Pet food particle forming quality detection method and system based on image recognition

Abstract

The invention discloses a pet food particle forming quality detection method and system based on image recognition, and relates to the technical field of image recognition. The method comprises the steps that particles are conveyed through a conveying unit, and a unique ID is determined for each particle; acquiring three-dimensional geometric data, multispectral data and polarized light data of the particles on line through a three-dimensional geometric detection station, a multispectral detection station and a polarized light detection station in sequence; and the processing control module binds the multi-dimensional data with the unique identifier ID and inputs the data to a two-stage analysis model. According to the model, whether particles have defects or not is identified through a multi-branch deep neural network, and then correlation analysis is carried out on the identified defects and upstream production process parameters through a causal inference model. According to the invention, through multi-modal data fusion and deep analysis, not only can unqualified products be accurately removed, but also data support can be provided for real-time optimization of a production process, and the depth and efficiency of product quality control are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method and system for detecting the molding quality of pet food particles based on image recognition. Background Art

[0002] The quality of pet food pellets is directly related to the product's palatability, nutritional balance, and the health and safety of pets. Therefore, strict quality testing during the production process is crucial.

[0003] Currently, the mainstream quality inspection methods in the industry fall into two main categories. The first is traditional manual sampling and visual inspection. This method involves quality inspection workers randomly selecting samples from the production line and visually inspecting them to determine if the particles exhibit size deviations, color anomalies, breakage, or adhesion. The drawbacks of this method are significant: first, low inspection efficiency and the inability to achieve full inspection, posing a significant risk of missed inspections. Second, test results are significantly influenced by subjective human factors, making it difficult to unify the judgment criteria of different inspectors, or even the same inspector at different times, resulting in poor consistency and reliability in test results. Finally, labor costs are high, making it unsuitable for the needs of modern large-scale production.

[0004] The second category is existing automated image inspection methods. These methods typically use a 2D industrial camera (color or monochrome) to capture particle images and identify color, size, and basic shape defects using traditional image processing algorithms or simple deep learning models. While these methods offer improvements in efficiency and consistency compared to manual inspection, they still have significant limitations:

[0005] Single detection dimension: Only two-dimensional plane information of particles can be obtained, and three-dimensional geometric features such as volume, surface roughness, and depth of tiny cracks cannot be detected.

[0006] Inability to perform intrinsic quality analysis: Traditional image inspection methods are unable to assess key intrinsic quality indicators, such as the uniform distribution of moisture or fat within the pellets and the presence of internal stress caused by process issues (such as excessive cooling). These intrinsic indicators are crucial factors affecting a product's shelf life, taste, and physical stability.

[0007] Lack of effective integration with production processes: Existing inspection systems eliminate defective products at the end of production. While they can indicate the number of defective products, they cannot accurately and immediately answer the reasons for these defects. Their inspection results are often macroscopic, lagging statistical data, making it difficult to quickly and accurately guide and adjust upstream production processes. Summary of the Invention

[0008] The present invention provides a method and system for detecting the molding quality of pet food pellets based on image recognition to address the existing problems of the inability to conduct in-depth, non-destructive online assessment of the intrinsic quality of individual pellets during online quality inspection of pet food pellets, and the inability to directly and directly attribute identified microscopic quality anomalies to specific upstream production process deviations in real time.

[0009] In view of the above problems, in a first aspect, the present invention provides a method for detecting the molding quality of pet food particles based on image recognition, comprising:

[0010] The pet food particles are conveyed through a conveying unit, and a unique particle ID is determined for each pet food particle entering a detection area based on real-time position information output by a rotary encoder coupled to the conveying unit;

[0011] The pet food particles carrying the unique ID of the particles are sequentially transported to the three-dimensional geometry detection station, the multi-spectral detection station and the polarized light detection station in the visual inspection unit;

[0012] Collecting three-dimensional geometric data, multispectral data, and polarization data of the pet food particles respectively, and binding each type of collected data to the unique identification ID of the particles;

[0013] The processing control module analyzes the data bound to the unique ID of the particle to generate a quality detection decision about the pet food particle;

[0014] According to the quality inspection decision, unqualified pet food particles are rejected in combination with the position information contained in the particle unique identification ID and the real-time speed of the conveying unit.

[0015] In a second aspect, the present invention further provides a pet food pellet molding quality detection system based on image recognition, characterized in that the system comprises:

[0016] A conveying unit, configured to carry and transport the pet food particles to be tested, wherein the conveying unit is provided with a rotary encoder configured to output a pulse signal representing the movement displacement of the conveying unit;

[0017] A visual inspection unit, wherein a three-dimensional geometric inspection station, a multi-spectral inspection station and a polarized light inspection station are sequentially arranged along the conveying path of the conveying unit;

[0018] a processing control module, electrically connected to the rotary encoder and each inspection station of the visual inspection unit, configured to determine a unique particle ID for each particle based on the pulse signal output by the rotary encoder, receive and process three-dimensional geometric data, multispectral data, and polarization data collected by each inspection station and bound to the unique particle ID, and generate a quality inspection decision;

[0019] A rejection unit is electrically connected to the processing control module and is used to remove unqualified pet food particles from the conveying unit based on the quality detection decision.

[0020] The technical solution provided by this application has at least the following technical effects or advantages:

[0021] This invention combines a 3D geometric detection station, a multispectral detection station, and a polarized light detection station to achieve multimodal information collection for individual particles. This not only detects surface defects such as color and shape that can be identified by traditional methods, but also obtains deeper depth information, expanding the breadth and depth of online quality inspection.

[0022] The multi-branch deep neural network employed in this paper is capable of fusing and processing feature data from three different physical dimensions. Compared to traditional models that rely on a single image, this multimodal data fusion provides the model with richer and more discriminatory judgment criteria, enabling it to more accurately identify and classify complex, combined, or subtle defects, effectively reducing false positives and missed detections, and improving the overall accuracy and reliability of detection.

[0023] This invention introduces a causal inference model. This model can perform real-time probabilistic correlation analysis between microscopic defect information output by the quality feature recognition model and upstream production process parameters. This enables instant, microscopic feedback and guidance of the production process.

[0024] This invention generates a multidimensional digital profile for every particle that passes inspection, whether it's qualified or unqualified, containing information such as geometry, chemistry, and physical stress. This quality data constitutes a new, traceable data asset. It not only enables precise product quality traceability but also provides powerful data support for new product development, formulation optimization, and long-term process improvement. Its value far exceeds the simple statistical reports generated by traditional inspection systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the method for detecting the molding quality of pet food particles based on image recognition in the present invention;

[0026] Figure 2 This is the architecture diagram of the pet food pellet molding quality detection system based on image recognition. DETAILED DESCRIPTION

[0027] The present invention relates to a method and system for detecting the molding quality of pet food particles based on image recognition, aiming to address the existing problems in the online quality inspection of pet food particles, which include the inability to conduct in-depth, non-destructive online evaluation of the intrinsic quality of individual particles, and the inability to directly and directly attribute identified microscopic quality anomalies to specific upstream production process deviations in real time.

[0028] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the drawings.

[0029] like Figure 1 The pet food pellet molding quality detection system based on image recognition is shown in the architecture diagram. The system is used to perform any of the following methods, and the system includes:

[0030] A conveying unit, configured to carry and transport the pet food particles to be tested, wherein the conveying unit is provided with a rotary encoder configured to output a pulse signal representing the movement displacement of the conveying unit;

[0031] A visual inspection unit, wherein a three-dimensional geometric inspection station, a multi-spectral inspection station and a polarized light inspection station are sequentially arranged along the conveying path of the conveying unit;

[0032] a processing control module, electrically connected to the rotary encoder and each inspection station of the visual inspection unit, configured to determine a unique particle ID for each particle based on the pulse signal output by the rotary encoder, receive and process three-dimensional geometric data, multispectral data, and polarization data collected by each inspection station and bound to the unique particle ID, and generate a quality inspection decision;

[0033] A rejection unit is electrically connected to the processing control module and is used to remove unqualified pet food particles from the conveying unit based on the quality detection decision.

[0034] Furthermore, the conveyor unit includes a conveyor belt driven by a servo motor. Specifically, the conveyor belt is made of food-grade polyurethane with a matte blue finish. This color is designed to provide high optical contrast for the pet food particles being tested, which are typically yellow or brown, facilitating subsequent image segmentation. The servo motor's speed can be precisely set and controlled by the processing control module, within a range of 0.5 m / s to 2.0 m / s.

[0035] Furthermore, the rotary encoder is an incremental encoder, the shaft of which is mechanically coupled to the drive shaft of the conveyor belt. Specifically, the rotary encoder has a high resolution of 4096 pulses per revolution and can continuously output high-frequency pulse signals to the processing control module, which uses these signals to accurately calculate the real-time position of the particles in the conveying direction (defined as the X-axis).

[0036] Furthermore, the visual inspection unit is a fully enclosed metal box. Specifically, its interior surface is coated with a high-absorbency paint to isolate interference from ambient light, providing a stable, non-reflective optical environment for internal image acquisition. Within this unit, along the conveyor belt's direction, the 3D geometry inspection station, the multispectral inspection station, and the polarized light inspection station are sequentially positioned.

[0037] Furthermore, the processing control module is an industrial computer. Specifically, the industrial computer is equipped with a multi-core central processing unit, a graphics processor supporting parallel computing, a large-capacity random access memory, and a multi-channel digital input / output card for high-speed signal exchange with external devices.

[0038] Furthermore, the rejection unit is a nozzle array bracket that spans the conveyor belt. Specifically, 16 high-speed response solenoid valves are evenly installed on the bracket, each of which is connected to a nozzle with a diameter of 1.5 mm and is connected to a compressed air source.

[0039] like Figure 2 The flow chart of the method for detecting the molding quality of pet food particles based on image recognition is shown in FIG. The method includes:

[0040] The pet food particles are conveyed through a conveying unit, and a unique particle ID is determined for each pet food particle entering a detection area based on real-time position information output by a rotary encoder coupled to the conveying unit;

[0041] The pet food particles carrying the unique ID of the particles are sequentially transported to the three-dimensional geometry detection station, the multi-spectral detection station and the polarized light detection station in the visual inspection unit;

[0042] collecting three-dimensional geometric data, multispectral data, and polarization data of the pet food particles, respectively, and binding each type of collected data to a unique ID of the particle; and analyzing, by a processing control module, the data bound to the unique ID of the particle to generate a quality inspection decision regarding the pet food particles;

[0043] According to the quality inspection decision, unqualified pet food particles are rejected in combination with the position information contained in the particle unique identification ID and the real-time speed of the conveying unit.

[0044] Specifically, the workflow of the present invention is as follows: pet food particles are evenly spread on the conveyor belt by upstream equipment and enter the visual inspection unit along the conveyor belt. When a particle enters the field of view, the system determines the unique identification ID of the particle for it. The particles pass through three detection stations in sequence to complete the collection of multi-dimensional data, and all collected data are bound to the particle's unique identification ID. The bound data is sent to the processing control module, and the software in the module analyzes the data and makes a "qualified" or "unqualified" decision. For particles that are judged to be unqualified, the system calculates the exact time when they arrive at the rejection unit based on their unique identification ID and the real-time speed of the conveyor belt, and drives the unit to remove the unqualified particles from the conveyor belt.

[0045] Furthermore, the particle unique identification ID is generated based on a first position coordinate of the particle in the conveying direction determined by the pulse count value output by the rotary encoder, and a second position coordinate determined by the imaging position of the particle on the first image sensor or the second image sensor.

[0046] Specifically, the first position coordinate (X-axis coordinate) is determined by the processing control module by reading and accumulating the output pulse count of the rotary encoder in real time. It accurately reflects the distance the particle has traveled along the conveyor belt. The second position coordinate (Y-axis coordinate) is determined by the pixel row number on the photosensitive chip of the first or second image sensor where the particle's image centroid falls. The combination of these two coordinates provides a stable and unique identity for each particle throughout the entire inspection process.

[0047] Furthermore, the collecting of three-dimensional geometric data includes: using a line laser light source and a first image sensor to collect a laser line image that changes due to the particle contour; and performing a laser triangulation operation on the laser line image to reconstruct a three-dimensional point cloud model of the pet food particles, and then calculating a geometric feature vector including at least one of the particle volume, surface area, crack depth, and total crack length as the three-dimensional geometric data.

[0048] Specifically, in the 3D geometry inspection station, the line laser light source emits red laser light with a wavelength of 660 nanometers, projecting a fine laser line less than 0.1 mm thick onto the conveyor belt surface. The first image sensor is a monochrome industrial camera with a resolution of 2048x1088 pixels. Its optical axis is angled at a 45-degree angle to the conveyor belt plane and precisely aligned with the laser line projection area. The sensor is configured to capture 2000 frames per second. The captured image sequence is fed to a field-programmable gate array (FPGA) processing board. Edge computing using the FPGA enables real-time, high-load point cloud processing and feature extraction at the front end, reducing the computational burden on the processing and control modules and data transmission bandwidth, thereby ensuring the real-time performance of the entire system at high throughput. The processing board incorporates a laser triangulation algorithm that reconstructs the 2D image sequence into a 3D point cloud model of the particles in real time. To reduce data transmission bandwidth, the processing board further analyzes the point cloud model locally, directly calculating and outputting a geometric feature vector consisting of 9 floating-point numbers. The vector specifically includes: particle volume (in cubic millimeters), surface area (in square millimeters), maximum caliper diameter (in millimeters), minimum caliper diameter (in millimeters), height (in millimeters), sphericity (a dimensionless number between 0 and 1), surface roughness (Ra, in microns), maximum crack depth detected (in millimeters), and total length of all cracks (in millimeters).

[0049] Furthermore, the collection of multispectral data includes: within a preset extremely short time window, using a pulsed multispectral light source to sequentially perform pulse irradiation with four different bands of blue light, red light, first near-infrared light, and second near-infrared light, and using a second image sensor to synchronize exposure with each pulse irradiation to collect four monochrome images as the multispectral data.

[0050] Specifically, in the multispectral detection station, the second image sensor is a monochrome CMOS industrial camera with a resolution of 5 megapixels (2448x2048). The pulsed multispectral light source is a custom coaxial light source that integrates four independently and precisely controlled LED chip arrays with central wavelengths of 450 nanometers (blue light), 660 nanometers (red light), 850 nanometers (first near-infrared light), and 940 nanometers (second near-infrared light). When a particle moves to the center of the station's field of view, the processing and control module sends a series of synchronized trigger signals via the digital I / O interface. Within a very short time window, such as 1 millisecond, this signal causes the four LED chips to flash rapidly in sequence. The second image sensor is exposed synchronously with each flash, thereby obtaining four consecutive monochrome images reflecting the particle's reflectivity under different spectral conditions. Using a single black-and-white industrial camera with a pulsed multispectral light source offers significant cost advantages and compactness compared to using multiple cameras or expensive color or hyperspectral cameras. The microsecond-level pulsed flashes effectively "freeze" fast-moving particles, avoiding motion blur and ensuring image quality. At the same time, the design of the coaxial light source can minimize shadows and provide high-quality original image data for subsequent defect analysis.

[0051] Furthermore, the method also includes performing pixel-level operations on the four monochrome images to generate at least two feature maps: generating a moisture distribution map representing moisture distribution by performing a normalized difference operation on the first near-infrared image and the second near-infrared image; and generating a surface defect map highlighting mold spots or minor scratches by applying a Laplace high-pass filter to the blue light image. Furthermore, the graphics processor within the processing control module performs the following pixel-level operations on the four monochrome images to generate feature maps with specific physical meaning.

[0052] Specifically, the moisture distribution map is calculated using the formula (first near-infrared image - second near-infrared image) / (first near-infrared image + second near-infrared image), which utilizes the difference in water absorption in the two near-infrared bands. The fat distribution map is calculated using the formula (red light image - second near-infrared image) / (red light image + second near-infrared image). The basic chromaticity map directly uses the red light image as the image to characterize the basic color of the particles. The surface defect map applies a 3x3 Laplace high-pass filter to the blue light image to enhance and highlight high-frequency details such as mold spots and tiny scratches.

[0053] Furthermore, the collecting of polarized light data includes: emitting polarized light through a light source and a linear polarizer fixed in the light path of the light source, and switching between two mutually orthogonal polarization directions through a liquid crystal polarization rotator installed in front of an image sensor lens, so that the image sensor respectively collects a parallel polarized light image and an orthogonal polarized light image as the polarized light data, and generating an internal stress map representing the internal stress distribution by calculating the pixel ratio of the two images.

[0054] Specifically, at the polarization detection station, immediately after multispectral data collection, the processing control module again sends two sets of trigger signals. This station uses the same second image sensor and 660nm red LED light source as the multispectral detection station.

[0055] The first group of signals drives the liquid crystal polarization rotator to a 0 degree direction (parallel polarization), and simultaneously triggers the red LED to flash, thereby acquiring a parallel polarized light image.

[0056] The second set of signals drives the liquid crystal polarization rotator to a 90-degree direction (orthogonal polarization), triggering the red LED to flash again and acquiring an orthogonal polarization image.

[0057] The graphics processor within the processing control module generates an internal stress map by calculating the pixel ratio of the orthogonal polarized light image to the parallel polarized light image. Regions of abnormal pixel value variation in the map indicate areas of stress concentration within the pellets caused by improper production processes (such as extrusion and drying).

[0058] This method utilizes the principle of the photoelastic effect. When particles experience internal stress due to uneven cooling or drying, their material exhibits temporary birefringence. Orthogonally polarized light cannot penetrate the stress-free, isotropic region. However, in the stress-concentrated region with birefringence, the light's polarization state changes, allowing it to pass through the second polarizer and be captured by the image sensor as bright spots or streaks. Therefore, this internal stress map can non-destructively and intuitively reveal the potential fragility of particles.

[0059] Furthermore, generating a quality inspection decision includes processing the three-dimensional geometric data, the multispectral data, and the polarized light data through a quality feature recognition model, wherein the quality feature recognition model is a deep neural network with a multi-branch structure, and the network includes:

[0060] A branch A configured to receive and process the three-dimensional geometric data;

[0061] a branch B configured to receive and process the multispectral data;

[0062] A branch C configured to receive and process the polarized light data;

[0063] The deep neural network fuses the feature vectors extracted by branches A, B, and C, and inputs them into a fully connected layer to output the occurrence probabilities corresponding to multiple predefined defect types.

[0064] Specifically, Branch A of the quality feature recognition model is a multi-layer perceptron network whose input is the 9-dimensional geometric feature vector output by the 3D geometry detection station. Branch B is a convolutional neural network (e.g., using the ResNet18 architecture) whose input is a 4-channel image composed of the stacked four feature maps generated by the multispectral detection station. Branch C is another convolutional neural network whose input is the internal stress map generated by the polarization detection station.

[0065] Fusion and Output: The feature vectors extracted at the end of the three branch networks are concatenated and input into one or more fully connected layers. Finally, a 12-node Sigmoid activation function layer outputs a multidimensional defect recognition result vector. The 12 dimensions of this vector correspond to the probability of occurrence of 12 predefined defect types (such as "breakage," "adhesion," "out-of-tolerance," "burning," "mildew," and "excessive moisture." For example, for a particle with a broken edge and slightly burnt texture, the output 12-dimensional vector might be [breakage: 0.95, adhesion: 0.02, ..., burnt: 0.65, ...], with the probability values ​​of most other dimensions close to 0.

[0066] In one embodiment, the multi-branch deep neural network is trained using a labeled dataset containing tens of thousands of particle samples covering all defect types. The training process utilizes the Adam optimizer and a weighted cross-entropy loss function to account for the imbalance in the number of samples of different defect types in the dataset. Additionally, online data augmentation techniques, such as random rotation, flipping, and brightness adjustment, are employed to improve the model's generalization capabilities.

[0067] Furthermore, the generation of quality inspection decisions also includes: through a causal inference model, using the defect type output by the quality feature identification model and the probability of occurrence exceeding a preset threshold as reasoning evidence to infer the upstream process root cause of the defect type and its probability of occurrence.

[0068] Furthermore, the causal inference model is a Bayesian network; the network topology of the Bayesian network includes a root node defined as representing upstream process parameters, and a leaf node defined as representing the type of quality defects. Specifically, the causal inference model is a pre-trained Bayesian network. The topology of the network is pre-defined by process experts based on production experience, and includes root nodes representing upstream process parameters (such as "raw material mixing ratio", "extruder screw speed", "drying zone temperature") and leaf nodes representing the type of quality defects (such as "burning" and "breakage"). The network uses the maximum likelihood estimation algorithm to perform offline training on the conditional probability table in the network by changing the process parameters under controlled conditions and collecting a large amount of sample data. In real-time operation, the defect types output by the quality feature recognition model with probability values ​​exceeding a preset threshold (e.g., 0.5) are input into the Bayesian network as evidence, thereby inferring the most likely process root cause and its probability of causing the defect, and presenting the results on the monitoring interface. For example, when the system continuously detects abnormalities in the "internal stress map" of multiple particles, and the causal inference model indicates that the probability of "cooling too fast" is continuously higher than 80%, the system can automatically generate an early warning message, suggesting that the fan speed in the cooling zone be checked. This enables process adjustments to shift from macro-statistics based on hours to micro-insights based on seconds, thereby enabling earlier identification of hair quality issues.

[0069] Finally, the decision-making and control logic in the processing control module is a rule engine based on an editable configuration file (such as XML format). The engine compares the probability value in the multi-dimensional defect recognition result vector with the quality inspection standard set in the configuration file. When the decision is "unqualified", the logic module calculates the precise trigger delay time of the rejection action based on the particle's unique identification ID, combined with the current conveyor belt speed and the physical position of the rejection unit, and sends a DC pulse with a duration of 5 to 10 milliseconds to the solenoid valve located directly above the unqualified particle in the rejection unit through the digital I / O card. The pulse drives the solenoid valve to open instantly, and a small stream of filtered compressed air is ejected through the nozzle, blowing the target particles from the conveyor belt into the defective product collection box on the side, completing the entire closed-loop detection and rejection process.

[0070] The present invention also provides a human-computer interaction interface matched with the processing control module, and the interface runs on an industrial computer or a display in a remote central control room.

[0071] The main monitoring dashboard displays the real-time video stream captured by the second image sensor on the left side of the interface. Colored rectangles (e.g., green for pass, red for fail) are superimposed on the particles to visually demonstrate inspection results. The right side of the interface displays key production indicators in numerical and graphical form, such as real-time pass rate, cumulative reject count, and a dynamic pie chart showing the percentage of each defect type.

[0072] Process Insight: The core area of ​​the interface is a causal flow graph. The left side of the graph represents all detected defect types, while the right side represents the inferred process root causes. The width of the stripe flowing between nodes is proportional to the probability of the causal relationship. This allows the operator to clearly identify which process root cause (e.g., unstable extrusion pressure) is primarily responsible for the current major defect (e.g., breakage).

[0073] Drill-down analysis of defect samples: Users can click on any defect type node in the causal flow diagram. A new window will pop up, displaying digital twins of the last 100 particles classified as that defect. Each profile includes a 3D model (rotatable 360 ​​degrees), moisture distribution, fat distribution, internal stress maps, and other comprehensive information, enabling engineers to conduct in-depth morphological analysis.

[0074] In summary, the present invention discloses a method and system for inspecting the molding quality of pet food pellets based on image recognition. This method sequentially collects three-dimensional geometric data, multispectral data, and polarization data from individual pet food pellets online to obtain multidimensional information characterizing their physical properties. This collected multidimensional data is then input into a two-stage analysis framework consisting of a quality feature recognition model and a causal inference model. This framework first identifies and probabilistically assesses the presence of predefined quality defects in the pellets. It then correlates the identified defect types with upstream production process parameters.

[0075] By fusing and analyzing data from multiple detection modalities, it aims to provide a richer basis for judgment than a single detection modality. The system can perform online detection and screening of particles on the production line. At the same time, its analysis results can also provide reference data for adjusting the production process. The embodiments described in this specification are only preferred embodiments of the present invention. Those skilled in the art may make various modifications and variations without departing from the core idea of ​​the present invention, and these modifications and variations should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for detecting the molding quality of pet food particles based on image recognition, characterized in that: include: The pet food particles are conveyed through a conveying unit, and a unique particle ID is determined for each pet food particle entering a detection area based on real-time position information output by a rotary encoder coupled to the conveying unit; The pet food particles carrying the unique ID of the particles are sequentially transported to the three-dimensional geometry detection station, the multi-spectral detection station and the polarized light detection station in the visual inspection unit; Collecting three-dimensional geometric data, multispectral data, and polarization data of the pet food particles respectively, and binding each type of collected data to the unique identification ID of the particles; The processing control module analyzes the data bound to the unique ID of the particle to generate a quality detection decision about the pet food particle; According to the quality inspection decision, unqualified pet food particles are rejected in combination with the position information contained in the particle unique identification ID and the real-time speed of the conveying unit.

2. The method for detecting the molding quality of pet food particles based on image recognition according to claim 1, wherein: The collecting of three-dimensional geometric data includes: using a line laser light source and a first image sensor to collect a laser line image that changes due to the particle contour; and performing a laser triangulation operation on the laser line image to reconstruct a three-dimensional point cloud model of the pet food particles, thereby calculating a geometric feature vector including at least one of the particle volume, surface area, crack depth, and total crack length as the three-dimensional geometric data.

3. The method for detecting the molding quality of pet food particles based on image recognition according to claim 1, wherein: The multispectral data collection includes: within a preset extremely short time window, using a pulsed multispectral light source to sequentially perform pulse irradiation with four different wavelength bands of blue light, red light, first near-infrared light, and second near-infrared light, and using a second image sensor to synchronize exposure with each pulse irradiation to collect four monochrome images as the multispectral data.

4. The method for detecting the molding quality of pet food particles based on image recognition according to claim 3, wherein: The method further includes performing pixel-level operations on the four monochrome images to generate at least two feature maps as follows: generating a moisture distribution map representing moisture distribution by performing a normalized difference operation on the first near-infrared light image and the second near-infrared light image; By applying a Laplacian high-pass filter to the blue light image, a surface defect map for highlighting mildew spots or tiny scratches is generated.

5. The method for detecting the molding quality of pet food particles based on image recognition according to claim 1, wherein: The polarized light data collection includes: emitting polarized light through a light source and a linear polarizer fixed in the light path of the light source, and switching between two mutually orthogonal polarization directions through a liquid crystal polarization rotator installed in front of the image sensor lens, so that the image sensor respectively collects a parallel polarized light image and an orthogonal polarized light image as the polarized light data, and generating an internal stress map representing the internal stress distribution by calculating the pixel ratio of the two images.

6. The method for detecting the molding quality of pet food particles based on image recognition according to claim 1, wherein: Generating a quality inspection decision includes processing the three-dimensional geometric data, the multispectral data, and the polarized light data through a quality feature recognition model, wherein the quality feature recognition model is a deep neural network with a multi-branch structure, and the network includes: A branch A configured to receive and process the three-dimensional geometric data; a branch B configured to receive and process the multispectral data; A branch C configured to receive and process the polarized light data; The deep neural network fuses the feature vectors extracted by branches A, B, and C, and inputs them into a fully connected layer to output the occurrence probabilities corresponding to multiple predefined defect types.

7. The method for detecting the molding quality of pet food particles based on image recognition according to claim 6, wherein: The generating of quality inspection decisions also includes: using a causal inference model to use the defect type output by the quality feature identification model and the probability of occurrence exceeding a preset threshold as reasoning evidence to infer the upstream process root cause of the defect type and its probability of occurrence.

8. The method for detecting the molding quality of pet food particles based on image recognition according to claim 7, wherein: The causal inference model is a Bayesian network; the network topology of the Bayesian network includes a root node defined as representing an upstream process parameter and a leaf node defined as representing a quality defect type.

9. The method for detecting the molding quality of pet food particles based on image recognition according to claims 2 and 3, characterized in that: The particle unique identification ID is generated based on the first position coordinates of the particle in the conveying direction determined by the pulse count value output by the rotary encoder, and the second position coordinates determined by the imaging position of the particle on the first image sensor or the second image sensor.

10. The pet food pellet molding quality detection system based on image recognition is characterized by: The system is used to perform the method according to any one of claims 1 to 9, and the system comprises: A conveying unit, configured to carry and transport the pet food particles to be tested, wherein the conveying unit is provided with a rotary encoder configured to output a pulse signal representing the movement displacement of the conveying unit; A visual inspection unit, wherein a three-dimensional geometric inspection station, a multi-spectral inspection station and a polarized light inspection station are sequentially arranged along the conveying path of the conveying unit; a processing control module, electrically connected to the rotary encoder and each inspection station of the visual inspection unit, configured to determine a unique particle ID for each particle based on the pulse signal output by the rotary encoder, receive and process three-dimensional geometric data, multispectral data, and polarization data collected by each inspection station and bound to the unique particle ID, and generate a quality inspection decision; A rejection unit is electrically connected to the processing control module and is used to remove unqualified pet food particles from the conveying unit based on the quality detection decision.

Images