Deep learning visual defect detection system for high-end encasement injection molding
Patent Information
- Application Number
- CN202610445899.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-03
Smart Images
Figure CN122336252A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine vision inspection, specifically relating to a deep learning vision defect detection system for injection-molded housings of high-end equipment. Background Technology
[0002] With the intelligent transformation of the high-end equipment manufacturing industry, fields such as aerospace precision components and medical imaging equipment have placed higher demands on the surface quality of injection-molded housings. Automated visual inspection technology, as a key means to ensure product quality consistency, plays a crucial role in improving production line automation levels and defect detection rates. In the inspection of precision housings involving high-gloss, dark, or complex curved surfaces, achieving accurate identification of minute surface defects has become an important foundation for ensuring the high-reliability delivery of high-end equipment.
[0003] Deep learning-based visual defect detection systems integrate high-performance imaging components and advanced computer vision algorithms to achieve automatic localization and refined classification of defects on the surface of injection-molded housings. These systems typically utilize multi-layer neural networks to extract image features and employ deep learning models to perform real-time identification of defects such as scratches, stress marks, and orange peel texture. Due to the extremely complex surface treatment processes for high-end equipment housings, the system must maintain extremely high detection sensitivity and environmental robustness in highly dynamic production environments.
[0004] Existing technologies have limitations when processing injection-molded shells with complex geometries and optical properties. Dynamic optical noise generated by specular surfaces is highly coupled with minute defect features. Traditional fixed lighting and preprocessing methods struggle to eliminate reflective interference, causing weak defect features to be obscured by optical overflow. Minor fluctuations in process parameters during injection molding cause dynamic drift in defect morphology. Existing detection models rely excessively on static sample annotations, lacking adaptability to changes in production conditions. This leads to increased false positive rates and frequent downtime for retraining during long-term operation. Traditional serial detection architectures face a resource mismatch between global field-of-view positioning efficiency and local microscopic detail recognition accuracy when processing large-size, high-precision shells, making it difficult to achieve the technical goal of zero false negatives while meeting production line cycle time. Summary of the Invention
[0005] The purpose of this invention is to provide a deep learning visual defect detection system for injection-molded shells of high-end equipment, which can solve the problems in the background art such as dynamic surface optical distortion and micro-defect feature submersion, model adaptation failure caused by production state drift, and mismatch between detection efficiency and accuracy resources.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based visual defect detection system for injection-molded shells of high-end equipment includes a biomimetic compound eye programmable optical imaging subsystem, a production process perception module, a cloud-based adaptive learning center, and an edge-side multi-task inference terminal, as follows: The biomimetic compound eye programmable optical imaging subsystem is used to detect the surface morphology of the injection-molded shell to be inspected, dynamically generate a reverse compensation light field based on the acquired surface curvature data, and achieve contrast enhancement of defect features and background noise removal at the physical imaging level. The production process sensing module is used to monitor the physical environment parameters and equipment operating status in the injection molding production line in real time, extract feature vectors that can characterize the fluctuation of production status, and perform time stamp alignment processing on the feature vectors and image data. The cloud-based adaptive learning center is used to construct a defect feature evolution model based on meta-learning. By analyzing the morphological feature change patterns of historical defect samples under different process parameters, it generates network adaptation parameters for the current production state. The edge-side multi-task inference terminal is used to run a deep neural network that has undergone pruning and quantization, receive dynamic knowledge distillation results from the cloud-based adaptive learning center, and accurately locate and segment defect areas in the enhanced physical feature image at the pixel level.
[0007] Preferably, the biomimetic compound eye programmable optical imaging subsystem includes a micro-programmable light-emitting unit array, a multispectral sensing component, and a reverse light field controller. The reverse light field controller calculates the reflected energy distribution of each light-emitting point on the sensor surface based on the digital elevation model of the injection-molded shell under test. The micro-programmable light-emitting unit array includes thousands of independently controlled semiconductor point light sources, whose distribution pattern references the structure of an insect compound eye, enabling full-coverage illumination within a spherical or hemispherical space. The reverse light field controller adjusts the luminous intensity, spectral frequency, and flicker phase of each light-emitting point light source. In areas with high curvature on the injection-molded shell surface and prone to high-light overflow, it reduces the output power of the corresponding angle light source; in concave areas with weaker light, it enhances the intensity of multi-angle cross-illumination; and through physical light field superposition, it eliminates high-light noise in the image.
[0008] Furthermore, the multispectral sensing component in the biomimetic compound eye programmable optical imaging subsystem has wavelength-selective acquisition capabilities. When detecting dark or highly reflective injection-molded shells, the reverse light field controller instructs the micro-programmable light-emitting unit array to switch to a specific wavelength band of spectral output. Simultaneously, the multispectral sensing component activates the corresponding narrowband filtering logic to physically filter out ambient light interference. The specific wavelength band includes, but is not limited to, the near-infrared band, utilizing the scattering characteristics of long-wavelength light on micro-defects to enhance the optical characterization capabilities of micron-level scratches, stress marks, and orange peel texture on the surface.
[0009] Furthermore, the production process sensing module includes multiple physical sensor interfaces for real-time access to data on mold temperature, melt pressure, holding time, cooling rate, and ambient humidity from the injection molding machine. The module utilizes a preset dimensionality reduction algorithm to transform high-dimensional process parameters into state vectors representing production stability. When the offset of the process state vector exceeds a preset offset threshold, the module immediately triggers a data upload mechanism, transmitting typical samples of the current state to a cloud-based adaptive learning center.
[0010] Furthermore, the cloud-based adaptive learning center runs a deep neural network model based on meta-learning. This meta-learning model is not trained for a single type of defect, but rather learns how to extract the mapping function between defect features and process parameters. The cloud-based adaptive learning center maintains a dynamically updated knowledge base that stores defect evolution sequences under different production cycles. By analyzing these sequences, the meta-learning model can predict the impact of injection molding process fluctuations on defect edge sharpness, contrast, and geometry.
[0011] Furthermore, the cloud-based adaptive learning center utilizes a dynamic knowledge distillation algorithm to transform the knowledge learned by the complex teacher network into student network parameters that are easily deployed at the edge. During the distillation process, the teacher network generates a set of optimal convolutional kernel weight parameters or feature adaptation layer parameters within a preset feature space based on the current process state vector provided by the production process perception module. This dynamic knowledge distillation algorithm ensures that the lightweight edge model achieves discriminative capabilities similar to complex cloud models by minimizing the distribution differences between the teacher and student networks in the output and intermediate feature layers.
[0012] Furthermore, the edge-side multi-task inference terminal adopts a parallel processing architecture. The neural network running on the multi-task inference terminal includes a feature extraction backbone network, an object detection branch, and a semantic segmentation branch. The feature extraction backbone network processes the physically enhanced image output by the biomimetic compound eye programmable optical imaging subsystem, and uses multi-scale feature fusion logic to extract high-level semantic information of defects. The object detection branch scans the entire image, filters out regions of interest suspected of having defects through a preset confidence threshold, and outputs the corresponding bounding box coordinates.
[0013] Furthermore, the semantic segmentation branch performs local ultra-high resolution pixel-level calculations only on the region of interest output by the target detection branch. After acquiring the bounding box coordinates, the edge-side multi-task inference terminal crops the corresponding patches from the original high-resolution image and inputs them into the segmentation sub-network. The segmentation sub-network is used to classify each pixel and accurately depict the geometric contour of the defect. Through this step-by-step inference process from global coarse localization to local fine segmentation, the edge-side multi-task inference terminal can reduce the demand for computing resources while ensuring detection accuracy, and ensure that the processing speed can match the real-time cycle time of a high-speed injection molding production line.
[0014] Furthermore, the edge-side multi-task inference terminal also includes an inference performance monitor, used to evaluate the detection determinism of the current model. If the entropy value of the classification probability distribution for a certain sample is higher than a preset entropy threshold, it indicates that the current model has uncertainty in its judgment of that sample. The inference performance monitor will mark the sample as a difficult sample and feed it back to the cloud-based adaptive learning center for targeted reinforcement learning.
[0015] Furthermore, upon receiving challenging samples from the edge, the cloud-based adaptive learning center initiates an online incremental learning program. This program utilizes an active learning strategy to automatically select representative unlabeled data from the sample pool, guiding the cloud-based model to fine-tune its parameters. After fine-tuning, the system again distributes the updated weights to the edge-side multi-task inference terminal through dynamic knowledge distillation, achieving closed-loop optimization and continuous evolution of system performance.
[0016] Furthermore, all calculation processes, parameter comparisons, and logical judgments involved in this invention are described in textual form. For example, when determining the reverse light field control parameters, the system calculates the cosine of the angle between the incident direction vector of the light-emitting unit at that point and the surface normal vector at that point. If the cosine value is greater than a preset reflection risk threshold, the current supply intensity of the light-emitting unit is reduced proportionally. When performing defect determination, if the calculated defect area score and the preset classification threshold satisfy a predetermined condition, it is determined that the area has a defect of a certain type.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves deep synergy between physical imaging and digital algorithms. Through a biomimetic compound eye-like programmable optical imaging subsystem, it solves the problem of specular curvature obscuring defect features at the physical level. Traditional solutions rely on post-processing algorithms to remove noise, often leading to feature loss. This invention, however, utilizes active detection and reverse light field compensation technology to suppress optical overflow. This enables the system to reliably detect minute defects such as micron-level silver streaks and shallow scratches, which are difficult for traditional vision systems to capture and are only visible under specific light fields. This raises the quality threshold for injection-molded parts in high-end equipment and achieves perception capabilities exceeding the limits of the human eye.
[0018] 2. This invention solves the problem of adaptive detection models to fluctuations in injection molding production conditions by introducing meta-learning and dynamic knowledge distillation mechanisms. Temperature and pressure fluctuations in the injection molding process often cause defect morphology to drift, and traditional models will experience a rapid increase in false alarm rate and become difficult to maintain under such conditions. The cloud-edge collaborative architecture constructed in this invention enables edge models to update knowledge in real time through distillation parameters without retraining, reducing downtime caused by incompatibility of the detection system to a negligible level. This not only ensures the long-term stability of detection but also improves the overall equipment efficiency of the factory.
[0019] 3. This invention achieves an asymmetric advantage in both accuracy and efficiency. Through a stepped architecture integrating detection and segmentation, the system enhances the quality of the input signal at the hardware level using physical enhancements, and optimizes computational power allocation at the algorithm level using a lightweight multi-task network. This design avoids resource conflicts in traditional serial architectures when processing large-size, high-precision targets, enabling pixel-level depth analysis of the entire surface within an extremely short sampling period. Its processing efficiency is significantly improved compared to traditional solutions of the same precision, perfectly meeting the high-speed, high-precision, and high-stability detection requirements of modern high-end manufacturing.
[0020] 4. The end-to-end closed-loop design adopted in this invention enables the system to continuously evolve. Through real-time interaction between the production process perception module and the cloud center, the system can transform dynamic changes in the production environment into the driving force for model evolution. As production data accumulates, the meta-learning model's prediction of process deviations becomes more accurate, and the parameters of dynamic distillation become more efficient. This self-optimizing characteristic ensures that the system can maintain excellent performance in complex and ever-changing industrial environments over a long period, providing a solid technical guarantee for zero-defect manufacturing of high-end equipment. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the adaptive evolution of the deep learning model based on meta-learning and dynamic knowledge distillation in this invention; Figure 3 This is a schematic diagram illustrating the principle of the biomimetic compound eye programmable optical imaging subsystem in this invention to achieve reverse compensation of the optical field and enhancement of physical features. Figure 4 This is a flowchart illustrating the logical process framework of the production process sensing module in this invention, which performs multi-dimensional parameter dimensionality reduction and state vector alignment. Figure 5 This is a schematic diagram of the stepped inference architecture of the edge-side multi-task inference terminal in this invention, which ranges from global target detection to local semantic segmentation. Figure 6 This is a comparison diagram of the closed-loop optimization principle based on inference uncertainty assessment and online incremental learning in this invention. Detailed Implementation
[0022] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6 A deep learning-based visual defect detection system for injection-molded shells of high-end equipment includes a biomimetic compound eye programmable optical imaging subsystem, a production process perception module, a cloud-based adaptive learning center, and an edge-side multi-task inference terminal. The system constructs a closed-loop detection architecture with environmental adaptability and ultra-high resolution recognition accuracy through deep collaboration between optical modulation in the physical domain and meta-learning algorithms in the digital domain.
[0023] The biomimetic compound eye programmable optical imaging subsystem, serving as the physical sensing front-end of the entire system, focuses on simulating the non-uniform spatial sampling and multi-angle light field sensing characteristics of insect compound eyes in nature. This subsystem is configured to perform real-time detection of the three-dimensional surface morphology of the injection-molded shell under inspection. In specific engineering implementations, the biomimetic compound eye programmable optical imaging subsystem does not rely on traditional single structured light, but instead employs a dynamic three-dimensional reconstruction logic based on multi-point projection and phase shift analysis. By acquiring a digital elevation model of the surface of the injection-molded shell under inspection, the system can accurately identify the shell's curvature changes at the micrometer level, local depressions, and stress deformation zones formed by the injection molding process. Based on the digital elevation model, the integrated reverse light field controller within the subsystem can calculate the reflection vector at each spatial coordinate point in real time.
[0024] The calculation logic of the reverse light field controller is as follows: The system pre-stores the spatial position coordinates of the imaging sensor and the coordinate distribution of each point light source in the light-emitting unit array. When the injection-molded shell enters the inspection station, the controller calculates the incident direction vector of each light-emitting unit at a specific pixel point on the surface to be inspected, and calculates the cosine value of the angle between the two based on the obtained surface normal vector. If the cosine value of the angle, after conversion by Fresnel's law of reflection, is estimated to result in a reflection energy intensity on the sensor surface exceeding a preset pixel saturation threshold (i.e., a risk of high light overflow), the reverse light field controller will instruct the corresponding light-emitting unit to reduce its output current intensity, or shorten its emission duty cycle through pulse width modulation technology. This dynamically generated reverse compensation light field cancels out the specular reflection interference caused by complex curvature during the physical imaging stage, achieving contrast enhancement of defect features and deep stripping of background noise.
[0025] Specifically, in the biomimetic compound eye programmable optical imaging subsystem, the light intensity modulation logic executed by the reverse light field controller is as follows: For any pixel on the surface of the injection-molded housing to be inspected The system obtains the surface normal vector in its three-dimensional coordinates. and the The spatial coordinates of a semiconductor point light source are used to calculate the point position of the light source. incident direction vector at point The system calculates the cosine of the angle between the incident direction vector and the surface normal vector. .
[0026] Furthermore, the reverse light field controller constructs a reflection energy prediction model based on Fresnel's law of reflection to calculate the reflection energy of the light source at a point. Predicted value of specular reflection energy generated at the location : in, Indicates the first Each light source at the pixel Predicted value of reflected energy at the location; Indicates the first The initial driving power of each light source; The surface roughness index is determined in advance by the material of the injection-molded housing to be inspected. Indicated based on the material's refractive index and angle of incidence Fresnel reflection coefficient.
[0027] The reverse light field controller will calculate the With preset pixel saturation threshold Compare and calculate the first... The dynamic adjustment coefficient of light intensity for this pixel is applied to each light source. : Finally, the reverse light field controller sends the first... The actual driving current intensity of each light source The minimum adjustment factor is determined by all pixels within its illumination range, i.e. Suppressing highlight overflow at the physical imaging level.
[0028] The biomimetic compound eye-like programmable optical imaging subsystem is further refined to include a micro-programmable light-emitting unit array, a multispectral sensing component, and a reverse light field controller. The micro-programmable light-emitting unit array consists of thousands of independently controlled semiconductor light-emitting diodes (LEDs) arranged in a spherical canopy topology, ensuring that light can cover the inspected casing from a near-global angle. Each LED is equipped with an independent addressing and driving circuit, supporting spectral switching and intensity adjustment within milliseconds.
[0029] The multispectral sensing component features wavelength-selective acquisition. When detecting specific dark-colored, high-absorption or high-transparency, high-reflectivity materials, the multispectral sensing component forms a spectral closed loop with the micro-programmable light-emitting unit array through a built-in electronically tunable filter or multi-bandpass filter array. For example, in scratch detection of black high-gloss injection molded parts, the light-emitting array is configured to output near-infrared light of a specific frequency, while the multispectral sensing component only activates the photosensitive logic for that infrared band. Because near-infrared light has a stronger scattering effect on fine surface defects and is less affected by ambient visible light, the system can improve the optical characterization capabilities of micron-level scratches, silver streaks, and orange peel texture at the physical level, providing high-quality raw signals for subsequent algorithm processing.
[0030] The production process sensing module is connected to the central controller and sensor bus of the injection molding production line to monitor physical environmental parameters during production in real time. This module is not a simple information acquisition unit, but an intelligent gateway with real-time data fusion and feature engineering capabilities. It includes multiple physical sensor interfaces, capable of simultaneously accessing mold temperature monitoring data from the injection molding machine, melt pressure change curves over time, real-time pressure distribution during the holding pressure stage, cooling water flow rate and temperature, and ambient humidity data from the detection workshop.
[0031] The production process perception module internally runs a dimensionality reduction algorithm based on streaming computing. This algorithm maps the aforementioned multi-dimensional physical parameters with different sampling frequencies into a low-dimensional feature space, generating a state vector that can characterize the stability of the current production state. To ensure the consistency of process parameters and visual images in causal logic, the production process perception module also performs strict timestamp alignment processing. When the injection-molded shell completes demolding and enters the visual inspection area, the module automatically extracts the complete process features of the product during the molding cycle and encapsulates them into a meta-data packet bound to the image ID.
[0032] In one specific embodiment, the dimensionality reduction algorithm based on streaming computing running in the production process sensing module is specifically a streaming PCA logic that combines principal component analysis algorithm with sliding time window.
[0033] Set the length of the sliding time window to The multidimensional physical environment parameter sequences collected within this time window constitute the original high-dimensional observation matrix. ,in This represents the total number of sensor dimensions connected (including mold temperature, melt pressure, etc.). The system calculates the mean vector of the data within this time window in real time. With covariance matrix : For covariance matrix Perform eigenvalue decomposition and select the first... The eigenvectors corresponding to the largest eigenvalues constitute the projection matrix. The system will display the latest set of process parameters at the current moment. Input the projection matrix to generate a low-dimensional state vector representing the stability of the current production state. : in, This is the reduced-dimensional state vector. The value ranges from 3 to 10. The system calculates the current state vector. With reference state vector Euclidean distance between As an offset, when the offset is greater than a preset offset threshold, the data upload mechanism is triggered.
[0034] When the offset of the state vector (obtained by calculating the Euclidean distance between the current vector and the reference vector) exceeds a preset offset threshold, it indicates that the injection molding process has drifted, which typically leads to subtle changes in the visual morphology of defects. The production process perception module immediately triggers a data upload mechanism to synchronize typical defect samples in the current state and the associated process state vectors to the cloud-based adaptive learning center, providing necessary external feature support for the online evolution of the model.
[0035] The cloud-based adaptive learning center serves as the brain and knowledge repository of the entire system, used to construct and maintain a meta-learning-based defect feature evolution model. Unlike traditional static recognition models, the meta-learning model in this system does not memorize the pixel features of specific defects, but rather learns the mapping function of feature extraction logic as process parameters change. The cloud-based adaptive learning center internally maintains a large-scale, dynamically updated industrial defect knowledge base, storing defect evolution sequences under thousands of different process combinations.
[0036] By deeply analyzing these sequences, the meta-learning model can establish a predictive mechanism: for example, when the mold temperature increases by 5 degrees Celsius, the system can predict that the geometric features of surface shrinkage will become smoother, and automatically adjust the weights of the edge detection operators in the model. Utilizing this meta-learning strategy of learning how to learn, the cloud-based adaptive learning center can quickly generate a set of optimized network adaptation parameters for the specific production state of the current edge-side feedback.
[0037] As a preferred implementation, the defect feature evolution model based on meta-learning built inside the cloud-based adaptive learning center specifically adopts a hypernetwork architecture to realize the mapping of feature extraction logic as process parameters change.
[0038] The super network The state vector uploaded by the production process sensing module As input, the convolutional kernel weight parameters required by the feature extraction backbone network in the edge-side multi-task inference terminal are directly output through multiple fully connected layers. : in, This represents the generated set of adaptation parameters, where $M$ represents the total number of parameters that need to be adapted. This represents the learnable parameters of the hypernetwork itself.
[0039] During the cloud training phase, hypernetwork parameters The optimization objective is to minimize the loss of the inner loop task under various historical process states. For the first [item] extracted from the industrial defect knowledge base... A combination of process status samples Its inner loop loss function Defined as: in, Represents the parameters generated by the hypernetic network. The constructed feature extraction network, For defect images, For defect labels, For cross-entropy loss, These are the base network parameters for pre-training. This is the regularization coefficient, used to constrain the generated parameters from drastic drift. The cloud-based system minimizes the sum of all historical sample combinations through backpropagation. To update the sum of the numbers .
[0040] The cloud-based adaptive learning center further utilizes a dynamic knowledge distillation algorithm. In the specific distillation process, the teacher network possesses a large number of parameters and a global perspective, enabling it to capture complex defect correlations. Upon receiving challenging samples or production state vectors uploaded from the edge, the teacher network performs targeted reinforcement inference in the cloud and generates soft labels containing rich semantic information. The system compresses and injects the discriminative power of the teacher network into lightweight student network parameters by minimizing the differences between the teacher network and student network in the output layer classification distribution and intermediate feature map activation distribution (e.g., by calculating KL divergence). The finally generated dynamic convolutional kernel weights or feature adaptation layer parameters are distributed to the edge-side multi-task inference terminal via the Industrial Internet of Things, enabling the edge to achieve real-time adaptation to newly emerging defect forms without changing hardware computing power.
[0041] Specifically, the cloud-based adaptive learning center employs a dynamic knowledge distillation architecture, whose loss function is composed of three weighted parts: hard label classification loss, soft label distillation loss, and intermediate feature layer alignment loss.
[0042] For the input defective image sample Teacher network (parameters are) ) and student network (parameters are Output respectively Previous logical vector and The system incorporates temperature parameters. The logic vector is softened to obtain a soft probability distribution. and : in, Indicates the defect category index. Indicates the total number of categories. Temperature hyperparameters greater than 1 are used to smooth the probability distribution to expose inter-class similarity information.
[0043] The total loss function of dynamic knowledge distillation Represented as: Among them, the first item Predicting probabilities for student networks With real labels Cross-entropy loss; second term The relative entropy (KL divergence) between the soft label distributions of the teacher network and the student network, multiplied by This is to ensure the consistency of gradient magnitude; the third term is the output feature map of a certain feature layer between the teacher network and the student network. and The squared loss of the Frobenius norm between them; , , These are the preset balanced weight coefficients. By minimizing this total loss function, the dynamic convolutional kernel weights are distributed to the edge side.
[0044] The edge-side multi-task inference terminal, deployed near the production line, uses a high-performance edge computing board as its hardware carrier to run a multi-task convolutional neural network that has undergone deep pruning and quantization. To resolve the conflict between high-precision detection and high-cycle production, the edge-side multi-task inference terminal adopts a parallel and tiered processing architecture.
[0045] The neural network logic architecture of the multi-task inference terminal consists of a feature extraction backbone network, an object detection branch, and a semantic segmentation branch. The feature extraction backbone network processes the physically enhanced high-contrast image output by the biomimetic compound eye programmable optical imaging subsystem. This backbone network utilizes a multi-scale feature fusion mechanism to extract semantic information of defects in different receptive fields, ensuring sensitivity to flaws of varying sizes.
[0046] The target detection branch performs a rapid scan across the entire image, and its core task is to coarsely locate suspected defect areas. Using a preset confidence threshold judgment logic, this branch outputs a series of bounding boxes containing defect type labels and spatial coordinate information. Subsequent depth processing logic is only triggered when the target detection branch determines a region of interest that is suspected of containing a defect.
[0047] The semantic segmentation branch embodies the design concept of local focus. It performs pixel-level depth calculations only on the patches within the bounding boxes output by the object detection branch. After obtaining the bounding box coordinates, the edge-side multi-task inference terminal uses a hardware acceleration unit to crop the corresponding local slices from the original high-resolution cached image. The segmentation subnetwork then classifies each pixel within the slice, accurately depicting the geometric contours of the defects. Through this logical jump from global coarse localization to local fine segmentation, the system avoids the huge computational overhead of pixel-level calculations on the entire image, enabling the overall detection speed to match the real-time shooting frequency of high-speed injection molding production lines.
[0048] Furthermore, the edge-side multi-task inference terminal also integrates an inference performance monitor. This monitor calculates the information entropy of the model's output results in real time. If, for a certain sample, the classification probability distribution given by the model is too flat (i.e., the entropy value is higher than a preset entropy threshold), it indicates that the current model has high uncertainty on that sample. The inference performance monitor will mark the sample as a difficult sample and automatically retain its original image and corresponding process parameter vector, feeding them back to the cloud-based adaptive learning center for targeted online incremental learning.
[0049] In one specific embodiment, the inference performance monitor of the edge-side multi-task inference terminal quantifies the detection determinism by calculating the Shannon information entropy of the model prediction results.
[0050] For any input image region to be detected, the pixel-level classification probability distribution output by the semantic segmentation branch is as follows: ,in (0 represents a normal pixel, 1 represents a defective pixel). This is a measure of macroscopic uncertainty in the region. Calculated using the following formula: in, This represents the binary classification information entropy of the classification result for that region. When the model is highly confident in its defect determination (e.g., ...), ... (close to 1 or 0) Approaching 0; when the model cannot distinguish whether the region is a defect or normal texture (e.g. (close to 0.5) It reached its maximum value of 1.
[0051] The inference performance monitor will calculate the information entropy. Compared with the preset uncertainty measurement threshold (For example, set to 0.85) Perform a comparison. If satisfied... If the condition is met, the sample is determined to be a difficult sample, triggering the encapsulation and upload logic.
[0052] When performing logical judgments, the system adheres to strict textual description guidelines for all calculation processes. For example, when determining the compensation parameters of the light-emitting unit, the system does not simply use algebraic expressions, but describes it as follows: calculate the angular relationship between the incident light vector emitted by the light-emitting unit and the surface normal vector at the point to be inspected. If the consistency of the two vector directions meets the preset geometric conditions for reflection into the lens, then the power supply duty cycle of the light-emitting unit is reduced proportionally. During the final defect judgment, when the calculated semantic association score of the defect area is numerically greater than the preset quality and safety threshold, and the geometric area continuity of the area meets the minimum defect judgment area standard, the system determines that the injection-molded shell is a non-conforming product.
[0053] Example 2: Building upon Example 1, this example provides a deep learning-based visual defect detection system for injection-molded housings of high-end equipment, featuring a distributed collaborative architecture, specifically designed for large-scale clustered production environments. In this alternative architecture, the various modules of the system are further decoupled and optimized in terms of physical deployment and task allocation to accommodate the parallel inspection needs of multiple injection molding production lines.
[0054] A distributed deep learning visual defect detection system includes a clustered bionic compound eye imaging array, an industrial bus gateway module, a distributed computing resource pool, and a centralized cloud decision center.
[0055] The clustered bionic compound eye imaging array is deployed at the end of each independent injection molding production line. Unlike Embodiment 1, the imaging subsystem here adopts a modular design, with each imaging unit having an independent field-programmable gate array (FPGA) chip. The FPGA chip is configured to perform real-time image preprocessing and inverse light field preliminary compensation logic. This means that during the calculation of light field modulation, the local FPGA can autonomously adjust the flicker phase of the micro-LED array based on the shell curvature sensed in real time by the production line. This design reduces the system's real-time dependence on the central controller, ensuring that the optical enhancement effect at the physical imaging level can still be stably maintained even under network congestion.
[0056] The industrial bus gateway module acts as a production process sensing component. In this embodiment, the module connects to the factory-level Manufacturing Execution System (MES) via a high-performance real-time Ethernet protocol (such as EtherCAT or Profinet). It not only collects operating parameters of individual injection molding machines but also acquires raw material batch numbers, drying humidity data from the upstream material supply system and quality feedback data from the downstream assembly line. The industrial bus gateway module utilizes an anomaly detection logic based on a variational autoencoder to compress complex end-to-end production data into a global state vector characterizing the robustness of the production environment.
[0057] In the distributed architecture described in Example 2, the anomaly detection logic of the industrial bus gateway module based on the variational autoencoder (VAE) is specifically manifested as follows: The gateway module inputs end-to-end production data into the VAE's encoder network. , high-dimensional data Compressed into low-dimensional latent variables Subsequently, through the decoder network Reconstruct the original data The system constructs anomaly scores by calculating the mean squared error between the original and reconstructed data and combining it with the KL divergence in the latent space. : in, This is a reconstruction error term used to measure whether there are abrupt changes or anomalies in the data that deviate from normal process rules; The KL divergence between the latent variable distribution and the standard normal distribution is used to constrain the regularity of the latent space. and These are the network parameters for the encoder and decoder, respectively. When... When the dynamic baseline is exceeded, it is determined that the state vector has drifted.
[0058] When a state vector drift occurs on one of the multiple production lines, the industrial bus gateway module can identify whether this drift is due to hardware aging of a single device (such as screw wear) or fluctuations in the composition of the entire batch of raw materials (such as excessive moisture). Based on this differentiated identification result, the gateway module will issue different detection task priorities to the distributed computing resource pool.
[0059] The distributed computing resource pool integrates the computing power resources of multiple edge computing servers within the factory. In this embodiment, the edge-side multi-task inference terminal is no longer a standalone configuration, but a shared virtualized computing power pool. When the density of parts to be inspected on production line A is high, the system will automatically allocate computing power resources from the idle production line B, and execute the operations of the defect detection branch and semantic segmentation branch through load balancing technology.
[0060] The lightweight model running within the distributed computing resource pool possesses multi-scale feature alignment capabilities. Considering the potential slight differences in installation angles across different production lines, the geometric correction submodule within the resource pool automatically compares calibration plate images from different imaging units, calculates the perspective transformation matrix, and inversely applies it to the feature extraction process, ensuring that the same model parameters have consistent physical scale representation capabilities across different production lines.
[0061] The centralized cloud decision center performs higher-level meta-learning tasks. In the distributed architecture of this embodiment, the cloud center introduces a joint learning mechanism. It does not directly collect every production image, but rather collects the gradient increments of the models trained by each distributed node. By securely aggregating and globally optimizing these gradients in the cloud, the centralized cloud decision center can learn the defect evolution patterns across production lines. For example, when production line 1 experiences minor silver streaks due to increased environmental humidity, the adaptation parameters generated by the cloud center are not only distilled into the local model of production line 1, but also pre-pushed to production lines 3 and 5, which are under similar environmental conditions, achieving preventative model updates to address defects.
[0062] At the hardware implementation level, each inference node in the distributed computing resource pool is configured with a hardware-level model security isolation zone. During the dynamic knowledge distillation process, the issued parameter packets are encrypted and verified to ensure the security of sensitive information such as production state vectors during transmission. The inference node also includes redundancy detection logic: when the main model's judgment result for a key injection-molded shell (such as an aerospace-grade precision structural component) is at a critical value, the system will automatically call a backup model with completely different parameter distributions for cross-comparison. If the difference in classification results given by the two models (measured in the form of cross-entropy) is greater than a preset collaborative judgment threshold, the system will automatically suspend the automatic release permission for that product, switch to manual intervention mode, and simultaneously trigger the reinforcement learning program in the cloud center.
[0063] Through this distributed collaborative architecture, the system can achieve comprehensive coverage of ultra-large-scale injection molding plants with lower single-line deployment costs. It solves the problem of uneven resource allocation through computing power pooling and improves the model's resilience to sudden process fluctuations through cross-production line joint learning.
[0064] Example 3: This example provides a deep learning-based visual defect detection system for injection-molded shells with ultra-high transparency and multi-layered composite structures. Based on Example 1, this system focuses on enhancing the optical sensing subsystem's ability to capture depth information and the algorithm terminal's accuracy in identifying internal defects.
[0065] In this embodiment, the biomimetic compound eye programmable optical imaging subsystem adds a polarization modulation unit. The polarization modulation unit is located at the front end of the micro-programmable light-emitting unit array and at the front end of the lens of the multispectral sensing component. Because highly transparent injection-molded parts generate complex internal stress distributions during the injection molding process, these stresses exhibit interference fringes (i.e., stress birefringence) under polarized light.
[0066] In the detection scheme for high-transparency composite structures in Example 3, the polarization modulation unit acquires four images (typically 100 images) under different polarization directions. , , , (Linear polarization state), calculate the Stokes vector at each pixel. : in, , , , These represent the measured light intensity values under the corresponding polarization directions. (This involves circular polarization components; if only linear polarization detection is used in this scheme, then the values are set to zero).
[0067] Furthermore, the system calculates the degree of polarization based on the Stokes vector. (Degree of Polarization) to characterize the optical anisotropy caused by internal stress birefringence: in, The value range is [0, 1]. In regions with internal microcracks or residual stress, the polarization state changes, leading to changes in the polarization state of that region. The value is significantly different from that of the surrounding normal matrix material. The feature map, as an additional input channel, is stitched together with the regular image acquired by the multispectral sensing component in the channel dimension and then input into the 3D convolution branch of the edge-side multi-task inference terminal for subsequent inference.
[0068] In this embodiment, the reverse light field controller is further configured to perform polarization state switching. It can instruct the light-emitting array to sequentially output mutually orthogonal linearly polarized light or circularly polarized light. The multispectral sensing component acquires four images under different polarization directions by using a rotating polarization filter or a focal plane polarization imaging technique. By calculating the polarization degree component in the Stokes vector, the system can extract subsurface damage features caused by stress concentration, beyond the geometric information of the outer shell surface. This physical-level feature enhancement enables the system to identify tiny cracks or bubbles that are completely invisible under normal lighting and lurking inside the injection-molded part.
[0069] Accordingly, the neural network architecture in the edge-side multi-task inference terminal has undergone targeted topology improvements. The feature extraction backbone network has added a parallel branch specifically for processing polarization feature maps. This branch employs a three-dimensional convolutional structure, treating two-dimensional image sequences at different polarization angles as temporal signals for feature extraction. In this way, the neural network can learn the deep correlation between the morphological evolution of stress birefringence fringes and injection molding process parameters (such as holding pressure and cooling rate).
[0070] In this embodiment, the production process sensing module focuses on monitoring mold venting pressure and melt front velocity. For composite material injection molding, these parameters directly determine the probability of internal defect formation. The module integrates a physical auxiliary reasoning unit, which uses a preset fluid dynamics model to simulate the current production state. If the simulation results show that the probability of stagnation or gas entrapment under the current operating conditions is higher than a preset risk threshold, the physical auxiliary reasoning unit generates a high-risk guidance signal.
[0071] The guiding signal is transmitted to the edge-side multi-task inference terminal as input to the attention mechanism. This allows the neural network to automatically allocate more computational resources to high-risk areas predicted by the fluid dynamics model during full-image scanning. This physics-driven visual attention mechanism improves the system's capture rate of complex, hidden defects and maintains the system's efficient operating cycle by reducing redundant computations on low-risk areas.
[0072] In this embodiment, the cloud-based adaptive learning center performs the task of digital twin creation. It utilizes a vast amount of collected process parameters and polarization defect images to construct a digital mapping system between the injection molding process and internal stress defects. When a new model of casing product undergoes trial production, the cloud-based center can, based on knowledge accumulated through meta-learning, simulate and generate virtual defect sample features using a physics engine, even without real defect samples. These simulated features are then transformed into initial weights for the edge-side model using dynamic knowledge distillation technology. This enables the system to possess preliminary defect detection capabilities from the moment the production line starts, shortening the quality ramp-up period after the high-end new product is put into production.
[0073] Example 4: This example describes a deep learning-based visual defect detection system with self-diagnosis and environmental self-healing capabilities. This solution primarily addresses the long-term stability issues under extreme industrial environments (such as high temperature, vibration, and dust interference), and significantly enhances the interactive feedback mechanism of each module within the system.
[0074] In this scheme, the biomimetic compound eye programmable optical imaging subsystem incorporates an optical integrity monitoring unit. This unit utilizes reference light-emitting points in a micro-programmable light-emitting unit array to periodically project a standard calibration light field onto the multispectral sensing component. By analyzing the brightness distribution consistency and spectral characteristics of the calibration image output by the sensing component, the system can determine in real time whether the optical component is currently affected by external oil contamination, dust obstruction, or lens optical axis misalignment due to environmental vibration.
[0075] If optical performance degradation is detected, the reverse light field controller automatically executes a compensation procedure. For example, if it is determined that there is a decrease in light transmittance due to dust in the upper left corner of the image, the controller will proportionally increase the initial power of the light-emitting unit at the corresponding angle in that area to maintain the consistency of the feature contrast at the imaging level. This physical self-healing capability ensures that the maintenance cycle of the system is extended in harsh industrial environments.
[0076] The edge-side multi-task inference terminal introduces a robustness evaluation branch. While executing the main task, this branch calculates the perceptual quality score of the input image. If the image quality deteriorates due to external environmental factors (such as smoke generated by a depressurizing injection molding machine), causing a sudden surge in the uncertainty entropy value given by the inference performance monitor, the edge-side terminal immediately switches to a low-power / high-robust mode. In this mode, the system sacrifices some pixel-level segmentation accuracy by increasing the receptive field and enabling spatial correlation filtering to ensure that at least the defect detection capability of the core area is maintained, preventing large-scale missed detections on the production line due to environmental interference.
[0077] In this embodiment, the cloud-based adaptive learning center executes environment-related collaborative optimization logic. It analyzes the relationship between production environment data from different time periods and seasons and the performance of the detection model. Through this long-term data analysis, the cloud center can generate timely model parameter packages. For example, in high-humidity summer environments, the cloud center will push student network parameters enhanced with humidity features in advance to compensate for the impact of humidity fluctuations on optical contrast and sensor noise levels.
[0078] In this embodiment, the production process sensing module integrates vibration and acoustic sensors. By monitoring the acoustic signature of the injection molding machine during operation, it can predict the decreasing trend of molding accuracy due to mechanical wear. This advanced sensing signal is converted into prior weights for model inference. For example, when acoustic analysis determines that the mold closing pressure may be insufficient, the system automatically lowers the threshold for judging flash defects in the edge model, achieving extremely sensitive capture of potential quality risks.
[0079] The closed-loop architecture adopted in this embodiment organically combines the status monitoring of optical components, the environmental robustness adjustment of the algorithm, and the long-term regularity learning in the cloud. The system exhibits extremely high reliability and consistency under the harsh working conditions of high-end manufacturing, and can provide full life-cycle technical support for zero-defect manufacturing of high-end equipment.
[0080] Example 5: This example discloses a deep learning-based visual defect detection system for large-size, highly complex injection-molded shells, employing field-of-view fusion and logical correlation. For ultra-large areas of the surface to be inspected, this system uses multiple biomimetic compound-eye programmable optical imaging subsystems arranged spatially to form an omnidirectional, blind-spot-free sensing array.
[0081] Each imaging subsystem is responsible for a specific spatial partition of the shell to be inspected. To address the issue of continuity in detection at the joints between partitions, the edge-side multi-task inference terminal incorporates a spatial consistency alignment module. This module utilizes the overlapping fields of view between adjacent imaging subsystems to extract common geometric feature points (such as positioning holes and reinforcing rib edges). By performing affine transformations and spatial stitching on the feature layers, the system constructs an ultra-high-dimensional global feature space within its internal logic.
[0082] The target detection branch employs a hierarchical filtering and global fusion strategy when processing such a large-scale feature space. The system extracts primary features within each local field of view and uses lightweight cross-view attention logic to identify long-sized defects (such as long scratches) that span two or more fields of view. This design avoids the common problem of misjudging fracture defects in traditional block detection, ensuring consistent evaluation across the entire system.
[0083] In this embodiment, the semantic segmentation branch supports dynamic resolution scheduling. When the target detection branch determines the presence of minute, highly complex defects (such as micrometer-level stress cracks) in a certain region, the edge-side terminal instructs the imaging subsystem corresponding to that region to enter a microscopic scanning mode. In this mode, the micro-programmable light-emitting unit array performs high-frequency scintillation coding, while the multispectral sensing component acquires a super-resolution image exceeding the physical limits of the sensor through sub-pixel displacement technology. The semantic segmentation subnetwork performs high-precision contour drawing on this super-resolution slice, achieving an optimal balance between local detail recognition accuracy and global scanning efficiency.
[0084] In this embodiment, the production process sensing module establishes a mapping model between spatial location and process parameters, enabling precise tracing of the causes of specific defects. For example, when the detection system frequently detects shrinkage defects within a specific field of view, the production process sensing module automatically retrieves the hot runner temperature probe data corresponding to that area of the injection mold. Through dual correlation analysis in both time and spatial dimensions, the system can not only determine the product's conformity but also output process optimization suggestions to the production management system, such as suggesting a reduction in the heating power of the third hot runner nozzle.
[0085] The cloud-based adaptive learning center employs a distributed expert network architecture for such large and complex systems. Specialized testing models for different materials (such as ABS, PC, and carbon fiber reinforced composites) are stored in the cloud. When the production line switches production tasks, the cloud center automatically sends the most suitable expert model parameters to the edge based on the raw material codes uploaded by the production process perception module. This on-demand loading mode allows the system to flexibly adapt to multi-variety, small-batch production modes, improving the versatility and response speed of the testing system.
[0086] Through this multi-subsystem collaboration and field-of-view fusion design, the present invention can meet the growing demand for inspection of ultra-large and complex structural components in high-end equipment manufacturing, achieving efficient online full inspection of large workpieces while ensuring sub-micron level recognition accuracy.
[0087] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0088] At the software architecture level, the various neural network models, adaptive learning algorithms, and feature extraction logic involved in this invention can be flexibly deployed on different hardware platforms according to actual computing power conditions and detection accuracy requirements. All parameter thresholds, time constants, and weight adjustment step sizes in the system can be optimized through the system's built-in automatic calibration program during initial equipment installation or major process changes.
[0089] In actual operation, the detection accuracy and processing speed exhibited by this system depend not only on the advancement of the algorithm itself, but also on the biomimetic compound eye-like optical enhancement technology at the physical level. This integrated hardware and software, physical compensation design approach is the core guarantee that this invention can break through the limitations of traditional machine vision and achieve accurate perception of minute defects.
[0090] The communication logic between the cloud-based adaptive learning center and the edge-side multi-task inference terminal in this invention employs a highly secure data encapsulation and asynchronous transmission mechanism to ensure the real-time nature and privacy of production data. With the continuous development of industrial internet technology, this system can seamlessly connect to larger-scale industrial intelligent platforms, enabling cross-factory and cross-regional quality data collaboration and production process optimization.
[0091] In summary, this invention improves the sensitivity, robustness, and intelligence of visual inspection of injection-molded shells for high-end equipment by constructing an efficient feature transfer and feedback closed loop between the physical and digital worlds. This has significant engineering implications and application value for promoting quality upgrades in the precision manufacturing industry.
Claims
1. A deep learning-based visual defect detection system for injection-molded housings of high-end equipment, characterized in that, include: The biomimetic compound eye programmable optical imaging subsystem is used to detect the surface morphology of the injection-molded shell under inspection. It dynamically generates a reverse compensation light field based on the acquired surface curvature data and achieves contrast enhancement of defect features and background noise removal at the physical imaging level. The production process sensing module is used to monitor the physical environment parameters and equipment operating status in the injection molding production line in real time, extract feature vectors that can characterize the fluctuation of production status, and align the feature vectors with image data using timestamps. The cloud-based adaptive learning center is used to build a defect feature evolution model based on meta-learning. By analyzing the morphological feature changes of historical defect samples under different process parameters, it generates network adaptation parameters for the current production state. The edge-side multi-task inference terminal is used to run a deep neural network that has undergone pruning and quantization, receive dynamic knowledge distillation results from the cloud-based adaptive learning center, and perform accurate localization and pixel-level segmentation of defect regions on the enhanced physical feature image.
2. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 1, characterized in that: The biomimetic compound eye programmable optical imaging subsystem includes a micro-programmable light-emitting unit array, a multispectral sensing component, and a reverse light field controller. The micro-programmable light-emitting unit array consists of thousands of independently controlled semiconductor point light sources, and the semiconductor point light sources are non-uniformly spatially distributed according to the topological structure of a spherical canopy, so as to achieve full coverage illumination in a spherical or hemispherical space. The multispectral sensing component has wavelength selective acquisition function and has a built-in electronically tunable filter for forming a spectral closed loop with the micro-programmable light-emitting unit array. The reverse light field controller is connected to the micro-programmable light-emitting unit array and the multispectral sensing component, and is used to calculate the reflected energy distribution of each light-emitting point on the sensor surface according to the digital elevation model of the injection-molded shell to be inspected. The reverse light field controller independently adjusts the luminous intensity, spectral frequency, and flicker phase of each semiconductor point light source. In areas where the curvature of the injection-molded shell surface is greater than a preset curvature threshold and there is a risk of high light overflow, it reduces the output power of the corresponding angle light source. At the same time, it enhances the intensity of multi-angle cross illumination in concave areas where the light is weaker than a preset brightness threshold. It eliminates high light noise in the image through physical light field superposition.
3. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 2, characterized in that: The reverse light field controller in the bionic compound eye programmable optical imaging subsystem executes the following decision logic: calculates the incident direction vector of each light-emitting unit at the sampling point on the surface of the injection-molded shell to be inspected, and calculates the cosine value of the angle between the incident direction vector and the surface normal vector based on the obtained surface normal vector; If the predicted reflection energy value after conversion by Fresnel reflection law is greater than the preset pixel saturation threshold, the reverse light field controller commands the corresponding light-emitting unit to reduce its driving current intensity, or shortens its light-emitting duty cycle through pulse width modulation logic. When detecting dark or highly reflective injection-molded shells, the reverse light field controller instructs the micro-programmable light-emitting unit array to switch to the near-infrared band spectral output, and the multispectral sensing component simultaneously activates the corresponding narrowband filter logic. By utilizing the scattering characteristics of long-wavelength light on micro-defects, the optical characterization capability of surface micron-level scratches, stress marks, and orange peel texture is enhanced.
4. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 3, characterized in that: The production process sensing module includes multiple physical sensor interfaces for real-time access to data such as mold temperature, melt pressure change over time, pressure distribution during the holding pressure stage, cooling water flow rate, and ambient humidity. The production process sensing module runs a dimensionality reduction algorithm based on streaming computing, which maps the physical environment parameters and equipment operating status data to a low-dimensional feature space to generate a state vector characterizing the stability of the production state. The production process sensing module also executes timestamp alignment logic to encapsulate the process features of the injection-molded shell during the molding cycle into a metadata data package bound to the image identifier; When the spatial geometric distance between the state vector and the reference vector is greater than a preset offset threshold, the production process perception module triggers a data upload mechanism to transmit typical defect samples and associated process state vectors under the current production state to the cloud adaptive learning center.
5. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 4, characterized in that: The cloud-based adaptive learning center maintains a dynamically updated industrial defect knowledge base, which stores defect evolution sequences under different injection molding process combinations. The meta-learning model running in the cloud-based adaptive learning center is configured to learn the mapping function of feature extraction logic as process parameters change. By analyzing the defect evolution sequence, it predicts the impact of injection molding process fluctuations on the clarity, contrast, and geometry of defect edges. Upon receiving the state vector transmitted by the production process perception module, the meta-learning model generates a set of optimal convolution kernel weight parameters or feature adaptation layer parameters within a preset feature space.
6. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 5, characterized in that: The cloud-based adaptive learning center adopts a dynamic knowledge distillation architecture, including a teacher network with a higher number of parameters than the edge-side multi-task inference terminal. The teacher network performs reinforcement reasoning based on the state vector and generates soft labels containing semantic information. The cloud-based adaptive learning center compresses the discriminative power of the teacher network and injects it into the parameters of the student network by minimizing the relative entropy difference between the teacher network and the student network of the edge-side multi-task inference terminal in terms of output layer classification distribution and intermediate feature layer activation distribution. The final generated dynamic convolutional kernel weights are distributed to the edge-side multi-task inference terminal via the Industrial Internet of Things.
7. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 6, characterized in that: The edge-side multi-task inference terminal adopts a parallel and tiered processing architecture, and its running deep neural network includes a feature extraction backbone network, an object detection branch, and a semantic segmentation branch. The feature extraction backbone network is used to directly process the physically enhanced image output by the biomimetic compound eye programmable optical imaging subsystem, and to extract semantic information of defects in different receptive fields using multi-scale feature fusion logic. The target detection branch is used to perform scanning across the entire map. By comparing the calculated probability of the presence of defects with a preset confidence threshold, it filters out regions of interest with suspected defects and outputs the corresponding bounding box spatial coordinates. The semantic segmentation branch is configured to perform local processing only on the region of interest output by the object detection branch, in order to reduce computational resource consumption.
8. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 7, characterized in that: After obtaining the bounding box spatial coordinates, the edge-side multi-task inference terminal uses a hardware acceleration unit to crop out the corresponding local map patch from the original high-resolution cached image and inputs it into the semantic segmentation branch. The semantic segmentation branch includes a segmentation subnetwork for classifying each pixel within the local patch, thereby accurately depicting the geometric outline of the defect. When performing logical judgment, if the calculated semantic association score of the defect area is greater than the preset quality and safety threshold, and the geometric area continuity of the area meets the minimum defect judgment area standard, then the edge-side multi-task reasoning terminal determines that the injection molded shell under inspection has a defect of a certain type and outputs an unqualified signal.
9. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 8, characterized in that: The edge-side multi-task inference terminal also integrates an inference performance monitor to evaluate the detection determinism of the current model; The inference performance monitor calculates the information entropy of the model output in real time. If the uniformity of the classification probability distribution for a certain sample is greater than the preset uncertainty measurement threshold, the inference performance monitor marks the certain sample as a difficult sample and feeds back the difficult sample and its corresponding process state vector to the cloud adaptive learning center. Upon receiving challenging samples, the cloud-based adaptive learning center initiates an online incremental learning program. It then uses an active learning strategy to select representative unlabeled data from the sample pool and guides the cloud-based model to fine-tune its parameters.
10. The deep learning visual defect detection system for injection-molded housings of high-end equipment according to claim 9, characterized in that: The system also includes an optical integrity monitoring unit, which is used to project a standard calibration light field onto the multispectral sensing component using a reference light-emitting point in the micro-programmable light-emitting unit array; By analyzing the brightness distribution consistency and spectral characteristics of the calibration image output by the sensing component, it can be determined whether the optical component is affected by external contaminant adhesion or environmental vibration causing optical axis displacement. If the decrease in transmittance of the image area meets the preset compensation condition, the reverse light field controller increases the initial power of the light-emitting unit at the corresponding angle of the image area proportionally to maintain the consistency of feature contrast at the physical level. The system combines physical-level self-healing compensation, algorithm-level environmental robustness adjustment, and cloud-based long-term pattern learning to construct a closed-loop evolution architecture for fluctuations in the injection molding production environment.