CN120656168B - PE film surface electrostatic dust recognition system and method - Google Patents

Abstract

This invention relates to the field of surface defect detection technology, and in particular to a system and method for identifying electrostatic dust on the surface of PE film. The system includes: an image acquisition module that acquires multi-angle, multi-wavelength images of the PE film surface using a ring LED array and a multi-wavelength illumination system; a multi-manifold mapping background suppression module that treats the PE film surface as a composite manifold, separating background and dust features; a differential geometric feature extraction module that extracts geometric features such as curvature and shape index, calculates the geodesic distance matrix, and generates a noise-dust probability map; and a topology feature segmentation module that performs edge detection, morphological processing, and topology-preserving clustering to classify image regions and determine dust locations. This method, combining multi-manifold mapping and differential geometric features, effectively overcomes the detection difficulties caused by the highly reflective surface of PE film and can accurately identify dust particles at the 0.5-10 μm level.

Description

PE film surface electrostatic dust recognition system and method

Technical Field

The invention relates to the technical field of surface defect detection, in particular to a PE film surface electrostatic dust recognition system and method for accurately detecting tiny electrostatic adsorption dust particles on the PE film surface.

Background

The PE film is used as an important packaging material and a protective layer of a photoelectric element, and the surface quality of the PE film directly influences the performance and appearance of a product. In the production and application process of the PE film, tiny dust particles are easy to be adsorbed on the surface of the PE film due to the electrostatic effect, so that pollution which is difficult to remove is formed.

The existing PE film surface detection technology mainly comprises manual visual inspection, common image processing and a traditional machine vision method. The manual visual inspection is inefficient and subject to fatigue, the conventional image processing method has limited capability of detecting minute dust particles (especially on the order of 0.5-10 μm), traditional machine vision methods often have difficulty in effectively distinguishing motes from background noise when treating highly reflective surfaces such as PE films.

In addition, in the prior art, a simple threshold segmentation or basic morphology processing method is generally adopted, so that complex light reflection characteristics of the surface of the PE film are difficult to process, and particularly, the detection accuracy is obviously reduced for PE films with different illumination angles and different surface materials. Therefore, the development of a system and a method capable of accurately identifying micro electrostatic dust particles on the surface of a PE film has important practical value.

Disclosure of Invention

The invention aims to provide a PE film surface electrostatic dust identification system and a PE film surface electrostatic dust identification method, which can overcome the problems existing in the prior art and realize accurate identification of PE film surface dust with the level of 0.5-10 mu m.

The invention provides a PE film surface electrostatic dust recognition system, which comprises:

the image acquisition module is used for acquiring multi-angle and multi-wavelength images of the PE film surface;

The multi-manifold mapping background suppression module is connected with the image acquisition module and is used for receiving the multi-angle multi-wavelength image, constructing PE film surface manifold representation, separating PE film background and potential dust features and generating a background suppression image;

The differential geometric feature extraction module is connected with the multi-manifold mapping background suppression module and is used for receiving the background suppression image, extracting differential geometric features, calculating a geodesic distance matrix and generating a noise-dust probability map;

The topological feature segmentation module is connected with the differential geometric feature extraction module and is used for receiving the noise-dust probability map, executing edge detection and morphological processing, carrying out topological retention clustering, dividing an image area into speckle noise, fuzzy noise, dust and background, and determining the position and feature information of the dust on the surface of the PE film.

Preferably, the image acquisition module includes:

An annular LED array for illuminating the PE film surface from multiple angles;

the multi-wavelength illumination system comprises three groups of dark field illumination modules with different wavelengths, which are used for sequentially activating and generating illumination light rays with different wavelengths;

and the plurality of photoelectric imaging detectors are respectively connected with the multi-wavelength illumination system and are used for acquiring PE film surface images from different angles.

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

The local structure holding unit is used for constructing a local geometric structure of the image and generating a local geometric feature map;

The global structure mapping unit is connected with the local structure maintaining unit and is used for receiving the local geometric feature map, constructing a global similarity map, optimizing manifold embedding objective functions and generating low-dimensional representations of PE films and dust particles;

And the self-adaptive threshold adjusting unit is connected with the global structure mapping unit and is used for receiving the low-dimensional representation, calculating manifold entropy, determining an optimal segmentation threshold and generating a background suppression image.

Preferably, the differential geometric feature extraction module includes:

the differential geometric feature extractor is used for converting the image into a height function, calculating Gaussian curvature, average curvature and shape index and generating a differential geometric feature map;

the geodesic distance calculation unit is connected with the differential geometric feature extractor and is used for receiving the differential geometric feature graph, defining a feature-based Riemann metric tensor, calculating the geodesic distance by using a fast travelling method and constructing a distance matrix;

And the manifold probability distribution estimator is connected with the geodesic distance calculation unit and is used for receiving the distance matrix and the differential geometric feature map, estimating the probability density of the feature space, calculating the probability that each point belongs to a noise point or dust and generating a noise point-dust probability map.

Preferably, the topological feature segmentation module comprises:

the edge detector is used for calculating a Gaussian Laplace operator, detecting zero crossing points and extracting an edge feature map;

a morphological processing unit, connected to the edge detector, for receiving the edge feature map, performing multi-scale morphological dilation and erosion operations, and generating a morphological enhancement feature map;

The topology maintenance clustering device is connected with the morphology processing unit and is used for receiving the morphology enhancement feature map, constructing a high-dimensional feature vector, executing topology maintenance clustering and clustering pixel points into speckle noise points, fuzzy noise points, tiny dust and background;

The feature space fusion device is connected with the topology maintenance clustering device and is used for receiving multi-level features and clustering results, calculating feature weights, performing nonlinear feature fusion and generating final dust position and feature information.

Preferably, the local structure holding unit constructs the local geometry by determining a k-nearest neighbor set of each pixel, calculating a similarity between the pixel and the nearest neighbor, constructing a local similarity matrix, and extracting feature vectors to represent the local geometry.

Preferably, the differential geometry features extracted by the differential geometry feature extractor include a principal curvature, a gaussian curvature, an average curvature, a shape index, and a curvature rate of change, and generate a multi-scale curvature representation over a plurality of scale spaces.

Preferably, the topology preserving clustering device performs topology preserving clustering by constructing high-dimensional feature vectors containing space, morphology and statistical features, defining a topology preserving objective function, preserving a local neighborhood structure through gradient descent iterative optimization, automatically determining the optimal category number and dividing pixel points into four categories.

Preferably, the feature space fusion device performs nonlinear feature fusion by calculating the reliability of each feature layer, distributing self-adaptive weights, designing nonlinear combination functions, integrating background suppression layer features, noise detection layer features and morphological processing features, and generating a final mote representation.

The PE film surface electrostatic dust identification method comprises the following steps:

collecting multi-angle multi-wavelength images of the surface of the PE film;

constructing PE film surface manifold representation, separating PE film background and potential dust feature, and generating background inhibition image;

extracting differential geometric features of the background suppression image, calculating a geodesic distance matrix, and generating a noise-dust probability map;

performing edge detection and morphological processing on the noise-dust probability map, performing topology maintenance clustering, and dividing an image area into speckle noise, fuzzy noise, dust and background;

and (5) fusing multi-level characteristics, and determining the position and characteristic information of the dust on the surface of the PE film.

The beneficial effects of the invention include:

1. The detection precision of the dust is improved, namely the dust with the level of 0.5 mu m which is difficult to find by the traditional method can be detected by multi-manifold mapping and differential geometric feature extraction technology, the detection rate reaches 85% in the range of 0.5-2 mu m, and the detection rate reaches 98% in the range of 2-10 mu m.

2. The anti-interference capability is enhanced, namely the distinguishing mechanism based on topological characteristics essentially distinguishes tiny dust and other interference factors (such as light reflection, scratches and the like) on the surface of the PE film, and the false alarm rate is reduced to below 3%.

3. The adaptability of the system is improved, the system can automatically adapt to PE films with different materials and surface characteristics by a self-adaptive parameter adjustment mechanism, and the universality is remarkably improved.

4. The production efficiency is improved, the system can process 1080p resolution ratio images at 15 frames/second, the detection efficiency is greatly improved, and the manual detection links are reduced.

5. The production cost is reduced, the defective rate is reduced, the loss of the subsequent processing link is reduced, and the overall economic benefit is improved.

Drawings

FIG. 1 is a schematic diagram of the overall structure of a PE film surface electrostatic dust recognition system according to the present invention;

FIG. 2 is a schematic diagram of a multi-manifold mapping background suppression module according to the present invention;

FIG. 3 is a schematic diagram of a differential geometry extraction module according to the present invention;

FIG. 4 is a schematic diagram of a topology feature segmentation module according to the present invention;

FIG. 5 is a flow chart of the method for identifying electrostatic dust on the surface of the PE film.

Detailed Description

Referring to fig. 1-5, the present invention will be described in further detail with reference to the accompanying drawings and specific embodiments. These examples are only for illustrating the present invention and are not intended to limit the scope of the present invention.

As shown in FIG. 1, the PE film surface electrostatic dust identification system provided by the invention comprises an image acquisition module 1, a multi-manifold mapping background suppression module 2, a differential geometric feature extraction module 3 and a topological feature segmentation module 4.

The image acquisition module 1 is used for acquiring multi-angle and multi-wavelength images of the PE film surface. The multi-manifold mapping background suppression module 2 is connected with the image acquisition module 1 and is used for receiving the multi-angle multi-wavelength image, constructing PE film surface manifold representation, separating PE film background and potential dust features and generating a background suppression image. The differential geometric feature extraction module 3 is connected with the multi-manifold mapping background suppression module 2 and is used for receiving the background suppression image, extracting differential geometric features, calculating a geodesic distance matrix and generating a noise-dust probability map. The topological feature segmentation module 4 is connected with the differential geometric feature extraction module 3 and is used for receiving the noise point-dust probability map, executing edge detection and morphological processing, carrying out topological retention clustering, dividing an image area into speckle noise points, fuzzy noise points, dust and background, and determining the position and feature information of the dust on the PE film surface.

As shown in fig. 1, in a preferred embodiment of the present invention, the image acquisition module 1 includes an annular LED array 11, a multi-wavelength illumination system 12, and a plurality of photo-imaging detectors 13.

The annular LED array 11 is used to illuminate the PE film surface from multiple angles. Preferably, the annular LED array 11 comprises 16 LED light sources which are uniformly distributed in an annular shape, and can irradiate the surface of the PE film from different angles to generate a multi-angle illumination effect. The wavelength range of the LED light source is preferably 400-700 nm, and the visible spectrum range is covered, so that abundant spectrum information can be obtained.

The multi-wavelength illumination system 12 comprises three sets of dark field illumination modules of different wavelengths for sequentially activating illumination light rays of different wavelengths. Preferably, the central wavelength of the three groups of dark field illumination modules is respectively 450nm (blue light), 550nm (green light) and 650nm (red light), and the wavelength interval is 100nm, so as to obtain the reflection characteristics of the PE film surface under different wavelengths. Each group of dark field illumination modules cooperates with the annular LED array 11 to ensure illumination effects of multiple angles and multiple wavelengths.

A plurality of photo imaging detectors 13 are respectively connected with the multi-wavelength illumination system 12 for acquiring PE film surface images from different angles. Preferably, the system is configured with three high resolution cameras (at least 1080p resolution) uniformly distributed at 120 ° angles to acquire an omnidirectional image of the PE film surface. The exposure time of the camera can be adjusted to be in the range of 1ms to 50ms so as to adapt to different illumination conditions.

In actual operation, the annular LED array 11 firstly illuminates the PE film surface from multiple angles, then the multi-wavelength illumination system 12 sequentially activates three groups of dark field illumination modules with different wavelengths, and finally the multiple photo-imaging detectors 13 synchronously acquire the PE film surface images. The multi-angle multi-wavelength acquisition mode can acquire richer image information on the surface of the PE film, and provides sufficient data support for subsequent processing.

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

The local structure holding unit 21 is used to construct a local geometry of the image, generating a local geometry map. Specifically, the local structure holding unit 21 constructs a local geometry by determining a k-nearest neighbor set of each pixel, calculating the similarity between the pixel and the nearest neighbor, constructing a local similarity matrix, and extracting feature vectors representing the local geometry.

In a practical implementation, for each pixel point p in an image, its k-nearest neighbor set is first determinedThe k value is preferably set to 8-12 to balance the integrity of the local information and the computational complexity. Then, the similarity between the point p and the adjacent point is calculated, and a Gaussian kernel function is adopted:

,

wherein: representing the similarity between pixel points i and j; And Feature vectors (including position coordinates and gray values) respectively representing pixel points i and j; The unit is pixel value, and the optimal value is 0.1-0.5 times of the standard deviation of the gray scale of the image; Representing feature vectors AndEuclidean distance between them.

In the practical application of PE film surface electrostatic dust identification, when detecting smaller dust particles (0.5-2 μm),The value may be set to a smaller value, e.g. 0.2 standard deviation, to enhance local detail, when larger dust particles (5-10 μm) are detected,The value may be set to a larger value (e.g., 0.4 standard deviation) to obtain a more stable representation of the feature.

Based on the similarity matrix W, constructing a feature representation of the local geometry:

,

Wherein L is Laplace matrix, dimension is n×n, n is total number of image pixel points, D is diagonal matrix, dimension is n×n, diagonal element And representing the sum of the similarity of the ith pixel point and all the adjacent points.

By solving the generalized eigenvalue problem:

,

f is a feature vector, and the dimension is n multiplied by 1; Is a characteristic value. The feature vector corresponding to the smallest non-zero feature value is taken as a representation of the local geometry.

The global structure mapping unit 22 is connected to the local structure holding unit 21 for receiving the local geometrical feature map, constructing a global similarity map, optimizing the manifold embedding objective function, and generating a low-dimensional representation of the PE film and the mote.

In a practical implementation, the global structure mapping unit 22 first builds a global adjacency graph G based on local features, and then defines an optimization objective function:

,

wherein: And Representing the representation of pixel points i and j in a low-dimensional space respectively, wherein the dimension is d multiplied by 1 (d is usually 2 or 3); a similarity matrix element calculated for the foregoing; Representing the euclidean distance between points i and j in the low dimensional space.

This optimization problem can be translated into a solution:

,

Wherein Y is a low-dimensional representation matrix of all points, and the dimension is nxd; Representing the transpose of matrix Y.

In PE film surface mote identification applications, the optimization objective is to keep similar points (as belonging to the PE film background or to the mote) close in low-dimensional space, while separating different classes of points (such as PE film background and mote points). By solving this optimization problem, a low-dimensional representation that maintains the topological relationship can be obtained, separating the PE film background and the mote in a low-dimensional space.

The adaptive threshold adjustment unit 23 is connected to the global structure mapping unit 22 for receiving the low-dimensional representation, calculating the manifold entropy, determining the optimal segmentation threshold, and generating the background suppression image.

In a practical implementation, the adaptive threshold adjustment unit 23 calculates the manifold entropy at different thresholds t:

,

wherein: Representing the manifold entropy at threshold t; the probability distribution of the ith area under the threshold t is represented, and the probability distribution is calculated as the number of pixel points of the area divided by the total number of pixel points; Representing the natural logarithm.

While taking into account contrast measures:

,

wherein: representing a contrast measure at a threshold t; And Average gray values of foreground (dust) and background (PE film) are represented respectively; And Respectively representing gray standard deviation of the foreground and the background; representing the absolute value of the difference between the foreground and background average gray values.

Constructing a comprehensive evaluation function:

,

wherein: is a comprehensive evaluation function; And The preferred values are 0.6 and 0.4 for the weight parameters, dimensionless, to balance the effects of entropy and contrast.

By searchingMaximum threshold valueThe method is applied to low-dimensional representation and generates a background suppression image. In the application of detecting the dust on the surface of the PE film,Generally corresponds to the optimal separation point of the PE film background and the mote, and the value of the separation point is related to PE film material, surface smoothness and mote characteristics. For example, for PE films with high smoothness,Typically between the low-dimensional representation of the two peaks of the gray histogram and for PE films with a rougher surface,It may be desirable to bias the background peaks to reduce false positives.

As shown in fig. 3, in a preferred embodiment of the present invention, the differential geometry extracting module 3 includes a differential geometry extractor 31, a geodesic distance calculating unit 32, and a manifold probability distribution estimator 33.

The differential geometry extractor 31 is configured to convert the image into a height function, calculate a gaussian curvature, an average curvature, and a shape index, and generate a differential geometry map. Specifically, the differential geometry features extracted by the differential geometry feature extractor 31 include a principal curvature, a gaussian curvature, an average curvature, a shape index, and a curvature change rate, and generate a multi-scale curvature representation in a plurality of scale spaces.

In a practical implementation, the background-suppressed image I (x, y) is regarded as a height function h (x, y) =i (x, y) on a two-dimensional manifold. Calculate first derivative (gradient):

,

wherein: A gradient vector representing a height function h, the dimension being 2 x 1; And The partial derivatives of h for x and y, respectively, can be calculated by finite difference methods.

Calculate the second derivative (Hessian matrix):

,

h is a Hessian matrix, and the dimension is 2 multiplied by 2; And Respectively representing the second partial derivatives of h to x and y; And Represents the cross partial derivatives of h, which are equal according to the nature of the continuous function.

Based on the Hessian matrix, the principal curvature is calculatedAnd(I.e., eigenvalues of the Hessian matrix). Then, the gaussian curvature K and the average curvature H are calculated:

,

,

Wherein K is Gaussian curvature and represents the inner curvature of the curved surface at the point, and H is average curvature and represents the outer curvature of the curved surface at the point; And Is the principal curvature, i.e. the eigenvalue of the Hessian matrix.

Calculating a shape index S:

,

wherein S is a shape index, and the value range is [ -1,1]; Representing the ratio of the sum of the principal curvatures to the difference between the principal curvatures, when And taking a limit value.

Calculating a curvature change rate C:

,

wherein C is curvature change rate; a gradient vector representing a gaussian curvature K; representing the dot product of the gradient vector, i.e. the modulo square of the gradient vector.

In PE film surface mote detection applications, these differential geometric features can effectively distinguish between different types of surface structures. For example, motes typically exhibit localized protrusions or depressions with a larger average curvature H and smaller shape index S variations, while PE film surface textures typically exhibit undulations with a smaller average curvature H and larger shape index S variations. By being in multiple dimensionsThe above calculations are repeated below, generating a multi-scale curvature representation. The preferred dimensions are selected asI=0, 1,2, covering different sized dust features.

The geodesic distance calculation unit 32 is connected to the differential geometric feature extractor 31 for receiving the differential geometric feature map, defining a feature-based Riemann metric tensor, calculating geodesic distances using a fast-marching method, and constructing a distance matrix.

In a practical implementation, a feature-based Riemann metric tensor M is first defined:

,

Wherein M (p) is a Riemann metric tensor at point p, and the dimension is 2× 2;I and is a 2×2 identity matrix; The dimension is 2×1 for the feature gradient vector at point p; alpha and beta are weight parameters, dimensionless, and optimal values are 0.2 and 0.8 respectively, and are used for balancing the influence of isotropy and anisotropy; representing the outer product of the vectors, the result being a 2 x 2 matrix.

In PE film surface mote detection applications, the Riemann metric tensor M defines a distance metric in a feature space, and can adaptively adjust distance calculation according to a feature gradient. For example, in the region of the dust edge, the feature gradient is large, M is mainly composed of anisotropic portionsContribution to increase the geodesic distance along the gradient direction, and in the flat region, the characteristic gradient is smaller, M is mainly formed from isotropic partContribution, close to euclidean distance.

Based on the Riemann metric tensor, a fast marching method is used to calculate the geodesic distance. The fast travelling method comprises the following basic steps:

1. Initializing, namely setting the distance of a starting point to be 0, setting the distance of other points to be infinity, and constructing a candidate point priority queue

2. Iterative calculation, namely, each time the point with the smallest distance in the queue is processed, the distance between the neighborhood points is updated, and the priority queue is readjusted

3. Termination condition, all points are processed or reach a preset distance threshold

Finally, a complete geodesic distance matrix D is constructed, whereinRepresenting the geodesic distance from point i to point j.

The manifold probability distribution estimator 33 is connected to the geodesic distance calculating unit 32, and is configured to receive the distance matrix and the differential geometric feature map, estimate probability density of the feature space, calculate probability that each point belongs to a noise point or dust, and generate a noise point-dust probability map.

In a practical implementation, representative noise and dust samples are first selected, and a probability density function is estimated based on these samples. The kernel density estimation method is used, but using geodesic distances instead of euclidean distances:

,

wherein: Representing that feature F belongs to a category Probability density of (c); Is of the category Is a sample number of (a); For kernel functions (preferably using Gaussian kernels ) H is a bandwidth parameter; The distance is measured for the ground wire; Is of the category Is a sample feature of (2); Representing pair categories Is summed up.

For each pixel point p, calculating the probability of belonging to noiseAnd probability of belonging to dust:

,

,

Wherein: the posterior probability that the point p belongs to the noise point is represented; the posterior probability that the point p belongs to the dust; And Respectively representing the conditional probability density of the characteristics of the point p under the noise point and the dust category; And The prior probabilities of the noise point and the tiny dust are respectively, the dimensionless, can be set according to experience, and the optimal values are respectively 0.7 and 0.3, and reflect the fact that the noise point is generally more common than the tiny dust on the surface of the PE film; For normalization factors, the following is calculated:

,

In PE film surface dust detection application, the probability that each point belongs to a noise point or dust can be estimated in a self-adaptive mode according to sample distribution in a feature space by a nuclear density estimation method, and a specific parameter distribution form is not required to be assumed. Meanwhile, by using the geodesic distance instead of the Euclidean distance, the nonlinear structure in the feature space can be captured better, and the classification accuracy is improved. Finally, a noise-dust probability map is generated, wherein the value of each pixel point represents the probability of belonging to dust The value range is [0,1].

As shown in fig. 4, in a preferred embodiment of the present invention, the topology feature segmentation module 4 includes an edge detector 41, a morphology processing unit 42, a topology preserving cluster 43, and a feature space fusion 44.

The edge detector 41 is configured to calculate a laplacian of gaussian, detect zero-crossing points, and extract an edge feature map.

In a practical implementation, gaussian smoothing is first applied to the noise-dust probability map P:

$,

wherein: Representing the gaussian smoothed image; represents a convolution operation; the standard deviation of Gaussian kernel is expressed in pixels, and the optimal value is 1.5-2.5 pixels so as to balance the accuracy and noise resistance of edge detection; Is a two-dimensional Gaussian function; for normalizing the coefficients, the sum of gaussian kernels is ensured to be 1.

The laplace operator is then calculated:

,

wherein: Representation of Is a laplace operator of (c); And Respectively representSecond partial derivatives of x and y.

The zero crossing point, i.e. the position at which the laplace operator value changes from positive to negative or vice versa, is detected, which positions correspond to the edges of the image. Meanwhile, using the gradient amplitude as a threshold condition, screening the significant edges:

,

wherein: Representation of Is calculated as the gradient amplitude of (1);The preferred value is 1.5 times the average value of the image gradients for the gradient threshold.

In PE film surface mote detection applications, the Gaussian Laplace operator can effectively detect mote edges, particularly for circular or elliptical motes, the edges of which appear as closed zero-crossing loops in the Laplace operator image. By adjustingThe value can be suitable for tiny dust with different sizes, namely smallerThe value (e.g. 1.5 pixels) is suitable for detecting the edges of small dust particles (0.5-2 μm), largerThe value (e.g. 2.5 pixels) is suitable for detecting the edges of large dust particles (5-10 μm). And finally, generating an edge characteristic diagram E.

The morphology processing unit 42 is connected to the edge detector 41 for receiving the edge signature, performing a multi-scale morphological dilation and erosion operation, generating a morphological enhancement signature.

In practical implementation, a multi-scale structure element set { is designedAnd (3) structural elements with different sizes and shapes are included to adapt to tiny dust with different sizes. Performing a morphological operation sequence on the edge feature map E:

expansion operation:

The expansion operation is defined as:

,

wherein: Representing the inflated image; Representing an edge feature map; representing a structural element; Representing a maximum value taking operation; representing the coordinate offset in the structural element. The expansion operation expands the edges, filling the small gaps.

And (3) corrosion operation:

The etching operation is defined as:

,

wherein: Representing the image after corrosion; Representing the inflated image; representing a structural element; Representing a minimum taking operation. The etching operation removes fine noise points and retains the main structure.

Opening and closing operation: the closing operation is defined as first expansion followed by corrosion: The opening operation is defined as corrosion followed by expansion: wherein Close and Open represent a closed operation and an Open operation, respectively; And Representing structural elements. The opening and closing operation further smoothes the edges, enhancing the area integrity.

In the PE film surface dust detection application, morphological operation can effectively enhance the morphological characteristics of dust, fill the edge gaps, remove stray noise points and enable the dust area to be more complete and outstanding. The preferred structural elements are 3x 3 to 7 x 7 pixels in size to cover the different size of the fine dust features, 3x 3 structural elements are suitable for handling small dust particles (0.5-2 μm) and 5 x 5 and 7 x 7 structural elements are suitable for handling medium and large dust particles (2-10 μm). By combining the multi-scale results, a morphologically enhanced feature map M is generated in which the mote region appears as a connected highlight region.

The topology preserving clustering unit 43 is connected to the morphology processing unit 42, and is configured to receive the morphology enhancement feature map, construct a high-dimensional feature vector, perform topology preserving clustering, and cluster the pixel points into speckle noise points, blur noise points, dust particles and background. Specifically, the topology preserving clustering device 43 performs topology preserving clustering by constructing high-dimensional feature vectors including spatial, morphological and statistical features, defining a topology preserving objective function, preserving a local neighborhood structure by iterative optimization of gradient descent, automatically determining the optimal number of classes, and classifying pixels into four classes.

In a practical implementation, first a high-dimensional feature vector F is constructed:

,

wherein F (p) represents a high-dimensional eigenvector of the point p; is a spatial feature, including pixel location (x, y) and local neighborhood statistics; Is a morphological feature including edge strength, direction, and curvature; is a statistical feature including local mean, variance, and histogram features; including the gaussian curvature calculated previously, the average curvature, etc., for the differential geometry.

Defining a topology preservation objective function:

,

Wherein J is an objective function value; And Representing the representation of points i and j in the clustering space respectively, with dimension d×1 (d usually takes 2 or 3); lambda is a balance parameter, dimensionless, preferably 0.3-0.7, K is a class number; Represents the kth category; Is of the category Is defined by a center of (a); representing the summation of all pairs of points (i, j); representing summing all categories; Representing pair categories Is added up.

The first term of the objective function ensures the maintenance of a local neighborhood structure, so that similar points are kept close in a clustering space, and the second term promotes the aggregation of class clusters, so that similar points are close to the center of the class.

Iterative optimization by gradient descent:

,

wherein: And The cluster space representation of the t time and the t+1st time iteration are respectively represented, wherein eta is the learning rate, and the optimal value is 0.01-0.05; Representing the objective function J pair Is a gradient of (a).

In PE film surface dust detection application, topology preserving clustering can divide pixel points into different categories according to multidimensional features, and meanwhile, the topology structure of data is preserved. In the iterative process, the category number K is automatically adjusted, the optimal range is 3-5, and finally, the optimal range is determined to be 4, and the optimal range corresponds to four categories of speckle noise points, fuzzy noise points, tiny dust and background. Speckle noise is usually represented as small and bright isolated dots, blurred noise is represented as diffuse areas with unclear boundaries, motes are represented as connected areas with clear boundaries and uniform interiors, and backgrounds are represented as flat areas with lower gray scales.

The feature space fusion device 44 is connected with the topology maintenance clustering device 43, and is used for receiving multi-level features and clustering results, calculating feature weights, performing nonlinear feature fusion, and generating final dust position and feature information. Specifically, feature space fusion engine 44 performs nonlinear feature fusion by computing the reliability of each feature layer, assigning adaptive weights, designing nonlinear combining functions, integrating background suppression layer features, noise detection layer features, and morphological processing features, and generating a final mote representation.

In practical implementation, the reliability R of each feature layer is first calculated:

,

wherein: Indicating the reliability of the ith feature layer, inter-CLASSVARIANCE indicating the inter-class variance, intra-CLASSVARIANCE indicating the intra-class variance; Is of the category Is a sample number of (a); Is of the category Is defined by a center of (a); as a global center, calculating as an average value of all samples; Is characteristic of sample j; representing summing all categories; Representing pair categories Summation of all samples in (a).

Based on reliability, adaptive weights are assigned:

,

Wherein: The weight of the ith feature layer is dimensionless; representing the sum of all feature layer reliabilities.

Designing a nonlinear combination function, and integrating multi-level characteristics:

,

wherein: a final feature representation representing point p; The weight of the ith feature layer; A representation of point p at the ith feature layer; as the nonlinear mapping function, a sigmoid function is preferably adopted; representing summing all feature layers.

The sigmoid function is defined as:

,

wherein: the value range is (0, 1) for sigmoid function value; And For the adjustable parameters, respectively controlling the steepness and the offset of the function, and adaptively determining a preferred value according to the characteristic distribution; Is the base of natural logarithms.

In PE film surface dust detection application, the feature fusion device can adaptively adjust weights according to the reliability of each feature layer, and integrate multi-level feature information. Through the nonlinear mapping function, complementarity among different feature layers can be enhanced, and overall detection performance is improved. The location and characteristic information of the mote is finally generated, including location coordinates (x, y), size (area and perimeter), shape (circularity and eccentricity), reliability score, and the like of the mote.

As shown in fig. 5, the invention further provides a method for identifying electrostatic dust on the surface of a PE film, which comprises the following steps:

S1, collecting multi-angle and multi-wavelength images of the surface of a PE film;

s2, constructing PE film surface manifold representation, separating PE film background and potential dust feature, and generating background inhibition image;

s3, extracting differential geometric features of the background suppression image, calculating a geodesic distance matrix, and generating a noise-dust probability map;

S4, performing edge detection and morphological processing on the noise-dust probability map, and performing topology maintenance clustering to divide an image area into speckle noise, fuzzy noise, dust and background;

And S5, fusing the multi-level characteristics and determining the position and characteristic information of the dust on the surface of the PE film.

In practical application, 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 geometric feature extraction module 3, and steps S4 and S5 are implemented by the topological feature segmentation module 4.

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

Preferably, in step S3, the process of extracting the differential geometry includes converting the image into a height function, calculating a first derivative (gradient) and a second derivative (Hessian matrix), calculating a principal curvature, a Gaussian curvature, an average curvature, a shape index, and a curvature change rate based on the Hessian matrix, and generating a multi-scale curvature representation in a plurality of scale spaces.

Preferably, in step S4, the process of topology preserving clustering comprises constructing high-dimensional feature vectors containing spatial, morphological, statistical and differential geometric features, defining a topology preserving objective function, iteratively optimizing through gradient descent while preserving a local neighborhood structure, automatically determining an optimal class number, and classifying pixel points into four classes of speckle noise, blur noise, dust and background.

Preferably, in step S5, the multi-level feature fusion process includes calculating reliability of each feature layer, distributing self-adaptive weights, designing a nonlinear combination function, integrating background suppression layer features, noise detection layer features and morphological processing features, and generating a final mote representation including position coordinates, size, shape, reliability scores and the like of motes.

The technical scheme of the present invention is described below by specific examples, but the scope of the present invention is not limited thereto.

In this embodiment, the PE film surface electrostatic dust recognition system of the present invention is used to detect PE film samples in a batch of production. The PE film had a gauge of 50 μm thick and dimensions 300mm by 400mm.

Firstly, collecting multi-angle multi-wavelength images of the PE film surface through an image collecting module 1. The annular LED array 11 contains 16 LED light sources with a wavelength of 550nm, illuminating the PE film surface from different angles. The multi-wavelength illumination system 12 sequentially activates three sets of dark field illumination modules having center wavelengths of 450nm, 550nm, and 650nm, respectively. The three photo imaging detectors 13 are uniformly distributed at an angle of 120 deg., with a resolution of 1920×1080 pixels and an exposure time set to 10ms.

The multi-manifold mapping background suppression module 2 then processes the acquired images. The local structure holding unit 21 determines a 10-neighbor set of each pixel, calculates the similarity using a gaussian kernel function, and the bandwidth parameter of the gaussian kernelIs set to 0.3 times the standard deviation of the gray scale of the image. The global structure mapping unit 22 builds a global adjacency graph based on the local features and solves the optimization problem to obtain a low-dimensional representation. The adaptive threshold adjustment unit 23 calculates manifold entropy and contrast measure at different thresholds, weight parametersAndAnd respectively setting the optimal segmentation threshold value to 0.6 and 0.4, and determining the optimal segmentation threshold value to generate a background suppression image.

Next, the differential geometry extraction module 3 processes the background suppressed image. The differential geometry extractor 31 converts the image into a height function, calculates a first derivative and a second derivative, and further calculates a principal curvature, a gaussian curvature, an average curvature, a shape index, and a curvature change rate. At three dimensionsRepeating the calculations at (i=0, 1, 2) to generate a multi-scale curvature representation. The geodesic distance calculation unit 32 defines a feature-based Riemann metric tensor, a weight parameterAndAnd respectively setting the distance to be 0.2 and 0.8, calculating the geodesic distance by using a fast travelling method, and constructing a distance matrix. Manifold probability distribution estimator 33 estimates probability density functions based on representative noise and mote samples, the prior probabilities of the noise and mote are set to 0.7 and 0.3, respectively, and the probability that each point belongs to a noise or mote is calculated, thus generating a noise-mote probability map.

Finally, the topological feature segmentation module 4 processes the noise-dust probability map. The edge detector 41 applies gaussian smoothing to the probability map, the standard deviation of the gaussian kernelSetting to 2.0 pixels, calculating Laplacian, detecting zero crossing point, gradient thresholdAnd setting the image gradient average value to be 1.5 times of the image gradient average value, and extracting an edge characteristic image. The morphology processing unit 42 performs expansion, etching, and opening and closing operations using the multi-scale structural elements of 3×3 to 7×7 pixels, generating a morphology-enhanced feature map. The topology preserving clustering means 43 constructs high-dimensional feature vectors comprising spatial, morphological, statistical and differential geometric features defining topology preserving objective functions, balancing parametersSet to 0.5, learning rateThe pixel point is set to be 0.03, and is divided into four types of speckle noise point, fuzzy noise point, tiny dust and background. The feature space fusion device 44 calculates the reliability of each feature layer, distributes adaptive weights, uses a sigmoid function as a nonlinear mapping function, integrates multi-level features, and generates final dust position and feature information.

In this example, the system successfully detected the mote on the PE film surface, including mote particles of various sizes in the range of 0.5-10 μm. The detection result shows that the detection rate reaches 86.3% in the range of 0.5-2 mu m, the detection rate reaches 98.7% in the range of 2-10 mu m, and the false alarm rate is only 2.1%. The processing speed is 14.5 frames/second (1080 p resolution), and the real-time detection requirement is met.

In this embodiment, a batch of PE film samples having different surface properties are detected using the PE film surface electrostatic dust identification method of the present invention. The PE film comprises three types of common PE film, antistatic PE film and optical PE film.

Firstly, step S1 is executed to collect multi-angle multi-wavelength images of the surfaces of three types of PE films. The PE film surface is irradiated from different angles by adopting 16 LED light sources with the wavelength of 550nm, three groups of dark field illumination modules with the central wavelengths of 450nm, 550nm and 650nm are sequentially activated, images are acquired from different angles by using three cameras with the resolution of 1920 multiplied by 1080 pixels, and the exposure time is adaptively adjusted within the range of 5-15ms according to the type of the PE film.

Then, step S2 is performed to construct a PE membrane surface manifold representation. For each type of PE film, determining a 10 neighbor point set of each pixel point, calculating similarity, constructing a local geometric structure, constructing a global adjacency graph based on local features, optimizing manifold embedding objective functions to generate low-dimensional representation, calculating manifold entropy and contrast measure, determining an optimal segmentation threshold, and generating a background suppression image. For the surface characteristics of different PE films, the self-adaptive adjustment parameters are Gaussian nuclear bandwidth parameters of common PE filmsThe gray standard deviation of the image is set to be 0.3 times, the antistatic PE film is set to be 0.4 times, and the optical PE film is set to be 0.25 times.

Next, step S3 is performed to extract the differential geometric feature. Converting the background inhibition image into a height function, calculating a first derivative and a second derivative, further calculating differential geometric characteristics, repeatedly calculating under three scales to generate multi-scale representation, defining a Riemann metric tensor based on the characteristics, calculating geodesic distance, estimating a probability density function, calculating the probability that each point belongs to a noise point or dust, and generating a noise point-dust probability map. For the surface characteristics of different PE films, the self-adaptive adjustment parameters are weight parameters of common PE filmsAndSet to 0.2 and 0.8, respectively, antistatic PE films set to 0.25 and 0.75, and optical grade PE films set to 0.15 and 0.85.

Then, step S4 is executed to perform topology preserving clustering. The method comprises the steps of applying Gaussian smoothing to a probability map, calculating a Laplacian, detecting zero crossing points, extracting an edge feature map, executing morphological operation to generate a morphological enhancement feature map, constructing a high-dimensional feature vector, defining a topology maintenance objective function, clustering, and classifying pixel points into four types. Aiming at the surface characteristics of different PE films, the self-adaptive adjustment parameters are that the standard deviation of Gaussian nuclei of the common PE filmSet to 2.0 pixels, antistatic PE film set to 2.2 pixels, optical grade PE film set to 1.8 pixels.

Finally, step S5 is executed to fuse the multi-level features. The reliability of each characteristic layer is calculated, self-adaptive weight is distributed, a nonlinear combination function is designed, multi-level characteristics are integrated, and final dust position and characteristic information are generated. For the detected motes, information such as position coordinates, size, shape, reliability score and the like is recorded.

In the embodiment, the method successfully adapts to the surface characteristics of three different types of PE films, and detection results show that the detection rate of a common PE film is 85.7% in the range of 0.5-2 mu m, 98.2% in the range of 2-10 mu m, 84.1% in the range of 0.5-2 mu m, 97.8% in the range of 2-10 mu m and 87.9% in the range of 0.5-2 mu m, and 99.3% in the range of 2-10 mu m. The false alarm rate of the three types of PE films is controlled below 3%, the processing speed is in the range of 12-15 frames/second, and the real-time detection requirement is met.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (8)

1. PE film surface static dust identification system, characterized by comprising:

   the image acquisition module is used for acquiring multi-angle and multi-wavelength images of the PE film surface;

   The multi-manifold mapping background suppression module is connected with the image acquisition module and is used for receiving the multi-angle multi-wavelength image, constructing PE film surface manifold representation, separating PE film background and potential dust features and generating a background suppression image;

   The differential geometric feature extraction module is connected with the multi-manifold mapping background suppression module and is used for receiving the background suppression image, extracting differential geometric features, calculating a geodesic distance matrix and generating a noise-dust probability map;

   The topological feature segmentation module is connected with the differential geometric feature extraction module and is used for receiving the noise-dust probability map, executing edge detection and morphological processing, carrying out topological retention clustering, dividing an image area into speckle noise, fuzzy noise, dust and background, and determining the position and feature information of the dust on the surface of the PE film;

   the multi-manifold mapping background suppression module comprises:

   The local structure holding unit is used for constructing a local geometric structure of the image and generating a local geometric feature map;

   The global structure mapping unit is connected with the local structure maintaining unit and is used for receiving the local geometric feature map, constructing a global similarity map, optimizing manifold embedding objective functions and generating low-dimensional representations of PE films and dust particles;

   The self-adaptive threshold value adjusting unit is connected with the global structure mapping unit and is used for receiving the low-dimensional representation, calculating manifold entropy, determining an optimal segmentation threshold value and generating a background suppression image;

   the differential geometry extraction module comprises:

   the differential geometric feature extractor is used for converting the image into a height function, calculating Gaussian curvature, average curvature and shape index and generating a differential geometric feature map;

   the geodesic distance calculation unit is connected with the differential geometric feature extractor and is used for receiving the differential geometric feature graph, defining a feature-based Riemann metric tensor, calculating the geodesic distance by using a fast travelling method and constructing a distance matrix;

   And the manifold probability distribution estimator is connected with the geodesic distance calculation unit and is used for receiving the distance matrix and the differential geometric feature map, estimating the probability density of the feature space, calculating the probability that each point belongs to a noise point or dust and generating a noise point-dust probability map.
2. 2. The PE film surface electrostatic mote identification system of claim 1, wherein the image acquisition module comprises:

   An annular LED array for illuminating the PE film surface from multiple angles;

   the multi-wavelength illumination system comprises three groups of dark field illumination modules with different wavelengths, which are used for sequentially activating and generating illumination light rays with different wavelengths;

   and the plurality of photoelectric imaging detectors are respectively connected with the multi-wavelength illumination system and are used for acquiring PE film surface images from different angles.
3. 3. The PE film surface electrostatic mote identification system of claim 1, wherein the topological feature segmentation module comprises:

   the edge detector is used for calculating a Gaussian Laplace operator, detecting zero crossing points and extracting an edge feature map;

   a morphological processing unit, connected to the edge detector, for receiving the edge feature map, performing multi-scale morphological dilation and erosion operations, and generating a morphological enhancement feature map;

   The topology maintenance clustering device is connected with the morphology processing unit and is used for receiving the morphology enhancement feature map, constructing a high-dimensional feature vector, executing topology maintenance clustering and clustering pixel points into speckle noise points, fuzzy noise points, tiny dust and background;

   The feature space fusion device is connected with the topology maintenance clustering device and is used for receiving multi-level features and clustering results, calculating feature weights, performing nonlinear feature fusion and generating final dust position and feature information.
4. 4. The PE film surface electrostatic mote identification system according to claim 1, wherein the local structure holding unit constructs the local geometry by determining a k-nearest neighbor set of each pixel, calculating a similarity between the pixel and the nearest neighbor, constructing a local similarity matrix, and extracting feature vectors to represent the local geometry.
5. 5. The PE film surface electrostatic mote identification system of claim 1, wherein the derivative geometry extracted by the derivative geometry extractor comprises a principal curvature, a Gaussian curvature, an average curvature, a shape index, and a curvature rate of change, and wherein the multi-scale curvature representation is generated in a plurality of scale spaces.
6. 6. The PE film surface electrostatic mote recognition system according to claim 3, wherein the topology preserving clustering device performs topology preserving clustering by constructing high-dimensional feature vectors including spatial, morphological and statistical features, defining a topology preserving objective function, preserving a local neighborhood structure through gradient descent iterative optimization, automatically determining an optimal number of classes, and classifying pixels into four classes.
7. 7. The PE film surface electrostatic mote identification system of claim 3, wherein the feature space fusion engine performs the nonlinear feature fusion by calculating the reliability of each feature layer, assigning adaptive weights, designing a nonlinear combining function, integrating background suppression layer features, noise detection layer features, and morphological processing features, and generating a final mote representation.
8. A method for identifying electrostatic dust on the surface of a PE film, which adopts the system for identifying electrostatic dust on the surface of the PE film according to any one of claims 1 to 7, and is characterized by comprising the following steps:

   collecting multi-angle multi-wavelength images of the surface of the PE film;

   constructing PE film surface manifold representation, separating PE film background and potential dust feature, and generating background inhibition image;

   extracting differential geometric features of the background suppression image, calculating a geodesic distance matrix, and generating a noise-dust probability map;

   performing edge detection and morphological processing on the noise-dust probability map, performing topology maintenance clustering, and dividing an image area into speckle noise, fuzzy noise, dust and background;

   and (5) fusing multi-level characteristics, and determining the position and characteristic information of the dust on the surface of the PE film.