Artificial intelligence industrial visual inspection method, system and equipment

Abstract

The invention belongs to the technical field of artificial intelligence, and provides an artificial intelligence industrial visual detection method, system and equipment, and the detection method comprises the steps: image collection and multi-modal data acquisition, defect sample generation, dynamic preprocessing and noise suppression, multi-modal feature fusion, and dynamic balance defect detection. According to the method, 2D HDR images, 3D point cloud and polarized light data of a product are synchronously collected, reflective interference is inhibited, microdefect contrast is enhanced, multi-modal data fusion is matched with cross-scale analysis, inter-modal feature alignment and intra-modal multi-scale defect capture are achieved, the limitation of a traditional single-modal method is solved, the defect omission ratio and the defect false detection rate are reduced, and the product quality is improved. Diversified samples are generated based on real defect data, physical simulation is combined to verify defect rationality, rare defect training data volume is increased, data distribution of different production lines is aligned through adversarial training, cross-line detection accuracy fluctuates, synthetic data cooperation domain self-adaption is achieved, and the detection rate of rare defects in a small sample scene is increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an artificial intelligence industrial visual detection method, system and device. BACKGROUND

[0002] Industrial visual detection is a core link of manufacturing quality control. For example, in the manufacturing of lithium batteries, the lithium battery pole piece is a core component of the lithium ion battery, and the surface quality of the lithium battery pole piece directly affects the performance and safety of the battery. In a high-speed production line with a speed of 150 m / min, the pole piece needs to go through processes such as coating, rolling, and slitting, and may produce micron-level defects such as pores, scratches, and coating cracks with a size of <50 μm. These product defects need to be accurately identified, and then the product is subjected to corresponding post-processing.

[0003] At present, traditional methods rely on manual inspection or intelligent visual detection. Manual inspection is not only low in efficiency, but also has a high rate of missed detection. A single manual detection station can only process 200-300 pole pieces per hour, and the missed detection rate of artificial detection of small defects such as scratches with a width of ≤50 μm is as high as 15%-20%. Intelligent visual detection generally uses fixed rule-based image processing algorithms for defect detection, such as edge detection and threshold segmentation.

[0004] However, the existing defect visual detection method still has some problems. Traditional methods mostly rely on a single type of image data, such as 2D brightness images. Such single-modal input information has limited dimensions: when facing a highly reflective metal surface, 2D images are prone to lose defect textures due to overexposure or specular reflection; for small defects with weak background contrast or concave-convex defects with obvious three-dimensional topographic features, single 2D information is also difficult to provide sufficient discrimination basis, resulting in missed detection and false detection. In addition, most products in industrial production are good products, and real defect samples are already scarce. Rare defects such as micro-cracks have a very low frequency of occurrence in the production line, which leads to a serious shortage of labeled sample quantity. The model is difficult to learn effective features of such defects in the training process, resulting in a high rate of missed detection of small sample defects. SUMMARY

[0005] The purpose of the embodiments of the present application is to propose an artificial intelligence industrial visual detection method, system and device to solve the technical problems of missed detection and false detection of single-modal vision and high missed detection rate of small sample defects raised in the background art.

[0006] To solve the above technical problems, the embodiments of the present application provide an artificial intelligence industrial visual detection method, which adopts the technical solutions as follows: An artificial intelligence industrial visual detection method, comprising the following steps: Real-time acquisition of product surface images on the production line yields HDR images, 3D point clouds, and polarization images. Defect data, including real defect images, defect types, defect locations, process parameters, and metadata, is stored in a defect database. This data is used as input to generate diverse defect samples and synthetic defect images. Dynamic preprocessing of the HDR and synthetic defect images eliminates uneven lighting and noise, standardizes image contrast, and produces a normalized preprocessed image. Multimodal feature fusion is performed on the preprocessed image, 3D point cloud, and polarization image to achieve feature dimension matching, resulting in a fused feature map. The synthetic defect image is then mixed with the real defect image to obtain a mixed training set, which is used for offline training to generate a detection model. The detection model processes the fused feature map in real-time, performs defect detection, and outputs the defect detection results, including defect location, category, and confidence level. The confidence level is used to determine the presence and category of a defect. The system state space and action space are constructed. The state space includes system latency, GPU power consumption and defect detection accuracy, while the action space includes computing power allocation ratio. A scheduling strategy model is generated based on the system state space and action space. The system computing power is dynamically allocated through the scheduling strategy model. Online training is performed based on the defect detection results and the system state space to update the detection model and the scheduling strategy model.

[0007] Preferably, during real-time acquisition, HDR images are synthesized by fusing multiple exposure images, 3D point clouds are generated by structured light decoding and 3D reconstruction, and parallel polarization images and vertical polarization images are acquired by time-division triggering of linear polarization light sources, and polarization images are obtained after processing.

[0008] Preferably, the generation of the synthetic defect image is as follows: the noise vector of the real defect image is converted into an initial synthetic image by the generator of the generative adversarial network, and the rationality of the defect in the synthetic image is verified by physical simulation and evaluated and updated to obtain the synthetic defect image. During physical simulation verification, the defect morphology is simulated based on the finite element stress model, and synthetic images that do not conform to physical laws are eliminated.

[0009] Preferably, in the dynamic preprocessing, the HDR image and the synthesized defect image are mixed according to a preset ratio, and the mixed image is uniformly subjected to global illumination compensation and adaptive filtering. After pixel normalization, the preprocessed image is output.

[0010] Preferably, in the multimodal feature fusion, the features of the preprocessed image, 3D point cloud and polarization image are dynamically aligned through cross-modal attention, and the defect features at different scales are extracted through adaptive feature pyramid extraction, and the fused feature map is output after fusion.

[0011] Preferably, in defect detection, the detection model outputs the defect location through a regression branch and the defect category through a classification branch, dynamically adjusts the defect detection weights, calculates the defect confidence based on the positioning accuracy of the defect location and the classification accuracy of the category, and outputs the defect detection result based on the confidence level.

[0012] To address the aforementioned technical problems, this application also provides an artificial intelligence industrial vision inspection system, comprising: an image data acquisition module for acquiring original product images, obtaining HDR images, 3D point clouds, and polarization images of the product; the HDR image is output to a dynamic preprocessing module, and the 3D point cloud and polarization image are output to a multimodal feature fusion module; a defect sample generation module for offline training to generate synthetic defect images based on real defect images in a defect database, and outputting them to the dynamic preprocessing module; a dynamic preprocessing module for proportionally mixing the HDR image and the synthetic defect image, normalizing the mixture to obtain a preprocessed image, and outputting it to the multimodal feature fusion module; a multimodal feature fusion module for fusing the 3D point cloud, polarization image, and preprocessed image to obtain a fused feature map, and outputting it to a dynamically balanced defect detection module; a dynamically balanced defect detection module for analyzing and detecting the fused feature map and outputting defect detection results; and a real-time optimization module for online training based on defect detection results and system state space, updating computing power allocation strategies and model parameters to optimize the system in real time.

[0013] Preferably, the dynamic preprocessing module performs global illumination compensation and adaptive filtering on the image, the multimodal feature fusion module aligns multimodal features through an attention mechanism, and further extracts defect features at different scales through an adaptive feature pyramid, fusing 2D texture, 3D morphology and polarization characteristics, and the dynamic balance defect detection module dynamically adjusts the weight of rare defects and evaluates the confidence level based on the defect classification probability and positioning accuracy.

[0014] To address the aforementioned technical problems, this application also provides an electronic device, comprising: at least one processor; a memory communicatively connected to at least one of the processors; and a network interface; the memory storing instructions executable by the processor, the instructions being executed by the processor to implement the aforementioned artificial intelligence industrial vision inspection method.

[0015] The beneficial effects of this invention are as follows: This application provides an artificial intelligence-based industrial vision inspection method that constructs a multi-dimensional perception system by simultaneously acquiring and fusing HDR images, 3D point clouds, and polarization images of a product. HDR images effectively suppress metallic reflections while preserving details of dark-field defects, 3D point clouds accurately depict the micron-level morphology of the surface, and polarization images enhance the contrast of defects with specific orientations. Through cross-modal attention mechanisms and multi-scale feature extraction, the method achieves complementary advantages and deep fusion between different modal information, fundamentally overcoming the perception limitations of single-modal vision in complex industrial scenarios and significantly improving the detection capability of micro-defects and morphological defects. By combining generative adversarial networks and physical simulation to synthesize defect samples and dynamically adjusting the weights of rare defects during training, the method effectively solves the problem of weak model recognition ability caused by the scarcity of real rare defect samples.

[0016] This method differs from traditional training models that rely solely on real samples. It uses real defect data from a defect database as a foundation, transforms noise vectors into initial synthetic images through GAN, and then introduces physical simulation to verify the rationality of the defects. This ensures that the synthetic samples conform to the characteristics of actual industrial defects, avoiding the problem of synthetic samples being out of touch with real-world scenarios. At the same time, the construction of the hybrid training set takes into account both the reliability of real samples and the diversity of synthetic samples, enabling the model to fully learn the characteristics of different types and forms of defects, significantly improving the model's generalization ability. Compared with traditional simple data augmentation sample expansion methods, it is more adaptable to the complex inspection scenarios of industrial production lines.

[0017] This method constructs a state space that includes system latency, GPU power consumption, and rare defect detection accuracy, combined with an action space composed of computing power allocation ratios, to form a dynamic scheduling strategy model. This design overcomes the limitation of fixed computing power allocation in traditional detection systems, enabling real-time adjustment of computing power allocation based on detection results and system operating status. While ensuring detection accuracy, it reduces system latency and GPU power consumption, achieving global optimization of detection performance and resource consumption. This allows the system to better adapt to the real-time detection needs of high-speed industrial production lines, balancing detection quality and operational efficiency. Attached Figure Description

[0018] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the method in Embodiment 1 of this application; Figure 2 This is a system architecture diagram of Embodiment 2 of this application; Figure 3 This is a structural schematic diagram of Embodiment 3 of this application. Detailed Implementation

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings of this application are intended to cover non-exclusive inclusion.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0023] Example 1: like Figure 1 As shown, an artificial intelligence-based industrial vision inspection method includes steps 1 to 7: Step 1: Image Acquisition and Multimodal Data Acquisition: Real-time acquisition of product surface images on the production line to obtain the product's HDR (High Dynamic Range) image I. HDR 3D point cloud P(x,y,z) and polarization image I pol .

[0024] Furthermore, during the process of acquiring product surface images, a high dynamic range image (I) is synthesized through multi-exposure image fusion. HDR To eliminate metallic reflections, structured light decoding is used to output high-precision 3D point cloud data P={x,y,z} for 3D reconstruction, in order to measure micron-level concave and convex defects.

[0025] During image acquisition, the entire surface of the product is imaged, such as the full-width area of ​​a lithium battery electrode. The objects acquired include: Target characteristics: Lithium battery electrode surface coating, thickness 80-150μm; micron-level defects in the substrate, including: pores, diameter 10-50μm; scratches, length >100μm, depth >5μm; and coating cracks, width <5μm, extension length >200μm. Physical characteristics: Highly reflective metallic surface and multi-scale texture.

[0026] The hardware configuration for image acquisition is as follows: Line scan camera: 4096 pixels resolution, 80kHz line frequency, 120m / min line speed, and 5μm×5μm pixel size.

[0027] Multispectral light source: bright field illumination, wavelength 450-650nm, for global surface imaging; dark field ring light, wavelength 780nm, to enhance the contrast of edge defects; polarized light source, linear polarization degree >100:1, to suppress metal reflection.

[0028] Structured light projector: projects Gray code + phase shift grating to generate 3D point cloud data.

[0029] Image acquisition employs a synchronous control method: encoder-triggered acquisition with a pulse interval Δt=1ms to ensure strict synchronization between the image and the production line speed; multi-source time-sharing exposure with alternating triggering of bright field, dark field, and polarized light to avoid crosstalk.

[0030] The HDR imaging method is as follows: Image sequences with different exposure times are captured by a line-scan camera, and a weighted fusion algorithm is used to generate HDR images, thus solving the problem of localized overexposure caused by metallic reflections. This is achieved using Formula 1:

[0031] Generate HDR image I HDR Where x, y: represent the pixel coordinates of the image; T=3: the number of exposures, with short exposure, medium exposure, and long exposure times being t1=1ms, t2=5ms, and t3=20ms respectively; I t (x,y): Pixel value of the t-th exposure, 12-bit RAW data, range 0-4095; w t (x,y): The weighting coefficient of exposure time t, which is calculated by local contrast adaptively. The local contrast weight is 5×5 window standard deviation.

[0032] w t (x,y) can be expressed by Formula 2:

[0033] The calculation shows that, σ t (x,y): The standard deviation of pixel values ​​within a 5×5 window centered at (x,y), reflecting local contrast. t High-contrast regions in (x,y) are preferentially used for cross-modal attention computation to enhance defect edge features.

[0034] Multi-exposure image fusion eliminates overexposure of metallic reflections by weighted fusion of images with different exposure times, increasing the dynamic range to 14-bit, expanding the dynamic range, and improving the visibility of dark-field defects such as black spots. High-contrast areas, such as defect edges, are preferentially treated with short-exposure images to avoid overexposure; low-contrast areas, such as uniform coatings, are treated with long-exposure images to improve the signal-to-noise ratio. The output synthesized HDR image I... HDR ∈R 4096×2160 The dynamic range is extended to 14 bits, providing high-fidelity input for subsequent preprocessing modules and reducing false detections caused by uneven illumination.

[0035] The structured light 3D cloud reconstruction method is as follows: Multiple modulated images are acquired using structured light projection, Gray code, and phase-shifting grating, as shown in Formula 3:

[0036] Calculate the absolute phase φ(x,y), where N=4: the number of steps in the phase-shifting grating; I n (x,y): Image intensity acquired after the nth phase-shift projection; φ(x,y): Calculated phase value, ranging from -π to +π; Sub-pixel level surface topography information is obtained through phase calculation. Surface height change is calculated through phase difference, according to Formula 4:

[0037] The surface is converted to three-dimensional coordinates z(x,y), where z(x,y) is the surface height, λ=532nm is the laser wavelength, and Δφ(x,y) is the phase difference relative to the reference plane. High-precision 3D point cloud data P(x,y,z) with an accuracy of ±2μm is generated for detecting micron-level defects such as burrs and pits.

[0038] It should be noted that the accuracy of 3D point cloud data can be further improved by introducing reflectivity R, resulting in P={x,y,z,R}.

[0039] The method for generating polarization images is as follows: Time-division triggering of a linearly polarized light source at 0° and 90° directions, acquiring parallel polarized I... / / and vertical polarization I ⊥ Image, via Formula 5:

[0040] Generate polarization difference image I pol Polarization differential imaging suppresses metallic reflections, such as specular reflections, preserves diffuse reflection defect signals, and enhances the contrast of defects such as scratches.

[0041] Example of data acquisition and processing workflow: Multi-source time-division triggering, input encoder pulse signal, frequency 2kHz, synchronized with production line speed, in the light source control command, bright field, dark field, and polarized light are triggered alternately, bright field light source is triggered on the rising edge of the pulse, exposure time 1ms; dark field light source is triggered on the falling edge of the pulse, exposure time 5ms; structured light projection and image acquisition are strictly synchronized, error <1μs; multiple sets of RAW images are output {I bright ,I dark ,I phasel-4};I bright ,I dark ,I phasel-4 These are respectively a brightness image, a darkness image, and a phase-shifted raster image; HDR fusion and polarization processing, input multi-exposure image I bright t1 I bright t2 I bright t3 ; Calculate the weight w of each pixel t (x,y); Synthesized HDR image I HDR ; Generate polarization difference image I pol =I / / -I ⊥ Output the enhanced 2D image I HDR I pol ; 3D point cloud reconstruction, input phase-shifting grating image I phasel-4 Phase calculation yields φ(x,y), phase expansion eliminates 2π jumps, converts to height map z(x,y), and outputs 3D point cloud data P(x,y,z).

[0042] Through w t The (x,y) method replaces traditional standard deviation calculations, improving fusion efficiency and enabling dynamic adjustment of HDR weights. A 2D-3D joint calibration algorithm ensures pixel-level alignment with an error of <1 pixel, enabling cross-module calibration. In actual testing on a lithium battery electrode production line, the algorithm achieves 100% capture of micron-level defects and reduces false triggering rates.

[0043] Image acquisition and multimodal data acquisition are performed on product surfaces, such as lithium battery electrode coatings and metal substrates. 2D images are acquired through multispectral imaging, 3D point cloud data is generated through structured light projection, and high dynamic range images are synthesized through multi-exposure HDR fusion processing. Aligned 2D HDR images and 3D point clouds are output with an accuracy of ±2μm. A time-division triggering linearly polarized light source is used to acquire parallel and vertically polarized images, which are then processed and output as polarized images. pol 2D image I HDR Subsequent dynamic preprocessing is performed on the 3D point cloud P(x,y,z) and polarization image I. pol Multimodal feature fusion will then be performed.

[0044] Step 2: Defect Sample Generation: Store defect data in the defect database. The defect data includes real defect images I. real Defect data, including defect type, defect location, process parameters, and metadata, is used as input to synthesize diverse defect samples, resulting in a synthetic defect image I. syn ; Furthermore, the real defect image I real The noise vector is converted into an initial synthetic image G(z), which is used to simulate the distribution of real defects. The synthetic image G(z) is then evaluated and updated to obtain the synthetic defect image I. syn .

[0045] Taking the detection of micro-defects on the surface of lithium battery electrodes in the production line as an example, micron-level defects on the surface of lithium battery electrodes can be divided into the following four categories: Point defects include pores formed by the bursting of slurry bubbles, foreign matter shrinkage cavities caused by particulate contamination leading to liquid film migration, and agglomerate particles caused by uneven dispersion of conductive agents. Linear defects include scratches such as foreign objects stuck in the mold or damage to the substrate, vertical streaks such as poor slurry leveling, and horizontal lines such as mechanical vibration or slurry fluctuations. Regional defects include thick edges such as those caused by surface tension-driven material migration, cracks caused by incomplete drying of the coating leading to roll forming, and roll forming cracks in incompletely dried areas. Other defects include exposed foil due to insufficient coating, i.e., black and white spots caused by impurities or process abnormalities.

[0046] The classification criteria for surface defects of lithium battery electrodes are as follows: Morphological characteristics: based on the shape, size and location of defects, the shape including dots, lines and regions, the size including micron-sized pores and millimeter-sized scratches, and the location including edges or centers; Optical properties: Differentiating reflective properties through multispectral imaging, such as metal particle dirt and coating cracks; Cause correlation: Determine the type of defect based on abnormal process parameters. For example, vertical streaks are often caused by poor slurry leveling, and process parameters such as coating speed and slurry viscosity are also relevant.

[0047] Real Defect Image I real This includes the results of inspection and manual labeling, as well as the labeled pores and scratches. The process parameters include process data extracted from the MES system, such as coating speed and rolling pressure, which are used for physical simulation of defect generation. The metadata includes inspection time, production line number, confidence level, and manual review results.

[0048] By combining real defect images and process parameters, such as coating speed, defect samples are generated and data augmented. A diverse range of defect samples, such as microcracks and decarburization, are synthesized using a Generative Adversarial Network (GAN). Physical simulation is used to model the stress-induced defect morphology, and the resulting synthesized defect image I is output. syn Add it to the training set and perform dynamic preprocessing afterwards; Real Defect Image I real Defects are collected through historical production line inspections or manually labeled after online sorting. They are stored in a defect database as a training benchmark for training the initial defect detection model. They also serve as input for the GAN generator and physical simulation to synthesize diverse defect samples, which are then used as validation criteria to evaluate the quality of the generated samples and the performance of the detection model.

[0049] The image with simulated defects generated by generator G, I syn Mixed with real images, it is used to train the defect detection model and improve the model's generalization ability; A generator G transforms a random noise vector z into a synthetic image G(z), simulating the distribution of real defects. The generator can be a deep convolutional network, such as ResNet or U-Net. The input is a 256-dimensional noise vector, and the output is a high-resolution image, such as 4096×2160.

[0050] Distinguish real defect images I using discriminator D. real The generated image G(z) is output, and the discrimination probability D(x)∈[0,1] is output. The discriminator can be a convolutional neural network, such as PatchGAN, which extracts image features and classifies them through multiple convolutions.

[0051] Through formula 6:

[0052] Calculate the generator G loss function L G ,in: E z～Pz [log(1-D(G(z)))] is used to combat the loss, forcing the generator G to generate an image G(z) that can deceive the discriminator D; λ rec ·‖G(z)-I real ||1 represents the reconstruction loss, which constrains the generated image to be pixel-level similar to the real image, thus avoiding the generation of invalid samples that deviate from the true distribution. E z～Pz To determine the noise distribution P z The expectation is obtained by averaging over all possible noise vectors; z: Latent spatial noise vector, dimension = 256, sampled from a normal distribution; D: Discriminator network, ResNet-18 architecture; λrec =10: Reconstruction loss weights, determined through grid search, balancing generation realism and structural consistency; I real : Real defect images, from the defect database, are taken from defective products during the production line quality inspection process, and are stored in the database after noise reduction and labeling; G(z): The synthesized image output by the generator; D(G(z)): The probability of the discriminator classifying the generated image, 0~1; The generator objective is to minimize L G This makes D(G(z)) approach 1, deceiving the discriminator and thus generating defect samples consistent with the real data distribution, such as microcracks and decarburization. G The optimization directly drives the generator to produce a realistic image G(z) that conforms to the characteristics of real defects.

[0053] Through formula 7:

[0054] Calculate the discriminator D loss function L D ,in: E x～Pdata : For the true data distribution P data The expectation is represented by the average of all real samples; D(x): The probability of the discriminator classifying the real image, ideal value = 1; D(G(z)): The probability of the discriminator classifying the generated image; ideal value = 0.

[0055] The discriminator aims to minimize L. D Distinguish between real images I real With the synthesized image G(z), L D The optimization forces the discriminator to more accurately distinguish between real and generated images, thereby driving the generator to improve the quality of G(z).

[0056] L D The optimization forces the discriminator to more accurately distinguish between real and generated images, thereby driving the generator to improve the quality of G(z).

[0057] It should be noted that finite element analysis of G(z) can also be performed through physical simulation constraints to verify whether the defect morphology conforms to physical laws, such as the crack propagation direction being consistent with the stress distribution.

[0058] The physical simulation of defects uses a finite element stress model to simulate the stress distribution of the electrode during processes such as rolling and coating, and generates defect morphologies that conform to mechanical laws, such as crack propagation paths.

[0059] Through formula 8:

[0060] Verification was performed, where ▽ is the divergence operator, representing the degree of divergence of the field quantity, σ is the material stress tensor, and F is the external load, simulating the stress of the rolling process. The above formula shows that within a stationary material, the sum of the stress divergence and the external force is zero. The verification process is as follows: Based on the material's geometry, boundary conditions, and loading, a mathematical model is established. The continuous mathematical model is discretized into a finite number of small element units using the finite element method. These elements are connected by nodes. The discretized model is solved using finite element software to obtain the stress distribution σ within each element. The divergence ∠σ of the stress field is calculated, which can be approximated on the finite element mesh using numerical methods. The results are then verified to ensure that the desired stress distribution is met at each node. If the conditions are not met, the model or mesh needs to be adjusted until the equilibrium conditions are met; analyze the obtained stress distribution and deformation to ensure that they conform to the expected physical behavior. If defects exist, observe whether their morphology and effects are consistent with reality.

[0061] This process verifies whether the generated defect morphology conforms to physical laws, thus ensuring the accuracy and reliability of the simulation. It generates crack propagation morphologies that conform to mechanical laws, provides physical constraints to ensure that the generated defects match the morphological characteristics under actual manufacturing conditions, and supplements the data generation capabilities of GANs, solving complex defects that are difficult to generate through adversarial training alone, such as microcracks caused by stress concentration.

[0062] Physical simulation data, as prior knowledge, can be injected into the generator's input. For example, by concatenating stress distribution maps with noise z, the generator can be guided to produce physically plausible defects. The discriminator D's loss L... D It will simultaneously evaluate whether the generated image conforms to the real data distribution and physical laws.

[0063] The quality of the generated image is quantized by FID. The lower the value, the higher the quality of the generated image. FID<15 means that the generated image is very close to the real data.

[0064] Through formula 9:

[0065] Calculate the FID score, where μ real μ syn : Mean feature value of real / synthetic image in the Inception-v3 model, ∑ real ,∑ syn : Covariance matrix.

[0066] Processing result: Synthesis is effective if the FID score is <15; then it is added to the training set. Synthesis failure: Triggers GAN retraining, probability <5%; Physical simulation: Simulating the crack propagation morphology caused by stress based on the finite element model.

[0067] By generating synthetic defect image I syn This addresses the problem of insufficient small sample data, covers rare defect types, increases the sample size of rare defects, and reduces the model's false negative rate.

[0068] Synthetic Defect Image I syn The generation process is as follows: The input consists of a noise vector z ~ N(0,1) and process parameters, such as coating speed. These are then concatenated as the generator input. The generator propagates forward, and the generator G generates an initial synthesized image G(z) based on the input; Physical simulation constraints are used to verify G(z) through physical simulation and to eliminate samples that do not conform to the laws of mechanics, such as cracks in non-stress concentration areas. Discriminator evaluation, computation of D(G(z)), and updating of the parameters of G and D through backpropagation; Iterative optimization, repeating the above steps, continues until the synthesized image G(z) satisfies FID < 15 and is valid for manual evaluation; this is then the synthesized defect image I. syn .

[0069] Step 3: Dynamic preprocessing and noise suppression: For the original HDR image I HDR and synthetic defect image I syn Dynamic preprocessing is performed to eliminate uneven illumination and noise, improve the image signal-to-noise ratio, and standardize the image contrast. The resulting normalized preprocessed image I is obtained after preprocessing. ’ ; For the original HDR image I HDR and defect image I syn Dynamic preprocessing and noise suppression are performed. Pixel values ​​are normalized through illumination compensation, and noise is suppressed using an adaptive filtering dynamic Gaussian kernel. The output is a normalized image I. ’ I ’ Subsequent multimodal feature fusion will be performed; Furthermore, in the dynamic preprocessing, the original HDR image I HDR and synthetic defect image I syn The images are mixed at a preset ratio of 1:1, and the resulting images are uniformly normalized to pixels. The preprocessed image I is then output. ’ and noise mask.

[0070] Noise masks can be used to mark interference areas: identifying non-defect noise areas such as dust and reflections in an image. In the defect detection module, noise masks are used to suppress false detections, such as ignoring the detection results of the masked marked areas.

[0071] During the training phase, the synthesized images and real images undergo preprocessing such as normalization and filtering, and are used as input to the training set. Dynamic processing steps include illumination compensation and adaptive Gaussian filtering.

[0072] Illumination compensation performs pixel normalization on the blended image using formula 10:

[0073] The pixel value I at position (x,y) of the image after illumination compensation is calculated. comp (x,y), where I(x,y) is the pixel value of the original image at position (x,y), μ is the global average brightness of the image, σ is the standard deviation of the image, α=0.8, β=128 are the illumination compensation parameters, which were determined through 100 sets of industrial scene calibration experiments.

[0074] μ and σ are expressed by the following formula 11:

[0075] and formula 12:

[0076] The calculation shows that H is the height of the image, i.e. the number of pixels in the vertical direction, and W is the width of the image, i.e. the number of pixels in the horizontal direction.

[0077] Normalization is used to eliminate uneven lighting, eliminate reflection interference, and standardize image contrast. to I comp Dynamic Gaussian filtering is performed on (x,y) using Equation 13:

[0078] The filtered pixel value I is calculated. ’ (x,y), where σ d The local window dynamic standard deviation is given by G(i,j,σ). ​​The local window is 3×3, k=2: the radius of the filter kernel, covering a 5×5 neighborhood. d ): Gaussian kernel function, 5×5 neighborhood.

[0079] Dynamic standard deviation σ d Through formula 14:

[0080] The calculation shows that, μ localIt is the mean of the local region surrounding the pixel located at coordinates (x,y) in the image.

[0081] After adaptive filtering, the output I ’ With a PSNR > 40dB, it can suppress salt-and-pepper noise while preserving edge details, such as burr contours.

[0082] Step 4: Multimodal feature fusion: For the preprocessed image I ’ 3D point cloud P(x,y,z) and polarization image I pol Multimodal feature fusion is performed to achieve feature dimension matching, resulting in a fused feature map F. out ; For preprocessed image I ’ 3D point cloud P(x,y,z), polarization image I pol Multimodal feature fusion and enhancement are performed, dynamically fusing 2D texture and 3D morphology features. Defect features of 50μm~200μm are extracted through dilated convolution, and the fused feature map F is output. out F out Subsequent dynamic balancing defect detection will be conducted; Furthermore, in multimodal feature fusion, after dynamically aligning features from different modalities through cross-modal attention, defect features at different scales are extracted, and the fused feature map F is output. out .

[0083] The input for multimodal feature fusion includes: preprocessed image I ’ 3D point cloud P(x,y,z) and polarization image I pol First, multimodal features are aligned through an attention mechanism, and then multi-scale pyramids are used to refine the representation of defects of different sizes. 2D texture, 3D morphology and polarization characteristics are integrated to enhance the response of small defects.

[0084] When aligning multimodal features, use Equation 15:

[0085] Calculate the cross-modal attention weights Attention(Q,K,V), output the weighted feature vector, and dynamically fuse 2D texture and 3D shape features, where: Q=W q F 2D : 2D image feature projection, such as texture information in HDR images; K=W k P 3D 3D point cloud feature projection, such as surface height information; V=W v F pol Features derived from polarization images, such as defect scattering characteristics; W q W k W v Learnable weight matrix; d k =512: Feature dimension scaling factor to prevent gradient explosion; Cross-modal attention reflects the correlation between different modalities, enhancing the feature responses in 3D point clouds that are associated with 2D defect regions. For example, if a region in a 2D image shows a suspected crack with high gradient changes, while the 3D point cloud shows that the region is highly abnormal, with depressions or bulges, cross-modal attention will enhance the fusion weight of this region.

[0086] Based on the fused cross-modal features, defect features at different scales are further extracted using the adaptive feature pyramid (ASPP) to enhance sensitivity to minute defects such as microcracks and burrs, as shown in Formula 16:

[0087] By fusing features at different scales, such as 16×16, 32×32, and 64×64, cross-scale defects ranging from 50μm to 200μm are captured, and the fused feature map F is output. out , of which F n×n Feature maps at different scales.

[0088] Step 5: Offline training to generate the detection model: synthesize the defect image I syn Compared with real defect image I real The training set is then mixed to obtain a mixed training set, which is then trained offline to generate a detection model. The detection model is generated offline through training on real and synthetic defect data, and the initial parameters are stored in the defect detection module.

[0089] Step 6: Dynamic Equilibrium Defect Detection: The detection model processes and fuses feature maps F in real time. out Perform defect detection and output defect detection results, including defect location B, category C and confidence level S. Determine whether a defect exists and its category based on the detection results. Furthermore, in defect detection, the defect location B is output through the regression branch, and the defect category C is output through the classification branch. The defect detection weights are dynamically adjusted, and the defect confidence S is calculated based on the accuracy of location B and category C. The defect is then determined based on the confidence S.

[0090] According to the fusion feature map F out The defect category distribution data is dynamically balanced for defect detection and classification, increasing the weight of rare defects and improving bounding box regression to obtain defect location B, category C and confidence S, and then the detection results are determined and output.

[0091] The defect detection module analyzes the entire image to locate and classify defects. The input for defect detection includes the fused feature map F. out And defect category distribution, the sources of defect category distribution include: Dynamic statistics: Calculated in real-time based on the frequency of occurrence of various defects in the training data. Example: If porosity accounts for 70%, scratches 20%, and decarburization 10% in the training set, then the initial distribution D... cts =[0.7,0.2,0.1].

[0092] Balancing strategy: dynamically adjust weights, with rare defect weights α. t For example, the decarbonization weight increases to 0.316, and the porosity decreases to 0.032.

[0093] The bounding box position B=(x,y,w,h) is determined as follows: Regression branch output: Model predicted bounding box center offsets Δx, Δy and size scaling Δw, Δh. Decoding formulas are shown in formula group 17:

[0094] Where x and y are the coordinates of the predicted bounding box center, and c x c y : Grid cell coordinates, s w s h : Scale factor, used to adjust the scaling of width w and height h, p w p h : Preset anchor frame size, σ: Sigmoid function.

[0095] Through formula 18:

[0096] The location loss L of the defect was calculated. loc ,in: loU: Intersection over Union (IoU) ratio between predicted bounding boxes and ground truth bounding boxes; ρ: Distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; d: Length of the diagonal of the bounding box; v: Aspect ratio consistency measure.

[0097]

[0098] Through L loc This method measures the positional deviation between the model-predicted bounding boxes and the true bounding boxes, improving bounding box regression accuracy. It optimizes the defect location B=(x,y,w,h), improving the localization accuracy of small defects, achieving a 12% improvement in mAP@0.5 by minimizing L. loc The model outputs more accurate bounding box coordinates.

[0099] The determination of category C is as follows: Classification branch output: The Softmax layer generates the probability of each category, and selects the category with the highest probability.

[0100] When determining the defect category, use formula 19:

[0101] Dynamically calculate the classification loss L of defects cls ,in: t: Category index, representing the specific defect category being calculated. For example, if category C is crack, then when calculating the loss associated with the crack category, t represents the crack category. p t : The model's predicted probability for class t, output by Softmax; γ=2: Focusing factor, increases the weight of difficult-to-classify samples, such as rare defects; α t : Dynamic weights for rare defects, expressed by formula 20:

[0102] Calculations show that The number of samples in category t is dynamically counted; synthetic data are used. Statistics, adjustment of L cls α t Weighting improves the detection of rare classes.

[0103] Through L cls This measures the model's accuracy in classifying the target category, addresses class imbalance, and dynamically adjusts weights. When detecting rare defects, the weight α for rare defects is used. t Improvement can enhance the classification accuracy of rare defects by minimizing L. cls The model optimizes the classification branch, outputting a more accurate class probability C.

[0104] The confidence level S is determined as follows: According to formula 21:

[0105] Combining the precision of B and C, calculate the defect confidence level S, where p c : Classification probability, such as stoma probability = 0.95; CloU: Positioning accuracy, range 0~1; α=0.7, β=0.3: Weighting coefficients, used to adjust the contribution of the two parts to the final confidence score.

[0106] The reliability of the detection is quantified by combining the classification and localization results with the confidence level S.

[0107] Judge the defect according to the confidence level S, and the judgment criteria are as follows: High confidence level, S > 0.9: Directly determine it as a defect and trigger the sorting mechanism; Medium confidence level, 0.5 < S ≤ 0.9: Manual review; Low confidence level, S ≤ 0.5: Mark as qualified.

[0108] Defect category determination: Select the category C with the largest classification probability.

[0109] Step 7: Real-time resource optimization and feedback: According to the defect detection results and the system state space, perform online training to update the detection model and the scheduling policy model. Among them, the scheduling policy model is generated according to the state space and the action space, and the state space includes the system delay T, the GPU power consumption P, and the defect detection accuracy Acc rare , and the action space includes the computing power allocation ratio.

[0110] Furthermore, in real-time resource optimization, dynamically allocate the system computing power, update the parameters of the detection model according to the defect detection results of manual review, output the computing power allocation strategy based on the scheduling policy model, and allocate the system computing power to each link of multi-modal data acquisition, multi-modal feature fusion, and defect detection according to the computing power allocation strategy.

[0111] Perform real-time resource optimization and model iteration according to the detection results, system delay T, and GPU power consumption P, dynamically allocate computing power, and update the model parameters online.

[0112] The input of real-time resource optimization includes the detection results B, C, S, the system delay T, and the GPU power consumption P; the detection model is generated through offline training, and the updated parameters of the detection model include the classifier weight and the regressor offset, which are updated through backpropagation. The scheduling policy model is generated through online training, the initial Q-value table is initialized by experience, and the update policy of the scheduling policy model is the state-action value in the Q-value table, which is updated through the TD error.

[0113] Output: Model parameter update instruction △θ, GPU computing power allocation strategy P.

[0114] The calculation method of the system delay T is: Record the processing timestamps of each module, such as the preprocessing time t1 and the fusion time t2; the end-to-end delay T = t1 + t2 + … t n .

[0115] The acquisition method of the GPU power consumption P is to directly read the real-time power consumption of the GPU through NVIDIA-SMI or a hardware sensor, unit: watt.

[0116] The computing power allocation strategy is a priority strategy that dynamically allocates GPU computing resources to each module. This strategy is generated by learning the state-action value function Q(s,a) using Q-learning, selecting the action 'a' that maximizes the long-term reward. Example strategies are as follows: Under high load: allocate 60% of computing power to the feature fusion module, 30% to the detection module, and 10% to preprocessing. Under low load: distribute evenly across all modules.

[0117] The sources of model parameter update instructions include: incremental learning: fine-tuning model parameters based on suspected defect samples reviewed by humans; reinforcement learning: updating the parameters of the resource scheduling strategy, i.e., the Q-value table, through Q-Learning.

[0118] The updated parameter types include: Detection model: Network parameters of the dynamically balanced defect detection module, such as convolutional layer weights and classifier parameters. Classifier weights and regressor offsets are updated via backpropagation; Scheduling strategy model: Q-Learning's Q-value table and state-action value matrix. The state-action values ​​in the Q-value table are updated via TD error.

[0119] Update methods include: Detection model: The loss function is optimized through backpropagation, Equation 22: ;

[0120] Scheduling strategy: Update the Q value through time-differential scheduling (TD), Equation 23: .

[0121] The parameter update example is as follows: Detection model fine-tuning: For manually confirmed defect samples, such as 100 newly added decarbonized images, LoRA (low-rank adaptation) technology is used to update only some parameters of the classification layer, avoiding the computational overhead of global training.

[0122] Through formula 24:

[0123] Calculate the policy probability, where P(a|s) represents the probability of taking action a in state s, Q(s,a) represents the Q value of state s and action a, i.e. the value estimate of the combination of state and action, and τ is a parameter that controls the randomness of the choice. When τ is small, the probability of choosing an action with a high Q value is higher; when τ is large, the probability of choosing each action is more even.

[0124] Through formula 25:

[0125] Iteratively update Q(s,a), where η=0.01: learning rate, dynamically adjusted based on historical data; γ=0.9: future reward discount factor, balancing immediate and long-term returns; s ’ Indicates the next state, a ’ Indicates the possible actions in the next state; r: reward function that combines latency, power consumption, and rare defect accuracy; Through formula 26:

[0126] The value of r is calculated, where Acc is the value of r. rare To assess the system's accuracy in detecting rare defects, for example, in lithium battery electrode inspection, the system needs to detect common pores and rare decarburization. In the last 100 inspections, there were actually 10 decarburization defects; the system correctly detected 8 and missed 2. Therefore, the Acc rate is... rare =8 / (8+2)=80%, Acc rare As a reward, the computing power is tilted towards the feature fusion module. The strategy can optimize resource allocation and significantly improve the detection rate of rare defects. By introducing Q-value calculation and the comprehensive latency-power consumption-rare defect accuracy function r, the global optimization of resource allocation and detection performance is achieved.

[0127] An example of a dynamic GPU computing power allocation strategy is as follows: Define states and actions. State s represents the system's current load, GPU usage, and other relevant performance metrics. Action a represents different GPU resource allocation schemes. Design a reward function, r, that reflects the effect of resource allocation, such as reducing latency, lowering power consumption, and increasing task completion speed; Train the model, collect data through simulation or online experiments, and update the Q value using the above formula until the model converges to a stable policy. Application strategy: Once the model is trained, the probability distribution of each possible action a can be calculated using the first formula based on the current state s. Then, an action, i.e. a resource allocation scheme, can be selected to execute. Continuous optimization: In actual deployment, the system's performance can be continuously monitored, and the model's parameters can be adjusted or the model can be retrained as needed to adapt to new workloads and environmental changes.

[0128] Based on the same inventive concept as the artificial intelligence industrial vision inspection method provided in the embodiments of this application, the embodiments of this application also provide an artificial intelligence industrial vision inspection system. If there is anything unclear about the content in the system embodiments, please refer to the corresponding content in the method embodiments.

[0129] Example 2: likeFigure 2 As shown, an artificial intelligence industrial vision inspection system includes: The image data acquisition module is used to acquire original product images and obtain HDR images of the product. HDR 3D point cloud P(x,y,z) and polarization image I pol HDR image I HDR Output to the dynamic preprocessing module: 3D point cloud P(x,y,z) and polarization image I pol Output to the multimodal feature fusion module; The defect sample generation module generates samples based on real defect images from the defect database. real Offline training generates synthetic defect images I syn The output is sent to the dynamic preprocessing module; The dynamic preprocessing module converts HDR images into I HDR and synthetic defect image I syn After proportional mixing and normalization, the preprocessed image I is obtained. ’ Output to the multimodal feature fusion module; The multimodal feature fusion module integrates 3D point cloud P(x,y,z) and polarization image I. pol and preprocessed image I ’ The fusion is performed to obtain the fused feature map F. out The output is sent to the dynamic equilibrium defect detection module; The dynamic balance defect detection module performs a fusion feature map F out Perform analysis and detection, and output defect detection results; The real-time optimization module trains online and updates the computing power allocation strategy and model parameters based on defect detection results and system state space to optimize the system in real time.

[0130] The dynamic preprocessing module performs global illumination compensation and adaptive filtering on the image. The multimodal feature fusion module aligns multimodal features through an attention mechanism and further extracts defect features at different scales using an adaptive feature pyramid, fusing 2D texture, 3D morphology, and polarization characteristics. The dynamic balance defect detection module dynamically adjusts the weights of rare defects and evaluates confidence based on defect classification probability and positioning accuracy.

[0131] Based on the same inventive concept as the artificial intelligence industrial vision inspection method provided in the embodiments of this application, the embodiments of this application also provide an artificial intelligence industrial vision inspection device. If there is anything unclear about the content in the device embodiment, please refer to the corresponding content in the method embodiment.

[0132] Example 3: like Figure 3 As shown, an electronic device is provided, comprising: At least one processor 200; Memory 100 communicatively connected to at least one of the processors; and Network interface 300; The memory 100 stores instructions that can be executed by the processor 200 to implement the aforementioned artificial intelligence industrial vision inspection method.

[0133] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control. The memory 100, processor 200, and network interface 300 are interconnected via a system bus.

[0134] The memory 100 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory, random access memory, static random access memory, read-only memory, electrically erasable programmable read-only memory, programmable read-only memory, magnetic memory, magnetic disk, and optical disk. In some embodiments, the memory 100 may be an internal storage unit of the electronic device, such as the hard disk or memory of the electronic device. In other embodiments, the memory 100 may also be an external storage device of the electronic device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the electronic device. Of course, the memory 100 may include both internal storage units and external storage devices of the electronic device. In the embodiments of this application, the memory 100 is typically used to store the operating system and various application software installed on the electronic device, such as computer-readable instructions of methods. In addition, the memory 100 may also be used to temporarily store various types of data that have been output or will be output.

[0135] The processor 200 may be a central processing unit, a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 200 is typically used to control the overall operation of the electronic device. In this embodiment, the processor 200 is used to execute computer-readable instructions stored in the memory 100 or to process data, for example, to execute computer-readable instructions of the methods provided in this embodiment.

[0136] The network interface 300 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the electronic device and other electronic devices.

[0137] It should be noted that only electronic devices with 100 to 300 components are shown in the figure. However, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0138] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. An artificial intelligence-based industrial visual inspection method, characterized in that, Includes the following steps: Real-time acquisition of product surface images on the production line to obtain HDR images, 3D point clouds and polarization images of the products; Defect data is stored in the defect database, which includes real defect images, defect types, defect locations, process parameters, and metadata. The defect data is used as input to generate diverse defect samples and synthetic defect images. Dynamic preprocessing is performed on HDR images and synthetic defect images to eliminate uneven lighting and noise, and to standardize image contrast. The preprocessed images are then normalized. Multimodal feature fusion is performed on preprocessed images, 3D point clouds, and polarization images to achieve feature dimension matching and obtain a fused feature map; The synthetic defect image is mixed with the real defect image to obtain the mixed training set. The training set is then trained offline to generate the detection model. The detection model processes the fused feature map in real time, performs defect detection, and outputs the defect detection results. The defect detection results include the defect location, category, and confidence level. The existence and category of the defect are determined based on the confidence level. The system state space and action space are constructed. The state space includes system latency, GPU power consumption and defect detection accuracy, while the action space includes computing power allocation ratio. A scheduling strategy model is generated based on the system state space and action space. The system computing power is dynamically allocated through the scheduling strategy model. Online training is performed based on the defect detection results and the system state space to update the detection model and the scheduling strategy model.

2. The artificial intelligence industrial vision inspection method according to claim 1, characterized in that, During the real-time acquisition process, HDR images are synthesized by fusing multiple exposure images, 3D point clouds are generated by structured light decoding and 3D reconstruction, and parallel polarization images and vertical polarization images are acquired by time-division triggering of linear polarization light sources, which are then processed to obtain polarization images.

3. The artificial intelligence industrial vision inspection method according to claim 2, characterized in that, The generation of the synthetic defect image is as follows: the noise vector of the real defect image is converted into an initial synthetic image by the generator of the generative adversarial network. The rationality of the defect in the synthetic image is verified by physical simulation and evaluated and updated to obtain the synthetic defect image. During the physical simulation verification, the defect morphology is simulated based on the finite element stress model, and synthetic images that do not conform to physical laws are eliminated.

4. The artificial intelligence industrial vision inspection method according to claim 1, characterized in that, In the dynamic preprocessing, the HDR image and the synthesized defect image are mixed according to a preset ratio. The mixed image is then subjected to global illumination compensation and adaptive filtering. After pixel normalization, the preprocessed image is output.

5. The artificial intelligence industrial vision inspection method according to claim 4, characterized in that, In the multimodal feature fusion, features of the preprocessed image, 3D point cloud and polarization image are dynamically aligned through cross-modal attention, and defect features at different scales are extracted through adaptive feature pyramid extraction. After fusion, a fused feature map is output.

6. The artificial intelligence industrial vision inspection method according to claim 1, characterized in that, In the defect detection process, the detection model outputs the defect location through a regression branch and the defect category through a classification branch, dynamically adjusts the defect detection weights, calculates the defect confidence based on the location accuracy and classification accuracy, and outputs the defect detection result based on the confidence level.

7. The artificial intelligence industrial vision inspection method according to claim 6, characterized in that, In the dynamic allocation system computing power, the parameters of the detection model are updated based on the defect detection results reviewed by humans, and the computing power allocation strategy is output based on the scheduling strategy model to allocate the system computing power according to the computing power allocation strategy.

8. An artificial intelligence industrial vision inspection system, used in the artificial intelligence industrial vision inspection method according to any one of claims 1 to 7, characterized in that, include: The image data acquisition module is used to acquire original product images and obtain HDR images, 3D point clouds and polarization images of the product. The HDR images are output to the dynamic preprocessing module, and the 3D point clouds and polarization images are output to the multimodal feature fusion module. The defect sample generation module generates synthetic defect images offline based on real defect images in the defect database and outputs them to the dynamic preprocessing module. The dynamic preprocessing module mixes the HDR image and the synthetic defect image in proportion, normalizes them to obtain the preprocessed image, and outputs it to the multimodal feature fusion module. The multimodal feature fusion module fuses 3D point clouds, polarization images, and preprocessed images to obtain a fused feature map, which is then output to the dynamic equilibrium defect detection module. The dynamic balance defect detection module analyzes and detects the fused feature map and outputs the defect detection results. The real-time optimization module trains online and updates the computing power allocation strategy and model parameters based on defect detection results and system state space to optimize the system in real time.

9. The artificial intelligence industrial vision inspection system according to claim 8, characterized in that, The dynamic preprocessing module performs global illumination compensation and adaptive filtering on the image. The multimodal feature fusion module aligns multimodal features through an attention mechanism and further extracts defect features at different scales using an adaptive feature pyramid, fusing 2D texture, 3D morphology, and polarization characteristics. The dynamic balance defect detection module dynamically adjusts the weights of rare defects and evaluates confidence based on defect classification probability and positioning accuracy.

10. An electronic device for use in the artificial intelligence industrial vision inspection method according to any one of claims 1 to 7, characterized in that, include: At least one processor; A memory that is communicatively connected to at least one of the processors; as well as Network interface; The memory stores instructions that can be executed by the processor to implement the artificial intelligence industrial vision inspection method according to any one of claims 1 to 7.

Images