integrates a position detection sensor, which is used to confirm whether the shrimp is within the optimal field of view of the three-dimensional acquisition module.

3. The machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The structured light projection device includes a dual-wavelength laser structured light module, which projects an encoded light pattern onto the shrimp body. The wavelength of the encoded light pattern switches between 660nm and 850nm, and the projection resolution is ≤0.1mm.

4. The machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The binocular stereo camera and the time-of-flight depth sensor work simultaneously with a synchronization error Δt ≤ 1ms. The binocular camera has a baseline distance B ∈ [100mm, 200mm], a focal length f ≥ 8mm, and a resolution ≥ 1920×1080. The time-of-flight depth sensor has a ranging range of 0.2m–2m and a depth resolution ≤ 1mm.

5. A machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The operation of the 3D reconstruction module includes the following three steps: 5.1 Preliminary registration: Rigid registration is performed using the output data from the binocular stereo camera and the time-of-flight depth sensor; 5.2 Weighted Fusion: Weighted fusion of cloud source data for each point based on color consistency and surface normal similarity; 5.3 Surface Reconstruction: A smooth 3D model was generated using the Poisson reconstruction method, and the reconstruction error of the smooth 3D model was <0.5mm.

6. The machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The end-to-end deep regression network adopts a multilayer perceptron structure, and the model formula is as follows: L=α·MSE(y p ,y t )+β·Var(σ p ), Where L represents the total network loss, y p y represents the shrimp body mass predicted by the network. t This indicates the actual mass measured using a precision electronic balance. σ represents the mean squared error of all training samples. p This represents the uncertainty estimate for each predicted value in the network output. The sample variance represents the uncertainty estimate. α and β are weighting coefficients used to balance the quality error term and the uncertainty constraint term, and are set through cross-validation.

7. A machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The error self-correction includes adjusting the structured light projection parameters, adjusting the binocular stereo camera acquisition parameters, and adjusting the time-of-flight depth sensor parameters.

8. A machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The level determination module includes: Integrity Analysis Submodule: Based on the shrimp body shape features and edge integrity scores, calculate the integrity score I, ranging from [0,1]. Color analysis submodule: Extracts the color histogram in HSV space and calculates the color score C, ranging from [0,1]. Quality score submodule: maps predicted quality values ​​to normalized scores M, ranging from [0,1]; The formula for the grade determination module is as follows: S=ω1I+ω2C+ω3M Where: ω1 is the integrity score weight coefficient, ω2 is the color score weight coefficient, ω3 is the normalized score weight coefficient, and ω1+ω2+ω3=1, and S is the final comprehensive score.

9. A machine vision-based quality grading system for giant freshwater prawns according to claim 1, characterized in that: The level information includes; Excellent grade: Predicted mass M > 100g, overall score S ≥ 0.90; Excellent grade: Predicted weight 50g < M ≤ 100g, overall score 0.80 ≤ S < 0.90; Good grade: Predicted weight 20g < M ≤ 50g, comprehensive score 0.70 ≤ S < 0.80; Acceptable grade: Predicted mass M≤20g, comprehensive score 0.60≤S<0.70; Unsatisfactory: Overall score S < 0.60.

Images