A method and system for detecting defects in a polymer foam product based on machine vision

Abstract

This invention discloses a method and system for defect detection in polymer foam products based on machine vision, belonging to the field of image processing and analysis technology. It includes acquiring optical parameters and surface images of the product; calculating and generating structured light source commands and controlling the light source array imaging to obtain multimodal images; decoupling and generating surface reflection and internal transmission feature maps; fusing the light source commands and feature maps to locate suspicious areas; applying micro-vibrations and acquiring high-speed image sequences; extracting dynamic response data, aligning and stitching it with the decoupled feature maps to generate a fused defect feature vector; and analyzing and outputting the defect type, location, and three-dimensional size parameters. This invention employs an adaptive structured light source imaging and optomechanical multimodal feature fusion technology, enabling accurate identification and three-dimensional quantitative analysis of surface and internal defects in polymer foam products.

Description

Technical Field

[0001] This invention relates to the field of image processing and analysis technology, and in particular to a method and system for detecting defects in polymer foam products based on machine vision. Background Technology

[0002] Polymer foam products are widely used in aerospace, transportation, packaging materials, and other fields due to their excellent properties such as lightweight, high strength, heat insulation, and shock absorption. As structural or functional components, the integrity of their internal structure and surface directly affects the performance and safety of the final product. Therefore, efficient and accurate defect detection during the production process is crucial. Machine vision technology, as a non-contact, highly efficient, and automated inspection method, plays an increasingly important role in industrial product quality control, with image processing and analysis at its core.

[0003] In existing technologies, machine vision-based defect detection methods typically employ fixed lighting schemes, such as using a ring light source or backlight to uniformly illuminate the product, and then acquiring images of the product's surface using an industrial camera. After acquiring the images, image processing algorithms, such as threshold segmentation, edge detection, template matching, or the recently developed deep learning-based image segmentation and classification models, are used to analyze the images to identify potential defects such as scratches, dents, bubbles, and impurities.

[0004] However, the aforementioned existing technologies have significant shortcomings when applied to polymer foam products. These products are semi-transparent with uneven surface gloss, making it difficult for fixed lighting to simultaneously image both the surface and the interior. This often results in strong surface spots and weak internal defect signals, masking even minor defects. Furthermore, normal textures and defect information are often mixed in a single image, making it difficult for algorithms to distinguish them, leading to high false positive and false negative rates. These methods are mostly limited to two-dimensional images and cannot acquire three-dimensional defect information, thus failing to meet the needs for accurate quantitative assessment of product quality. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for detecting defects in polymer foam products based on machine vision. It employs an adaptive structured light source imaging and optomechanical multimodal feature fusion technology, which enables accurate identification and three-dimensional quantitative analysis of surface and internal defects in polymer foam products.

[0006] The above objectives can be achieved through the following approach:

[0007] A machine vision-based defect detection method for polymer foam products includes: acquiring basic optical parameters of the polymer foam product and acquiring an initial surface image of the target product; performing optical calculations on the initial surface image based on the basic optical parameters to generate a structured light source control command; controlling a light source array to image the target product according to the structured light source control command to acquire a multimodal image set; performing feature decoupling processing on the multimodal image set based on the basic optical parameters to generate a surface reflection feature map and an internal transmission feature map; performing parameter-aware fusion processing on the structured light source control command, the surface reflection feature map, and the internal transmission feature map to generate a decoupling enhancement map and suspicious area location information; applying controllable micro-vibration to the target product according to the suspicious area location information and simultaneously acquiring a high-speed image sequence; extracting dynamic response feature data from the high-speed image sequence and spatially aligning and stitching it with the decoupling enhancement map to generate a fused defect feature vector; performing defect classification and quantitative analysis based on the fused defect feature vector and outputting defect type, location, and three-dimensional size parameters.

[0008] Optionally, the generation of structured light source control instructions includes: acquiring the material polarization optical parameters, spectral transmittance, light scattering coefficient, and material absorption coefficient of the polymer foam product to generate basic optical parameters; acquiring an initial surface image of the target product and performing region identification to obtain a highly reflective region and an internal structure of interest region; combining the basic optical parameters to perform polarization optical calculations on the highly reflective region to obtain a first polarization angle combination; performing spectral transmittance and light scattering calculations on the internal structure of interest region to generate a second spectral band and incident angle combination; and integrating the first polarization angle combination and the second spectral band and incident angle combination to generate structured light source control instructions.

[0009] Optionally, the feature decoupling process includes: acquiring a multimodal image set containing specific polarization and spectral information; calculating the surface reflection component of the multimodal image set based on the basic optical parameters and the principle of polarization reflection to generate a surface reflection contribution component; extracting the internal transmission component of the multimodal image set based on the basic optical parameters and the physical laws of light transmission to generate an internal transmission contribution component; performing linear normalization processing on the surface reflection contribution component to generate a surface reflection feature map; and performing edge-preserving noise suppression processing on the internal transmission contribution component to generate an internal transmission feature map.

[0010] Optionally, generating the decoupling enhancement map and suspicious region location information includes: encoding and converting the structured light source control command to generate a parameter feature vector; calculating dynamic fusion weights based on the parameter feature vectors, performing weighted fusion on the surface reflection feature map and the internal transmission feature map to generate a fused feature map; and training and optimizing the fused feature map based on a physical constraint loss function to generate a decoupling enhancement map containing the surface dominant map and the internal dominant map, as well as suspicious region location information.

[0011] Optionally, the training and optimization of the fused feature map based on the physical constraint loss function includes: calculating the gradient distribution constraint of the fused feature map based on the light intensity gradient distribution law of surface defects to generate a surface defect gradient constraint loss; calculating the transmission intensity attenuation constraint of the fused feature map based on the transmission intensity attenuation law of internal defects to generate an internal defect attenuation constraint loss; and weighting the surface defect gradient constraint loss and the internal defect attenuation constraint loss to generate a physical constraint loss function.

[0012] Optionally, the extraction of dynamic response feature data from the high-speed image sequence includes: based on the location information of the suspected area, performing micro-vibration excitation and continuous imaging on the surface of the target product to generate a high-speed image sequence; performing motion estimation calculation on the high-speed image sequence to generate a local deformation displacement matrix; extracting dynamic parameters from the local deformation displacement matrix to generate a vibration frequency spectrum and a damping attenuation coefficient; and calculating dynamic response features based on the vibration frequency spectrum and the damping attenuation coefficient to generate dynamic response feature data.

[0013] Optionally, the spatial alignment and feature stitching includes: performing coordinate transformation processing on the dynamic response feature data based on the location information of the suspicious region to generate a dynamic feature map spatially aligned with the decoupling enhancement map; performing channel concatenation operation on the decoupling enhancement map and the dynamic feature map to generate a multi-channel feature tensor; and performing weighted fusion processing on the multi-channel feature tensor based on an attention mechanism to generate a fused defect feature vector.

[0014] Optionally, the step of performing defect classification and quantification analysis based on the fused defect feature vector includes: determining the defect type of the fused defect feature vector and generating defect type information; performing transmission intensity-depth mapping calculation on the internal defects based on the defect type information and generating a defect depth distribution map; and performing volume correction calculation on the defect depth distribution map to generate defect volume parameters.

[0015] Optionally, the transmission intensity-depth mapping calculation includes: performing an initial depth calculation on the internal dominant map to generate an initial depth distribution map; calibrating the effective path length of the initial depth distribution map to generate a calibration path length parameter; and performing iterative optimization calculation based on the calibration path length parameter to generate a converged defect depth distribution map.

[0016] Based on the same inventive concept, this invention also provides a machine vision-based defect detection system for polymer foam products. The system includes: a data acquisition module for acquiring basic optical parameters of the polymer foam product and acquiring an initial surface image of the target product; a light source strategy generation module for performing optical calculations on the initial surface image based on the basic optical parameters to generate structured light source control commands; a multimodal imaging module for controlling a light source array to image the target product according to the structured light source control commands, acquiring a multimodal image set; and a physical feature decoupling module for performing feature decoupling processing on the multimodal image set based on the basic optical parameters to generate surface reflection feature maps and internal... The system includes: a transmission feature map; a parameter-aware fusion module for performing parameter-aware fusion processing on the structured light source control command, the surface reflection feature map, and the internal transmission feature map to generate a decoupled enhancement map and suspicious area location information; a vibration-assisted fine inspection module for applying controllable micro-vibration to the target product based on the suspicious area location information and simultaneously acquiring high-speed image sequences; a multimodal feature fusion module for extracting dynamic response feature data from the high-speed image sequences and spatially aligning and stitching them with the decoupled enhancement map to generate a fused defect feature vector; and a defect three-dimensional quantization module for performing defect classification and quantification analysis based on the fused defect feature vector, outputting defect type, location, and three-dimensional size parameters.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] This invention achieves high-precision and high-reliability detection of defects in polymer foam products by constructing a multi-stage, multi-modal detection framework. First, the method generates an adaptive structured light source strategy based on the material's inherent optical properties, optimizing illumination for highly reflective areas and internal structural regions. This improves image acquisition quality from the source, effectively suppresses strong surface reflection interference, enhances the ability to penetrate and detect minute internal defects, and ensures a solid data foundation for subsequent analysis.

[0019] This invention effectively separates aliased optical information by combining physical models with data-driven approaches. Based on physical laws such as polarization reflection and transmission attenuation, surface and internal features are decoupled, and optimization is performed using a physical constraint loss function. This ensures the physical authenticity of the feature map, avoids misjudgments and missed detections caused by the mutual interference between surface texture and internal structure in traditional methods, and significantly improves the accuracy of defect identification.

[0020] This invention innovatively introduces a multimodal information fusion mechanism combining optical and mechanical data, enabling a deep understanding of defect properties. By applying micro-vibrations to suspicious areas identified through initial optical screening and analyzing their dynamic response characteristics, the static morphological information of the defect is combined with its dynamic mechanical properties. This cross-modal verification method can accurately distinguish between genuine structural defects and benign optical artifacts, greatly enhancing the confidence of the detection results and providing a novel basis for determining the type of complex internal defects.

[0021] This invention overcomes the limitations of traditional two-dimensional inspection, achieving precise three-dimensional quantification of defects. This method not only determines the type and location of defects but also iteratively optimizes the calculation of defect depth and volume by integrating optical transmission intensity and dynamic mechanical response data. This method overcomes the errors of single-modal measurement, making the analysis results of three-dimensional dimensional parameters closer to physical reality, providing reliable quantitative data support for refined product quality assessment and digital improvement of production processes.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic flowchart of a defect detection method for polymer foam products based on machine vision, according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the optical feature decoupling process according to an embodiment of the present invention.

[0026] Figure 3 This is a flowchart of the dynamic response feature extraction process according to an embodiment of the present invention.

[0027] Figure 4 This is a light intensity depth calibration curve diagram according to an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of a defect detection system for polymer foam products based on machine vision, according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Reference Figure 1 One embodiment of the present invention proposes a defect detection method and system for polymer foam products based on machine vision. It adopts a technical solution of adaptive structured light source imaging and optical-mechanical multimodal feature fusion, which can realize accurate identification and three-dimensional quantitative analysis of surface and internal defects of polymer foam products.

[0031] The method described in this embodiment specifically includes:

[0032] Obtain the basic optical parameters of the polymer foam product and acquire the initial surface image of the target product;

[0033] Based on the aforementioned basic optical parameters, optical calculations are performed on the initial surface image to generate structured light source control commands;

[0034] The structured light source control command controls the light source array to image the target product, thereby acquiring a multimodal image set;

[0035] Based on the aforementioned fundamental optical parameters, the multimodal image set is subjected to feature decoupling processing to generate surface reflection feature maps and internal transmission feature maps;

[0036] The structured light source control command, the surface reflection feature map, and the internal transmission feature map are subjected to parameter-aware fusion processing to generate a decoupled enhancement map and suspicious area location information;

[0037] Based on the location information of the suspicious area, a controllable micro-vibration is applied to the target product and a high-speed image sequence is acquired simultaneously.

[0038] The dynamic response feature data extracted from the high-speed image sequence is spatially aligned and feature-stitched with the decoupled enhancement map to generate a fused defect feature vector.

[0039] Based on the fused defect feature vector, defect classification and quantification analysis are performed, and the defect type, location, and three-dimensional size parameters are output.

[0040] Specifically, firstly, based on the inherent optical parameters of polymer foam products, an adaptive structured light source control command is generated through physical calculations to acquire initial multimodal images containing surface and internal information in an optimized manner. Subsequently, a physical model is used to decouple the mixed signals in the images, separating and generating independent surface reflection and internal transmission feature maps. Building upon this, a parameter-aware fusion mechanism is introduced to intelligently fuse the illumination strategy parameters during imaging with the separated feature maps, thereby enhancing defect features and accurately locating suspicious areas. For these suspicious areas, the method switches from the optical detection domain to the mechanical detection domain, applying controllable micro-vibrations and acquiring their dynamic responses to obtain dynamic data characterizing the local structural stiffness and damping properties. Finally, the spatially aligned optical enhancement map and dynamic response features are deeply stitched and fused to generate a fused defect feature vector containing both optical and mechanical information, which serves as the basis for final defect classification and three-dimensional dimensional quantification analysis. This method can not only determine the existence and type of defects, but also break through the limitations of two-dimensional images and output the three-dimensional size parameters of defects, providing unprecedented quantitative and high-precision data support for product quality control, and realizing a leap from simple defect detection to comprehensive defect quantitative analysis.

[0041] Optionally, the generated structured light source control instructions include:

[0042] Obtain the material polarization optical parameters, spectral transmittance, light scattering coefficient, and material absorption coefficient of polymer foam products to generate basic optical parameters;

[0043] Acquire initial surface images of the target product and perform region identification to obtain highly reflective areas and areas of interest in the internal structure;

[0044] Based on the aforementioned basic optical parameters, polarization optics calculations are performed on the highly reflective region to obtain the first polarization angle combination;

[0045] Spectral transmittance and light scattering are calculated for the region of interest in the internal structure to generate a second spectral band and incident angle combination.

[0046] By integrating the first polarization angle combination with the second spectral band and incident angle combination, a structured light source control command is generated.

[0047] Specifically, this method first measures the basic optical parameters of the polymer foam product using specialized instruments. For example, an ellipsometer or goniometer is used to determine the material's polarization optical parameters, which describe the change in polarization state of light as it reflects off the product's surface. A spectrophotometer is used to measure the material's spectral transmittance at different wavelengths to determine the optimal wavelength range for light penetration. Simultaneously, by analyzing the intensity attenuation of light after passing through a sample of known thickness, the material's light scattering coefficient and absorption coefficient are calculated. These two coefficients together determine the degree of light attenuation as it propagates within the material. After acquiring the basic optical parameters, the system uses a standard industrial camera to acquire an initial surface image of the target product under uniform illumination. This initial surface image is then analyzed using an image processing algorithm for region identification. The algorithm segments the image based on pixel brightness values. Regions with brightness exceeding a preset threshold (obtained dynamically by statistically analyzing the mean and standard deviation of image brightness in a qualified sample library, combined with material gloss and local contrast) are identified as high-reflectivity areas. These areas are prone to specular reflection, interfering with the observation of surface defects. Other areas in the image that require focused detection of internal structures based on product design or texture features are identified as areas of interest for internal structure analysis. The system performs targeted optical calculations for the two different types of identified regions. For highly reflective regions, the system calls upon the material polarization optical parameters from the basic optical parameters to perform polarization optical calculations. A specular reflection suppression model is established using Fresnel's law of reflection, and the first polarization angle combination is solved analytically.

[0048]

[0049] Among them, I spec (θ p To minimize the intensity of specularly reflected light, I0 is the intensity of incident light, and θ is... p Let θ be the polarization angle of the light source. t Let n be the angle of refraction, and n1 and n2 be the refractive indices of air and the material, respectively. Differentiate the objective function. Analytical calculation makes I spec smallest and the corresponding camera analyzer angle Generate the first polarization angle combination For the region of interest in the internal structure, the system calculates based on the spectral transmittance and light scattering coefficient from the fundamental optical parameters, aiming to maximize the light penetration depth and reduce the impact of internal scattering on image quality. The formula is:

[0050]

[0051] Where, d eff (λ,θ i To maximize the effective penetration depth, λ is the spectral wavelength, and θ iLet θ be the incident angle of the light source, α(λ) be the material absorption coefficient, β(λ) be the material scattering coefficient, and g(θ) be the scattering coefficient. i ) is the scattering phase function. The constraint condition is: λ min ≤λ≤λ max (System available spectral range), 0°≤θ i ≤70° (avoid total internal reflection). Solve analytically within the constraints. d eff (λ,θ i Generate the optimal spectral band λ * and angle of incidence Generate a second spectral band and incident angle combination Finally, the system integrates the optimal parameters calculated for the two types of regions. The integration process combines the first polarization angle combination with the second spectral band and incident angle combination into a unified, detailed control sequence, namely, a structured light source control command. This command is sent to a programmable light source array, precisely controlling the output of each light source unit in the array, including its emission wavelength, polarization state, and illumination angle on the target product. In this way, different regions of the target product will be subjected to optimized, non-uniform structured illumination according to their own characteristics. This method abandons the fixed, universal lighting method used in traditional inspections, instead dynamically generating customized lighting schemes based on the optical properties of the material itself and the specific conditions of the surface of the inspected product.

[0052] Optionally, the feature decoupling process includes:

[0053] Acquire a collection of multimodal images containing specific polarization and spectral information;

[0054] Based on the aforementioned fundamental optical parameters and the principle of polarization reflection, surface reflection components are calculated for the multimodal image set to generate surface reflection contribution components.

[0055] Based on the aforementioned fundamental optical parameters and the physical laws of light transmission, the internal transmission components of the multimodal image set are extracted to generate internal transmission contribution components.

[0056] The surface reflection contribution components are linearly normalized to generate a surface reflection feature map;

[0057] The internal transmission contribution component is subjected to edge-preserving noise suppression processing to generate an internal transmission feature map.

[0058] Specifically, the first step is to acquire a set of multimodal images containing specific polarization and spectral information. Each image in this set was captured under previously generated structured light source control commands, thus containing mixed optical information of surface reflection and internal transmission. To separate these two types of information, the surface reflection component is calculated first. This calculation is based on the principle of polarized reflection, which states that specular reflection from a material surface typically retains its polarization, while light that enters the material and undergoes multiple scatterings before being transmitted mostly loses its polarization. This method utilizes images with the same polarization configuration and orthogonal polarization configuration from the multimodal image set, combined with basic optical parameters, for calculation. The formula for calculating the surface reflection contribution component is:

[0059] I s =I || -I ⊥ ,

[0060] Among them, I s I represents the intensity of the calculated surface reflection contribution component; || I represents the image intensity acquired under the same polarization configuration; ⊥ This represents the image intensity acquired under orthogonal polarization configuration. After obtaining the surface reflection contribution component, the internal transmission component can be extracted. Since the internal transmission component V is unpolarized, its energy is evenly distributed in both the co-directional and orthogonal polarization channels. Under orthogonal polarization configuration, ideal surface reflection light is completely filtered out, therefore the acquired image intensity I... ⊥ It only represents half of the internal transmission component, therefore the total intensity I of the internal transmission contribution component is... t =2*I ⊥ Subsequently, linear normalization is performed on the surface reflection contribution components, scaling their pixel intensity values ​​to a standardized range to generate a high-contrast, easily analyzable surface reflection feature map. The calculation process can be represented as follows:

[0061] I norm =(I s -I min ) / (I max -I min ),

[0062] Among them, I norm I represents the normalized pixel intensity. s I contributes pixel intensity to the component image of the original surface reflection. max and I minThese represent the maximum and minimum pixel intensities in the image, respectively. For the internal transmission contribution component, since the scattering process of light within the foam material introduces a large amount of random noise, edge-preserving noise suppression is employed, such as using a bilateral filter. This effectively protects the contour information of internal defects while smoothing out noise, ultimately generating a clear internal transmission feature map. For example... Figure 2 As shown, this invention successfully separates surface reflection feature maps and internal transmission feature maps by performing feature decoupling processing on multimodal optical images, effectively solving the technical problem of overlapping surface features and internal defects in the original mixed images. This separation of feature sources fundamentally avoids the interference of strong surface reflection on internal defect detection and the confusion of internal structural shadows on surface texture judgment, providing pure and high-contrast input data for subsequent accurate defect classification and quantitative analysis, greatly improving the accuracy and reliability of detection.

[0063] Optionally, the generation of the decoupling enhancement map and the location information of the suspected area includes:

[0064] The structured light source control commands are encoded and converted to generate parameter feature vectors;

[0065] Based on the parameter feature vector, a dynamic fusion weight is calculated, and a weighted fusion is performed on the surface reflection feature map and the internal transmission feature map to generate a fused feature map.

[0066] The fused feature map is trained and optimized based on a physical constraint loss function to generate a decoupled enhanced map containing a surface dominant map, an internal dominant map, and location information of suspicious regions.

[0067] Specifically, the first step is to encode and convert the control commands for the structured light source used for imaging. These commands include the polarization angle θ. pol Spectral band λ spec and the incident angle θ inc Specific physical parameters, etc. The encoding and conversion process is as follows: θ pol , λ spec and θ inc After normalization, the vectors are concatenated into a numerical vector, namely the parameter feature vector V. p For example, for each parameter x, according to its physically feasible range [x min ,x max Perform min-max normalization: x norm =(xx) min ) / (x max -x min ), then V p =[θ pol,norm ,λ spec,norm ,θ inc,normThis vector can be directly processed by subsequent algorithms. Then, based on this vector, dynamic fusion weights are calculated using a pre-defined mapping function. This "pre-defined mapping function" is implemented by a small, trainable fully connected neural network. It takes the encoded parameter feature vector as input and learns to non-linearly map these parameters through hidden layers. Finally, its output layer generates two scalar values: surface feature weights w. s and internal feature weights w t The design principle of this network is that when the parameters in the illumination command (such as orthogonal polarization) are intended to prioritize the detection of surface defects, the model learns to output higher surface feature weights w. s When the parameters in the instruction (such as using high-transmittance spectroscopy) are designed to prioritize the detection of internal defects, the model learns to output higher internal feature weights w. t Next, based on the calculated dynamic fusion weights, the system performs weighted fusion on the previously generated surface reflection feature map and internal transmission feature map to generate a preliminary fused feature map. This weighted fusion process can be represented as:

[0068] F fusion =w s *M surface +w t *M internal ,

[0069] Among them, F fusion M represents the fused feature map. surface M represents the surface reflection feature map. internal This represents the internal transmission feature map. This fused feature map is not the final result, but rather an intermediate product input into a physically constrained iterative optimization process, trained and optimized using a physically constrained loss function. This optimization process is performed by a lightweight convolutional neural network (CNN), specifically employing a U-Net architecture. The U-Net network takes the fused feature map as input and, through its encoder-decoder structure and skip connections, learns to deeply separate and refine the mixed features. The network's final output consists of two key results: first, a decoupled enhancement map, which contains two channels: a surface dominant map refined from surface features and an internal dominant map refined from internal features; second, location information of suspicious regions, typically represented as a binary mask image or a set of coordinates, precisely marking the location of potential defects. This method is not merely a simple image overlay, but a deep reconstruction of the fusion result, ensuring that the final decoupled enhancement map enhances defect features while conforming to optical physics, effectively avoiding artifacts and misinformation.

[0070] Optionally, the step of training and optimizing the fused feature map based on the physical constraint loss function includes:

[0071] Based on the light intensity gradient distribution law of surface defects, gradient distribution constraint calculation is performed on the fused feature map to generate surface defect gradient constraint loss.

[0072] Based on the transmission intensity attenuation law of internal defects, the transmission intensity attenuation constraint is calculated on the fused feature map to generate internal defect attenuation constraint loss.

[0073] The physical constraint loss function is generated by weighting and combining the surface defect gradient constraint loss and the internal defect attenuation constraint loss.

[0074] Specifically, the construction of the physical constraint loss function in this method aims to guide and optimize the fused feature map by introducing prior physical knowledge of defects. This process is first based on the universal characteristic of surface defects in optical imaging, namely that they cause abrupt changes in local surface normals, leading to drastic changes in reflected light intensity. This characteristic is summarized as the light intensity gradient distribution law of surface defects. To utilize this law, the system optimizes the surface dominant map S output by the U-Net network. pred Perform gradient distribution constraint calculation. Surface defect gradient constraint loss L grad The design goal is to maximize S pred The gradient response, its mathematical expression is:

[0075]

[0076] Where * represents the convolution operation, G x G y , where 1 represents the pixel index and N represents the total number of pixels. This loss function motivates the network to generate S values ​​with high gradient responses at surface defects by minimizing the negative mean of the gradient magnitude. pred Simultaneously, the system constrains internal defects. Internal defects, such as voids or impurities, have optical properties different from the matrix material, causing additional absorption or scattering of transmitted light along its propagation path. This results in a significant reduction in the light intensity transmitted to the camera sensor compared to the surrounding normal area. This is defined as the transmission intensity attenuation law of internal defects. Based on this law, the system defines the internal dominant graph I of the U-Net network output. pred Perform transmission intensity attenuation constraint calculation. Internal defect attenuation constraint loss L. atten Aimed at enhancing the contrast between internal defect areas (low intensity) and the surrounding background (high intensity), its mathematical expression can be defined as:

[0077]

[0078] Where j is the pixel index predicted as an internal defect region, and M is the total number of pixels in that region. This loss function minimizes the average light intensity of the predicted defect region, thereby causing the network output to show a significant attenuation of light intensity at the internal defect location. pred This prediction region can be determined by analyzing I. pred Low-threshold segmentation is dynamically obtained. Finally, the system weights and combines the surface defect gradient constraint loss and internal defect attenuation constraint loss calculated based on the two physical priors mentioned above to form the final physical constraint loss function. Its expression is:

[0079] L phys =λ s *L grad +λ i *L atten ,

[0080] Among them, L phys L represents the final physical constraint loss function. grad It is the calculated surface defect gradient constraint loss, L atten This is the calculated internal defect attenuation constraint loss. λ s and λ i Two hyperparameters, or weighting coefficients, are dimensionless scalars used to balance the relative importance of the two types of losses in the total loss function. Their values ​​can be adjusted according to the different emphases on surface and internal defects in the detection task. This physical constraint significantly improves the model's ability to identify real defects, effectively suppresses artifacts and false alarms caused by factors such as uneven lighting or normal material surface textures, and ultimately generates a physically more reliable and feature-pure decoupled enhancement map, improving the robustness and accuracy of defect detection.

[0081] Optionally, extracting dynamic response feature data from the high-speed image sequence includes:

[0082] Based on the location information of the suspicious area, the surface of the target product is subjected to micro-vibration excitation and continuous imaging to generate a high-speed image sequence.

[0083] Motion estimation calculations are performed on the high-speed image sequence to generate a local deformation displacement matrix;

[0084] Dynamic parameters are extracted from the local deformation displacement matrix to generate the vibration frequency spectrum and damping attenuation coefficient;

[0085] Dynamic response characteristics are calculated based on the vibration frequency spectrum and the damping attenuation coefficient to generate dynamic response characteristic data.

[0086] Specifically, the process of extracting dynamic response feature data in this method begins with utilizing previously determined location information of suspicious areas. Based on this location information, the system precisely aligns a micro-vibration excitation source, such as a piezoelectric actuator, with the suspicious area on the surface of the target artifact. The excitation source applies a preset, energy-controllable mechanical pulse or frequency sweep vibration, causing a small forced vibration in the local area. Simultaneously, a high-speed camera positioned above is activated, continuously imaging the area at a frame rate far exceeding that of conventional imaging, thereby capturing the entire process of vibration from generation to decay and generating a high-speed image sequence containing rich temporal dynamic information. After acquiring the high-speed image sequence, the system performs motion estimation calculations. This calculation employs optical flow methods (such as the Lucas-Kanadee method or...). The method involves analyzing the assumption of constant pixel brightness between two consecutive frames I(x,y,t) and I(x,y,t+1) in a sequence, and calculating the displacement vector d(x,y,t) of each pixel (x,y) in the suspected region at time t. The set of displacement vectors of all pixels over the entire time series constitutes the local deformation displacement matrix M. disp This matrix records in detail the trajectory of each point within the suspected region during the vibration process. Next, dynamic parameters are extracted from this matrix. This process involves M... disp Each pixel (x, y) in the image is analyzed individually: to extract the vibrational frequency spectrum (S). freq The system extracts the displacement signal d(x,y,t) of the point in the time dimension and applies a Fast Fourier Transform (FFT) to it, transforming it from the time domain to the frequency domain.

[0087]

[0088] Among them, S freq (f) represents the vibration amplitude at frequency f, which constitutes the vibration frequency spectrum. This represents the Fast Fourier Transform operator. By analyzing this spectrum, the natural vibration frequency and response amplitude at this point can be identified. Simultaneously, by analyzing the envelope of the vibration amplitude decaying over time, the damping coefficient describing the energy dissipation rate can be calculated. The system first extracts the vibration envelope A(t) of the displacement signal d(x,y,t), which describes the decay process of the vibration amplitude over time. Subsequently, an exponential decay curve is fitted to this envelope:

[0089]

[0090] Where A(t) represents the vibration amplitude envelope at time t, A0 represents the initial vibration amplitude, ζ represents the damping attenuation coefficient to be solved, which is a dimensionless parameter, and w nLet represent the system's natural angular frequency determined by the frequency spectrum, and t be time. Using fitting algorithms such as the least squares method, the damping attenuation coefficient ζ can be accurately calculated from the experimental data; this coefficient directly reflects the rate of energy dissipation. Finally, based on the extracted vibration frequency spectrum S... freq The dynamic response characteristics are calculated using the damping attenuation coefficient ζ. This calculation extracts the corresponding vibration frequency spectrum S for each coordinate point (x, y) within the suspected region. freq The damping attenuation coefficient ζ is organized into a two-dimensional characteristic matrix S. dyn (x,y)=[S freq (x,y),ζ(x,y)]. For example... Figure 3 The diagram shows the flowchart for dynamic response feature extraction. This method can effectively confirm whether a suspicious area initially identified by optical methods is a real structural defect using dynamic response feature data. It can also distinguish between different types of defects, greatly enhancing the confidence of detection and the ability to identify complex internal defects, thus achieving a leap from "seeing" defects to "perceiving their physical properties".

[0091] Optionally, the spatial alignment and feature stitching include:

[0092] Based on the location information of the suspected area, the dynamic response feature data is subjected to coordinate transformation processing to generate a dynamic feature map that is aligned with the decoupling enhancement graph space;

[0093] The decoupled enhancement map and the dynamic feature map are cascaded through channels to generate a multi-channel feature tensor.

[0094] The multi-channel feature tensor is weighted and fused based on an attention mechanism to generate a fusion defect feature vector.

[0095] Specifically, this method first performs coordinate transformation on the dynamic response feature data. This process utilizes existing location information of suspected areas to precisely map the dynamic response feature data obtained from vibration tests, represented in physical coordinates, such as the vibration frequency spectrum and damping attenuation coefficient, to the same image pixel coordinate system as the decoupling enhancement map. This is equivalent to assigning each pixel of the decoupling enhancement map its corresponding dynamic mechanical properties, thereby generating one or more dynamic feature maps that are spatially perfectly aligned with the decoupling enhancement map, with each channel representing a dynamic parameter. After spatial alignment, the system integrates these two feature maps from different sources. This integration is achieved through channel cascading. Specifically, the system stacks the decoupling enhancement map and the newly generated dynamic feature map along the dimension of the feature channels. If the original decoupling enhancement map is a three-dimensional data block with a specific height, width, and several optical feature channels, then the cascading operation forms a thicker "data block" with the same height and width but more channels, containing both optical and mechanical information—that is, a multi-channel feature tensor. Finally, the system employs a more intelligent fusion strategy: a weighted fusion processing of the multi-channel feature tensor based on a channel attention mechanism (specifically, the Squeeze-and-Excitation module). The core of this mechanism is learning the importance weights of each feature channel, generating a set of dynamic weights, and using these weights to perform weighted summation or more complex operations on the multi-channel feature tensor, thereby highlighting key features and suppressing irrelevant or redundant information. After this adaptive weighted fusion processing, the most discriminative information from both optical and mechanical modalities is extracted and compressed, ultimately generating a highly condensed information carrier—the fused defect feature vector. This vector provides the most direct input for subsequent classification and quantitative analysis. This collaborative analysis mechanism greatly enhances the robustness and representational ability of the features, enabling the final fused defect feature vector to more comprehensively and accurately describe the essence of the defect, laying a solid foundation for achieving high-precision defect classification and quantitative analysis.

[0096] Optionally, performing defect classification and quantification analysis based on the fused defect feature vector includes:

[0097] The fused defect feature vector is used to determine the defect type and generate defect type information;

[0098] Based on the defect type information, a transmission intensity-depth mapping calculation is performed on the internal defects to generate a defect depth distribution map.

[0099] The volume correction calculation is performed on the defect depth distribution map to generate defect volume parameters.

[0100] Specifically, the defect classification and quantification analysis process of this method begins with the analysis of the fused defect feature vector. This vector, as a high-dimensional data carrier integrating optical and mechanical information, is first fed into a pre-trained classification model for defect type determination. This model is preferably a multilayer perceptron (MLP) neural network. The input layer receives the fused defect feature vector, where each dimension represents a specific physical quantity. The MLP model performs nonlinear transformations and combinations on these input features through at least one hidden layer, learning the unique distribution patterns of different defect types in the high-dimensional feature space. The output layer of the model uses the Softmax activation function, outputting a probability distribution vector. The category corresponding to the maximum value is determined as the final defect type information. This model is trained through supervised learning on a dataset containing a large number of known defect type samples. When an internal defect is determined, the system initiates the transmission intensity-depth mapping calculation. This calculation is specifically used to process the previously generated internal dominant map, which clearly reflects the intensity changes of light after penetrating the material. The system, based on the physical relationship model between light intensity and defect depth (i.e., the light intensity attenuation law), converts the intensity value of each pixel in the image into a specific depth value, thereby generating a defect depth distribution map. This map visually displays the morphology of the defect in the depth direction. However, depth estimation based solely on optical data may have biases. Therefore, the system introduces the damping attenuation coefficient from the dynamic response feature data for volume correction calculation. The system uses this damping attenuation coefficient ζ(x,y) as a correction factor to correct the defect depth distribution map. The correction formula can be expressed as:

[0101] D corr (x,y)=D raw (x,y)*(1+C vol ·ζ(x,y)),

[0102] Among them, D raw (x,y) represents the original pixel depth values ​​in the defect depth distribution map, D corr (x,y) represents the corrected pixel depth value, C volThis is a material volume correction coefficient obtained through experimental calibration. It is derived by quantitatively analyzing a series of defect samples with known true volumes and damping characteristics, and then through regression analysis or optimization algorithms. After this correction calculation, the system integrates the depth distribution map and finally outputs accurate defect volume parameters. Combining the above steps, the system completes a comprehensive analysis of defect type, location, and three-dimensional size parameters. The significant technical advantage of this method lies in its leap from two-dimensional defect identification to three-dimensional defect quantification, and its significant improvement in the accuracy of the analysis results. By analyzing the fused defect feature vector that integrates optical and mechanical properties, it can not only accurately distinguish between surface and internal defects, but also make more detailed judgments on defect types, greatly improving the accuracy and reliability of the classification results.

[0103] Optionally, the transmission intensity-depth mapping calculation includes:

[0104] The initial depth of the internal dominant graph is calculated to generate an initial depth distribution map;

[0105] The initial depth distribution map is calibrated for effective path length to generate calibration path length parameters;

[0106] Iterative optimization calculations are performed based on the calibration path length parameter to generate a converged defect depth distribution map.

[0107] Specifically, the transmission intensity-depth mapping calculation in this method begins with the establishment of a fundamental physical model. The system first establishes a transmission intensity attenuation equation describing the attenuation of light propagating through the material, based on the measured material absorption coefficient; this equation is typically a variant of the Beer-Lambert law. Subsequently, the system uses this equation to calculate the initial depth of the internal dominant map. In this calculation, the light intensity value of each pixel in the internal dominant map is considered as the transmitted light intensity after passing through the defect, while the average light intensity of its neighboring normal region is used as a reference for the incident light intensity. By solving this equation, the initial depth value corresponding to each pixel can be obtained, and these values ​​are aggregated to form the initial depth distribution map. This calculation can be expressed as:

[0108] z init =-(1 / n)*ln(I defect / I normal ),

[0109] Among them, z init The initial depth is calculated, n is the known material absorption coefficient, and I is... defect I represents the pixel light intensity at the defect location in the internal dominant image. normal The reference light intensity is set to the normal region adjacent to the defect. Next, to correct for initial calculation errors, the system extracts the defect deformation from the high-speed image sequence, specifically the maximum vibration amplitude A on the surface of the defect region. maxThe system uses a preset calibration function to adjust parameter A. max This is transformed into a correction of the initial optical depth, generating a calibration path length parameter z that is closer to the physical reality. calib :

[0110] z calib =z init *(1+p*A max ),

[0111] Where p is the material property calibration coefficient, which is determined by fitting the linear relationship between the vibration amplitude and the optical depth deviation of a sample with known depth defect through a comparative experiment of vibration testing and optical measurement. It characterizes the depth correction ratio corresponding to a unit vibration amplitude. Figure 4 The depth calibration process is illustrated, with the dashed line representing the defect light intensity distribution, the solid line representing the initial optical depth, and the dotted-dash line representing the corrected depth after calibration. Finally, iterative optimization calculations are performed based on the calibration path length parameter. This process uses the calibrated z-axis... calib Starting with z0 as the iteration starting point, the depth value is continuously optimized by minimizing the error between the theoretical and measured light intensity. First, a target error function L(z) is established to quantify the theoretical light intensity I corresponding to the current depth estimate z. normal ·e -αz The actual measured defect light intensity I defect The squared difference between them, L(z) = (I defect -I normal ·e -αz ) 2 Then, gradient descent is used to iteratively update the depth value z along the negative gradient direction of the error function until convergence. The update rule for the (k+1)th iteration is:

[0112]

[0113] Among them, z k Let be the depth estimate for the k-th iteration, and η be the learning rate (a hyperparameter that controls the update step size). Let the error function L(z) be in z kThe gradient at the point is determined. This iterative process ensures that the final generated defect depth distribution map simultaneously satisfies both optical attenuation laws and mechanical response constraints, thus obtaining a physically more reliable and convergent defect depth distribution map. This method innovatively introduces the mechanical parameter of defect deformation, which provides crucial structural constraint information for defect depth from a completely independent physical dimension. By calibrating the initial depth and performing iterative optimization, this method effectively compensates for the inherent limitations of the optical model, making the final defect depth distribution map no longer a simple optical calculation result, but a more reliable physical measure verified by both optics and mechanics, laying a solid foundation for achieving high-precision 3D defect size reconstruction.

[0114] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides a defect detection system for polymer foam products based on machine vision, the system comprising:

[0115] The data acquisition module is used to obtain the basic optical parameters of polymer foam products and to acquire the initial surface image of the target product.

[0116] The light source strategy generation module is used to perform optical calculations on the initial surface image based on the basic optical parameters and generate structured light source control instructions.

[0117] A multimodal imaging module is used to control a light source array to image the target product according to the structured light source control command, thereby acquiring a multimodal image set;

[0118] The physical feature decoupling module is used to perform feature decoupling processing on the multimodal image set based on the basic optical parameters to generate surface reflection feature map and internal transmission feature map;

[0119] The parameter-aware fusion module is used to perform parameter-aware fusion processing on the structured light source control command, the surface reflection feature map, and the internal transmission feature map to generate a decoupled enhancement map and suspicious area location information;

[0120] The vibration-assisted precision inspection module is used to apply controllable micro-vibration to the target product based on the location information of the suspicious area and simultaneously acquire high-speed image sequences.

[0121] The multimodal feature fusion module is used to extract dynamic response feature data from the high-speed image sequence and spatially align and stitch the decoupled enhancement map to generate a fused defect feature vector;

[0122] The defect 3D quantization module is used to perform defect classification and quantification analysis based on the fused defect feature vector, and output the defect type, location and 3D size parameters.

[0123] To verify the feasibility of this invention in practice, it was applied to the online defect detection stage of a polymer foam product production line. This production line mainly produces polyurethane foam sound-absorbing panels for automotive interiors and polystyrene foam cushioning materials for electronic product packaging.

[0124] This embodiment selected two typical polymer foam products from the production line: polyurethane sound-absorbing panels (with surface scratches, pits, and uneven internal pores) and polystyrene cushioning materials (with internal voids, impurities, and delamination) for a three-month online testing experiment. Simultaneously, a testing system equipped with a traditional fixed light source and a single vision algorithm was set up as a control group.

[0125] First, the system uses specialized equipment to pre-measure the fundamental optical parameters of polyurethane and polystyrene foam, including material polarization optical parameters, spectral transmittance, light scattering coefficient, and material absorption coefficient, and acquires initial surface images of the target product. Based on these fundamental optical parameters and feature recognition of the initial image regions, the system dynamically generates structured light source control commands, precisely controlling the light source array to perform customized multi-angle, multi-wavelength, and multi-polarization state imaging of the product, thereby acquiring an initial multimodal image set containing rich optical information.

[0126] Secondly, for the acquired multimodal image set, the system performs image feature decoupling processing based on fundamental optical parameters, polarization reflection principles, and the physical laws of light transmission. This successfully separates and generates clean surface reflection feature maps and internal transmission feature maps, effectively suppressing the mutual interference between strong surface reflection and internal scattering. Subsequently, the system performs parameter-aware fusion processing on the parameter feature vectors of the structured light source control commands, the surface reflection feature map, and the internal transmission feature map. Through training optimization based on a physical constraint loss function, a decoupled enhanced map containing the surface dominant map and the internal dominant map is generated, accurately identifying the location information of suspicious areas.

[0127] Furthermore, for suspicious areas identified in the initial optical screening, the system applies controllable micro-vibrations via a vibration-assisted fine inspection module (e.g., using a piezoelectric ceramic actuator to perform micro-vibrations at a frequency of 100Hz and an amplitude of 20μm in the suspicious area), while simultaneously acquiring high-speed image sequences (1000fps). The system performs motion estimation calculations on the high-speed image sequences, extracts the local deformation displacement matrix, and further obtains dynamic response feature data such as vibration frequency spectrum and damping attenuation coefficient. Subsequently, the multimodal feature fusion module performs spatial alignment and channel cascading operations on these dynamic response feature data with the decoupled enhancement map, and performs weighted fusion through an attention mechanism to generate the final fused defect feature vector.

[0128] Finally, the 3D defect quantification module performs defect classification and quantification analysis based on the fused defect feature vector. The system utilizes optical and mechanical information from the fused vector to accurately determine the defect type (e.g., scratches, dents, bubbles, delamination), and performs transmission intensity-depth mapping calculations for internal defects (e.g., bubbles, delamination) to generate a defect depth distribution map. Then, it combines mechanical dynamic response characteristics (e.g., damping attenuation coefficient) for volume correction calculations, ultimately outputting the defect type, location, and centimeter-level accuracy (3D dimensional parameters). All detection results are recorded and visualized in real time through the production line control system.

[0129] Data comparison shows that the detection system of this invention has significant advantages in response speed, defect identification, and three-dimensional quantization. Taking polyurethane sound-absorbing panels as an example, the average detection time for a single (30cm x 30cm) sound-absorbing panel is reduced from 4.5 seconds in the traditional system to 1.8 seconds, and the throughput of the detection line is increased by approximately 150%. The overall false alarm rate is reduced from over 8% in the traditional system to below 0.5%, and the false alarm rate is reduced from over 15% to below 2%. The system's confidence in identifying complex defects is significantly improved.

[0130] Table 1 Comparison of Defect Detection Accuracy for Foam Products

[0131] Surface scratch (0.1mm wide) Traditional system 82.0 18.0 8.0 This invention 99.2 1.5 0.8 Surface pit (0.08mm deep) Traditional system 75.0 22.0 10.0 This invention 98.5 1.8 1.2 Internal air bubble (0.3mm diameter) Traditional system 55.0 25.0 30.0 This invention 97.8 2.5 2.2 Inner layer (0.2mm thick) Traditional system 30.0 40.0 60.0 This invention 96.5 3.0 3.5

[0132] Table 2 Examples of 3D Quantization Accuracy of Internal Defects

[0133]

[0134] Table 3 Comparison of System Overall Performance

[0135]

[0136] As can be seen from the data in Tables 1-3 above, the machine vision defect detection system of this invention exhibits significant advantages when applied to polymer foam product production lines. Through adaptive structured light sources and multimodal feature fusion, the system effectively overcomes the complex optical properties of materials, significantly improving the detection accuracy of surface and internal defects and substantially reducing false alarm and missed detection rates. Simultaneously, it innovatively achieves three-dimensional quantification of defects, providing unprecedented quantitative evidence for refined product quality control and production process optimization. The system also shows significant improvements in detection speed and overall efficiency, demonstrating its enormous application potential and beneficial effects in the field of industrial online inspection.

[0137] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0138] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for defect detection in polymer foam products based on machine vision, characterized in that, The method includes: Obtain the basic optical parameters of the polymer foam product and acquire the initial surface image of the target product; Based on the aforementioned basic optical parameters, optical calculations are performed on the initial surface image to generate structured light source control commands. The generation of these commands includes: acquiring the material polarization optical parameters, spectral transmittance, light scattering coefficient, and material absorption coefficient of the polymer foam product to generate basic optical parameters; acquiring an initial surface image of the target product and performing region identification to obtain highly reflective regions and internal structural interest regions; combining the basic optical parameters, performing polarization optical calculations on the highly reflective regions to obtain a first polarization angle combination; performing spectral transmittance and light scattering calculations on the internal structural interest regions to generate a second spectral band and incident angle combination; and integrating the first polarization angle combination and the second spectral band and incident angle combination to generate structured light source control commands. The structured light source control command controls the light source array to image the target product, thereby acquiring a multimodal image set; Based on the fundamental optical parameters, the multimodal image set is subjected to feature decoupling processing to generate a surface reflection feature map and an internal transmission feature map. The feature decoupling processing includes: acquiring a multimodal image set containing specific polarization and spectral information; calculating the surface reflection component of the multimodal image set based on the fundamental optical parameters and the principle of polarization reflection to generate a surface reflection contribution component; extracting the internal transmission component of the multimodal image set based on the fundamental optical parameters and the physical laws of light transmission to generate an internal transmission contribution component; performing linear normalization processing on the surface reflection contribution component to generate a surface reflection feature map; and performing edge-preserving noise suppression processing on the internal transmission contribution component to generate an internal transmission feature map. The structured light source control command, the surface reflection feature map, and the internal transmission feature map are subjected to parameter-aware fusion processing to generate a decoupled enhancement map and suspicious area location information; Based on the location information of the suspicious area, a controllable micro-vibration is applied to the target product and a high-speed image sequence is acquired simultaneously. The dynamic response feature data extracted from the high-speed image sequence is spatially aligned and feature-stitched with the decoupled enhancement map to generate a fused defect feature vector. Based on the fused defect feature vector, defect classification and quantification analysis are performed, and the defect type, location, and three-dimensional size parameters are output.

2. The method according to claim 1, wherein The generated decoupling enhancement map and suspicious region location information include: The structured light source control commands are encoded and converted to generate parameter feature vectors; Based on the parameter feature vector, a dynamic fusion weight is calculated, and a weighted fusion is performed on the surface reflection feature map and the internal transmission feature map to generate a fused feature map. The fused feature map is trained and optimized based on a physical constraint loss function to generate a decoupled enhanced map containing a surface dominant map, an internal dominant map, and location information of suspicious regions.

3. The method according to claim 2, wherein the method is characterized by, The training and optimization of the fused feature map based on the physical constraint loss function includes: Based on the light intensity gradient distribution law of surface defects, gradient distribution constraint calculation is performed on the fused feature map to generate surface defect gradient constraint loss. Based on the transmission intensity attenuation law of internal defects, the transmission intensity attenuation constraint is calculated on the fused feature map to generate internal defect attenuation constraint loss. The physical constraint loss function is generated by weighting and combining the surface defect gradient constraint loss and the internal defect attenuation constraint loss. 4.The method of claim 2, wherein the method further comprises: The extraction of dynamic response feature data from the high-speed image sequence includes: Based on the location information of the suspicious area, the surface of the target product is subjected to micro-vibration excitation and continuous imaging to generate a high-speed image sequence. Motion estimation calculations are performed on the high-speed image sequence to generate a local deformation displacement matrix; Dynamic parameters are extracted from the local deformation displacement matrix to generate the vibration frequency spectrum and damping attenuation coefficient; Dynamic response characteristics are calculated based on the vibration frequency spectrum and the damping attenuation coefficient to generate dynamic response characteristic data.

5. The method of claim 4, wherein the method further comprises: The spatial alignment and feature stitching process includes: Based on the location information of the suspected area, the dynamic response feature data is subjected to coordinate transformation processing to generate a dynamic feature map that is aligned with the decoupling enhancement graph space; The decoupled enhancement map and the dynamic feature map are cascaded through channels to generate a multi-channel feature tensor. The multi-channel feature tensor is weighted and fused based on an attention mechanism to generate a fusion defect feature vector.

6. The method according to claim 5, wherein The step of performing defect classification and quantitative analysis based on the fused defect feature vector includes: The fused defect feature vector is used to determine the defect type and generate defect type information; Based on the defect type information, a transmission intensity-depth mapping calculation is performed on the internal defects to generate a defect depth distribution map. The volume correction calculation is performed on the defect depth distribution map to generate defect volume parameters.

7. The method of claim 6, wherein the method further comprises: The transmission intensity-depth mapping calculation includes: The initial depth of the internal dominant graph is calculated to generate an initial depth distribution map; The initial depth distribution map is calibrated for effective path length to generate calibration path length parameters; Iterative optimization calculations are performed based on the calibration path length parameter to generate a converged defect depth distribution map.

8. A machine vision-based polymer foam product defect detection system for use in a machine vision-based polymer foam product defect detection method as claimed in any one of claims 1-7, characterized by, The system includes: The data acquisition module is used to obtain the basic optical parameters of polymer foam products and to acquire the initial surface image of the target product. A light source strategy generation module is used to perform optical calculations on the initial surface image based on the basic optical parameters to generate structured light source control instructions. The generation of structured light source control instructions includes: acquiring the material polarization optical parameters, spectral transmittance, light scattering coefficient, and material absorption coefficient of the polymer foam product to generate basic optical parameters; acquiring the initial surface image of the target product and performing region identification to obtain highly reflective regions and internal structural interest regions; combining the basic optical parameters to perform polarization optical calculations on the highly reflective regions to obtain a first polarization angle combination; performing spectral transmittance and light scattering calculations on the internal structural interest regions to generate a second spectral band and incident angle combination; and integrating the first polarization angle combination and the second spectral band and incident angle combination to generate structured light source control instructions. A multimodal imaging module is used to control a light source array to image the target product according to the structured light source control command, thereby acquiring a multimodal image set; A physical feature decoupling module is used to perform feature decoupling processing on the multimodal image set based on the fundamental optical parameters, generating a surface reflection feature map and an internal transmission feature map. The feature decoupling processing includes: acquiring a multimodal image set containing specific polarization and spectral information; calculating the surface reflection component of the multimodal image set based on the fundamental optical parameters and the principle of polarization reflection, generating a surface reflection contribution component; extracting the internal transmission component of the multimodal image set based on the fundamental optical parameters and the physical laws of light transmission, generating an internal transmission contribution component; performing linear normalization processing on the surface reflection contribution component, generating a surface reflection feature map; and performing edge-preserving noise suppression processing on the internal transmission contribution component, generating an internal transmission feature map. The parameter-aware fusion module is used to perform parameter-aware fusion processing on the structured light source control command, the surface reflection feature map, and the internal transmission feature map to generate a decoupled enhancement map and suspicious area location information; The vibration-assisted precision inspection module is used to apply controllable micro-vibration to the target product based on the location information of the suspicious area and simultaneously acquire high-speed image sequences. The multimodal feature fusion module is used to extract dynamic response feature data from the high-speed image sequence and spatially align and stitch the decoupled enhancement map to generate a fused defect feature vector; The defect 3D quantization module is used to perform defect classification and quantification analysis based on the fused defect feature vector, and output the defect type, location and 3D size parameters.

Images