Foaming diffusion plate bubble uniformity analysis system and method based on image analysis
By combining multi-view synchronous imaging and deep learning networks, the problem of non-destructive quantitative uniformity assessment of the three-dimensional distribution of bubbles in foamed diffuser plates was solved, achieving efficient and accurate bubble uniformity analysis and avoiding the use of expensive equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-13
Smart Images
Figure CN121660992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical material manufacturing quality inspection technology, specifically to an image analysis-based system and method for analyzing the uniformity of bubbles in foamed diffuser plates. Background Technology
[0002] As a core light-diffusing material in LCD backlight modules and high-end lighting devices, the optical performance of foamed diffuser plates directly depends on the uniformity of the internal bubble distribution. Traditional inspection methods rely on manual visual sampling or local imaging with a two-dimensional microscope, which has three major drawbacks: First, manual assessment is highly subjective and cannot quantify the bubble size distribution and spatial density fluctuations; second, two-dimensional images cannot reflect the true volume of the bubbles and their three-dimensional spatial relationship, leading to projection overlap errors; and third, destructive slicing inspection is inefficient and cannot perform full inspection.
[0003] Existing image analysis techniques attempt to address this problem, but they still encounter bottlenecks in practical applications: imaging of semi-transparent materials often results in blurred bubble boundaries due to light scattering; dense bubble adhesion renders traditional segmentation algorithms ineffective; while X-ray tomography can acquire three-dimensional structures, the equipment is expensive and the detection speed is slow. Academic research over the past five years shows that deep learning-based image segmentation is highly effective in biological cell detection, but it is not adapted to the optical properties of foamed materials; multi-view 3D reconstruction technology has applications in industrial part measurement, but it has not solved the matching problem of variable bubble morphology and lack of surface texture. The industry urgently needs an automated bubble uniformity analysis solution that balances accuracy, efficiency, and cost to meet the quality control requirements of mass production of high-end optical devices. Summary of the Invention
[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide an image analysis-based system and method for analyzing the uniformity of bubbles in foamed diffuser plates, which solves the problem of non-destructive quantitative evaluation of the three-dimensional distribution of bubbles in foamed diffuser plates. This invention overcomes the imaging blurring of semi-transparent materials through multi-view synchronous imaging, accurately segments adhered bubbles using an improved deep learning network, reconstructs the three-dimensional spatial distribution of bubbles based on multi-view geometric relationships through back projection, and generates a heat map by fusing size statistics and topological density field. Finally, it outputs a comprehensive uniformity score and spatial location of defects. The entire process does not require sample destruction and avoids reliance on expensive CT equipment.
[0005] This invention provides an image analysis-based bubble uniformity analysis system for foamed diffuser plates, comprising: The acquisition module performs simultaneous image acquisition on the foamed diffusion plate sample from at least two different perspectives to form a multi-view image signal set. The bubble segmentation module receives a multi-view image signal set, identifies bubble boundaries and adhesion regions through a pre-trained deep learning segmentation network, and generates segmentation mask data containing bubble positions and contours. The 3D feature reconstruction module receives segmentation mask data, performs 3D coordinate back projection calculation of bubbles based on the spatial geometric relationship of multi-view images, and fuses the segmentation results of each view to generate a 3D feature dataset of bubble equivalent volume, spatial coordinates and neighborhood topology. The uniformity assessment module receives a three-dimensional feature dataset, first calculates the coefficient of variation of bubble size distribution and the standard deviation of local density as basic uniformity indicators, then constructs a spatially weighted density field based on the topological relationship of bubble neighborhood, generates a global uniformity heat map of the diffuser plate, and finally integrates the basic indicators and heat map data to output a comprehensive uniformity score and defect location marking signal.
[0006] In one embodiment of the present invention, the acquisition module includes a multispectral imaging unit and a polarization filtering unit. The multispectral imaging unit captures the surface reflection image and internal transmission image of the foamed diffuser by switching light sources of different wavelengths, forming a multi-view image signal set with superimposed spectral channels. The polarization filtering unit sets a rotatable polarizer group in the imaging optical path to enhance the contrast of the bubble boundary by eliminating specular reflection interference from the material surface. The multispectral imaging unit and the polarization filtering unit work together to ensure that the image signal set received by the bubble segmentation module contains depth information and anti-interference features.
[0007] In one embodiment of the present invention, the training process of the deep learning segmentation network adopts a synthetic data augmentation strategy: first, a basic bubble model library is generated, which contains virtual bubbles with different diameters, ellipticity and surface textures; then, the virtual bubbles are embedded into a simulated foamed diffuser material substrate according to random spatial distribution, and a set of labeled training images is generated by ray tracing rendering; finally, material defect noise and optical scattering noise are introduced to simulate actual imaging conditions, so that the segmentation network has the ability to handle bubble boundary blurring and local deformation.
[0008] In one embodiment of the present invention, when the three-dimensional feature reconstruction module performs the three-dimensional coordinate back projection calculation of the bubble, it establishes an epipolar geometric constraint model between viewpoints: first, it calculates the fundamental matrix between viewpoints based on the calibration parameters of the binocular imaging device, and then generates an initial depth map by disparity matching of bubble contour points in the segmented mask data; then, it optimizes the depth value by utilizing the bubble morphology continuity constraint to eliminate coordinate ambiguity caused by projection overlap; finally, it calculates the three-dimensional spatial coordinates of the bubble centroid and the equivalent sphere radius by using the triangulation method.
[0009] In one embodiment of the present invention, the method for constructing the neighborhood topology relationship is as follows: a bubble spatial index structure is established in the three-dimensional feature dataset, and a set of neighboring bubbles within a preset radius is searched with the target bubble as the center; the shortest Euclidean distance between the target bubble and the surface of each neighboring bubble is calculated as the adjacency strength weight, and the equivalent volume ratio of the neighboring bubbles is recorded at the same time; a topological vector describing the local density is generated based on the adjacency strength weight and the volume ratio, and this vector is used for the calculation of the spatial weighted density field.
[0010] In one embodiment of the present invention, the generation process of the spatially weighted density field includes dynamic kernel function configuration: the sensing radius of the density calculation kernel function is automatically adjusted according to the bubble size distribution, so that the large-size bubble area adopts the extended kernel function to avoid the density value being artificially high, and the small-size bubble dense area adopts the compact kernel function to improve the spatial resolution; the kernel function weight is adaptively scaled with the equivalent volume of the bubble to ensure that the bubble volume ratio per unit volume is accurately mapped to the heat map grid node.
[0011] In one embodiment of the present invention, the fusion algorithm for the comprehensive uniformity score adopts a two-level weighting mechanism: the first level assigns a weight of 0.6 to the coefficient of variation of size distribution in the basic uniformity index and a weight of 0.4 to the local density standard deviation, generating a primary score; the second level weights and superimposes the primary score with the entropy value of the uniformity heatmap, wherein the entropy value of the heatmap reflects the spatial disorder of the density field, and its weight coefficient increases linearly with the increase of the thickness of the foamed diffuser plate, and finally outputs a normalized comprehensive score.
[0012] In one embodiment of the present invention, the generation logic of the defect location marking signal is as follows: an abnormal region with a density value exceeding a dynamic threshold is identified in the uniformity thermal map. The dynamic threshold is determined by combining the average bubble density and the material optical performance requirements. Morphological closing operations are performed on the abnormal region to connect adjacent discrete points, and the boundary of the connected domain is extracted as the defect marking region. At the same time, the three-dimensional feature data of the bubbles in the region are associated, and the maximum equivalent bubble diameter and the average neighborhood spacing are marked as auxiliary descriptions of the defect.
[0013] In one embodiment of the present invention, a process feedback interface module is also included. The process feedback interface module receives the comprehensive uniformity score output by the uniformity evaluation module. When the score is lower than a preset warning value, a parameter adjustment command is triggered. The command matches a pre-stored process rule library according to the direction of the deviation of the bubble size distribution variation coefficient. If the variation coefficient is too high, a control suggestion to reduce the foaming temperature is generated. If the local density standard deviation is too high, a control suggestion to adjust the pressure curve is generated. The control suggestion is transmitted to the production equipment controller through an industrial communication protocol.
[0014] The present invention also includes a method for analyzing the uniformity of bubbles in a foamed diffuser plate based on image analysis, comprising: S1: Simultaneous image acquisition of the foamed diffuser plate sample from at least two different perspectives to form a multi-view image signal set; S2: Receive multi-view image signal set, identify bubble boundaries and adhesion areas through a pre-trained deep learning segmentation network, and generate segmentation mask data containing bubble position and contour; S3: Receive segmentation mask data, perform 3D coordinate back projection calculation of bubbles based on the spatial geometric relationship of multi-view images, and fuse the segmentation results of each view to generate a 3D feature dataset of bubble equivalent volume, spatial coordinates and neighborhood topology. S4: Receives a three-dimensional feature dataset, first calculates the coefficient of variation of bubble size distribution and the standard deviation of local density as basic uniformity indicators, then constructs a spatially weighted density field based on the topological relationship of bubble neighborhood, generates a global uniformity heat map of the diffuser plate, and finally integrates the basic indicators and heat map data to output a comprehensive uniformity score and defect location marking signal.
[0015] The present invention provides an image analysis-based bubble uniformity analysis system and method for foamed diffuser plates. It overcomes the imaging blur of semi-transparent materials through multi-view synchronous imaging, accurately segments adhered bubbles using an improved deep learning network, reconstructs the three-dimensional spatial distribution of bubbles based on multi-view geometric relationships through back projection, generates a heat map by fusing size statistics and topological density field, and finally outputs a comprehensive uniformity score and defect spatial location. The entire process does not require destroying the sample and avoids reliance on expensive CT equipment. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a system architecture diagram of a foam diffusion plate bubble uniformity analysis system based on image analysis. Figure 2 A schematic diagram illustrating the workflow of an image analysis-based bubble uniformity analysis system for foamed diffuser plates; Figure 3 This is a flowchart of a method for analyzing the uniformity of bubbles in a foamed diffuser plate based on image analysis. Detailed Implementation
[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.
[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.
[0021] Please see Figure 1-3 The image shows the bubble uniformity analysis system and method for foamed diffuser plates based on image analysis according to the present invention. The bubble uniformity analysis system for foamed diffuser plates based on image analysis of the present invention includes: an acquisition module, which simultaneously acquires images of a foamed diffuser plate sample from at least two different perspectives to form a multi-view image signal set; a bubble segmentation module, which receives the multi-view image signal set and identifies bubble boundaries and adhesion regions through a pre-trained deep learning segmentation network to generate segmentation mask data containing bubble positions and contours; a 3D feature reconstruction module, which receives the segmentation mask data, performs 3D coordinate back-projection calculation of bubbles based on the spatial geometric relationships of the multi-view images, and fuses the segmentation results from each perspective to generate a 3D feature dataset containing the equivalent volume, spatial coordinates, and neighborhood topological relationships of the bubbles; and a uniformity evaluation module, which receives the 3D feature dataset, first calculates the coefficient of variation of bubble size distribution and the standard deviation of local density as basic uniformity indicators, then constructs a spatially weighted density field based on the neighborhood topological relationships of the bubbles to generate a global uniformity heatmap of the diffuser plate, and finally fuses the basic indicators and heatmap data to output a comprehensive uniformity score and defect location marker signal.
[0022] like Figure 1As shown, the system consists of four core modules forming a closed-loop data flow. The acquisition module employs a rigidly mounted binocular industrial camera group with an adjustable optical axis angle of 25 to 60 degrees, combined with a coaxial pulsed light source and a diffused backlight unit to achieve multimodal illumination. Under the control of a trigger signal, the two cameras perform millisecond-level synchronous exposure on moving or stationary foamed diffuser samples, generating a set of paired image signals with parallax information. This module ensures that the samples are within the depth of field through a mechanical positioning device and integrates a vibration isolation base to suppress external interference. After receiving the original image signal set, the bubble segmentation module performs a four-order preprocessing process: first, dark field flat field correction is performed to eliminate uneven illumination; second, adaptive histogram stretching is applied to enhance low-contrast areas; then, anisotropic diffusion filtering is used to preserve edges and suppress shot noise; finally, the dual-view images are stacked into a four-channel tensor input to a pre-trained segmentation network. The deep learning segmentation network uses a residual U-Net as its backbone. In the encoder stage, dilated convolutional layers with increasing dilation rates are embedded to expand the receptive field and capture bubble morphological features with a size span of up to 100 times. In the decoder stage, a spatial attention gate mechanism is added before each upsampling layer to enhance the boundary response of adhered bubbles through spatial weight recalibration of the feature maps. The network output is a binary mask and boundary confidence map with the same resolution as the input. After morphological post-processing, a segmentation dataset containing bubble centroid pixel coordinates, contour chain codes, and adhesion markers is generated. The 3D feature reconstruction module establishes a projection model based on the camera group intrinsic and extrinsic parameter calibration matrix and performs a three-step reconstruction for each bubble contour point set: the first step uses epipolar geometry constraints to match corresponding bubble contour points in both views; the second step calculates the 3D point cloud of the contour points through triangulation and fits a minimum bounding ellipsoid; the third step fuses the segmentation results from both views to correct projection ambiguities, ultimately generating a data structure with bubbles as units, including the equivalent sphere radius, ellipsoid eccentricity, centroid spatial coordinates, and a list of neighboring bubble indices. The uniformity assessment module first traverses the 3D bubble dataset, divides the sample space into sliding cubic grids, counts the number and volume ratio of bubbles in each grid cell, and calculates the coefficient of variation of size distribution and the standard deviation of local density as basic indicators. Second, it constructs a spatially weighted density field, using the bubble centroid as the node, equivalent volume as the weight, and neighborhood topological distance as the kernel function bandwidth, and generates a global 3D heatmap through Gaussian density estimation. Finally, it normalizes the basic indicators to values in the range of zero to one hundred, and performs weighted fusion with the entropy value and gradient features of the heatmap to output a comprehensive uniformity score and defect spatial coordinate encoding based on the extreme value region of the heatmap.
[0023] Furthermore, the multispectral imaging unit of the acquisition module includes a beam splitting design: the surface reflection imaging beam uses a combination of a ring-shaped high-brightness white LED array and an industrial polarization camera. The incident angle of the LED array is controlled within the range of 55 to 65 degrees to suppress specular reflection. A linear polarizer is placed in front of the polarization camera, and its direction is orthogonal to the polarization direction of the LED, effectively eliminating the interference of glare from the resin material surface on the identification of bubble boundaries. The internal transmission imaging beam uses a combination of a collimated infrared backlight and a monochrome camera. The infrared wavelength is selected in the 750-950 nm band to match the transmission window of the foaming material. The monochrome camera is equipped with a bandpass filter to shield ambient stray light. The dual beam paths achieve spectral coupling through dichroic mirrors to ensure that the surface reflection image and the internal transmission image are strictly aligned in time and space. The polarization filtering unit deploys a rotatable polarizer group in each imaging beam path, and its rotation angle is dynamically adjusted by a feedback control algorithm: first, in the pre-scanning stage, the surface reflection characteristic distribution map of the sample is obtained to identify the position and intensity of the highlight area; then, the optimal extinction angle combination of the polarizers is solved to reduce the pixel intensity of the highlight area by more than 70% without losing bubble feature details. The mechanical structure employs a modular design, with the surface reflection and transmission optical paths separately mounted on a shock-resistant optical platform, and independent optical path calibration achieved via a precision lead screw. The sample stage integrates a vacuum adsorption device to ensure the flatness error of the material is less than 0.1 millimeters, and works in conjunction with an XYZ three-axis fine-tuning platform to achieve millimeter-level positioning repeatability. This module ultimately outputs a six-channel image signal set: including three RGB channels for surface reflection, a single near-infrared channel for internal transmission, and a dual-polarization differential image channel, providing depth information and material penetration capability for subsequent segmentation.
[0024] Specifically, the training process of the segmentation network employs a five-stage synthetic data augmentation strategy. In the basic model library construction stage, hundreds of thousands of virtual bubble models are generated based on physical rules: the bubble morphology parameters follow a Weibull distribution, the diameter ranges from 5 micrometers to 500 micrometers, the ellipticity parameter is randomly generated between 1.0 and 2.5, and the surface texture simulates the light scattering characteristics of a real bubble film using a Berlin noise function. In the material substrate simulation stage, a Monte Carlo ray tracing engine is used to simulate the optical behavior of the foamed board: the substrate refractive index is set to be variable between 1.4 and 1.6, a titanium dioxide scattering particle concentration gradient is added, and the surface diffuse reflection characteristics of different foaming processes are simulated using a bidirectional scattering distribution function. In the synthetic rendering stage, virtual bubbles are embedded into the material substrate according to a random Poisson distribution, and a path tracing algorithm is used to render dual-view images, while simultaneously generating pixel-level accurate mask annotation maps; to address the bubble adhesion phenomenon, bubble clusters with an overlap of 30% to 70% are forcibly generated, and the topological relationships of the contact points are recorded. The noise injection stage simulates four types of real-world noise: Gaussian-Poisson mixture noise is added to simulate camera noise; point spread function blur is introduced to simulate optical defocus; Rayleigh scattering noise is superimposed to simulate material turbidity; and scratches and impurities are randomly generated as interference. The network training stage employs a transfer learning strategy: based on ImageNet pre-trained weights, a focus loss function is used to address the pixel imbalance between bubbles and the background, with its adjustment factor set to 2.0; the spatial attention gate uses a channel separation training mechanism, first freezing the encoder weights and training the attention gate separately for three months, then fine-tuning the entire network end-to-end; adaptive learning rate decay is used during training, reducing the learning rate to one-third of its original value when the Dice coefficient on the validation set does not improve for five consecutive rounds. The network inference stage deploys dual-path optimization: during forward propagation, test-time enhancement techniques are used, applying random rotation and mirroring to the input image to generate eight copies, and taking the probability mean of the output mask as the final result; in the post-processing stage, a conditional random field is applied to optimize boundary smoothness, with its binary potential parameter adaptively correlated with the gradient strength of the segmentation boundary. This scheme improves the Dice coefficient of the segmentation network to over 0.92 in scenarios with blurred bubble boundaries, partial occlusion, and heavy adhesion, while keeping the missegmentation rate below 1.5%.
[0025] In one embodiment of the present invention, when the 3D feature reconstruction module performs 3D coordinate backprojection of the bubble, a hierarchical optimization framework based on multi-view geometry is established. In the initial depth map generation stage, the intrinsic and extrinsic rotation and translation matrices calibrated by the binocular camera are used to calculate the essential matrix between the left and right view images, constructing an epipolar constraint model to match bubble contour points in the segmentation mask: first, contour curvature feature points are extracted as a candidate matching point set, and initial matching pairs are calculated using normalized cross-correlation coefficients; then, a random sampling consensus algorithm is applied to eliminate mismatched points, retaining reliable point pairs that satisfy the epipolar constraints. In the triangulation stage, based on the pixel coordinates of reliable point pairs and the camera projection matrix, the least squares method is used to solve for the 3D coordinates of spatial points, forming an initial bubble surface point cloud. To address the point cloud sparsity problem, a morphological interpolation algorithm is introduced: the point density of the missing region is predicted based on the equivalent diameter of adjacent bubbles, and interpolation points are generated along the normal direction of the bubble surface. The deep optimization stage incorporates three constraints in the energy function design: a bubble surface smoothness constraint forces continuous changes in the normal vectors of adjacent points; a projection consistency constraint requires that the overlap with the segmentation mask exceeds 95% when reprojected onto the original image; and a volume conservation constraint, based on the bubble's equivalent sphere assumption, limits the range of point cloud volume fluctuations. The energy function is minimized using the Gauss-Newton iterative method, outputting an optimized dense point cloud. Finally, principal component analysis is used to fit the point cloud to generate bubble-oriented bounding boxes, calculating the lengths and direction vectors of the major, median, and minor axes, representing non-spherical bubbles as ellipsoids. Output parameters include the ellipsoid's center coordinates, axial length ratio, Euler angles, and equivalent sphere radius conversion. The neighborhood topology index is stored using an octree spatial data structure, with each bubble recording a list of neighboring bubble identifiers and centroid spacing within a radius five times its size.
[0026] like Figure 2As shown, the construction of neighborhood topology adopts an adaptive radius search strategy. In the spatial index construction stage, the 3D bubble dataset is divided into spatial grids, with the grid size dynamically adjusted according to the average bubble diameter. The preset basic grid side length is 1.5 to 3 times the average bubble diameter. A KD tree of bubble centroid coordinates is established to accelerate the neighbor search, with the initial search radius set to three times the radius of the bubble's equivalent sphere. In the adjacency strength calculation stage, a range query is performed for each target bubble: first, all neighboring bubbles within the initial radius are retrieved, and the minimum Euclidean distance from the target bubble surface to each neighboring bubble surface is calculated as the basic adjacency strength value; when the number of neighboring bubbles is below a threshold, the search radius is gradually expanded until the upper limit radius is reached. A volume weighting factor is introduced, scaling the adjacency strength according to the ratio of the equivalent volume of neighboring bubbles to the volume of the target bubble. When the volume ratio is greater than one, the reciprocal is used to avoid large bubbles dominating. The topology vector generation stage aggregates three features: a one-dimensional distance feature, which is the arithmetic mean and standard deviation of the minimum surface distances of all neighboring bubbles; a two-dimensional distribution feature, which calculates the skewness and kurtosis of the centroid distribution of neighboring bubbles through principal component analysis; and a three-dimensional volume feature, which calculates the ratio of the sum of the volumes of the target bubble and its neighboring bubbles to the volume of the enclosing space. The resulting seven-dimensional topology vector is stored in the three-dimensional feature dataset and includes the minimum distance mean, minimum distance standard deviation, distribution skewness, distribution kurtosis, volume density, maximum volume ratio, and number of neighboring bubbles. This vector is used for configuring the kernel function parameters of the spatially weighted density field, where the minimum distance mean directly determines the sensing range for density calculation.
[0027] Furthermore, the generation of the spatially weighted density field includes a dynamic kernel function configuration system. In the kernel function selection stage, the basic kernel type is chosen based on the dispersion coefficient of the bubble size distribution: when the dispersion coefficient is less than 0.2, a Gaussian kernel is used to ensure smoothness; when the dispersion coefficient is greater than 0.5, an Epanechnikov kernel is used to suppress the weight of large bubbles; for intermediate values, a bicubic kernel is used to balance sensitivity and stability. In the sensing radius decision stage, a radius-volume correlation model is established: with the target bubble as the center, the initial value of the sensing radius is set to four times the equivalent diameter of the bubble; when the volume of the target bubble is greater than the average volume of neighboring bubbles, the sensing radius is linearly expanded proportionally to the volume excess, with an upper limit of six times the average diameter; when the volume of the target bubble is less than the average, the sensing radius is reduced to three times the diameter. In the weight allocation stage, a dual weighting factor is designed: the basic weight is proportional to the equivalent volume of the bubble; the distance weight decays with Euclidean distance according to the kernel function curve, and the decay rate is adjusted by the minimum distance mean in the topological vector. Density calculation employs a moving window method: the three-dimensional space of the foamed diffuser is discretized into a voxel grid, with the grid resolution preset according to system accuracy requirements; for each voxel center point, all bubbles within the sensing range are retrieved, and the density value at that point is calculated using a weighted superposition formula. In the density field post-processing stage, a three-dimensional bilateral filter is applied to smooth noise while preserving density transition edges, ultimately generating a global density scalar field. In the heatmap mapping stage, the density field is projected onto a two-dimensional plane along the thickness direction, using a nonlinear color mapping function: a gradient color level is used for the density mean range, and high-density areas exceeding the threshold and low-density areas are mapped to red and blue warning colors, respectively. The heatmap output is a pseudo-color image with a spatial coordinate reference system, where pixel intensity strictly corresponds to local density values.
[0028] like Figure 3 The diagram illustrates the image analysis-based bubble uniformity analysis method for foamed diffuser plates according to the present invention. S1: Simultaneous image acquisition from at least two different perspectives is performed on the foamed diffuser plate sample to form a multi-view image signal set; S2: The multi-view image signal set is received, and bubble boundaries and adhesion regions are identified through a pre-trained deep learning segmentation network to generate segmentation mask data containing bubble positions and contours; S3: The segmentation mask data is received, and the three-dimensional coordinate back-projection of bubbles is calculated based on the spatial geometric relationships of the multi-view images. The segmentation results from each perspective are fused to generate a three-dimensional feature dataset containing the equivalent volume, spatial coordinates, and neighborhood topological relationships of the bubbles; S4: The three-dimensional feature dataset is received, and the coefficient of variation of bubble size distribution and the standard deviation of local density are first calculated as basic uniformity indicators. Then, a spatially weighted density field is constructed based on the neighborhood topological relationships of the bubbles to generate a global uniformity heatmap of the diffuser plate. Finally, the basic indicators and heatmap data are fused to output a comprehensive uniformity score and defect location marker signal.
[0029] Furthermore, the fusion algorithm for the comprehensive uniformity score adopts a hierarchical feature aggregation architecture. In the initial score calculation stage, the basic uniformity indicators are normalized and weighted: the coefficient of variation of the size distribution is logarithmically transformed to compress the range to the 0-100 interval and assigned a weight of 0.6; the local density standard deviation is smoothed by cubic spline interpolation and mapped to the same interval, assigned a weight of 0.4. The weighted sum of these two constitutes the initial score. In the heatmap feature extraction stage, three sets of derived parameters are calculated: the spatial entropy value is used to calculate the texture disorder of the density field through the gray-level co-occurrence matrix, reflecting the randomness of the density distribution; the average gradient magnitude characterizes the severity of density changes; and the number of high-density connected regions is used to count the number of clusters whose density exceeds twice the overall mean. In the second-level fusion stage, a dynamic weight allocator is designed: the spatial entropy value of the heatmap accounts for 50% of the second-level weight base, the average gradient magnitude accounts for 30%, and the number of high-density connected regions accounts for 20%. The second-level weight base itself is adjusted by the thickness of the foamed diffuser plate; when the thickness is less than one millimeter, the base is set to 0.3; when the thickness exceeds three millimeters, the base is increased to 0.7, with intermediate values linearly interpolated. The final comprehensive score calculation formula is: Primary score × (1 - Secondary weight base) + (Heatmap entropy value × 0.5 + Gradient mean × 0.3 + Number of connected components × 0.2) × Secondary weight base × 100. During the defect location marking stage, adaptive threshold segmentation is performed on the heatmap: Threshold T = Overall density mean + K × Standard deviation, where the coefficient K is preset according to the material's optical grade: K = 1.5 for ordinary grade and K = 2.0 for optical grade. Three-dimensional connected component analysis is performed on regions exceeding the threshold to extract defect clusters containing volumes greater than 100,000 cubic micrometers. The coordinates of the center of the smallest bounding cube and its axial range are encoded as defect location signals. Each defect cluster is associated with a quality analysis report, including the maximum equivalent bubble diameter, average adjacency distance, volume percentage, and suggested process adjustment direction codes.
[0030] The present invention provides an image analysis-based bubble uniformity analysis system and method for foamed diffuser plates. This system overcomes the imaging blur of semi-transparent materials through multi-view synchronous imaging, accurately segments adhered bubbles using an improved deep learning network, reconstructs the three-dimensional spatial distribution of bubbles based on multi-view geometric relationships through back projection, generates a heat map by fusing size statistics and topological density field, and finally outputs a comprehensive uniformity score and defect spatial location. The entire process does not require destroying the sample and avoids reliance on expensive CT equipment.
[0031] Therefore, the image analysis-based bubble uniformity analysis system and method for foamed diffuser plates of the present invention solves the problem of non-destructive quantitative uniformity assessment of the three-dimensional distribution of bubbles in foamed diffuser plates.
[0032] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.
Claims
1. A bubble uniformity analysis system for foamed diffuser plates based on image analysis, characterized in that, include: The acquisition module performs simultaneous image acquisition on the foamed diffusion plate sample from at least two different perspectives to form a multi-view image signal set. A bubble segmentation module receives the multi-view image signal set, identifies bubble boundaries and adhesion regions through a pre-trained deep learning segmentation network, and generates segmentation mask data containing bubble positions and contours. The three-dimensional feature reconstruction module receives the segmentation mask data, performs three-dimensional coordinate back projection calculation of the bubble based on the spatial geometric relationship of the multi-view images, and fuses the segmentation results of each view to generate a three-dimensional feature dataset of the bubble's equivalent volume, spatial coordinates, and neighborhood topological relationship. The uniformity assessment module receives the three-dimensional feature dataset, first calculates the coefficient of variation of bubble size distribution and the standard deviation of local density as basic uniformity indicators, then constructs a spatially weighted density field based on the topological relationship of bubble neighborhood, generates a global uniformity heat map of the diffuser plate, and finally integrates the basic indicators and heat map data to output a comprehensive uniformity score and defect location marking signal.
2. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The acquisition module includes a multispectral imaging unit and a polarization filtering unit. The multispectral imaging unit captures the surface reflection image and internal transmission image of the foamed diffuser by switching light sources of different wavelengths, forming a multi-view image signal set with superimposed spectral channels. The polarization filtering unit sets a rotatable polarizer group in the imaging optical path to enhance the contrast of the bubble boundary by eliminating specular reflection interference from the material surface. The multispectral imaging unit and the polarization filtering unit work together to ensure that the image signal set received by the bubble segmentation module contains depth information and anti-interference features.
3. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The training process of the deep learning segmentation network adopts a synthetic data augmentation strategy: first, a basic bubble model library is generated, which contains virtual bubbles with different diameters, ellipticity and surface textures; then, the virtual bubbles are embedded into a simulated foamed diffuser material substrate according to random spatial distribution, and a set of labeled training images is generated by ray tracing rendering. Finally, material defect noise and optical scattering noise are introduced to simulate actual imaging conditions, enabling the segmentation network to handle bubble boundary blurring and local deformation.
4. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, When the three-dimensional feature reconstruction module performs the three-dimensional coordinate back projection calculation of the bubble, it establishes an epipolar geometric constraint model between viewpoints: first, it calculates the fundamental matrix between viewpoints based on the calibration parameters of the binocular imaging device, and then generates an initial depth map by disparity matching of bubble contour points in the segmented mask data; then, it optimizes the depth value by utilizing the bubble morphology continuity constraint to eliminate coordinate ambiguity caused by projection overlap; finally, it calculates the three-dimensional spatial coordinates of the bubble centroid and the equivalent sphere radius by using the triangulation method.
5. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The method for constructing the neighborhood topology is as follows: a bubble spatial index structure is established in the three-dimensional feature dataset, and a set of neighboring bubbles within a preset radius is searched with the target bubble as the center; the shortest Euclidean distance between the target bubble and the surface of each neighboring bubble is calculated as the adjacency strength weight, and the equivalent volume ratio of the neighboring bubbles is recorded at the same time; a topological vector describing the local density is generated based on the adjacency strength weight and the volume ratio, and this vector is used to calculate the spatial weighted density field.
6. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The generation process of the spatially weighted density field includes dynamic kernel function configuration: the sensing radius of the density calculation kernel function is automatically adjusted according to the bubble size distribution, so that the large-sized bubble area adopts the extended kernel function to avoid the density value being artificially high, and the small-sized bubble dense area adopts the compact kernel function to improve the spatial resolution; the kernel function weight is adaptively scaled with the equivalent volume of the bubble to ensure that the bubble volume ratio per unit volume is accurately mapped to the heat map grid node.
7. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The fusion algorithm for the comprehensive uniformity score adopts a two-level weighting mechanism: the first level assigns a weight of 0.6 to the coefficient of variation of size distribution in the basic uniformity index and a weight of 0.4 to the local density standard deviation, generating a primary score; the second level weights and superimposes the primary score with the entropy value of the uniformity heatmap, where the heatmap entropy value reflects the spatial disorder of the density field, and its weight coefficient increases linearly with the increase of the thickness of the foamed diffuser plate, finally outputting a normalized comprehensive score.
8. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, The generation logic of the defect location marking signal is as follows: in the uniformity thermal map, abnormal regions with density values exceeding the dynamic threshold are identified. The dynamic threshold is determined by combining the average bubble density and the material optical performance requirements. Morphological closing operations are performed on the abnormal regions to connect adjacent discrete points, and the boundaries of the connected domains are extracted as defect marking regions. At the same time, the three-dimensional feature data of the bubbles in the region are associated, and the maximum equivalent bubble diameter and the average neighborhood spacing are marked as auxiliary descriptions of the defects.
9. The image analysis-based bubble uniformity analysis system for foamed diffuser plates according to claim 1, characterized in that, It also includes a process feedback interface module, which receives the comprehensive uniformity score output by the uniformity evaluation module. When the score is lower than a preset warning value, a parameter adjustment command is triggered. The command matches a pre-stored process rule library according to the direction of the deviation of the bubble size distribution variation coefficient. If the variation coefficient is too high, a control suggestion to reduce the foaming temperature is generated. If the local density standard deviation is too high, a control suggestion to adjust the pressure curve is generated. The control suggestion is transmitted to the production equipment controller through an industrial communication protocol.
10. The intelligent suppression management method for the bubble uniformity analysis system of the foamed diffuser plate based on image analysis according to claims 1-9, comprising: S1: Simultaneous image acquisition of the foamed diffuser plate sample from at least two different perspectives to form a multi-view image signal set; S2: Receive the multi-view image signal set, identify the bubble boundary and adhesion region through a pre-trained deep learning segmentation network, and generate segmentation mask data containing the bubble position and contour. S3: Receive the segmentation mask data, perform back projection calculation of the three-dimensional coordinates of the bubble based on the spatial geometric relationship of the multi-view images, and fuse the segmentation results of each view to generate a three-dimensional feature dataset of the bubble's equivalent volume, spatial coordinates, and neighborhood topological relationship. S4: Receive the three-dimensional feature dataset, first calculate the bubble size distribution variation coefficient and local density standard deviation as basic uniformity indicators, then construct a spatial weighted density field based on the bubble neighborhood topology, generate a global uniformity heat map of the diffuser plate, and finally fuse the basic indicators and heat map data to output a comprehensive uniformity score and defect location marking signal.