System and method for identifying electrostatic dust on surface of PE (Poly Ethylene) film

Abstract

The invention relates to the technical field of surface defect detection, in particular to a PE film surface electrostatic dust recognition system and method.The method comprises the steps that an image collection module collects multi-angle and multi-wavelength images of the surface of a PE film through an annular LED array and a multi-wavelength illumination system; the multi-manifold mapping background suppression module regards the surface of the PE film as a composite manifold and separates background and dust characteristics; the differential geometric feature extraction module extracts geometric features such as curvature and shape indexes, calculates a geodesic distance matrix and generates a noisy point-tiny dust probability graph; the topological feature segmentation module executes edge detection, morphological processing and topology preserving clustering, classifies image areas and determines the position of the tiny dust, and the method combines multi-manifold mapping and differential geometric features, so that the detection difficulty caused by the high-reflective surface of the PE film is effectively overcome, and the tiny dust of the level of 0.5-10 microns can be accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of surface defect detection, and in particular to a system and method for identifying electrostatic dust on the surface of a PE film, which is used to accurately detect tiny electrostatically adsorbed dust particles on the surface of a PE film. Background Art

[0002] As an important packaging material and protective layer for optoelectronic components, the surface quality of PE film directly impacts product performance and appearance. During PE film production and application, tiny dust particles are easily adsorbed on the film surface due to electrostatic effects, forming difficult-to-remove contamination.

[0003] Existing PE film surface inspection technologies primarily include manual visual inspection, general image processing, and traditional machine vision methods. Manual visual inspection is inefficient and prone to fatigue. General image processing methods have limited detection capabilities for tiny dust particles (especially those in the 0.5-10μm range). Traditional machine vision methods often struggle to effectively distinguish between fine dust and background noise when working with highly reflective surfaces like PE film.

[0004] Furthermore, existing technologies typically employ simple threshold segmentation or basic morphological processing methods, which struggle to account for the complex light reflection characteristics of PE film surfaces. Detection accuracy decreases significantly, especially for PE films with varying illumination angles and surface materials. Therefore, developing a system and method that can accurately identify tiny electrostatic dust particles on PE film surfaces is of great practical value. Summary of the Invention

[0005] The purpose of the present invention is to provide a system and method for identifying electrostatic dust on the surface of PE film, which can overcome the problems existing in the prior art and achieve accurate identification of 0.5-10μm level dust on the surface of PE film.

[0006] The present invention proposes a PE film surface electrostatic dust identification system, comprising:

[0007] Image acquisition module, used to collect multi-angle and multi-wavelength images of the PE film surface;

[0008] a multi-manifold mapping background suppression module, connected to the image acquisition module, for receiving the multi-angle and multi-wavelength images, constructing a manifold representation of the PE film surface, separating the PE film background and potential dust features, and generating a background suppressed image;

[0009] a differential geometry feature extraction module connected to the multi-manifold mapping background suppression module, configured to receive the background suppression image, extract differential geometry features, calculate a geodesic distance matrix, and generate a noise-dust probability map;

[0010] The topological feature segmentation module is connected to the differential geometry feature extraction module, and is used to receive the noise-dust probability map, perform edge detection and morphological processing, and perform topology-preserving clustering to divide the image area into speckled noise, blurred noise, dust and background, and determine the location and feature information of the dust on the PE film surface.

[0011] Preferably, the image acquisition module includes:

[0012] Ring-shaped LED array for illuminating the PE film surface from multiple angles;

[0013] A multi-wavelength illumination system includes three sets of dark field illumination modules with different wavelengths, which are used to activate and generate illumination light of different wavelengths in sequence;

[0014] A plurality of photoelectric imaging detectors are respectively connected to the multi-wavelength illumination system and are used to collect images of the PE film surface from different angles.

[0015] Preferably, the multi-manifold mapping background suppression module includes:

[0016] The local structure preserving unit is used to construct the local geometric structure of the image and generate a local geometric feature map;

[0017] A global structure mapping unit, connected to the local structure preserving unit, is used to receive the local geometric feature map, construct a global similarity map, optimize the manifold embedding objective function, and generate a low-dimensional representation of the PE film and the dust particles;

[0018] The adaptive threshold adjustment unit is connected to the global structure mapping unit and is used to receive the low-dimensional representation, calculate the manifold entropy, determine the optimal segmentation threshold, and generate a background suppression image.

[0019] Preferably, the differential geometry feature extraction module includes:

[0020] Differential geometry feature extractor, used to convert the image into a height function, calculate Gaussian curvature, mean curvature and shape index, and generate differential geometry feature map;

[0021] a geodesic distance calculation unit, connected to the differential geometry feature extractor, configured to receive the differential geometry feature map, define a feature-based Riemannian metric tensor, calculate the geodesic distance using a fast marching method, and construct a distance matrix;

[0022] A manifold probability distribution estimator is connected to the geodesic distance calculation unit, and is used to receive the distance matrix and the differential geometry feature map, estimate the probability density of the feature space, calculate the probability of each point belonging to a noise point or a dust particle, and generate a noise point-dust particle probability map.

[0023] Preferably, the topological feature segmentation module includes:

[0024] Edge detector, used to calculate the Gaussian Laplacian operator, detect zero crossing points, and extract edge feature maps;

[0025] a morphological processing unit connected to the edge detector, configured to receive the edge feature map, perform multi-scale morphological dilation and erosion operations, and generate a morphological enhancement feature map;

[0026] a topology preserving clusterer connected to the morphological processing unit, configured to receive the morphologically enhanced feature map, construct a high-dimensional feature vector, and perform topology preserving clustering to cluster pixels into speckled noise, blurred noise, dust particles, and background;

[0027] The feature space fuser is connected to the topology preserving clusterer and is used to receive multi-level features and clustering results, calculate feature weights, perform nonlinear feature fusion, and generate final dust particle positions and feature information.

[0028] Preferably, the local structure preserving unit constructs the local geometric structure by determining the k-nearest neighbor point set of each pixel point, calculating the similarity between the pixel point and the neighbor points, constructing a local similarity matrix, and extracting a feature vector to represent the local geometric structure.

[0029] Preferably, the differential geometric features extracted by the differential geometric feature extractor include: principal curvature, Gaussian curvature, mean curvature, shape index and curvature change rate, and generate multi-scale curvature representation in multiple scale spaces.

[0030] Preferably, the topology preserving clusterer performs topology preserving clustering in the following manner: constructing a high-dimensional feature vector containing spatial, morphological and statistical features, defining a topology preserving objective function, maintaining the local neighborhood structure through gradient descent iterative optimization, automatically determining the optimal number of categories, and dividing the pixels into four categories.

[0031] Preferably, the feature space fuser performs nonlinear feature fusion in the following manner: calculating the reliability of each feature layer, assigning adaptive weights, designing a nonlinear combination function, integrating background suppression layer features, noise detection layer features and morphological processing features, and generating a final dust representation.

[0032] The method for identifying electrostatic dust on the surface of a PE film comprises the following steps:

[0033] Collect multi-angle and multi-wavelength images of the PE film surface;

[0034] Constructing a manifold representation of the PE film surface, separating the PE film background and potential dust features, and generating a background suppressed image;

[0035] Extract the differential geometric features of the background suppressed image, calculate the geodesic distance matrix, and generate the noise-dust probability map;

[0036] Edge detection and morphological processing are performed on the noise-dust probability map, and topology-preserving clustering is performed to divide the image area into speckled noise, blurred noise, dust and background;

[0037] Multi-level features are integrated to determine the location and characteristic information of dust particles on the surface of PE film.

[0038] The beneficial effects of the present invention include:

[0039] 1. Improved dust detection accuracy: Through multi-manifold mapping and differential geometry feature extraction technology, it can detect 0.5μm-level dust particles that are difficult to detect with traditional methods. The detection rate in the 0.5-2μm range is 85%, and the detection rate in the 2-10μm range is 98%.

[0040] 2. Enhanced anti-interference capability: The topological feature-based differentiation mechanism essentially distinguishes between dust and other interference factors on the PE film surface (such as reflections, scratches, etc.), reducing the false alarm rate to below 3%.

[0041] 3. Improved system adaptability: The adaptive parameter adjustment mechanism enables the system to automatically adapt to PE films of different materials and surface characteristics, significantly improving versatility.

[0042] 4. Improved production efficiency: The system can process 1080p resolution images at a speed of 15 frames per second, greatly improving detection efficiency and reducing manual detection steps.

[0043] 5. Reduced production costs: High-precision dust detection reduces the defective rate, reduces losses in subsequent processing links, and improves overall economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the overall structure of the electrostatic dust identification system on the PE film surface of the present invention;

[0045] Figure 2 Schematic diagram of the structure of the multi-manifold mapping background suppression module of the present invention;

[0046] Figure 3 Schematic diagram of the structure of the differential geometry feature extraction module of the present invention;

[0047] Figure 4 Schematic diagram of the structure of the topological feature segmentation module of the present invention;

[0048] Figure 5 The figure is a flow chart of the method for identifying electrostatic dust on the surface of PE film of the present invention. DETAILED DESCRIPTION

[0049] Please refer to Figure 1-Figure 5 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0050] like Figure 1 As shown, the PE film surface electrostatic dust recognition system provided by the present invention includes an image acquisition module 1, a multi-manifold mapping background suppression module 2, a differential geometry feature extraction module 3 and a topological feature segmentation module 4.

[0051] Image acquisition module 1 is used to capture multi-angle, multi-wavelength images of the PE film surface. Multi-manifold mapping background suppression module 2 is connected to image acquisition module 1 and is used to receive multi-angle, multi-wavelength images, construct a manifold representation of the PE film surface, separate the PE film background and potential dust features, and generate a background-suppressed image. Differential geometry feature extraction module 3 is connected to multi-manifold mapping background suppression module 2 and is used to receive the background-suppressed image, extract differential geometry features, calculate the geodesic distance matrix, and generate a noise-dust probability map. Topological feature segmentation module 4 is connected to differential geometry feature extraction module 3 and is used to receive the noise-dust probability map, perform edge detection and morphological processing, and perform topology-preserving clustering to divide the image area into speckled noise, blurred noise, dust, and background, and determine the location and feature information of dust on the PE film surface.

[0052] like Figure 1 As shown, in a preferred embodiment of the present invention, the image acquisition module 1 includes a ring-shaped LED array 11 , a multi-wavelength illumination system 12 and a plurality of photoelectric imaging detectors 13 .

[0053] The annular LED array 11 is used to illuminate the PE film surface from multiple angles. Preferably, the annular LED array 11 includes 16 LED light sources evenly arranged in a circular pattern, capable of illuminating the PE film surface from different angles, producing a multi-angle lighting effect. The wavelength range of the LED light sources is preferably 400nm-700nm, covering the visible spectrum to obtain rich spectral information.

[0054] The multi-wavelength illumination system 12 includes three sets of dark-field illumination modules with different wavelengths, which are activated sequentially to generate illumination light of different wavelengths. Preferably, the center wavelengths of the three dark-field illumination modules are 450nm (blue light), 550nm (green light), and 650nm (red light), respectively, with wavelength intervals of 100nm. This allows the reflection characteristics of the PE film surface to be captured at different wavelengths. Each dark-field illumination module works in conjunction with the annular LED array 11 to ensure multi-angle, multi-wavelength illumination.

[0055] Multiple photoelectric imaging detectors 13 are connected to the multi-wavelength illumination system 12 to capture images of the PE film surface from various angles. Preferably, the system is equipped with three high-resolution cameras (at least 1080p resolution) evenly spaced at a 120° angle to capture a comprehensive image of the PE film surface. The camera exposure time is adjustable from 1ms to 50ms to accommodate varying lighting conditions.

[0056] In actual operation, the ring-shaped LED array 11 first illuminates the PE film surface from multiple angles. Then, the multi-wavelength illumination system 12 sequentially activates three sets of dark-field illumination modules with different wavelengths. Finally, multiple photoelectric imaging detectors 13 synchronously capture images of the PE film surface. This multi-angle, multi-wavelength acquisition method can obtain richer image information from the PE film surface, providing sufficient data support for subsequent processing.

[0057] like Figure 2 As shown, in a preferred embodiment of the present invention, the multi-manifold mapping background suppression module 2 includes a local structure preserving unit 21 , a global structure mapping unit 22 and an adaptive threshold adjustment unit 23 .

[0058] The local structure preserving unit 21 is used to construct the local geometric structure of the image and generate a local geometric feature map. Specifically, the local structure preserving unit 21 constructs the local geometric structure by determining the set of k nearest neighbors for each pixel, calculating the similarity between the pixel and its neighbors, constructing a local similarity matrix, and extracting a feature vector representing the local geometric structure.

[0059] In actual implementation, for each pixel point p in the image, first determine its k nearest neighbor point set , the k value is preferably set to 8-12 to balance the integrity of local information and computational complexity. Then the similarity between point p and its neighboring points is calculated using the Gaussian kernel function:

[0060] ,

[0061] in: Represents the similarity between pixels i and j; and Represent the feature vectors of pixel points i and j respectively (including position coordinates and grayscale values); is the bandwidth parameter of the Gaussian kernel, in pixel value, and the preferred value is 0.1-0.5 times the standard deviation of the image grayscale; Represents the feature vector and The Euclidean distance between .

[0062] In the practical application of electrostatic dust identification on the surface of PE film, when detecting smaller dust particles (0.5-2μm), The value can be set to a smaller value (such as 0.2 times the standard deviation) to enhance local details; when detecting larger dust particles (5-10μm), The value can be set to a larger value (such as 0.4 times the standard deviation) to obtain a more stable feature representation.

[0063] Based on the similarity matrix W, the feature representation of the local geometric structure is constructed:

[0064] ,

[0065] Where: L is the Laplace matrix with dimension n×n, n is the total number of image pixels; D is the diagonal matrix with dimension n×n; the diagonal elements Represents the sum of similarities between the i-th pixel and all its neighboring points.

[0066] By solving the generalized eigenvalue problem:

[0067] ,

[0068] Where: f is the feature vector with dimension n×1; The eigenvalue is taken as the eigenvector corresponding to the smallest non-zero eigenvalue as the representation of the local geometric structure.

[0069] The global structure mapping unit 22 is connected to the local structure preserving unit 21 and is used to receive the local geometric feature map, construct a global similarity map, optimize the manifold embedding objective function, and generate a low-dimensional representation of the PE film and the dust.

[0070] In actual implementation, the global structure mapping unit 22 first constructs a global adjacency graph G based on local features, and then defines the optimization objective function:

[0071] ,

[0072] in: and Respectively represent the representation of pixel i and j in low-dimensional space, with a dimension of d×1 (d is usually 2 or 3); is the similarity matrix element calculated previously; represents the Euclidean distance between points i and j in low-dimensional space.

[0073] This optimization problem can be transformed into solving:

[0074] ,

[0075] Where: Y is the low-dimensional representation matrix of all points, with dimensions nxd; tr represents the trace operation of the matrix, that is, the sum of the main diagonal elements; Represents the transpose of matrix Y.

[0076] In the application of identifying dust particles on PE film surfaces, the optimization goal is to keep similar points (such as those belonging to the PE film background or to dust particles) close together in low-dimensional space, while separating points of different categories (such as PE film background points and dust particles). By solving this optimization problem, a low-dimensional representation that preserves topological relationships can be obtained, thereby separating the PE film background and dust particles in low-dimensional space.

[0077] The adaptive threshold adjustment unit 23 is connected to the global structure mapping unit 22 and is used to receive the low-dimensional representation, calculate the manifold entropy, determine the optimal segmentation threshold, and generate a background suppression image.

[0078] In actual implementation, the adaptive threshold adjustment unit 23 calculates the manifold entropy under different thresholds t:

[0079] ,

[0080] in: represents the manifold entropy under the threshold t; represents the probability distribution of the i-th region under the threshold t, which is calculated as the number of pixels in the region divided by the total number of pixels; Represents the natural logarithm.

[0081] Also consider the contrast measure:

[0082] ,

[0083] in: represents the contrast measure under threshold t; and Represent the average grayscale values ​​of the foreground (dust) and background (PE film); and Represents the grayscale standard deviation of foreground and background respectively; It represents the absolute value of the difference between the average grayscale values ​​of the foreground and background.

[0084] Construct a comprehensive evaluation function:

[0085] ,

[0086] in: is the comprehensive evaluation function; and is a dimensionless weight parameter with preferred values ​​of 0.6 and 0.4 to balance the effects of entropy and contrast.

[0087] By searching Maximum threshold , applied to low-dimensional representation to generate background suppressed images. In the application of dust detection on the surface of PE film, Usually corresponds to the best separation point between the PE film background and the dust, and its value is related to the PE film material, surface smoothness and dust characteristics. For example, for a PE film with high smoothness, Usually between the two peaks of the low-dimensional grayscale histogram; for PE films with rough surfaces, It may be necessary to bias towards background peaks to reduce false detections.

[0088] like Figure 3 As shown, in a preferred embodiment of the present invention, the differential geometry feature extraction module 3 includes a differential geometry feature extractor 31 , a geodesic distance calculation unit 32 and a manifold probability distribution estimator 33 .

[0089] The differential geometry feature extractor 31 is used to convert the image into a height function, calculate Gaussian curvature, mean curvature, and shape index, and generate a differential geometry feature map. Specifically, the differential geometry feature extractor 31 extracts differential geometry features including principal curvature, Gaussian curvature, mean curvature, shape index, and curvature change rate, and generates a multi-scale curvature representation in multiple scale spaces.

[0090] In actual implementation, the background suppressed image I(x,y) is considered as a height function h(x,y)=I(x,y) on a two-dimensional manifold. The first-order derivative (gradient) is calculated as:

[0091] ,

[0092] in: Represents the gradient vector of the height function h, with a dimension of 2×1; and denote the partial derivatives of h with respect to x and y, respectively, which can be calculated by the finite difference method.

[0093] Calculate the second-order derivative (Hessian matrix):

[0094] ,

[0095] Where: H is the Hessian matrix with a dimension of 2×2; and denote the second-order partial derivatives of h with respect to x and y respectively; and represents the cross partial derivatives of h, which are equal according to the properties of continuous functions.

[0096] Calculate the principal curvature based on the Hessian matrix and (That is, the eigenvalues ​​of the Hessian matrix). Then calculate the Gaussian curvature K and the mean curvature H:

[0097] ,

[0098] ,

[0099] Where: K is the Gaussian curvature, which represents the intrinsic curvature of the surface at that point; H is the mean curvature, which represents the extrinsic curvature of the surface at that point; and is the principal curvature, that is, the eigenvalue of the Hessian matrix.

[0100] Calculate the shape index S:

[0101] ,

[0102] Where: S is the shape index, ranging from [-1,1]; arctan is the inverse tangent function; It represents the ratio of the sum of the principal curvatures to the difference of the principal curvatures. When , take the limit value.

[0103] Calculate the curvature change rate C:

[0104] ,

[0105] Where: C is the curvature change rate; represents the gradient vector of Gaussian curvature K; Represents the dot product of the gradient vector, that is, the square of the magnitude of the gradient vector.

[0106] In the application of dust detection on PE film surface, these differential geometric features can effectively distinguish different types of surface structures. For example, dust usually appears as local convex or concave, with a large average curvature H and a small shape index S variation; while the texture of PE film surface usually appears as undulating, with a small average curvature H and a large shape index S variation. Repeat the above calculations to generate a multi-scale curvature representation. The preferred scale is , i=0,1,2, covering dust features of different sizes.

[0107] The geodesic distance calculation unit 32 is connected to the differential geometry feature extractor 31, and is used to receive the differential geometry feature map, define a feature-based Riemann metric tensor, calculate the geodesic distance using the fast marching method, and construct a distance matrix.

[0108] In actual implementation, we first define the feature-based Riemann metric tensor M:

[0109] ,

[0110] Where: M(p) is the Riemann metric tensor at point p, with dimension 2×2; I is the 2×2 identity matrix; is the feature gradient vector at point p, with a dimension of 2×1; is the transpose of the feature gradient vector, with a dimension of 1×2; α and β are weight parameters, dimensionless, with preferred values ​​of 0.2 and 0.8, respectively, used to balance the effects of isotropy and anisotropy; Represents the outer product of vectors, resulting in a 2×2 matrix.

[0111] In the application of dust detection on the surface of PE film, the Riemann metric tensor M defines the distance metric in the feature space and can adaptively adjust the distance calculation according to the feature gradient. For example, in the edge area of ​​dust, the feature gradient is large, and M is mainly composed of the anisotropic part. Contribution, so that the geodesic distance increases along the gradient direction; in the flat area, the characteristic gradient is small, and M is mainly composed of the isotropic part Contribution, close to the Euclidean distance.

[0112] Based on the Riemann metric tensor, the geodesic distance is calculated using the fast marching method. The basic steps of the fast marching method are:

[0113] 1. Initialization: Set the starting point distance to 0, the distances of other points to infinity, and build a priority queue of candidate points

[0114] 2. Iterative calculation: Each time the point with the smallest distance in the queue is processed, the distance of its neighboring points is updated and the priority queue is readjusted

[0115] 3. Termination condition: All points are processed or the preset distance threshold is reached

[0116] Finally, the complete geodesic distance matrix D is constructed, where Represents the geodesic distance from point i to point j.

[0117] The manifold probability distribution estimator 33 is connected to the geodesic distance calculation unit 32, and is used to receive the distance matrix and the differential geometry feature map, estimate the probability density of the feature space, calculate the probability of each point belonging to a noise point or a dust particle, and generate a noise point-dust particle probability map.

[0118] In the actual implementation, we first select representative noise and dust samples and estimate the probability density function based on these samples. We use the kernel density estimation method, but use geodesic distance instead of Euclidean distance:

[0119] ,

[0120] in: Indicates that feature F belongs to the category The probability density of For category The number of samples; is the kernel function (preferably using a Gaussian kernel ), h is the bandwidth parameter; is the geodesic distance; For category Sample characteristics; Indicates the category Sum all samples of .

[0121] For each pixel p, calculate the probability that it belongs to noise and the probability of belonging to dust particles :

[0122] ,

[0123] ,

[0124] in: Indicates the posterior probability that point p belongs to a noise point; represents the posterior probability that point p belongs to a mote; and Represent the conditional probability density of the feature of point p under the noise and dust categories respectively; and are the prior probabilities of noise and dust, respectively. They are dimensionless and can be set according to experience. The preferred values ​​are 0.7 and 0.3, respectively, reflecting the fact that noise is usually more common than dust on the surface of PE film; is the normalization factor, calculated as follows:

[0125] ,

[0126] In the application of dust detection on the surface of PE film, the kernel density estimation method can adaptively estimate the probability of each point belonging to noise or dust based on the sample distribution in the feature space, without assuming a specific parameter distribution form. At the same time, using geodesic distance instead of Euclidean distance can better capture the nonlinear structure in the feature space and improve classification accuracy. Finally, a noise-dust probability map is generated, in which the value of each pixel represents the probability of its belonging to dust. , the value range is [0,1].

[0127] like Figure 4 As shown, in a preferred embodiment of the present invention, the topological feature segmentation module 4 includes an edge detector 41, a morphological processing unit 42, a topology preserving clusterer 43 and a feature space fuser 44.

[0128] The edge detector 41 is used to calculate the Laplacian of Gaussian operator, detect zero crossing points, and extract edge feature maps.

[0129] In actual implementation, Gaussian smoothing is first applied to the noise-dust probability map P:

[0130] $ ,

[0131] in: Represents the image after Gaussian smoothing; * represents the convolution operation; The standard deviation of the Gaussian kernel is expressed in pixels, with a preferred value of 1.5-2.5 pixels to balance the accuracy of edge detection and noise immunity. is a two-dimensional Gaussian function; is the normalization coefficient, ensuring that the sum of the Gaussian kernel is 1.

[0132] Then calculate the Laplacian operator:

[0133] ,

[0134] in: express The Laplace operator of and Respectively Second-order partial derivatives with respect to x and y.

[0135] Detect zero crossing points, that is, the locations where the Laplace operator value changes from positive to negative or from negative to positive. These locations correspond to the edges of the image. At the same time, use the gradient magnitude as the threshold condition to screen significant edges:

[0136] ,

[0137] in: express The gradient magnitude is calculated as ; is the gradient threshold, and the preferred value is 1.5 times the average value of the image gradient.

[0138] In the application of dust detection on the surface of PE film, the Gaussian Laplacian operator can effectively detect the edge of dust, especially for circular or elliptical dust, whose edge appears as a closed zero-crossing ring in the Laplacian operator image. value, can adapt to different sizes of dust: smaller A value (such as 1.5 pixels) is suitable for detecting the edges of small dust particles (0.5-2μm); larger A value (e.g., 2.5 pixels) is suitable for detecting the edges of large dust particles (5-10 μm). Finally, an edge feature map E is generated.

[0139] The morphological processing unit 42 is connected to the edge detector 41 and is used to receive the edge feature map, perform multi-scale morphological dilation and erosion operations, and generate a morphological enhancement feature map.

[0140] In actual implementation, a set of multi-scale structural elements { }, including structural elements of different sizes and shapes to accommodate dust particles of different sizes. A sequence of morphological operations is performed on the edge feature map E:

[0141] Expansion operation:

[0142] The expansion operation is defined as:

[0143] ,

[0144] in: represents the image after expansion; Represents edge feature map; Represents a structural element; Indicates the maximum value operation; Represents the coordinate offset in the structuring element. The dilation operation expands the edges and fills small gaps.

[0145] Corrosion operation:

[0146] The erosion operation is defined as:

[0147] ,

[0148] in: Represents the image after corrosion; represents the image after expansion; Represents a structural element; Indicates the minimum operation. The erosion operation removes small noise points and retains the main structure.

[0149] Opening and closing operation: , the closing operation is defined as dilation followed by corrosion: , the opening operation is defined as corrosion followed by expansion: , where: Close and Open represent closing and opening operations respectively; and Represents structural elements. Opening and closing operations further smooth edges and enhance region integrity.

[0150] In the application of particle detection on PE film surfaces, morphological operations can effectively enhance the morphological features of particles, fill edge gaps, remove stray noise, and make the particle area more complete and prominent. The preferred structuring element size is 3×3 to 7×7 pixels to cover particle features of different sizes: 3×3 structuring elements are suitable for processing small particles (0.5-2μm); 5×5 and 7×7 structuring elements are suitable for processing medium-to-large particles (2-10μm). By combining multi-scale results, a morphologically enhanced feature map M is generated, in which the particle area appears as a connected, bright region.

[0151] The topology-preserving clusterer 43 is connected to the morphological processing unit 42 and is configured to receive the morphologically enhanced feature map, construct a high-dimensional feature vector, and perform topology-preserving clustering to cluster the pixels into speckled noise, blurred noise, dust particles, and background. Specifically, the topology-preserving clusterer 43 performs topology-preserving clustering by constructing a high-dimensional feature vector containing spatial, morphological, and statistical features, defining a topology-preserving objective function, and performing iterative optimization through gradient descent to maintain the local neighborhood structure, automatically determining the optimal number of categories, and classifying the pixels into four categories.

[0152] In actual implementation, we first construct a high-dimensional feature vector F:

[0153] ,

[0154] Where: F(p) represents the high-dimensional feature vector of point p; are spatial features, including pixel location (x, y) and local neighborhood statistics; are morphological features, including edge strength, direction, and curvature; Statistical features include local mean, variance and histogram features; is the differential geometry feature, including the Gaussian curvature and mean curvature calculated previously.

[0155] Define the topology preserving objective function:

[0156] ,

[0157] Where: J is the objective function value; and Respectively represent the representation of point i and j in the cluster space, with a dimension of d×1 (d is usually 2 or 3); is the similarity matrix element; λ is the balance parameter, dimensionless, with an optimal value of 0.3-0.7; K is the number of categories; represents the kth category; For category the center of represents the sum of all pairs of points (i, j); represents the sum of all categories; Indicates the category Sum all points in .

[0158] The first term of the objective function ensures the preservation of the local neighborhood structure, so that similar points remain close in the cluster space; the second term promotes intra-class aggregation, so that similar points are close to the class center.

[0159] Iterative optimization via gradient descent:

[0160] ,

[0161] in: and denote the cluster space representation of the tth and t+1th iterations respectively; η is the learning rate, with an optimal value of 0.01-0.05; Represents the objective function J gradient.

[0162] In the application of detecting fine dust on PE film surfaces, topology-preserving clustering can classify pixels into different categories based on multidimensional features while maintaining the topological structure of the data. During the iteration process, the number of categories, K, is automatically adjusted, preferably within a range of 3-5, and ultimately settled on 4, corresponding to the four categories of speckle noise, fuzzy noise, fine dust, and background. Speckle noise typically appears as small, bright, isolated points; fuzzy noise appears as diffuse areas with unclear boundaries; fine dust appears as connected areas with clear boundaries and uniform interiors; and background appears as flat areas with low grayscale.

[0163] Feature space fuser 44 is connected to topology-preserving clusterer 43 and is used to receive multi-level features and clustering results, calculate feature weights, perform nonlinear feature fusion, and generate the final dust particle position and feature information. Specifically, feature space fuser 44 performs nonlinear feature fusion by calculating the reliability of each feature layer, assigning adaptive weights, designing a nonlinear combination function, and integrating the background suppression layer features, noise detection layer features, and morphological processing features to generate the final dust particle representation.

[0164] In actual implementation, the reliability R of each feature layer is first calculated:

[0165] ,

[0166] in: Represents the reliability of the i-th feature layer; inter-classvariance represents the between-class variance; intra-classvariance represents the within-class variance; For category The number of samples; For category the center of is the global center, calculated as the average of all samples; is the feature of sample j; represents the sum of all categories; Indicates the category Sum all samples in .

[0167] Assign adaptive weights based on reliability :

[0168] ,

[0169] in: is the weight of the i-th feature layer, dimensionless; Represents the sum of the reliabilities of all feature layers.

[0170] Design nonlinear combination functions to integrate multi-level features:

[0171] ,

[0172] in: Represents the final feature representation of point p; is the weight of the i-th feature layer; is the representation of point p in the i-th feature layer; is a nonlinear mapping function, preferably a sigmoid function; represents the sum of all feature layers.

[0173] The sigmoid function is defined as:

[0174] ,

[0175] in: is the sigmoid function value, and its value range is (0,1); and are adjustable parameters that control the steepness and offset of the function respectively, and the optimal value is adaptively determined according to the characteristic distribution; is the base of natural logarithms.

[0176] In the application of particle detection on PE film surfaces, the feature fusion engine can adaptively adjust weights based on the reliability of each feature layer, integrating multi-level feature information. Through nonlinear mapping functions, it enhances the complementarity between different feature layers, improving overall detection performance. Ultimately, it generates particle location and feature information, including the particle's location coordinates (x, y), size (area and perimeter), shape (circularity and eccentricity), and reliability score.

[0177] like Figure 5 As shown, the present invention also provides a method for identifying electrostatic dust on the surface of a PE film, comprising the following steps:

[0178] Step S1: collecting multi-angle and multi-wavelength images of the PE film surface;

[0179] Step S2: constructing a manifold representation of the PE film surface, separating the PE film background and potential dust features, and generating a background suppressed image;

[0180] Step S3: extracting the differential geometric features of the background suppressed image, calculating the geodesic distance matrix, and generating a noise-dust probability map;

[0181] Step S4: performing edge detection and morphological processing on the noise-dust probability map, performing topology-preserving clustering, and dividing the image area into speckled noise, blurred noise, dust, and background;

[0182] Step S5: Fuse multi-level features to determine the location and feature information of dust particles on the surface of the PE film.

[0183] In practical applications, step S1 is implemented by the image acquisition module 1, step S2 is implemented by the multi-manifold mapping background suppression module 2, step S3 is implemented by the differential geometry feature extraction module 3, and steps S4 and S5 are implemented by the topological feature segmentation module 4.

[0184] Preferably, in step S2, the process of constructing a manifold representation of the PE film surface includes: constructing a local geometric structure of the image and generating a local geometric feature map; constructing a global similarity map based on the local geometric feature map, optimizing the manifold embedding objective function, and generating a low-dimensional representation of the PE film and dust; calculating the manifold entropy, determining the optimal segmentation threshold, and generating a background suppression image.

[0185] Preferably, in step S3, the process of extracting differential geometric features includes: converting the image into a height function, calculating the first-order derivative (gradient) and the second-order derivative (Hessian matrix); calculating the principal curvature, Gaussian curvature, mean curvature, shape index and curvature change rate based on the Hessian matrix; and generating multi-scale curvature representation in multiple scale spaces.

[0186] Preferably, in step S4, the process of topology preserving clustering includes: constructing a high-dimensional feature vector containing spatial, morphological, statistical and differential geometric features; defining a topology preserving objective function, iteratively optimizing it by gradient descent while maintaining the local neighborhood structure; automatically determining the optimal number of categories, and dividing the pixel points into four categories: speckled noise, blurred noise, dust and background.

[0187] Preferably, in step S5, the process of multi-level feature fusion includes: calculating the reliability of each feature layer and assigning adaptive weights; designing a nonlinear combination function to integrate the background suppression layer features, the noise detection layer features and the morphological processing features; generating the final dust particle representation, including the position coordinates, size, shape and reliability score of the dust particles.

[0188] The technical solutions of the present invention are described below through specific embodiments, but the protection scope of the present invention is not limited thereto.

[0189] In this embodiment, the PE film surface electrostatic dust identification system of the present invention is used to detect a batch of PE film samples in production. The PE film specifications are 50 μm thick and 300 mm × 400 mm in size.

[0190] First, the image acquisition module 1 captures multi-angle, multi-wavelength images of the PE film surface. A circular LED array 11, comprising 16 LEDs with a wavelength of 550 nm, illuminates the PE film surface from different angles. The multi-wavelength illumination system 12 sequentially activates three dark-field illumination modules with central wavelengths of 450 nm, 550 nm, and 650 nm, respectively. Three photoelectric imaging detectors 13 are evenly spaced at a 120° angle, with a resolution of 1920 × 1080 pixels and an exposure time of 10 ms.

[0191] Then, the multi-manifold mapping background suppression module 2 processes the collected image. The local structure preservation unit 21 determines the 10 nearest neighbor point sets for each pixel point and calculates the similarity using the Gaussian kernel function. The bandwidth parameter of the Gaussian kernel is The global structure mapping unit 22 constructs a global adjacency graph based on local features and solves the optimization problem to obtain a low-dimensional representation. The adaptive threshold adjustment unit 23 calculates the manifold entropy and contrast measure under different thresholds, and the weight parameter and They are set to 0.6 and 0.4 respectively to determine the optimal segmentation threshold and generate background suppressed images.

[0192] Next, the differential geometry feature extraction module 3 processes the background suppressed image. The differential geometry feature extractor 31 converts the image into a height function, calculates the first and second order derivatives, and then calculates the principal curvature, Gaussian curvature, mean curvature, shape index and curvature change rate. Repeat the calculation under (i=0,1,2) to generate a multi-scale curvature representation. The geodesic distance calculation unit 32 defines the feature-based Riemann metric tensor, and the weight parameter and The α values ​​are set to 0.2 and 0.8, respectively. The geodesic distance is calculated using the fast marching method and a distance matrix is ​​constructed. The manifold probability distribution estimator 33 estimates the probability density function based on representative noise and dust samples. The prior probabilities of noise and dust are set to 0.7 and 0.3, respectively. The probability of each point belonging to a noise point or dust point is calculated, generating a noise-dust point probability map.

[0193] Finally, the topological feature segmentation module 4 processes the noise-dust probability map. The edge detector 41 applies Gaussian smoothing to the probability map, and the standard deviation of the Gaussian kernel is Set to 2.0 pixels, calculate the Laplacian operator, detect zero crossing points, and gradient threshold The morphological processing unit 42 uses multi-scale structural elements of 3×3 to 7×7 pixels to perform dilation, erosion and opening and closing operations to generate morphological enhancement feature maps. The topology preserving clusterer 43 constructs a high-dimensional feature vector containing spatial, morphological, statistical and differential geometric features, defines the topology preserving objective function, and balances the parameters. Set to 0.5, the learning rate The value is set to 0.03, classifying pixels into four categories: speckled noise, blurred noise, dust particles, and background. The feature space fuser 44 calculates the reliability of each feature layer, assigns adaptive weights, and uses the sigmoid function as a nonlinear mapping function to integrate multi-level features to generate the final dust particle position and feature information.

[0194] In this example, the system successfully detected fine dust particles on the surface of PE film, including particles of various sizes ranging from 0.5 to 10 μm. The detection rate reached 86.3% in the 0.5-2 μm range and 98.7% in the 2-10 μm range, with a false alarm rate of only 2.1%. The system processed at a speed of 14.5 frames per second (1080p resolution), meeting real-time detection requirements.

[0195] In this embodiment, the method for identifying electrostatic dust on the surface of PE films of the present invention was used to detect a batch of PE film samples with different surface properties. PE films include three types: ordinary PE films, antistatic PE films, and optical-grade PE films.

[0196] First, step S1 is performed to capture multi-angle, multi-wavelength images of the three types of PE film surfaces. Sixteen LED light sources with a wavelength of 550 nm illuminate the PE film surfaces from different angles. Three dark-field illumination modules with center wavelengths of 450 nm, 550 nm, and 650 nm are sequentially activated. Three cameras with a resolution of 1920 × 1080 pixels capture images from different angles. The exposure time is adaptively adjusted within a range of 5–15 ms depending on the PE film type.

[0197] Then, step S2 is executed to construct a manifold representation of the PE film surface. For each type of PE film, the 10 nearest neighbor points of each pixel are determined, the similarity is calculated, and the local geometric structure is constructed; a global adjacency graph is constructed based on local features, the manifold embedding objective function is optimized, and a low-dimensional representation is generated; the manifold entropy and contrast measure are calculated, the optimal segmentation threshold is determined, and a background suppression image is generated. According to the surface characteristics of different types of PE films, the parameters are adaptively adjusted: the Gaussian kernel bandwidth parameter of ordinary PE film It is set to 0.3 times of the image grayscale standard deviation, 0.4 times for antistatic PE film, and 0.25 times for optical grade PE film.

[0198] Next, step S3 is executed to extract differential geometric features. The background suppressed image is converted into a height function, and the first and second derivatives are calculated to calculate the differential geometric features. The calculation is repeated at three scales to generate a multi-scale representation. The feature-based Riemannian metric tensor is defined to calculate the geodesic distance. The probability density function is estimated to calculate the probability of each point belonging to noise or dust, and a noise-dust probability map is generated. According to the surface characteristics of different types of PE films, the parameters are adaptively adjusted: the weight parameter of ordinary PE film and They are set to 0.2 and 0.8 respectively, the antistatic PE film is set to 0.25 and 0.75, and the optical grade PE film is set to 0.15 and 0.85.

[0199] Then, step S4 is executed to perform topology-preserving clustering. Gaussian smoothing is applied to the probability map, the Laplace operator is calculated, zero crossing points are detected, and edge feature maps are extracted; morphological operations are performed to generate morphologically enhanced feature maps; a high-dimensional feature vector is constructed, a topology-preserving objective function is defined, and clustering is performed to classify the pixels into four categories. According to the surface characteristics of different types of PE films, the parameters are adaptively adjusted: the standard deviation of the Gaussian kernel of ordinary PE films The default setting is 2.0 pixels, the anti-static PE film is set to 2.2 pixels, and the optical-grade PE film is set to 1.8 pixels.

[0200] Finally, step S5 is executed to fuse the multi-level features. The reliability of each feature layer is calculated and adaptive weights are assigned. A nonlinear combination function is designed to integrate the multi-level features. The final dust particle location and feature information is generated. For each detected dust particle, its location coordinates, size, shape, and reliability score are recorded.

[0201] In this example, the method successfully adapted to the surface characteristics of three different types of PE films. The test results showed that the detection rate for ordinary PE film was 85.7% in the 0.5-2μm range and 98.2% in the 2-10μm range; the detection rate for antistatic PE film was 84.1% in the 0.5-2μm range and 97.8% in the 2-10μm range; and the detection rate for optical-grade PE film was 87.9% in the 0.5-2μm range and 99.3% in the 2-10μm range. The false alarm rate for all three types of PE film was kept below 3%, and the processing speed was within the range of 12-15 frames per second, meeting the requirements of real-time detection.

[0202] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. PE film surface electrostatic dust identification system, characterized by: include: Image acquisition module, used to collect multi-angle and multi-wavelength images of the PE film surface; a multi-manifold mapping background suppression module, connected to the image acquisition module, for receiving the multi-angle and multi-wavelength images, constructing a manifold representation of the PE film surface, separating the PE film background and potential dust features, and generating a background suppressed image; a differential geometry feature extraction module connected to the multi-manifold mapping background suppression module, configured to receive the background suppression image, extract differential geometry features, calculate a geodesic distance matrix, and generate a noise-dust probability map; The topological feature segmentation module is connected to the differential geometry feature extraction module, and is used to receive the noise-dust probability map, perform edge detection and morphological processing, and perform topology-preserving clustering to divide the image area into speckled noise, blurred noise, dust and background, and determine the location and feature information of the dust on the PE film surface.

2. The PE film surface electrostatic dust identification system according to claim 1, characterized in that: The image acquisition module includes: Ring-shaped LED array for illuminating the PE film surface from multiple angles; A multi-wavelength illumination system includes three sets of dark field illumination modules with different wavelengths, which are used to activate and generate illumination light of different wavelengths in sequence; A plurality of photoelectric imaging detectors are respectively connected to the multi-wavelength illumination system and are used to collect images of the PE film surface from different angles.

3. The PE film surface electrostatic dust identification system according to claim 1, characterized in that: The multi-manifold mapping background suppression module includes: The local structure preserving unit is used to construct the local geometric structure of the image and generate a local geometric feature map; A global structure mapping unit, connected to the local structure preserving unit, is used to receive the local geometric feature map, construct a global similarity map, optimize the manifold embedding objective function, and generate a low-dimensional representation of the PE film and the dust particles; The adaptive threshold adjustment unit is connected to the global structure mapping unit and is used to receive the low-dimensional representation, calculate the manifold entropy, determine the optimal segmentation threshold, and generate a background suppression image.

4. The PE film surface electrostatic dust identification system according to claim 1, characterized in that: The differential geometry feature extraction module includes: Differential geometry feature extractor, used to convert the image into a height function, calculate Gaussian curvature, mean curvature and shape index, and generate differential geometry feature map; a geodesic distance calculation unit, connected to the differential geometry feature extractor, configured to receive the differential geometry feature map, define a feature-based Riemannian metric tensor, calculate the geodesic distance using a fast marching method, and construct a distance matrix; A manifold probability distribution estimator is connected to the geodesic distance calculation unit, and is used to receive the distance matrix and the differential geometry feature map, estimate the probability density of the feature space, calculate the probability of each point belonging to a noise point or a dust particle, and generate a noise point-dust particle probability map.

5. The PE film surface electrostatic dust identification system according to claim 1, characterized in that: The topological feature segmentation module includes: Edge detector, used to calculate the Gaussian Laplacian operator, detect zero crossing points, and extract edge feature maps; a morphological processing unit connected to the edge detector, configured to receive the edge feature map, perform multi-scale morphological dilation and erosion operations, and generate a morphological enhancement feature map; a topology preserving clusterer connected to the morphological processing unit, configured to receive the morphologically enhanced feature map, construct a high-dimensional feature vector, and perform topology preserving clustering to cluster pixels into speckled noise, blurred noise, dust particles, and background; The feature space fuser is connected to the topology preserving clusterer and is used to receive multi-level features and clustering results, calculate feature weights, perform nonlinear feature fusion, and generate final dust particle positions and feature information.

6. The PE film surface electrostatic dust identification system according to claim 3, characterized in that: The local structure preserving unit constructs the local geometric structure by determining the k-nearest neighbor point set of each pixel point, calculating the similarity between the pixel point and the nearest neighbor points, constructing a local similarity matrix, and extracting a feature vector to represent the local geometric structure.

7. The PE film surface electrostatic dust identification system according to claim 4, characterized in that: The differential geometric features extracted by the differential geometric feature extractor include principal curvature, Gaussian curvature, mean curvature, shape index and curvature change rate, and multi-scale curvature representation is generated in multiple scale spaces.

8. The PE film surface electrostatic dust identification system according to claim 5, characterized in that: The topology-preserving clusterer performs topology-preserving clustering by constructing a high-dimensional feature vector containing spatial, morphological and statistical features, defining a topology-preserving objective function, and classifying pixels into four categories by iterative optimization through gradient descent while maintaining the local neighborhood structure and automatically determining the optimal number of categories.

9. The PE film surface electrostatic dust identification system according to claim 5, characterized in that: The feature space fuser performs nonlinear feature fusion by calculating the reliability of each feature layer, assigning adaptive weights, designing a nonlinear combination function, integrating the background suppression layer features, the noise detection layer features and the morphological processing features to generate the final dust particle representation.

10. A method for identifying electrostatic dust on the surface of a PE film, characterized in that: The following steps are involved: Collect multi-angle and multi-wavelength images of the PE film surface; Constructing a manifold representation of the PE film surface, separating the PE film background and potential dust features, and generating a background suppressed image; Extract the differential geometric features of the background suppressed image, calculate the geodesic distance matrix, and generate the noise-dust probability map; Edge detection and morphological processing are performed on the noise-dust probability map, and topology-preserving clustering is performed to divide the image area into speckled noise, blurred noise, dust and background; Multi-level features are integrated to determine the location and characteristic information of dust particles on the surface of PE film.

Images