Industrial design appearance defect intelligent detection system based on computer vision

Abstract

The invention discloses an industrial design appearance defect intelligent detection system based on computer vision, and relates to the technical field of computer vision and industrial intelligent detection.During operation of the system, multi-angle space information and multispectral imaging data of the surface of a to-be-detected product are obtained through a preset structured light source module and an array type industrial camera cluster; the method comprises the following steps: reconstructing high-dimensional visual feature distribution on the surface of a product, executing noise suppression and fusion, generating an enhanced feature map with a high signal-to-noise ratio, receiving the enhanced feature map, identifying defects by utilizing an attention mechanism, generating defect candidate areas and a preliminary classification probability, reducing the calculated amount through knowledge distillation, and realizing few-sample migration; and deploying the optimized detection model to a high-parallel hardware acceleration platform, resolving a detection result in real time, and sending a sorting instruction, an alarm signal or a process parameter correction instruction to a production execution system according to a preset quality judgment threshold.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and industrial intelligent inspection technology, specifically to an intelligent inspection system for industrial design appearance defects based on computer vision. Background Technology

[0002] In the development of intelligent manufacturing and industrial automation, the appearance quality control of industrial products is a crucial component of the production process. With the evolution of computer vision technology, image processing-based automated inspection methods have gradually replaced traditional manual visual inspection, becoming the core technological support for ensuring product qualification rates. This field encompasses multiple technical branches, including image acquisition, feature extraction, pattern recognition, and automated execution, playing a vital role in improving the automation level of production lines, enhancing product consistency, and reducing labor costs in the manufacturing process.

[0003] Among them, the intelligent detection technology for industrial design appearance defects based on computer vision focuses on using high-resolution imaging systems and deep learning algorithms to achieve automated identification and refined classification of minute defects on product surfaces. The basic principle of this technology is to capture multi-dimensional visual information of the product surface using an industrial camera, and combine this with a predefined defect feature model or an end-to-end deep neural network architecture to quickly locate and quantitatively analyze abnormal states such as foreign objects, color differences, deformations, and structural damage in the images. Its goal is to replace the traditional manual sampling inspection mode with high-frequency, non-contact visual scanning, thereby establishing an intelligent quality monitoring system covering the entire production process.

[0004] Existing technologies still exhibit significant limitations in visual inspection under complex industrial environments. The drastic fluctuations in lighting conditions at industrial sites, along with the high reflectivity and interference characteristics of complex product surfaces, often result in low signal-to-noise ratios in acquired images, making even subtle defect features easily obscured by background noise. While deep learning technology introduces stronger feature representation capabilities, existing detection frameworks suffer from significant missed detections when handling extremely small defects and lack flexibility for multi-specification, multi-form product lines, necessitating extremely high costs for re-annotation and migration when facing product upgrades. Furthermore, there is a significant contradiction between the high computational complexity of large-scale deep learning algorithms and the extremely short cycle times required by industrial production lines. Coupled with the difficulty in acquiring defect samples during production, leading to extreme imbalances in sample distribution, it becomes difficult to achieve synergistic optimization of the over-detection rate and false detection rate under complex conditions, severely restricting the practical deployment effectiveness of defect detection systems in precision manufacturing. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent detection system for industrial design appearance defects based on computer vision, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: an intelligent detection system for industrial design appearance defects based on computer vision, comprising a multi-dimensional light field perception unit, an adaptive image enhancement unit, a multi-scale defect representation neural network, a knowledge transfer and model compression unit, and an edge computing and closed-loop control unit; The multi-dimensional light field sensing unit is used to acquire multi-angle spatial information and multi-spectral imaging data of the surface of the product to be inspected through a preset structured light source module and an array of industrial cameras. The structured light source module uses pulse width modulation technology to dynamically adjust the light intensity of different spectral bands at the microsecond level, and works with a synchronous trigger controller to drive the array of industrial cameras to perform synchronous acquisition at the nanosecond level, so as to reconstruct the high-dimensional visual feature distribution of the product surface. An adaptive image enhancement unit is used to perform a joint spatial and frequency domain transformation on the original image acquired by the multi-dimensional light field sensing unit to filter out high-frequency pulse noise, and to perform pixel-level fusion of multiple frames of images under different exposure parameters through a generative adversarial network. At the same time, it combines automatic gain control logic to compensate the average gray value of the image in real time to generate an enhanced feature map with a high signal-to-noise ratio. A multi-scale defect characterization neural network is used to receive the enhanced feature map, extract multi-scale features through a multi-level feature pyramid architecture with receptive field enhancement modules, identify and locate geometric distortions, surface scratches, color differences and structural damage on the product surface using a dual attention mechanism of space and channel, and enhance the identification of rare defect categories through a weighted cross-entropy loss function, thereby generating defect candidate regions and preliminary classification probabilities. The knowledge transfer and model compression unit is used to perform knowledge distillation processing on the multi-scale defect representation neural network by using a preset teacher network, which introduces intermediate feature mapping similarity loss, and compresses the number of model parameters to less than 15% of the original size with weight pruning technology. At the same time, it combines a few-shot learning algorithm based on contrastive learning to achieve domain adaptation of defect features of new product specifications. Edge computing and closed-loop control unit are used to deploy the optimized detection model to the heterogeneous computing acceleration platform. The floating-point operation is converted into 8-bit integer operation through model quantization technology to solve the detection results in real time. The process trend prediction algorithm based on Markov chain is used to calculate the process state transition probability. Then, according to the preset quality judgment threshold, sorting instructions, alarm signals or process parameter correction instructions are sent to the production execution system.

[0007] Preferably, the multi-dimensional light field sensing unit includes a ring array light source composed of multi-color temperature light-emitting diodes, a high-resolution area array camera located at the center of the axis, and several tilting observation cameras distributed around it; the ring array light source realizes dynamic adjustment of light intensity of different spectral bands through pulse width modulation technology to ensure that diffuse reflection supplementary lighting of highly reflective material surfaces is completed within an exposure time of 10 microseconds; the array-type industrial camera cluster adopts a synchronous trigger controller to realize synchronous acquisition of multiple cameras at the nanosecond level, thereby obtaining the complete envelope visual information of the product under inspection in 3D space.

[0008] Preferably, the adaptive image enhancement unit filters out high-frequency pulse noise in the acquired image by performing a joint transformation of the spatial and frequency domains; it uses a deep learning-based generative adversarial network structure to perform pixel-level fusion of multiple frames of images under different exposure parameters, preserving the details of defect edges while eliminating the light spot masking phenomenon caused by metal or glass materials; the image enhancement algorithm integrates automatic gain control logic, which can calculate the compensation coefficient in real time based on the feedback data of the ambient light sensor, and stabilize the average gray value of the image in the ideal range of about 128.

[0009] Preferably, the multi-scale defect representation neural network adopts a multi-level feature pyramid architecture, which effectively captures extremely small defects with a size of less than 0.1 mm by attaching receptive field enhancement modules to feature layers of different resolutions. The network embeds a dual spatial and channel attention mechanism, which guides the system to ignore the interference of background texture by assigning higher weight coefficients to pixel regions containing potential defects. In order to deal with the problem of extreme imbalance in the distribution of positive and negative samples, the network adopts a weighted cross-entropy loss function, which significantly improves the sensitivity of rare defect categories.

[0010] Preferably, the knowledge transfer and model compression unit removes redundant weight connections in the multi-scale defect representation neural network through pruning technology, compressing the number of model parameters to less than 15% of the original size. This unit utilizes a pre-training mechanism based on contrastive learning to extract general feature expressions on a small number of unlabeled samples, enabling the system to complete model fine-tuning and deployment within 4 hours when facing product upgrades by acquiring only 50 or fewer defect samples, greatly reducing the cost of flexible transformation of the production line.

[0011] Preferably, the edge computing and closed-loop control unit is based on a heterogeneous computing architecture of field-programmable gate arrays or graphics processors. It uses model quantization technology to convert floating-point operations into 8-bit integer operations, so that the single-piece inspection time is controlled within 30 milliseconds, which meets the high-speed cycle requirement of 2,000 pieces per minute of the production line. The unit also integrates a process trend prediction algorithm based on Markov chains. Through statistical analysis of continuous inspection results, when the probability of a specific defect occurs shows a monotonically increasing trend, it automatically sends compensation control instructions to the front-end process.

[0012] It also includes a centralized defect feature fingerprint database for storing encrypted defect image feature vectors; this database is connected to multiple production bases through an industrial internet platform to achieve cross-regional collaborative evolution and knowledge sharing of defect patterns, ensuring strict consistency of quality evaluation standards under a globalized production system.

[0013] Preferably, the multi-scale defect characterization neural network passes through an uncertainty estimation module before outputting the detection results. This module evaluates the confidence level of the detection conclusion using Monte Carlo sampling technology. When the system's confidence in judging a suspected defect is lower than 90%, it will automatically forward the image to a manual review terminal for remote calibration. The calibration results will be fed back into the model update process in real time as new evidence for online learning.

[0014] The hardware layer of the edge computing and closed-loop control unit adopts a dual-redundancy architecture. A heartbeat monitoring mechanism ensures that when the main computing node fails, the backup node can take over within 50 milliseconds, ensuring uninterrupted operation of the production line. A real-time operating system is deployed at the bottom layer of the system. Through a priority preemption scheduling mechanism, it ensures that visual acquisition tasks and defect judgment tasks have the highest execution authority, avoiding jitter caused by background processes affecting detection delays.

[0015] Preferably, when processing industrial products with highly repetitive textures, the adaptive image enhancement unit enables a texture removal algorithm based on Fourier transform. This algorithm identifies and blocks periodic frequency components representing normal textures in the spectral space, and recovers a residual image containing only non-periodic defects through inverse transform, thereby increasing the contrast to more than 5 times that of the initial state.

[0016] Preferably, when performing knowledge distillation, the knowledge transfer and model compression unit introduces soft label cross-entropy and intermediate feature mapping similarity loss. By simulating the probability distribution of the teacher network at the defect edge, the lightweight student network can obtain sub-pixel-level positioning capabilities similar to those of a large model. This process is executed asynchronously on an independent computing power cluster and does not occupy real-time detection resources in the production site.

[0017] Preferably, the structured light source module of the multi-dimensional light field sensing unit also includes a set of polarization illumination components. By adjusting the relative angle between the polarizer and the analyzer, the specular glare on the product surface caused by Fresnel reflection is eliminated, and the contrast characteristics of deep scratches and shallow foreign objects are significantly enhanced.

[0018] This invention provides an intelligent detection system for industrial design appearance defects based on computer vision, which has the following beneficial effects: (1) During system operation, by constructing an intelligent detection system that combines multi-dimensional light field perception and multi-scale defect characterization, high-precision identification and stable evaluation of complex industrial design appearance defects are achieved; by introducing a nonlinear comprehensive defect risk index and a secondary optimization analysis mechanism, the false detection rate caused by material reflection differences and light fluctuations is effectively reduced; at the same time, by combining knowledge transfer and edge computing deployment methods, the system response speed and cross-product specification adaptability are significantly improved while ensuring detection accuracy, thereby improving the automation level of production line appearance quality control and process closed-loop adjustment capability.

[0019] (2) This invention solves the imaging problem in industrial sites with high reflectivity and strong interference by working together with the multi-dimensional light field sensing unit and the adaptive image enhancement unit at both the physical and algorithmic levels. The fusion of multi-angle spatial information and multi-spectral data enables the system to break through the limitations of traditional two-dimensional vision, extract weak defect features from multiple dimensions, and improve the signal-to-noise ratio under complex working conditions by more than 3 times, laying a solid data foundation for subsequent high-precision recognition.

[0020] (3) The multi-scale defect representation neural network used in this invention combines a global attention mechanism and a weighted loss function, which fundamentally solves the problem of detection bias caused by the missed detection of extremely small defects and sample imbalance. By performing depth weighting at the pixel level and feature level, the system can achieve stable capture of defects at the 0.1 mm level. The over-detection rate and false detection rate are reduced by more than 40% compared with traditional deep learning schemes, which greatly improves the automation first pass rate of precision manufacturing production lines.

[0021] (4) This invention has built a strong flexible production adaptability through knowledge transfer and model compression units. Based on few-sample learning and knowledge distillation technology, when the system faces product replacement or specification adjustment, it does not need to carry out expensive re-labeling and ultra-large-scale model training. It can achieve the online operation of the new production line within a few hours with only a small amount of domain adaptive adjustment. This significantly reduces the technical deployment cost and operation and maintenance difficulty of enterprises in small batch and multi-variety production mode.

[0022] This invention utilizes edge computing and a closed-loop control unit to successfully transform sophisticated intelligent algorithms into extremely short real-time decision responses. Through heterogeneous acceleration and model quantization, the system maintains extremely high detection accuracy while achieving a single-piece processing latency of less than 30 milliseconds, perfectly matching the operating cycle of ultra-high-speed industrial production lines. Furthermore, the closed-loop control mechanism enables automated feedback from detection to process optimization, transforming quality control from simple end-point interception to preventative correction throughout the entire process, demonstrating extremely high industrial application value and economic benefits. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical architecture of the intelligent detection system for industrial design appearance defects based on computer vision proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the multi-scale defect characterization neural network in this invention; Figure 3 This is a logical flowchart of the knowledge transfer and model compression unit in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1

[0026] This invention provides an intelligent detection system for industrial design appearance defects based on computer vision. Please refer to [link / reference]. Figure 1 The intelligent detection system for industrial design appearance defects based on computer vision consists of a multi-dimensional light field perception unit, an adaptive image enhancement unit, a multi-scale defect representation neural network, a knowledge transfer and model compression unit, and an edge computing and closed-loop control unit. The units interact with each other in real time through the high-bandwidth industrial Ethernet protocol, thus building a closed-loop system from the underlying physical perception to the high-level logical decision-making.

[0027] The multi-dimensional light field sensing unit is responsible for acquiring the raw visual signals of the product under inspection at the physical level. Please refer to the appendix. Figure 1This unit is deployed at the inspection station of an industrial production line. Its core component includes a ring array light source composed of multi-color temperature light-emitting diodes (LEDs). This ring array light source is not a single-wavelength illumination device, but rather a hybrid light source module integrating red, green, blue, and near-infrared spectra, designed to provide differentiated lighting effects for product surfaces of different materials. The ring array light source uses pulse width modulation (PWM) technology to precisely control the current duty cycle of each LED at the microsecond level, thereby achieving dynamic adjustment of light intensity across different spectral bands. In actual operation, the system can, according to a preset task list, complete diffuse reflection supplementary lighting on highly reflective metal surfaces within an extremely short exposure time of 10 microseconds using high-frequency pulse drive, eliminating strong glare points caused by direct illumination from point sources. The unit is also equipped with a high-resolution area array camera located at the center of the axis and several surrounding tilting observation cameras. These cameras form an array-type industrial camera cluster, receiving encoder signals from high-precision motors through a synchronous trigger controller to achieve synchronous shutter triggering of multiple cameras at the nanosecond level. This multi-view arrangement ensures that the system can acquire the complete envelope visual information of the product under inspection in 3D space, capturing lateral scratches or minor geometric distortions that are easily obscured under traditional vertical views.

[0028] The multi-dimensional light field sensing unit further includes a polarization illumination assembly. This assembly consists of a polarizer positioned at the light source end and an analyzer positioned at the lens end. By adjusting the relative angle between the polarizer and analyzer via an electrically controlled rotating shaft, the system can effectively block specular reflection light directly entering the lens using Fresnel's law of reflection, allowing only diffusely reflected polarized light carrying the detailed features of the product surface to pass through. This mechanism is significant for detecting high-gloss painted surfaces, polished metal parts, or glass materials, significantly enhancing the contrast features of deep scratches and shallow foreign objects in the image, increasing the signal-to-noise ratio to more than three times the initial state.

[0029] The adaptive image enhancement unit receives the raw high-dimensional image data transmitted by the multi-dimensional light field sensing unit. This unit first performs joint spatial and frequency domain transformation processing, filtering out high-frequency impulse noise and periodic texture noise caused by industrial electromagnetic interference by constructing low-pass and band-stop filters in the frequency domain. As one of the core processing logics of this unit, it utilizes a deep learning-based generative adversarial network (GAN) structure to perform pixel-level fusion of multiple frames acquired by the multi-dimensional light field sensing unit under different exposure parameters. In this process, the generator of the GAN constructs a synthetic image that retains high dynamic range information, while the discriminator ensures the realism of the synthetic image in terms of defect edge details through continuous adversarial evolution. This fusion technique can completely eliminate the large-area light spot masking phenomenon caused by metallic materials while preserving the weak gradients at defect edges.

[0030] The adaptive image enhancement unit also integrates automatic gain control logic. This logic dynamically calculates the image compensation coefficient by retrieving real-time feedback data from ambient light sensors located inside the detection chamber. When the ambient light intensity fluctuates, the automatic gain control logic adjusts the lookup table parameters in the image processing pipeline to stabilize the average grayscale value of the output image within an ideal range of approximately 128, providing a consistent input benchmark for subsequent neural network recognition. When dealing with specific industrial products with highly repetitive textures, the adaptive image enhancement unit activates a texture removal algorithm based on Fourier transform. This algorithm converts the image to the spectral space using Fast Fourier Transform, automatically identifying high-energy frequency feature points representing normal, regular, and periodic textures in the spectrum, and blocking these components using frequency domain masking techniques. Subsequently, a residual image containing only non-periodic defects is recovered through inverse Fourier transform. Through this processing flow, the contrast of minor surface scratches or structural damage can be improved by more than five times after the background is removed.

[0031] The multi-scale defect characterization neural network, acting as the intelligent recognition center of the system, directly receives enhanced feature maps from the adaptive image enhancement unit. Please refer to the appendix. Figure 2 This network employs a deep, multi-level feature pyramid architecture. In this architecture, after initial feature extraction via low-level convolutional operators, the spatial resolution is progressively reduced while the channel depth is increased, resulting in feature representations at different scales. To capture extremely small defects smaller than 0.1 mm, the network attaches receptive field enhancement modules to multiple feature layers at different resolutions. Through dilated convolutions and cross-level feature fusion, the network expands the perceptual range of neurons without losing detailed information. The network incorporates a dual spatial and channel attention mechanism. The spatial attention module generates a weighted mask map by calculating long-range dependencies between pixels, assigning higher energy weights to suspected pixel regions containing potential defects. The channel attention module, on the other hand, uses global average pooling to statistically analyze the importance of each channel, guiding the system to ignore background texture features that lack discriminative significance.

[0032] Example 2

[0033] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 3Specifically, addressing the common problem of extreme imbalance in the distribution of positive and negative samples in industrial inspection—where the number of qualified products far exceeds the number of defective products—the multi-scale defect representation neural network employs a weighted cross-entropy loss function in model training and inference evaluation. This loss function, by introducing a class balancing factor, automatically increases the penalty for misidentification of rare defect categories, thereby significantly improving the system's sensitivity in identifying rare defect categories. Before outputting detection results, the multi-scale defect representation neural network also undergoes an uncertainty estimation module. This module uses Monte Carlo sampling technology to perform multiple forward inferences with random deactivation attributes on the input image to assess the confidence level of the detection conclusion. When the system's confidence in determining a suspected defect is below 90%, it does not force a conclusion but automatically forwards the image to a human review terminal for remote calibration. The results of the human calibration serve as new supervisory signals, fed back into the incremental update process of the model in real time through an online learning algorithm, ensuring that the system's recognition capabilities continuously evolve over production time.

[0034] The knowledge transfer and model compression unit aims to address the pain points of limited computing resources and frequent product specification changes in industrial scenarios. Please refer to the appendix. Figure 3 This unit first employs fine-grained pruning techniques to remove weight connections with low contribution from the multi-scale defect representation neural network. While keeping the model's inference accuracy decline to within 1%, this technique can compress the model's parameter count to less than 15% of the original size. To further optimize performance, the knowledge transfer and model compression unit introduces knowledge distillation. Within this framework, a pre-trained, high-performance but structurally bloated teacher network serves as the knowledge source, guiding the lightweight student network to learn by outputting soft-label cross-entropy and the similarity loss of intermediate feature maps. This teaching model enables the student network to simulate the complex probability distribution at defect edges of the teacher network, thereby achieving sub-pixel-level localization capabilities without bearing a massive parameter burden.

[0035] To address the need for flexible production line transformation, the knowledge transfer and model compression unit utilizes a pre-training mechanism based on contrastive learning to extract universal visual feature representations from massive amounts of unlabeled industrial images. When enterprises introduce new product specifications, the system no longer needs to conduct large-scale sample collection and model training from scratch. Relying on the learned universal features, the system only needs to acquire up to 50 on-site defect samples to complete rapid domain adaptation and deployment of the model within 4 hours through a few-shot learning algorithm. This process is typically executed asynchronously on a dedicated offline computing cluster, without consuming the extremely valuable real-time detection resources on the production line, ensuring continuous operation of the production line.

[0036] The edge computing and closed-loop control unit is the execution center of the entire system. Based on a heterogeneous computing architecture of field-programmable gate arrays (FPGAs) and graphics processing units (GPUs), this unit uses model quantization technology to convert 32-bit floating-point operations in neural networks into 8-bit integer operations. This quantization operation, combined with the underlying hardware pipeline parallel acceleration, ensures that the entire process inspection time for a single product is strictly controlled within 30 milliseconds. This performance level perfectly matches the high-speed operation rate of 2000 pieces per minute on the production line, preventing the inspection process from becoming a bottleneck in production efficiency. The edge computing and closed-loop control unit also integrates a process trend prediction algorithm based on Markov chains. This algorithm statistically analyzes the inspection results over a continuous time series. When it detects a specific type of defect, such as geometric distortion of a certain shape, with a monotonically increasing probability, the system predicts that mold wear or parameter deviation may have occurred in the upstream production process. At this time, the unit automatically sends compensation control commands to the upstream process, realizing a shift from end-point interception to source prevention.

[0037] In terms of hardware reliability, the edge computing and closed-loop control unit adopts a dual-redundancy architecture. This architecture includes a primary computing node and a backup computing node, which maintain synchronization through a heartbeat monitoring mechanism. If the primary node experiences a hardware failure or task timeout, the backup node can quickly take over all visual acquisition and processing tasks within 50 milliseconds, ensuring uninterrupted production line operation. A real-time operating system is deployed at the system's underlying layer, using a priority-based preemptive scheduling mechanism to ensure that visual acquisition and defect determination tasks always have the highest execution privileges, minimizing latency jitter caused by system task scheduling.

[0038] The complete system provided by this invention also includes a centralized defect feature fingerprint database. This database stores defect image feature vectors that have undergone hash encryption. Through an industrial internet platform, this database can be connected to multiple production bases. Detection systems in different factories can upload newly discovered defect patterns to the cloud for feature fusion, achieving cross-regional collaborative evolution and knowledge sharing of defect patterns. This mechanism ensures that the quality evaluation standards of each production line remain strictly consistent within the globalized production system of large group enterprises, eliminating standard drift caused by regional differences or human factors.

[0039] Example 3

[0040] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1Specifically, based on Embodiment 1, this embodiment provides a special implementation scheme for curved metal shells with ultra-high reflectivity. In the multi-dimensional light field sensing unit, in addition to the ring array light source, a stripe generation module based on structured light projection is further added. This module can project phase-shifted sinusoidal stripes onto the product surface. By acquiring stripe deformation images modulated by the product surface topography using an oblique observation camera, the system can use a phase unwrapping algorithm to reconstruct high-precision 3D point cloud data of the product surface. This supplementation of multi-dimensional information enables the system to achieve detection accuracy more than 10 times higher than traditional 2D vision when facing defects in the depth direction such as tiny depressions and protrusions.

[0041] In this embodiment, the adaptive image enhancement unit is specifically optimized for curved surface reflections. This unit utilizes binocular vision calibration parameters to perform perspective transformation and resampling on the acquired curved surface image, unfolding the complex curved surface into a flat feature plane. When performing background noise suppression, this unit introduces a combined algorithm of adaptive median filtering and non-local means denoising, aiming to accurately separate the brushed metal texture from defect signals. The multi-scale defect characterization neural network in this embodiment adds a dedicated depth feature input branch. This branch converts the reconstructed 3D point cloud data into a depth map and performs early fusion with the color enhancement map at the feature layer. Through this multi-modal data input, the network can more accurately characterize the physical properties of defects, such as distinguishing between surface black dirt and actual physical pits.

[0042] In this embodiment, the knowledge transfer and model compression unit introduces a meta-learning mechanism for few-shot learning algorithms. The system learns a learning strategy capable of rapidly adapting to new tasks through meta-training on multiple related industrial tasks. When faced with entirely new metal materials or painting processes, the system can reach convergence in a shorter time with only a minimal number of gradient update steps. Furthermore, for some extremely scarce defect samples, this unit utilizes a generative adversarial network to generate highly realistic virtual defect samples and injects them into the training set, mitigating the overfitting risk caused by extreme sample scarcity through data augmentation.

[0043] In this embodiment, edge computing and the closed-loop control unit further enhance the deep integration with the production line actuators. When continuous structural damage is detected, the unit not only sends sorting instructions to the pneumatic sorting device, but also communicates in real time with the front-end injection molding machine or stamping machine via standard industrial communication protocols to directly retrieve and correct its current pressure, temperature, or mold closing stroke parameters. This visual feedback-based closed-loop control enables self-correction and dynamic optimization of the production process.

[0044] Example 4

[0045] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, this embodiment focuses on intelligent management in a large-scale distributed factory environment. The multi-dimensional light field sensing unit incorporates a self-calibration mechanism during the acquisition process. By arranging high-precision standard color plates and geometric correction plates inside the detection chamber, the system automatically performs online calibration of camera intrinsic and extrinsic parameters during hourly intervals. This mechanism effectively compensates for physical position shifts caused by factory environmental temperature fluctuations and equipment vibrations, ensuring long-term stability of image acquisition.

[0046] The adaptive image enhancement unit employs a distributed computing framework when processing large-scale concurrent data streams. For ultra-high-definition images with a resolution of 8K or higher, the unit divides them into several overlapping sub-blocks and utilizes multi-core processors in edge computing nodes to execute enhancement algorithms in parallel. During the sub-block merging stage, a weighted stitching technique is used to eliminate the step effect at the stitching points, ensuring the continuity of global features.

[0047] In this embodiment, the multi-scale defect representation neural network achieves dynamic perception of the model's own decision boundaries by integrating an uncertainty estimation module. When the high-confidence region output by the uncertainty estimation module shifts, the system automatically triggers an alarm, alerting process engineers that unknown environmental interference or novel defects may have occurred. Furthermore, by introducing a federated learning framework, the system enables independent edge detection nodes to exchange model gradient parameters and collaboratively train a more robust global defect detection model without disclosing the original production image data.

[0048] In this embodiment, the knowledge transfer and model compression unit developed an automated model quantization sensitivity analysis tool to meet the needs of large-scale deployment. This tool can automatically identify the layers in a multi-scale defect representation neural network that are most sensitive to quantization errors, preserving higher computational accuracy for them while performing more aggressive quantization compression on insensitive layers. This hybrid accuracy strategy further taps into the computing potential of edge hardware while ensuring that detection accuracy does not collapse.

[0049] In this embodiment, the edge computing and closed-loop control unit, in addition to performing physical sorting, also undertakes the responsibility of quality big data analysis. This unit can automatically generate daily and weekly quality fluctuation reports by performing multi-dimensional cluster analysis on historical inspection data. These reports include spatial distribution heatmaps of defects, evolution trend diagrams of defect categories, and correlation analysis results between process parameters and defect incidence rates. This in-depth analysis data is synchronized to a centralized defect feature fingerprint database via an encrypted channel, providing precise data support for the company's top-level decision-making.

[0050] In this embodiment, the centralized defect fingerprint database employs asymmetric encryption and blockchain distributed storage technologies. Each defect fingerprint record is assigned a unique digital signature and timestamp upon generation, ensuring the immutability of data during transmission and storage. This provides legally binding evidence for product quality traceability. In the event of a global product recall or quality dispute, companies can quickly locate the affected batches, production lines, specific process parameters, and original inspection images through this database.

[0051] In summary, this embodiment constructs a comprehensive, deep-level, and self-evolving industrial inspection ecosystem through precise acquisition by the multi-dimensional light field sensing unit, signal purification by the adaptive image enhancement unit, intelligent recognition by the multi-scale defect characterization neural network, performance optimization by the knowledge transfer and model compression unit, and rapid response by the edge computing and closed-loop control unit. The organic synergy between these units not only significantly improves the accuracy and speed of appearance inspection of industrial products but also provides a solid technical guarantee for the digital transformation and intelligent upgrading of modern manufacturing through closed-loop feedback and knowledge sharing.

[0052] Example 5

[0053] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the multi-scale defect representation neural network in the system employs the following operational logic for its integrated attention mechanism when extracting defect features in complex backgrounds: First, it obtains channel-level statistical descriptors through a global average pooling layer. Then, it uses two fully connected layers to construct a non-linear mapping relationship between channels, generating channel attention vectors. Simultaneously, it performs max pooling and average pooling along the channel dimension on the feature map, concatenates them, and then uses a 7x7 convolutional kernel to generate a spatial attention map. Finally, it performs pixel-by-pixel multiplication of the input feature map, the channel attention vectors, and the spatial attention map, thereby achieving precise focusing on key defect regions. When handling extremely small defects, the system introduces a feature resampling module at the top layer of the feature pyramid, weightedly fusing deep semantic information with shallow high-resolution information to ensure that even minute features at the 0.1 mm level are not lost during downsampling.

[0054] In the actual deployment of edge computing and closed-loop control units, the model quantization technology adopts a symmetric quantization strategy. By performing saturation analysis on the validation set, the quantization scaling factor for each layer's activation value and weight is determined. By mapping floating-point numbers to the integer space from -128 to 127, the resource utilization of the hardware multiplier is reduced by more than 70%, significantly improving processor throughput. Simultaneously, the system's internal Markov chain process trend prediction algorithm establishes a state transition matrix containing normal, metastable, drifting, and fault states, calculating in real-time the probability of the current process being in different states. Once the probability of a drifting state exceeds a preset threshold of 0.6, the closed-loop control command is immediately triggered.

[0055] When faced with flexible production tasks involving multiple product types and small batches, the knowledge transfer and model compression unit demonstrates extremely high adaptability. The system utilizes a prototype-based metric learning method to establish a central prototype in the feature space for each known defect category. When minor defects appear in new product specifications, the system defines the new category using only a minimal number of support set samples by calculating the Euclidean distance or cosine similarity between the test sample and each prototype. This approach avoids the dependence of traditional deep learning on massive amounts of labeled data, reducing the debugging cycle of new production lines from weeks to hours.

[0056] To further enhance the system's anti-interference capability, the synchronization trigger controller of the multi-dimensional light field sensing unit adopts a hardware timer based on a field-programmable gate array (FPGA), with a clock frequency set to 100 MHz, ensuring that the time synchronization error between the cameras is less than 10 nanoseconds. On a high-speed production line, this extremely high synchronization accuracy can completely eliminate image ghosting or spatial matching errors caused by asynchronous camera triggering, providing a high-precision data source for subsequent 3D reconstruction.

[0057] When handling high dynamic range fusion, the adaptive image enhancement unit employs an exposure-weight-based fusion operator to address the specular reflection characteristics of metal surfaces. This operator assigns pixel-level weights to image frames with different exposure times by evaluating pixel brightness saturation, local contrast, and edge intensity. In areas obscured by strong light, the weight of the short-exposure image is automatically increased; in areas with dark details, the weight of the long-exposure image is increased. This refined fusion strategy ensures that the final enhanced feature map clearly presents all surface texture details in both bright and dark areas, significantly improving the recognition limit of the multi-scale defect representation neural network.

[0058] The entire system runs on a customized industrial computer platform, which integrates a multi-level thermal management system. Through a combination of heat pipe cooling and active air cooling, it ensures that the operating temperature of the core processor and graphics processor remains below 70 degrees Celsius even in a high-temperature workshop environment of 50 degrees Celsius, avoiding computing power throttling issues caused by overheating. The system's communication interface uses a ruggedized fiber optic adapter module, effectively resisting strong electromagnetic interference generated by high-power motors and frequency converters in industrial environments, ensuring a zero packet loss rate for the visual data stream transmission.

[0059] In the knowledge distillation process of the knowledge transfer and model compression unit, the similarity loss function between the teacher network and the student network further incorporates Gram matrix feature comparison. By comparing the second-order statistical properties of the feature maps generated by the two networks at different levels, the student network learns not only the classification results but also the teacher network's representation of surface texture, thus achieving deep suppression of complex textured backgrounds within a lightweight architecture. This deep feature alignment technique is key to the system's ability to maintain 30 milliseconds of real-time performance while achieving detection accuracy comparable to that of large-scale laboratory servers.

[0060] The real-time operating system of the edge computing and closed-loop control unit adopts a microkernel architecture, isolating the vision driver, algorithm logic, and communication stack. Even if a certain algorithm module crashes intermittently due to abnormal data, the system's kernel monitoring mechanism can detect and restart the module within 1 microsecond without affecting the operation of the entire system. This highly reliable design, combined with a dual-redundancy architecture, enables the entire detection system to meet the extremely stringent requirements of industrial applications for annual downtime of no more than 5 minutes.

[0061] In this embodiment, the centralized defect feature fingerprint database is also equipped with a deep mining tool. This tool, by performing association rule mining on hundreds of millions of defect feature vectors, can uncover deep-seated patterns hidden behind the production process. For example, the system may discover a significant correlation between the frequency of a certain scratch defect and the batch number of the raw material supplier, the workshop humidity on that day, and the shift of the operators. These insights are pushed to the mobile terminals of factory managers in real time through visualized reports, realizing a comprehensive transformation of factory management from experience-driven to data-driven.

[0062] When performing multispectral imaging, the multidimensional light field sensing unit utilizes narrowband filter technology to filter the broadband light emitted by the light-emitting diode into monochromatic light with center wavelengths of 460 nm, 525 nm, and 630 nm. For certain specific organic coating defects, the system switches to near-infrared mode, leveraging the penetrability of near-infrared light to detect whether there are cracks or corrosion in the substrate beneath the coating. This cross-spectral sensing capability enables the system proposed in this invention to solve many concealed quality problems that traditional single visible light solutions cannot address.

[0063] To ensure real-time processing, the adaptive image enhancement unit utilizes the shared memory architecture of the graphics processor during image enhancement. By oversampling image segments and temporarily storing the data in high-bandwidth registers, the system significantly reduces the frequency of accesses to video memory, thereby improving the processing speed of the texture culling algorithm by more than 2.5 times. Furthermore, the automatic gain control logic integrated within the enhancement algorithm can predict illumination changes and pre-estimate the gain coefficient for the next time step using a second-order Taylor expansion, achieving smooth and seamless gain adjustment.

[0064] The multi-scale defect representation neural network employs ensemble learning to output the final classification probability. The system incorporates multiple differentiated fully connected branches in the last few layers of the model, independently discriminating between geometric shape, color distribution, and texture anomalies, and finally arriving at a final conclusion through a weighted voting mechanism. This multi-criteria decision-making method effectively reduces the misclassification rate when dealing with complex and composite defects.

[0065] In summary, this embodiment not only delves deeply into the internal logic of each functional unit but also constructs a complete technical closed loop across multiple dimensions, including system integration, hardware assurance, data security, and intelligent decision-making. Through the collaborative operation of multi-dimensional light field perception, adaptive image enhancement, deep neural network recognition, knowledge transfer compression, and edge closed-loop control, this invention successfully transforms complex computer vision technology into a large-scale industrial automation solution, significantly improving the intelligence level and production efficiency of industrial appearance inspection.

[0066] In subsequent upgrades and maintenance, the knowledge transfer and model compression unit also supports remote firmware updates. Through the encrypted channel of the industrial internet platform, the R&D headquarters can push model weight files optimized for new defects to edge computing nodes around the world without entering the factory site. This cloud-based training and edge deployment model marks a new stage of cloud-edge collaboration in industrial quality control.

[0067] In practical operation, the polarization illumination component of the multi-dimensional light field sensing unit can also achieve non-mechanical angle switching through a high-speed liquid crystal polarization adjuster. This electrically controlled polarization technology can complete the polarization state switching within 1 millisecond, enabling the system to acquire multiple sets of images with different polarization directions in a single detection process. Furthermore, the extremely weak stress distribution characteristics can be extracted through a polarization difference algorithm, which plays an irreplaceable role in detecting structural defects inside transparent plastic parts or tempered glass.

[0068] The texture culling algorithm in the adaptive image enhancement unit is not limited to periodic textures; its enhanced version also utilizes the multi-scale decomposition capability of wavelet transform. By setting nonlinear thresholds on different wavelet sub-bands, the system can filter out non-stationary random background textures, such as leather or matte textures. This processing method achieves extreme background smoothing while maintaining the sharpness of defect edges, providing extremely clean feature input for the neural network.

[0069] In the specific defect assessment process, a multi-scale defect representation neural network generates a pixel-level defect mask. This mask not only identifies the location of the defect but also calculates its precise area, perimeter, and fractal dimension. These quantified parameters are fed into the decision engine of the edge computing and closed-loop control unit, automatically classifying products into grades such as qualified, second-class, rework, and scrap based on the detailed quality requirements of different customers. This automatic grading function greatly reduces the workload of manual re-inspection and improves the transparency of the entire quality chain.

[0070] The process trend prediction algorithm for edge computing and closed-loop control units incorporates a forgetting factor mechanism when processing Markov chains. This mechanism assigns different weights to historical data from different times, making the system more sensitive to recent quality fluctuations and thus responding more quickly to sudden process anomalies. Simultaneously, when the system determines that the current process is in a metastable state, it triggers an early warning indicator to remind operators to perform necessary preventative maintenance and avoid generating batches of defective products.

[0071] The centralized defect fingerprint database employs a differential privacy algorithm for knowledge sharing. This algorithm introduces a suitable amount of random noise before uploading feature vectors, ensuring that while it can extract statistical patterns from group characteristics, it cannot reverse-engineer specific production images from a particular factory. This design protects the trade secrets of each factory while promoting quality improvement across the entire industry. The database also provides a graphical interface that displays real-time trends in global production line pass rates, defect type distribution ratios, and performance improvement curves after model optimization.

[0072] Through the coordinated operation of the aforementioned units, the computer vision-based intelligent inspection system for industrial design appearance defects proposed in this invention not only achieves industry-leading levels in core indicators such as detection accuracy, real-time performance, and flexibility, but also provides modern factories with a new paradigm of intelligent quality management capable of self-sensing, self-diagnosing, and self-repairing through deep integration of closed-loop control and big data. This system demonstrates extremely high application value and promising prospects in fields such as consumer electronics, automotive precision parts, and high-end equipment manufacturing.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A computer vision-based intelligent detection system for industrial design appearance defects, characterized in that: include: Multi-dimensional light field sensing unit, adaptive image enhancement unit, multi-scale defect representation neural network, knowledge transfer and model compression unit, and edge computing and closed-loop control unit; The multi-dimensional light field sensing unit is used to acquire multi-angle spatial information and multi-spectral imaging data of the surface of the product to be inspected through a pre-set structured light source module and an array of industrial cameras, so as to reconstruct the high-dimensional visual feature distribution of the product surface. An adaptive image enhancement unit is used to perform background noise suppression, illumination normalization, and high dynamic range fusion processing on the original image acquired by the multi-dimensional light field sensing unit to generate an enhanced feature map with a high signal-to-noise ratio. A multi-scale defect characterization neural network is used to receive enhanced feature maps. Through deep feature extraction operators and global attention mechanisms, it identifies and locates geometric distortions, surface scratches, color differences and structural damage on the product surface at different spatial scales, and generates defect candidate regions and preliminary classification probabilities. The knowledge transfer and model compression unit is used to perform knowledge distillation on the multi-scale defect representation neural network using a pre-set teacher network, thereby reducing the computational load of the model while maintaining detection accuracy, and combining it with a few-shot learning algorithm to achieve fast domain adaptation of defect features of new product specifications. Edge computing and closed-loop control unit are used to deploy the optimized detection model to a high-parallel hardware acceleration platform, calculate the detection results in real time, and send sorting instructions, alarm signals or process parameter correction instructions to the production execution system according to the preset quality judgment threshold.

2. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The multi-dimensional light field sensing unit also includes a ring array light source composed of multi-color temperature light-emitting diodes and a polarization illumination component. The ring array light source integrates monochromatic light sources with wavelengths of 460 nm, 525 nm and 630 nm respectively. The polarization illumination assembly eliminates specular glare on the product surface caused by Fresnel reflection by adjusting the relative angle between the polarizer and the analyzer.

3. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The adaptive image enhancement unit is also used to perform a texture culling operation based on Fourier transform when processing industrial products with periodic textures, as follows: The image is converted to the spectral space using Fast Fourier Transform; periodic frequency components representing normal texture are identified and blocked in the spectral space; and a residual image containing only non-periodic defects is recovered using Inverse Fourier Transform.

4. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The specific logic of the attention mechanism integrated within the multi-scale defect characterization neural network is as follows: a channel-level statistical descriptor is obtained through a global average pooling layer, and a nonlinear mapping relationship between channels is constructed using two fully connected layers to generate channel attention vectors. The feature maps are subjected to max pooling and average pooling along the channel dimension, concatenated, and then a spatial attention map is generated using a 7x7 convolution kernel. The input feature map is multiplied pixel-by-pixel with the channel attention vector and the spatial attention map to focus on key defect regions.

5. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The multi-scale defect characterization neural network performs uncertainty estimation before outputting the detection results. The specific process includes: The input image is subjected to multiple forward inferences with random deactivation attributes using Monte Carlo sampling technology to assess the confidence level of the detection conclusion; when the system's confidence in the determination of a suspected defect is lower than 90%, the image is automatically forwarded to a manual review terminal for remote calibration. The calibration results are then fed back into the model update process in real time as new evidence for online learning.

6. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: When performing the knowledge distillation process, the knowledge transfer and model compression unit introduces soft label cross-entropy and intermediate feature mapping similarity loss to simulate the probability distribution of the teacher network at the defect edge, enabling the lightweight student network to obtain sub-pixel-level localization capability. The knowledge distillation process is executed asynchronously on an independent computing cluster.

7. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The specific process by which the edge computing and closed-loop control unit executes the process trend prediction algorithm is as follows: Establish a state transition matrix that includes normal state, metastable state, drift state and fault state; The probability values ​​of the current process being in different states are calculated in real time based on the detection results within a continuous time series. When the probability value of the drift state exceeds 0.6, the process parameter correction command is triggered.

8. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The edge computing and closed-loop control unit adopts a dual redundancy architecture. The heartbeat monitoring mechanism ensures that when the main computing node fails, the backup node takes over within 50 milliseconds. The edge computing and closed-loop control unit is equipped with a real-time operating system, which ensures that visual acquisition tasks and defect judgment tasks have the highest execution authority through a priority preemption scheduling mechanism.

9. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The system also includes a centralized defect feature fingerprint database for storing encrypted defect image feature vectors; The defect feature fingerprint database is connected to multiple production bases through an industrial internet platform, and random noise is introduced before uploading the feature vector using a differential privacy algorithm to achieve cross-regional collaborative evolution of defect patterns.

10. The intelligent detection system for industrial design appearance defects based on computer vision according to claim 1, characterized in that: The multi-dimensional light field sensing unit also includes a stripe generation module based on structured light projection, used to project phase-shifted sinusoidal stripes onto the product surface. The system is also used to acquire stripe deformation images modulated by the surface topography of the product through an oblique observation camera, and to restore the 3D point cloud data of the product surface using a phase unwrapping algorithm.

Images