A method for automatically identifying appearance defects of a petanque ball

Abstract

The application discloses a kind of Pickleball appearance defect automatic identification method, belong to machine vision detection technical field.There are problems such as hole easy misjudgment, micro crack difficult detection, gloss relies on contact type sampling inspection in prior art.The application includes: by multi-station cooperation acquisition synchronous acquisition Pickleball's topography, spectrum and weight data;U-Net segmentation model is constructed to generate hole and non-hole area binary mask, respectively execute hole edge burr detection and non-hole area defect preliminary screening;Using photometric stereo, Gabor filtering and gray statistics respectively extracts three-dimensional topography, microscopic stress and surface optical characteristics, generates 19-dimensional multi-modal feature vector;Input lightweight MobileNetV3 fusion model for grading determination, output competition level, training level or waste, and trigger corresponding alarm and rejection action.The application realizes the high-precision full inspection of Pickleball appearance defect, the detection accuracy reaches 98.7%, solves the hole misjudgment and micro crack identification problem, and can be widely applied in sports goods intelligent manufacturing field.

Description

Technical Field

[0001] This invention belongs to the field of machine vision inspection and intelligent manufacturing technology of sporting goods. Specifically, it relates to a detection method that uses multimodal sensor fusion and deep learning algorithms to achieve real-time identification and automatic grading of appearance defects in pickles. It is applicable to quality control work in various pickle production lines, sporting goods quality inspection centers and other scenarios. Background Technology

[0002] As a rapidly developing ball sport, the quality of the peak ball directly affects the athlete's hitting feel, flight trajectory stability, and the fairness of the game. The peak ball has a unique surface structure, typically featuring 26 or 40 circular holes and a distinctive "dimpled" texture (for outdoor balls). This structure significantly impacts aerodynamic performance. However, precisely because of this unique perforated structure, detecting surface defects in peak balls presents numerous technical challenges.

[0003] Currently, quality inspection in peak ball manufacturers still relies primarily on manual visual inspection. This method is greatly affected by factors such as worker eye fatigue, lighting conditions, and subjective judgment differences, resulting in poor consistency and a high rate of missed defects. In particular, manual visual inspection is insufficient to effectively identify minute defects such as burrs and microcracks at the edges of holes.

[0004] Regarding automated testing equipment, existing technologies, such as Chinese patent application CN119869958A, disclose a fully automated quality control device for pickles, which mainly uses an appearance inspector to inspect the appearance of the pickles and combines it with a weighing module for weight screening. However, this device has the following technical defects:

[0005] First, the detection dimension is limited, making it impossible to distinguish between process textures and actual defects. Existing equipment treats peaked balls as ordinary spheres, making it difficult to differentiate between "production-permissible dimple textures" and "real hole deformations and edge burrs." When the camera captures a hole area, the system easily misidentifies normal holes as black foreign objects or defects, resulting in a high false detection rate.

[0006] Second, there is a lack of ability to identify microscopic defects. During rotational molding or early use, peaked balls may have microcracks or stress concentration points on their surface. Ordinary optical inspection systems struggle to detect uncolored microcracks or structural stress anomalies that are about to crack using RGB images. Although international patent US2025 / 0073540A1 proposes using force-sensitive color-changing materials to make cracks "color-reveal," this falls under the category of material modification, is costly, and difficult to promote in existing product lines.

[0007] Third, the detection of surface optical properties relies on contact instruments. For key appearance indicators that affect flight trajectory, such as surface gloss and roughness, existing detection methods mostly use offline, contact-based roughness or gloss meters for spot checks, which cannot achieve 100% online inspection and are prone to scratching the surface of the sphere.

[0008] Fourth, multi-station workflow leads to data silos. Existing systems typically inspect weight, dimensions, and appearance at separate stations, without data fusion. This makes it impossible to achieve multimodal fusion judgment based on "visual features + physical quantities," hindering defect tracing and process feedback.

[0009] The existing technologies mentioned above suffer from problems such as low detection accuracy, difficulty in identifying micro-defects, reliance on manual visual inspection, and isolated data processing, which seriously restrict the improvement of quality control in peak ball production. Therefore, there is an urgent need for an automated detection method that can solve these problems. Summary of the Invention

[0010] This invention aims to overcome the technical shortcomings of existing methods for detecting appearance defects in pickles, such as false detection of holes, difficulty in identifying micro-defects, low efficiency in detecting surface optical properties, and data silos. It provides an automatic identification method for appearance defects in pickles based on multimodal data fusion and hierarchical alarm. Through multi-station collaborative acquisition, adaptive segmentation of aperture features, multi-dimensional feature extraction, and lightweight neural network fusion judgment, it achieves high-precision, real-time, and full inspection of appearance defects in pickles, providing technical support for intelligent manufacturing of sporting goods.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] An automatic method for identifying appearance defects in pickle balls includes a data transmission and processing flow. This method utilizes a collaborative approach involving a transmission and positioning module, a multi-station synchronous acquisition module, an image processing and feature extraction module, a multimodal fusion analysis module, and a graded alarm and rejection module. Data transmission and command interaction between these modules are achieved via a high-speed data bus and industrial communication protocols. Specific steps and technical solutions are as follows:

[0013] Step S1: Data Acquisition and Spatiotemporal Synchronization

[0014] The following multimodal data of the Peak ball under test are acquired synchronously through the multi-station collaborative acquisition module:

[0015] Topographic data: High-resolution images of the sphere's surface acquired by at least three line-scan cameras arranged at different angles. The cameras are arranged in a surround layout, with the optical axes of adjacent cameras at an angle of 120°, ensuring 360° coverage of the sphere's surface. The image resolution is 12 megapixels, and the acquisition frequency is 30 frames per second.

[0016] Spectral data: Multi-band reflectance images of the spherical surface acquired under illumination by a combination of multispectral light sources. The multispectral light sources include a coaxial light source, a grazing-effect light source, and a ring light source. The coaxial light source is used to eliminate surface reflection interference, the grazing-effect light source is used to enhance the diffraction effect of microcracks, and the ring light source is used for uniform illumination. The light source wavelengths cover the entire visible light spectrum (400nm-700nm), and the light source combination can be switched at different detection stations.

[0017] Physical quantity data: Real-time weight data of the sphere acquired by a dynamic weighing sensor. The dynamic weighing sensor adopts a strain gauge structure, with a sampling frequency of 100Hz, a measurement accuracy of ±0.1g, and achieves millisecond-level time synchronization with image acquisition.

[0018] The multi-station collaborative acquisition module achieves clock synchronization among its sub-modules via the IEEE 1588 precise time protocol, ensuring that data from different modalities can be subsequently fused and processed under the same time reference.

[0019] Step S2: Adaptive Region Segmentation Based on Aperture Features

[0020] A pre-trained deep learning segmentation model is constructed to identify the inherent hole regions and non-hole regions on the surface of a pickle and generate a binary mask. The segmentation model adopts the U-Net architecture. The encoder part contains 5 convolutional blocks, each containing two 3×3 convolutional layers and one 2×2 max pooling layer; the decoder part contains 5 upsampling blocks, which fuse the shallow features of the encoder and the deep features of the decoder through skip connections.

[0021] The model training dataset contains 5000 high-resolution pickle ball images with annotated hole boundaries, covering both 26-hole and 40-hole sizes, as well as image variations under different lighting conditions. The training process uses the cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, and 100 training epochs.

[0022] In the non-pore area, initial defect screening is performed, including scratch detection, pit detection, and abnormal gloss detection. In the pore area, specific detection of pore edge burrs and molten metal nodules is performed. The pore edge burr detection method is as follows: extract the edge contour line of the pore area and calculate the Hough circle gradient change rate of the contour line; when the local curvature change of the contour line exceeds a preset first threshold (curvature change rate > 15%), or when there is a gray scale bulge at the pore edge that is more than 0.2 mm higher than the nominal surface of the sphere, it is determined to be a "pore edge burr defect" or a "molten metal nodule defect".

[0023] Step S3: Multi-dimensional appearance feature extraction

[0024] Based on the data from steps S1 and S2, the following feature vectors are extracted and quantized respectively:

[0025] Three-dimensional topographic features include spherical roundness, local depression depth, and surface smoothness. Using photometric stereoscopic technology, the distribution of the sphere's surface normal vectors is inverted by utilizing image brightness variations under multi-angle light source illumination, thereby reconstructing the three-dimensional topography. The roundness calculation formula is: Roundness = 1 - (Maximum radius - Minimum radius) / Average radius; the local depression depth is calculated by fitting the normal distance between the spherical reference surface and the measurement point.

[0026] Microscopic stress characteristics: Based on the gray-level gradient changes in the grazing light image, texture anomalies characterizing microcracks or stress concentration areas are extracted. Specifically, the image acquired by grazing light is enhanced using Gabor filtering to extract high-frequency texture components. When continuous linear dark lines exceeding a preset second threshold (0.5 mm) appear in the filtered image, and the gray-level value of these lines is more than 30% lower than the surrounding background gray-level, they are identified as "potential microcrack defects." The Gabor filter parameters are set as follows: wavelength λ = 8 pixels, directions θ = 0°, 45°, 90°, and 135°, phase offset φ = 0°, aspect ratio γ = 0.5, and bandwidth σ = 1.5.

[0027] Surface optical characteristics: Based on multi-angle photometric stereo technology, the spherical normal vector distribution is inverted to calculate the global gloss uniformity index and surface roughness distribution map. The gloss uniformity index is calculated as follows: the spherical surface is divided into N equal-area regions, and the average gray value Gi of each region is calculated. Then, the gloss uniformity index U = 1 - (standard deviation / mean). When U is greater than 0.85, it is considered to have uniform gloss; when U is less than 0.70, it is considered to have uneven gloss. The surface roughness distribution map is obtained by analyzing the fluctuation of the normal vector of each pixel. The roughness value Rz = average peak-valley height difference.

[0028] Step S4: Multimodal data fusion and hierarchical determination

[0029] The various feature vectors extracted in step S3 are aligned pixel-level with the real-time weight data in step S1 according to spatial coordinates, and then input into the pre-trained lightweight neural network fusion model.

[0030] The lightweight neural network fusion model adopts an improved MobileNetV3 architecture, with an input layer dimension of (feature dimension number, 1). The input features include: a 3D topographic feature vector (8-dimensional), a micro-stress feature vector (4-dimensional), a surface optical feature vector (6-dimensional), and a weight bias (1-dimensional), totaling 19 features. The model contains 3 depthwise separable convolutional blocks, 2 fully connected layers, and 1 Softmax output layer. The model uses knowledge distillation technology for parameter compression, achieving a compression ratio of 50%, and combines TensorRT INT8 quantization technology to enable real-time inference on edge computing devices.

[0031] The model training dataset contains 100,000 pairs of pickle samples labeled with quality levels, covering three levels: competition level, training level, and scrap level, with a balanced number of samples in each level. During training, the Focal Loss function is used to address class imbalance, the AdamW optimizer is used, the learning rate is 0.0005, and the training epochs are 200. The model validation set mAP@0.5 reaches 94.7%.

[0032] The multimodal data fusion also includes the introduction of a process traceability feedback mechanism: when the fusion model simultaneously identifies the "hole edge burr" feature and the real-time weight data is lower than the third threshold of the standard weight (the standard weight is 26g, the third threshold is 25.5g, i.e. the deviation exceeds 0.5g), a process warning signal of "insufficient injection pressure" is automatically generated, and the signal is fed back to the host computer or injection molding machine control system in real time to realize closed-loop optimization of process parameters.

[0033] Step S5: Result Output and Tiered Alarm

[0034] The fusion model outputs multidimensional determination results, including:

[0035] Overall quality rating: Divided into at least three levels: "Competition Grade," "Training Grade," and "Scrap." Specific judgment rules are as follows:

[0036] "Competition Grade": Meets all of the following conditions: No scratches or pits on the surface (scratch length ≤ 0 mm, pit depth ≤ 0 mm); microcrack characteristic value is 0; global gloss uniformity index is higher than the fourth threshold (U ≥ 0.85); weight error is within ±0.5g (25.5g~26.5g); all hole edges are free of burrs (hole edge burr height ≤ 0.1mm); roundness ≥ 0.98.

[0037] "Training Grade": Meets the following conditions: The surface has shallow scratches (depth ≤ 0.2 mm) with a length less than the fifth threshold (5 mm) and not connected (interval between adjacent scratches ≥ 10 mm), or has slight uneven gloss (0.70 ≤ U < 0.85); weight error within ±1.0 g (25.0 g ~ 27.0 g); no microcrack features; no functionally impactful burrs on the hole edges (burr height ≤ 0.3 mm and does not affect the rolling of the ball).

[0038] "Scrap": meets any of the following conditions: has microcrack features (microcrack length ≥ 0.5 mm or number ≥ 2); has molten nodules or blockages at the hole edge (molten nodule height ≥ 0.3 mm or hole blockage area ≥ 20% of hole area); weight error exceeds ±1.0g (< 25.0g or > 27.0g); roundness < 0.95; has through-hole defects (such as cracks penetrating the wall thickness of the sphere).

[0039] Defect code matrix: Marks the specific defect type (burrs on hole edges, surface scratches, microcracks, uneven gloss, weight deviation) and spatial location. The defect code adopts an 8-digit encoding format, with the first 2 digits representing the defect type, the middle 4 digits representing the X coordinate (0-9999), and the last 2 digits representing the Y coordinate (0-99), which facilitates subsequent quality traceability.

[0040] Based on the overall quality level, different audible and visual alarm signals and physical rejection actions are triggered: competition-grade products output a green signal and enter the qualified product channel; training-grade products output a yellow signal and enter the substandard product channel; scrap products output a red signal and trigger a pneumatic rejection device to blow the scrap product into the scrap recycling bin. The alarm signals are transmitted to the production line central control system via industrial Ethernet. Attached Figure Description

[0041] In the attached diagram:

[0042] Figure 1 This is the overall flow control logic diagram of the method of the present invention;

[0043] Figure 2 This is a detailed flowchart of the adaptive region segmentation of aperture features in step S2 of the present invention;

[0044] Figure 3 This is a detailed flowchart of the multi-dimensional appearance feature extraction in step S3 of the present invention;

[0045] Figure 4 This is a detailed flowchart of the multimodal data fusion and hierarchical determination in step S4 of the present invention. Detailed Implementation

[0046] Example 1

[0047] This embodiment provides a specific implementation of an automatic identification method for appearance defects in pickles. This method is achieved collaboratively through the following modules: a transmission and positioning module, a multi-station synchronous acquisition module, an image processing and feature extraction module, a multimodal fusion analysis module, and a hierarchical alarm and rejection module. The specific configurations and parameters of each module are as follows:

[0048] Transmission and Positioning Module: A chain conveyor belt with a width of 300mm and a running speed of 0.5m / s is used. The conveyor belt is equipped with evenly spaced spherical positioning grooves, each with a diameter of 74mm (the standard diameter of a peaked ball is 74mm) and a depth of 30mm. The spacing between adjacent grooves is 200mm. The conveyor belt is driven by a servo motor, and an encoder provides position feedback, achieving a positioning accuracy of ±2mm.

[0049] Multi-station synchronous data acquisition module: Three data acquisition stations are arranged sequentially along the conveyor belt movement direction, and each station is configured as follows:

[0050] First acquisition station (top acquisition): Equipped with one 12-megapixel industrial camera (Brand: Basler, Model: aca4024-29um), with an 8mm fixed-focus lens, a working distance of 300mm, and a field of view of 100mm×100mm, to acquire images of the top and side areas of the sphere. The light source is configured as a coaxial light source (Brand: CCS, Model: LFL-300-SW), with a luminous area of ​​300mm×300mm and a color temperature of 6500K.

[0051] The second acquisition station (side acquisition): Equipped with two 12-megapixel industrial cameras, symmetrically arranged on both sides of the conveyor belt at a 60° angle, with a working distance of 250mm, to acquire images of the sphere's side. The light source is configured as a side-grazing light source, with an incident angle of 15° relative to the normal to the sphere's surface, and a wavelength of 620nm in the red light band to enhance the diffraction effect of the microcracks.

[0052] The third acquisition station (bottom acquisition): Equipped with a 12-megapixel industrial camera, installed below the conveyor belt, it acquires images of the bottom of the sphere through a transparent viewing window. The light source is a ring light source with an inner diameter of 80mm, an outer diameter of 120mm, and a color temperature of 5000K.

[0053] The dynamic weighing sensor (brand: Mettler Toledo, model: SWB505) is installed between the first and second data acquisition stations. It has a range of 0-200g, an accuracy of ±0.1g, a sampling frequency of 100Hz, and is synchronized with the camera data acquisition via a hardware trigger signal with a synchronization accuracy of ≤1ms.

[0054] Image processing and feature extraction module: Utilizing an industrial PC (Brand: Advantech, Model: ARK-3531), configured with an Intel Core i7-11700 processor, 32GB DDR4 memory, a 1TB NVMe SSD, and an NVIDIA RTX 3060 graphics card (12GB VRAM). The software environment consists of an Ubuntu 20.04 operating system, Python 3.8 programming language, and the PyTorch 1.10 deep learning framework.

[0055] The pre-trained aperture segmentation model uses the U-Net architecture with the following parameters: input image size 512×512 pixels, encoder channel count [64, 128, 256, 512, 512], decoder channel count [512, 256, 128, 64, 32], and output channel count 2 (aperture / non-aperture). Model training parameters: batch size 8, initial learning rate 0.001, cosine annealing learning rate scheduler, 150 training epochs, and a Dice coefficient of 0.96 on the validation set.

[0056] Multimodal fusion analysis module: Employs an edge computing device (NVIDIA, Jetson AGXOrin) with 200 TOPS of computing power and 32GB of memory. It features a lightweight neural network fusion model, with a 19-dimensional feature vector as input and a probability distribution for three categories as output. After knowledge distillation compression, the model has 1.2M parameters; after TensorRT INT8 quantization, the model size is 3.5MB, and the time for a single inference iteration is ≤5ms.

[0057] The tiered alarm and rejection module includes a three-color audible and visual alarm (Patlite, model LME-203-RYG) and three independently controlled physical rejection devices. The rejection devices employ a pneumatic blowing structure, with solenoid valves (SMC, model VQ110U-5M) controlling compressed air (0.5MPa pressure) to blow the spheres into the corresponding channels. Competition-grade, training-grade, and scrap channels correspond to three different collection bins.

[0058] The working process of this embodiment is as follows:

[0059] 1. Feeding and positioning: The test peak ball is automatically fed by a vibratory feeder and falls into the positioning groove of the conveyor belt. The conveyor belt moves at a constant speed of 0.5m / s, and the encoder provides real-time feedback of position information.

[0060] 2. Image Acquisition and Weighing: When the peak ball passes through the first acquisition station, the photoelectric sensor triggers the top camera to acquire an image, and the coaxial light source illuminates synchronously with an exposure time of 5ms. Subsequently, the peak ball passes through the dynamic weighing sensor, and the weighing data is transmitted to the industrial control computer via the RS485 interface. When the peak ball passes through the second acquisition station, the cameras on both sides trigger image acquisition simultaneously, and the side grazing light source illuminates with an exposure time of 8ms. When the peak ball passes through the third acquisition station, the bottom camera triggers image acquisition, and the ring light source illuminates with an exposure time of 5ms.

[0061] 3. Aperture Adaptive Segmentation: The industrial control computer receives images (4 images in total, 12 megapixels each) acquired by four cameras. Image preprocessing is performed first, including grayscale conversion (converting RGB images to single-channel grayscale images using the formula: Gray = 0.299R + 0.587G + 0.114B), Gaussian filtering (kernel size 5×5, standard deviation σ = 1.0), and histogram equalization to enhance contrast. The preprocessed images are input into the U-Net segmentation model, which outputs a hole region mask. The mask size is consistent with the original image, and each pixel is classified as either "hole" (label 1) or "non-hole" (label 0).

[0062] 4. Multi-dimensional feature extraction:

[0063] 3D topographic feature extraction: Photometric stereo technology is employed, using three images illuminated by light sources from different directions to solve for the surface normal vectors. The light source direction is calculated using pre-calibrated light source and camera positions. The normal vector map is then integrated to reconstruct a depth map with a resolution of 512×512 pixels and a depth accuracy of 0.05mm. Based on the depth map, eight-dimensional features, including roundness, depression depth, and flatness, are calculated.

[0064] Microscopic stress feature extraction: Gabor filtering was applied to the grazing light image for enhancement. Filter parameters: wavelength λ = 8 pixels, direction θ = 0°, 45°, 90°, 135°, spatial aspect ratio γ = 0.5, bandwidth σ = 1.5. The maximum value of the filtering results in the four directions was used as the enhanced image. Linear dark stripe regions were extracted using threshold segmentation (threshold = background gray level × 0.7). Connected component analysis was used to calculate the length of the dark stripes, and four-dimensional features (maximum length, total length, number, and average width) were extracted.

[0065] Surface optical feature extraction: The sphere surface is divided into 32 regions, the mean and standard deviation of gray level of each region are calculated, and the gloss uniformity index U=1-σ / μ is calculated; the Sobel operator is used to calculate the gradient magnitude of each pixel, the gradient distribution is statistically analyzed, the surface roughness features are calculated, and 6-dimensional features are extracted.

[0066] 5. Multimodal fusion judgment: The 19-dimensional feature vector is input into the lightweight neural network fusion model. The model passes through 3 depthwise separable convolutional blocks (3×3 kernel size, 1.0 depth multiplier), 2 fully connected layers (128 and 64 dimensions respectively), and a Softmax output layer. The output is the competition-level probability P1, the training-level probability P2, and the scrap probability P3. The category corresponding to the highest probability is taken as the judgment result.

[0067] 6. Process traceability feedback: When the judgment result is a defective product and the feature value of the hole edge burr and the feature value of the weight deviation in the feature vector both exceed the threshold (burr height ≥ 0.3mm and weight < 25.5g or > 26.5g), the system sends an "injection pressure optimization suggestion" signal to the injection molding machine control system through the ModbusTCP protocol, suggesting the following parameter adjustments: increase the injection pressure by 5% and extend the holding time by 0.5s.

[0068] 7. Graded Alarm and Rejection: A three-color audible and visual alarm displays the corresponding color based on the judgment result (competition grade - solid green, training grade - flashing yellow, scrap - solid red + buzzer). The pneumatic rejection device is triggered based on the encoder position delay. The delay time is calculated using the formula: T = L / v, where L is the distance from the detection position to the rejection position (0.8m), and v is the conveyor belt speed (0.5m / s), resulting in T = 1.6s. Scrap products are blown into the scrap recycling bin, training grade products are blown into the secondary product channel, and competition grade products naturally fall into the qualified product channel.

[0069] This embodiment underwent a three-month trial run on a peak ball production line, inspecting approximately 500,000 peak balls. The results, compared with manual re-inspection, show that the method of this invention achieves a detection accuracy of 98.7%, a false detection rate of 0.8%, and a missed detection rate of 0.5%. The average single-ball inspection time (from image acquisition to graded output) is 280ms, meeting the production line's cycle time requirements (approximately 12,000 balls per hour). Compared to traditional manual visual inspection, the inspection efficiency is increased by 5 times, and labor costs are reduced by 70%. Particularly noteworthy is the detection of burrs and microcracks at the hole edges; manual visual inspection has a detection rate of only 45% and 8%, respectively, while the method of this invention achieves detection rates of 96% and 91%, respectively, demonstrating significant effectiveness.

[0070] Example 2

[0071] The difference between this embodiment and Embodiment 1 is that the multi-station synchronous acquisition module adopts a higher-specification configuration. The camera is upgraded to a 20-megapixel global shutter camera (brand: FLIR, model: BFS-PGE-200S6C-C), and the acquisition frequency is increased to 60 frames / second; the light source is upgraded to a programmable multispectral light source, supporting 8 independent control bands (405nm, 460nm, 530nm, 590nm, 620nm, 660nm, 740nm, 850nm), which can automatically switch the optimal illumination band according to different defect types.

[0072] The incident angle of the side-grazing light source can be dynamically adjusted (10°-30°), automatically optimizing the incident angle for different types of microcracks. The system has added an intelligent light source control unit, which determines the optimal illumination parameters for different defect types through pre-experimentation, stores them in the light source parameter library, and automatically retrieves them based on the predicted defect type during the detection process.

[0073] The multispectral function of this embodiment achieves the following enhancements: using 405nm ultraviolet light can excite the fluorescence effect of fluorescent additives in the pickle material, enhancing the visibility of microcracks; using 850nm near-infrared light can penetrate the shallow layer of the sphere surface to detect structural anomalies in the subsurface layer.

[0074] The edge computing module has been upgraded to the NVIDIA Jetson AGX Orin 64GB version, increasing computing power to 275 TOPS and model inference speed to 3ms / frame. The multimodal fusion model introduces the Transformer attention mechanism, further improving the ability to identify minute defects.

[0075] The working process of this embodiment is basically the same as that of Embodiment 1. The newly added multispectral function improves the prediction accuracy of anthrax level (increased brittleness due to material aging) to 89%, enabling early warning before defects appear on the surface, further improving the foresight of quality control. The programmable light source makes the detection of different types of defects more accurate, and the overall detection accuracy is improved to 99.2%.

[0076] Example 3

[0077] The difference between this embodiment and Embodiment 1 is that the system adds a deep learning continuous learning function. The edge computing module has a built-in online learning unit that can continuously collect new sample data during production line operation and use an elastic weight consolidation algorithm to incrementally update the model, adapting to new defect types without forgetting old knowledge.

[0078] The specific implementation method is as follows: Every 1000 detection cycles, the system automatically collects sample data that has been manually verified and constructs an incremental training set. The EWC algorithm (Elastic Weight Consolidation) is used to calculate the importance weights of the model parameters. The learning rate of incremental learning is 1 / 10 of the original model training learning rate (i.e., 0.00005), which effectively prevents catastrophic forgetting.

[0079] This embodiment also adds a cloud-based collaborative analysis function. The edge computing module uploads the detection results and key sample images to the cloud server via a 5G network. The cloud server utilizes large-scale computing resources for data mining and model optimization. When data from multiple production lines converges in the cloud, a global optimization model is trained using federated learning technology and then distributed back to each edge computing module to achieve cross-production line collaborative optimization.

[0080] In this embodiment, the continuous learning function enables the model to quickly adapt to changes in defect characteristics brought about by new molds and materials, without the need for manual retraining. During the six-month trial period, the model automatically performed 15 incremental updates, successfully adapting to three new types of defects while consistently maintaining a detection accuracy rate of over 98%. The cloud-based collaboration function enabled data sharing and model unification across the three production bases, reducing the workload of repetitive annotation and shortening the model iteration cycle from four weeks to one week.

[0081] Example 4

[0082] The difference between this embodiment and Embodiment 1 is that the system adds a full lifecycle traceability function for the pickle. Each pickle is assigned a unique QR code identifier during the inspection process (marked on the inner wall of the sphere by a laser marking machine). The inspection system associates multimodal data, judgment results, defect code matrix with the QR code and stores them in a blockchain database.

[0083] The blockchain uses the Hyperledger Fabric framework. Each testing batch generates a block containing testing data, timestamps, operator information, and other information for all pickles in that batch. The immutability of the blockchain ensures the credibility of the quality data and facilitates tracing the source of subsequent quality disputes.

[0084] This embodiment applies to the production of high-end competition peaks. The full lifecycle traceability function makes the quality data of each peak verifiable and traceable, enhancing brand value. The open and transparent nature of blockchain data also facilitates quality audits by third-party testing agencies.

Claims

1. A method for automatic identification of appearance defects in pickles, characterized in that, Includes the following steps: Step S1: Data Acquisition and Spatiotemporal Synchronization The following multimodal data of the Peak ball under test are acquired synchronously through the multi-station collaborative acquisition module: Topographic data: High-resolution images of the sphere's surface acquired by at least three line-scan cameras arranged at different angles; Spectral data: Multi-band reflectance images of the sphere surface acquired under illumination by a combination of multispectral light sources (including coaxial light, grazing light, and ultraviolet light); Physical quantity data: Real-time weight data of the sphere acquired by a dynamic weighing sensor; Step S2: Adaptive Region Segmentation Based on Aperture Features A pre-trained deep learning segmentation model is constructed to identify the inherent porous and non-porous regions on the surface of a pickle and to generate a binary mask. Initial defect screening is performed in the non-porous regions, and specific detection of burrs and molten metal at the edge of the holes is performed in the porous regions. Step S3: Multi-dimensional appearance feature extraction Based on the data from steps S1 and S2, the following feature vectors are extracted and quantized respectively: Three-dimensional morphological features: including sphericity, depth of local depressions, and surface smoothness; Micro-stress characteristics: Based on the gray-level gradient changes in the side-grazing light image, texture anomalies that characterize microcracks or stress concentration areas are extracted; Surface optical characteristics: Based on multi-angle photometric stereo technology, the distribution of spherical normal vectors is inverted, and the global gloss uniformity index and surface roughness distribution map are calculated. Step S4: Multimodal data fusion and hierarchical determination The various feature vectors extracted in step S3 are aligned pixel-level with the real-time weight data in step S1 according to spatial coordinates and then input into the pre-trained lightweight neural network fusion model. Step S5: Result Output and Tiered Alarm The fusion model outputs multidimensional determination results, including: Overall quality rating: It is divided into at least three levels: "Competition Grade", "Training Grade", and "Scrap". Defect code matrix: Marks the specific defect type (burrs on hole edges, surface scratches, microcracks, uneven gloss, weight deviation) and spatial location; Based on the overall quality level, different audible and visual alarm signals and physical rejection actions are triggered.

2. The method for automatic identification of appearance defects in pickles according to claim 1, characterized in that, The specific inspection of burrs and weld beads at the hole edge in step S2 includes: Extract the edge contour of the hole area and calculate the Hough circle gradient change rate of the contour. When the local curvature change of the contour exceeds the preset first threshold, or when there is a gray scale bulge at the edge of the hole that is higher than the nominal surface of the sphere, it is determined to be a "hole edge burr defect" or a "melting defect".

3. The method for automatic identification of appearance defects in pickles according to claim 1, characterized in that, The method for extracting "micro-stress features" in step S3 includes: The image acquired by the side-grazing light is enhanced by Gabor filtering to extract high-frequency texture components. When a continuous linear dark line with a length exceeding a preset second threshold appears in the filtered image, and the gray value of the linear dark line is lower than a preset multiple of the gray value of the surrounding background, it is determined to be a "microcrack potential defect".

4. The method for automatic identification of appearance defects in pickles according to claim 1, characterized in that, The "multimodal data fusion" in step S4 also includes the introduction of a process traceability feedback mechanism: When the fusion model simultaneously identifies the "hole edge burr" feature and the real-time weight data is lower than the third threshold of the standard weight, it automatically generates a "insufficient injection pressure" process warning signal and feeds the signal back to the host computer or injection molding machine control system in real time.

5. The method for automatic identification of appearance defects in pickles according to claim 1, characterized in that, The specific rules for determining the "overall quality level" in step S5 are as follows: "Competition Grade": Simultaneously meets the following requirements: no scratches or dents on the surface; microcrack characteristic value of 0; global gloss uniformity index higher than the fourth threshold; weight error within ±0.5g; no burrs on the edges of all holes; "Training grade": meets the following requirements: the surface has shallow scratches with a length less than the fifth threshold and not connected, or slight gloss unevenness without functional impact; weight error within ±1.0g; "Scrap": meets any of the following conditions: has microcrack characteristics; There are weld beads or blockages at the edge of the hole; the weight error exceeds ±1.0g; the roundness is out of tolerance or there are through defects.

6. An automatic identification system for appearance defects of pickles for implementing the method of any one of claims 1 to 5, characterized in that, include: Transmission and positioning module: used to continuously transport the pickle to be tested and provide stable support; Multi-station synchronous acquisition module: includes at least three high-resolution industrial cameras arranged in a ring, a programmable multi-angle LED light source (coaxial light, side-grazing light, ultraviolet light) and a dynamic weighing sensor; Image processing and feature extraction module: Built-in pre-trained aperture segmentation model and photometric stereo reconstruction unit; Multimodal fusion analysis module: Equipped with a lightweight neural network processor, used to perform spatiotemporal data alignment and comprehensive judgment; Tiered alarm and rejection module: Includes an audible and visual alarm and at least two independently controlled physical rejection devices, corresponding to the "training level" feeding channel and the "waste" recycling channel, respectively.

7. The system according to claim 6, characterized in that, The side-grazing light in the "programmable multi-angle LED light source" has an incident angle of 15° to 30° relative to the normal of the sphere surface, and the light source wavelength is in the red light band of 600nm to 650nm, which is used to enhance the diffraction effect of microcracks.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images