A method and system for quality inspection of liquid coating products based on image recognition

Abstract

This invention discloses a method and system for quality inspection of liquid coating products based on image recognition, belonging to the field of image processing technology. The method includes acquiring a geometric topological model and extracting the interface curve; dynamically configuring the light source and adjusting the shooting direction according to the material reflectivity difference to acquire composite feature images; executing parallel analysis paths for the substrate, silicone parts, and interface bonding; the interface bonding path projects the interface curve onto the image and calculates the grayscale gradient modulus to quantify the bonding degree; inputting the indicators of each path into a multivariate nonlinear comprehensive quality evaluation function, and outputting the judgment result. This invention achieves simultaneous integrated detection of the substrate, silicone parts, and bonding interface for the first time. Adaptive imaging effectively eliminates interference from reflectivity differences, and the three-path parallel architecture meets online cycle time requirements. Combined with a graded evaluation of delamination degree and an adaptive threshold update mechanism, it possesses high detection rate, strong robustness, and good industrial adaptability.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically a method and system for quality inspection of liquid coating products based on image recognition. Background Technology

[0002] Liquid silicone overmolding technology integrates liquid silicone parts with metal or plastic materials through injection molding or compression molding, and is widely used in electronic seals, medical devices, automotive parts, and other fields. The core challenge of this process lies in the fact that under high temperature and pressure, the material itself may undergo radial compression deformation or surface heat melt damage, while the silicone part is prone to molding defects such as bubbles, insufficient adhesive, burrs, and cracks. Furthermore, the interface between the material and the silicone is susceptible to delamination or microscopic voids due to improper process parameters. Failure to detect any of these three types of defects can lead to product failure during use.

[0003] Existing vision-based automated inspection solutions mostly follow the inspection model for pure silicone parts, that is, after acquiring product appearance images through industrial cameras, algorithmic recognition is only performed on the image content of the silicone part area. This type of solution has three inherent drawbacks: First, the materials undergo the same thermal loads during the processing, and their geometric accuracy and surface integrity also need to be inspected, but existing systems completely ignore the material quality dimension; Second, there are significant differences between the materials and silicone parts in terms of material, color, and reflectivity. For example, the reflectivity of metallic materials can usually reach above 0.8, while the reflectivity of translucent silicone is often below 0.2. Fixed light sources and fixed camera layouts cannot simultaneously provide optimal imaging conditions for both materials, making it difficult to identify minute defects in low-contrast areas; Third, there is a lack of any quantitative detection methods for the bonding degree between the materials and silicone parts. Delamination and gaps, as latent defects, cannot be detected through conventional surface image analysis, becoming a major blind spot in quality control.

[0004] Therefore, there is an urgent need for a multi-dimensional detection technology that can adapt to the optical properties of heterogeneous materials and simultaneously cover the geometric morphology of the material, surface defects of silicone, and the quality of interfacial bonding. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for quality inspection of liquid coating products based on image recognition, so as to solve the problems of the prior art mentioned in the background art, such as the single detection dimension, inability to compensate for the difference in reflectivity of heterogeneous materials, and lack of quantitative evaluation of the degree of interface bonding.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-dimensional quality inspection method for liquid silicone overmolding products includes the following steps: Geometric modeling steps: Obtain the geometric topology model of the product to be inspected, and extract the 3D interface curves from the geometric topology model. Interface curve Defined as the contact boundary line between the outer surface of the product material and the inner surface of the silicone part; Adaptive imaging steps: Based on the difference in reflectivity between the substrate and the silicone part, the light source wavelength and intensity of the imaging system are dynamically configured, and the interface curve is used as a reference. Adjust the camera's shooting direction to acquire at least one composite feature image containing the material area, the silicone area, and the area where the two meet; Multi-path analysis steps: At least three analysis paths are executed in parallel on the composite feature image, including: Material quality analysis path, used to detect geometric deformation and surface damage of materials; The silicone part quality analysis path is used to detect bubbles and burr defects in silicone parts. The interface, combined with the quality analysis path, is used to display the interface curve. Projection curves are obtained by projecting onto the composite feature image. And calculate along the projection curve The image grayscale gradient magnitude is used to quantify the degree of interface bonding. Comprehensive scoring steps: Obtain the output indicators of the material quality analysis path, silicone part quality analysis path, and interface combination quality analysis path, and input them into a multivariate nonlinear comprehensive quality evaluation function to calculate a comprehensive quality score. Based on the comparison result of the comprehensive quality score and the preset qualified threshold, the product quality judgment result is output.

[0007] According to the above technical solution, the adaptive imaging step, which dynamically configures the light source wavelength and light intensity of the imaging system, further includes: Based on the reflectivity of the material Reflectivity of silicone parts The ratio of the wavelengths is used to select a light source with the corresponding wavelength. The output light intensity of the light source is dynamically configured based on the light intensity compensation function, as shown in the following formula: in, This is the contrast gain coefficient. Based on the light intensity.

[0008] According to the above technical solution, in the geometric modeling step, the three-dimensional interface curves are extracted. Further includes: For products of rotation, radial sections with equal angular intervals are selected on the outer surface of the material. The intersection points of each section with the outer surface of the material and the inner surface of the silicone part are obtained. All intersection points are connected and fitted in angular order to obtain the interface curve. ;or, For non-rotational products, interface points are extracted on the outer surface of the material using a constant arc length sampling method, and then connected and fitted to obtain the interface curve. .

[0009] Based on the above technical solution, the multivariate nonlinear comprehensive quality evaluation function in the comprehensive scoring step is: in, To calculate the overall quality score, These are the weighting coefficients for the quality of the material, silicone, and interface bonding, respectively. Sub-ratings for material quality, silicone quality, and interface quality. Through the Sigmoid function The calculation yielded that, The interface integration index. The threshold for determining interface debonding. This is the slope parameter.

[0010] According to the above technical solution, in the interface integration quality analysis path, quantifying the degree of interface integration further includes: Interface integration index The calculation formula is: in, This represents the total length of the projected curve. For the image at points grayscale gradient magnitude at that location The derivative of the arc length of the curve; Furthermore, by employing a multi-view fusion strategy, the interface integration index of images acquired from multiple viewpoints is calculated for each image. The results are then weighted and averaged to obtain the final interface bonding index. .

[0011] Based on the above technical solution, in the interface bonding quality analysis path, the threshold for determining interface debonding is... The method combines statistical analysis with destructive testing to determine: A batch of product samples confirmed to have good bonding through destructive testing were taken, and the interfacial bonding index of each sample was calculated under the same imaging conditions. This yields the sample set; Calculate the mean of the sample set. and standard deviation ; Set threshold .

[0012] According to the above technical solution, in the adaptive imaging step, based on the boundary interface curve... Adjusting the camera's shooting direction further includes: The boundary interface curve Discretize the material into multiple equally spaced points, and calculate the unit outward normal vector of the outer surface of the material or the inner surface of the silicone at each point. ; Adjust the camera's orientation so that its optical axis aligns with the unit outward normal vector. parallel.

[0013] According to the above technical solution, the detection of geometric deformation of the material in the material quality analysis path includes: The edge contours of the material are extracted from the composite feature image, and ellipse fitting is performed on the edge contours to obtain the actual major axis radius of the material. and minor axis radius ; Calculate the equivalent radius based on the actual major and minor axis radii. ; Calculate radial deformation ,in The standard radius is extracted from the geometric topology model; When radial deformation If the value exceeds the preset threshold, it is determined to be a deformation defect.

[0014] According to the above technical solution, the quality analysis path for silicone parts includes detecting bubbles and burrs, which are defects in silicone parts. For bubble defects, an adaptive threshold segmentation algorithm is used, which segments the silicone area based on the average grayscale value of the local window. and standard deviation Calculate the local threshold: in The empirical coefficients are used to determine the area of ​​the segmented connected components; For burr defects, the high-frequency components of the silicone edge are extracted using the difference of Gaussian operator, and a defect evaluation factor is defined: in It is the set of pixels in the neighborhood of the silicone edge. The response image is generated by convolving the image with the difference of Gaussian operator, and the defect evaluation factor is included. Thresholds obtained from statistical calibration Compare and determine whether there are burrs or cracks.

[0015] A multi-dimensional quality inspection system for liquid silicone overmolding products, the system comprising: The geometric modeling unit is used to obtain the geometric topology model of the product and extract the 3D interface curves. The adaptive imaging unit includes a programmable multispectral light source, multiple independently adjustable cameras, and a central control unit. The central control unit is used to configure the light source parameters according to the material reflectivity difference between the material and the silicone part, and to control the shooting direction of each camera according to the interface curve. The image analysis unit is configured with a multi-threaded parallel architecture to perform material quality analysis, silicone part quality analysis, and interface integration quality analysis in parallel. The comprehensive scoring unit receives the output from the image analysis unit, calculates the comprehensive quality score, and outputs the judgment result.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention fills a technological gap in the quantitative assessment of heterogeneous bonding by achieving simultaneous and integrated detection of the substrate, silicone parts, and their interface in liquid-coated products for the first time. Experiments have shown that it has an extremely high detection rate for debonding defects. At the same time, the proposed adaptive imaging configuration model (including a light intensity compensation function based on reflectivity ratio and a camera number function based on the product's outer diameter) effectively solves the imaging interference caused by the difference in reflectivity of heterogeneous materials and significantly improves the contrast of the interface area.

[0017] Based on this, the three-path parallel processing architecture makes full use of the multi-core processor capabilities to control the total time of a single inspection within an extremely short range, meeting the requirements of industrial online inspection cycle time; the constructed multivariate nonlinear comprehensive quality evaluation function integrates multiple heterogeneous inspection indicators into a single score, and supports custom weights and dynamic threshold adjustment, demonstrating good industrial adaptability and scalability.

[0018] Furthermore, statistical analysis of the interface gradient modulus enabled a graded assessment of the degree of degumming, providing a quantitative feedback basis for process optimization. The introduced two-level defect judgment logic and online adaptive threshold update mechanism further enhanced the robustness and long-term stability of the detection system. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the detection method of the present invention; Figure 2 This is a schematic diagram of the composition framework of the detection system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 like Figure 1 As shown, the detection method provided by this invention sequentially includes a geometric modeling and feature extraction stage, an adaptive imaging configuration stage, a multi-path parallel image analysis stage, and a comprehensive quality scoring stage. Each stage is described in detail below.

[0022] Step 1, Geometric Modeling and Feature Extraction Stage: This invention first obtains the geometric topological model of the liquid silicone-coated product to be tested. This model can be obtained by scanning the entire surface of a standard qualified product using a structured light 3D scanner, or it can be directly extracted from the product's computer-aided design document. For the point cloud data obtained by scanning, denoising, registration, and meshing processing are required to generate a closed triangular mesh model. For the CAD model, its boundary representation (B-Rep) structure is directly parsed.

[0023] Spatial analysis was performed on the geometric topology model to extract three key 3D geometric feature parameters: the centroid coordinates of the material, the thickness of the silicone part's covering, and the interface curve between the material and the silicone part. The centroid coordinates of the material are defined as the geometric center of all voxel points in the material portion, and the calculation formula is as follows: in, The coordinates of the centroid of the material (3D vector, unit: mm). Material volume (unit: mm) 3 ); Integration region Indicates the three-dimensional spatial area occupied by the material; This is the spatial position vector of the voxel point (unit: mm).

[0024] The silicone coating thickness is defined as the distance from the outer surface of the material along the normal direction to the outer surface of the silicone part. Due to the complexity of the product shape, this thickness may vary in different spatial locations. This invention uses the statistical average value along the circumference of the product as a characteristic parameter, and records the standard deviation of the thickness distribution to evaluate the uniformity of the process.

[0025] The interface curve is the core geometric feature of this invention. In three-dimensional space, this curve is defined as the contact boundary line between the outer surface of the material and the inner surface of the silicone part. For products of rotation (such as cylindrical materials coated with silicone), this interface curve appears as one or more closed spatial curves around the circumference of the product. The extraction method is as follows: a series of radial sections with equal angular intervals are selected on the outer surface of the material. Each section intersects the outer surface of the material to obtain a contour line. The point where this contour line intersects the inner surface of the silicone part is the interface point. All interface points are connected in angular order and fitted with cubic B-splines to obtain a smooth three-dimensional closed curve. For non-rotating products (such as rectangular or irregularly shaped materials), the interface curve is extracted using an equal arc length sampling method. The extracted interface curve is denoted as... This is one or more closed curves in three-dimensional space. In subsequent image processing, this curve will be projected onto a two-dimensional image plane according to the camera's perspective projection model, serving as a geometric constraint for interface bonding analysis.

[0026] Step 2, Adaptive Imaging Configuration Stage: In response to the significant difference in surface reflectivity between the material and the silicone part, this invention designs a dynamically configurable imaging system. This system includes a programmable multispectral light source (containing red, green, blue and white LED arrays, with independent adjustment of the light intensity of each channel), multiple independently oriented industrial cameras (each camera is equipped with an electric gimbal, which can achieve attitude adjustment with two degrees of freedom: pitch and yaw), and a central control unit.

[0027] The imaging configuration parameters are calculated based on the material properties of the substrate and silicone. Let the surface reflectance of the substrate be... The surface reflectivity of silicone parts is Both can be obtained through offline measurement with a spectrophotometer. During measurement, a narrow-band filter corresponding to the wavelength of the light source is used to eliminate ambient light interference. When the material is a highly reflective metal ( And the silicone parts are made of a semi-transparent material. When the material is a dark-colored plastic with low reflectivity, this invention selects a red light source with a wavelength range of 620nm to 650nm. The physical basis for choosing red light is that longer wavelength light has a lower scattering attenuation coefficient in silicone (Rayleigh scattering intensity is inversely proportional to the fourth power of the wavelength), which can effectively penetrate the silicone surface and reduce surface specular reflection interference. At the same time, the metal has a moderate reflectivity for red light and is not prone to saturation. When the silicone is light-colored, blue light (center wavelength 450nm) is selected to improve texture contrast. For materials with similar reflectivity to silicone, white light is used to preserve full-spectrum information.

[0028] The light intensity output by the light source is not a fixed value, but is dynamically compensated based on the reflectivity ratio of the material and the silicone. The light intensity compensation function is as follows: in, Light intensity output by the light source (unit: lux); This is the contrast gain coefficient, which is dimensionless and ranges from 0.5 to 2.0. The surface reflectance of the material is dimensionless. The surface reflectance of the silicone part is dimensionless. The basic light intensity (unit: lux) is calibrated as follows: Place a standard white board (reflectivity 0.99) at the testing station, adjust the light source output so that the grayscale response of the camera sensor reaches 80% of the saturation value (255 for an 8-bit camera), and record the light intensity value at this point as the baseline light intensity. .

[0029] If calculated according to the above formula Exceeding the physical maximum output light intensity of the light source (Unit: lux), then the actual output is taken as... and correspondingly reduce To ensure the dynamic range of the image, the exposure time can be increased or decreased. Specifically, the exposure time can be reduced first. To the point ,like If the exposure time is reduced to the lower limit of 0.5 and still cannot meet the requirements, the exposure time will be automatically extended to the original value. The total exposure is kept constant while the image quality is increased by a factor of 100. To further improve image quality, this invention also introduces real-time feedback closed-loop control: after acquiring the first frame image, the average grayscale values ​​of the material area and the silicone area are calculated respectively. and (Unit: grayscale levels, value range 0 to 255 in 8-bit image), if If the value exceeds a preset threshold (e.g., 30 gray levels), the value will be automatically adjusted according to the direction of the deviation until the two are close.

[0030] Regarding the camera layout, let the product's outer diameter be... (Unit: mm), working distance is (Distance from the front surface of the camera lens to the product surface, unit: mm), lens field of view is Then the field of view of a single lens is: in, The field of view of a single lens (unit: mm).

[0031] To ensure complete coverage of the product's entire circumferential surface and to prevent feature matching redundancy between adjacent images, an image overlap rate is introduced. Its value ranges from 0.05 to 0.25. The choice of overlap rate needs to balance detection accuracy and system cost: the higher the overlap rate, the smaller the detection blind zone, but the more cameras are required; the lower the overlap rate, the lower the system cost, but edge features may be lost. The minimum number of cameras required is determined by the following formula: in, The minimum number of cameras required (integer); This represents the floor function. In practical systems, redundant cameras are often used to eliminate detection blind spots, for example, when the calculated value... When the value is 2, the actual installation One camera ( The cameras are evenly distributed around the circumference of the product, with an angle between adjacent cameras. For non-rotational products (such as rectangular cross-sections), the above formula is no longer applicable. In this case, the perimeter of the product's maximum cross-section must be considered. (Unit: mm) Substitution Furthermore, the camera layout needs to consider the normal direction of each face, with at least one camera facing each face. The shooting direction of each camera is adjusted to point to the normal direction of the interface curve at the corresponding angle. The specific implementation method is as follows: the 3D interface curve... Discretized At each of the three equally spaced points, calculate the unit outward normal vector of the outer surface of the material (or the inner surface of the silicone), denoted as . Then, the camera's orientation is adjusted using a motorized gimbal so that its optical axis is aligned with... Parallelism. This configuration ensures that the area near the interface curve is free of perspective distortion in the image, keeping the pixel-to-spatial-size correspondence constant in subsequent gradient analysis.

[0032] Step 3, Multi-path Parallel Image Analysis Stage: After completing the above adaptive imaging configuration, the system synchronously triggers all cameras to acquire images, obtaining composite feature images from multiple perspectives. Each image contains the material region, the silicone region, and the interface region between the two. To improve detection efficiency, this invention adopts a multi-threaded parallel architecture, independently allocating three analysis paths to each image acquired by each camera. Each path executes simultaneously in different threads, with no data dependency between them. The following explanation uses the processing flow of a single image as an example. During image analysis, image pixel coordinates are used uniformly. (Unit: pixels) represents the position on the image plane, to distinguish it from three-dimensional world coordinates. .

[0033] The material quality analysis path first extracts the edge contours of the material from the image. Due to the significant grayscale difference between the material and the silicone (material is typically brighter, silicone is darker), the Canny edge detection operator is used to obtain an initial set of edge points. The high and low thresholds of the Canny operator are determined using an adaptive method: the high threshold is set to the bottom 20% quantile (i.e., the 80th quantile) of the image's grayscale histogram, and the low threshold is set to 0.4 times the high threshold. Subsequently, morphological closing operations (using a 5×5 pixel circular kernel as the structuring element) are used to connect broken edges, and isolated noise points with areas smaller than the preset threshold are removed. Ellipse fitting is then performed on the processed closed contours using an algebraic distance method based on least squares, while the RANSAC algorithm is used to remove outliers to improve the robustness of the fitting. The ellipse fitting result gives the actual major axis radius of the material. (Unit: mm) and minor axis radius (Unit: mm). For cylindrical products, take the equivalent radius: in, The equivalent radius of the material (unit: mm). Radial deformation is defined as: in, Radial deformation (unit: mm); The standard radius (extracted from the geometric topology model). When Defects exceeding a preset detection threshold (e.g., 0.05 mm) are marked as deformation defects. Furthermore, the residual standard deviation obtained through ellipse fitting can assess the local unevenness of the material's edges. This residual standard deviation is calculated in pixel coordinates and converted to millimeters using the spatial resolution (mm / pixel) obtained through camera calibration. Residuals exceeding 0.02 mm are marked as local indentations.

[0034] Meanwhile, surface material damage detection of the material was performed using the gray-level co-occurrence matrix method. Within the extracted material area, a step size was used... Pixels, orientation Gray-level co-occurrence matrices are constructed in four directions: 0°, 45°, 90°, and 135°. The average of these four matrices is then used as the final GLCM. (Gray-level number) The image is compressed to 64 levels to reduce computational load. Contrast eigenvalues ​​are calculated from the GLCM as shown in the following formula: in, These are contrast characteristic values, dimensionless; The grayscale level is 64 (here). Gray levels in the gray-level co-occurrence matrix and The normalized probability that occurs simultaneously is dimensionless. Gray-level index (value range 1 to...) ).

[0035] Energy eigenvalues: in, It is the energy characteristic value (second moment of the angle), which is dimensionless.

[0036] Correlation eigenvalues: in The value is a correlation characteristic, dimensionless. Let be the row marginal probability of the gray-level co-occurrence matrix; Let the marginal probability be columnar; Contrast is most sensitive to surface roughness; heat melting or abrasion can cause... The value increased significantly. The area to be tested... The value is compared with a standard template (the average value pre-calculated from a batch of qualified products), defining a relative offset: in, This represents the relative contrast offset (in %). The contrast characteristic value of the area to be tested; This is the average contrast value of the standard template. If =0, then directly determine =0 (no contrast change). When If the percentage exceeds 20%, it is determined that there is a hot melt or wear defect. If the ASM decreases while the CORR increases, it is further confirmed that it is a surface damage rather than an artifact caused by changes in lighting.

[0037] The quality analysis path for silicone parts distinguishes between two types of defects: regional defects (bubbles, impurities) and edge defects (burrs, cracks). For bubbles and impurities, an adaptive threshold segmentation algorithm is used. First, Otsu's method is used to perform global threshold segmentation on the entire image to coarsely separate the silicone region from the background. Then, within the silicone region, a local window is taken centered on each pixel, with the window size... Dynamically set based on typical defect dimensions of silicone parts: in Take the minimum acceptable defect diameter specified in the product specifications (unit: pixels, obtained from physical dimensions through camera calibration). Calculate the average grayscale value within the window. and standard deviation The local threshold is set as follows: in, Local threshold (grayscale level); The average gray level (grayscale value) within a local window. The standard deviation of gray levels within a local window; This is an empirical coefficient, ranging from 2 to 3. This threshold can accommodate uneven lighting caused by variations in thickness or curvature of the silicone parts. After binarization, an eight-connected-domain labeling algorithm is used to count the pixel area of ​​each connected domain, and the pixel area is converted into physical area using camera calibration parameters. This invention introduces a two-level judgment logic: the first level is the lower limit for defect detection. (Unit: mm) 2 For example, 0.05mm 2 Connected components exceeding this value are marked as potential defects and their areas are recorded; the second level is the upper limit for direct rejection. (Unit: mm) 2 For example, 0.2mm 2 When the area of ​​the potential defect exceeds The product is immediately deemed unqualified. For areas between... and The defects between them are accumulated and added to the total defect area in the comprehensive score. (Unit: mm) 2 However, it does not trigger an immediate rejection. For areas exceeding... For the connected components, further calculate their roundness factor: 4π·area / perimeter. 2 Those with a roundness close to 1 are bubbles, while those with a smaller roundness are irregular impurities.

[0038] For burrs and cracks, the difference of Gaussian (DG) operator is used to extract the high-frequency components of the silicone edge. First, the silicone region is segmented from the image. Morphological erosion is used to obtain the internal kernel of the silicone region, which is then subtracted from the original region to obtain the edge band. The DG operator is defined as the difference between two Gaussian kernels of different scales: in It is a Gaussian difference function; Image pixel coordinates (unit: pixels); Let be the scale parameter of the Gaussian kernel (unit: pixels), and ; The two-dimensional Gaussian kernel function is defined as follows: in, It is a natural exponential function. The choice of the scale parameter is related to the expected defect size: Take 0.5 pixels (corresponding to sub-pixel level edge). Take 2.0 pixels (to suppress texture noise). [Image] (Grayscale function) and Convolution yields the response image To focus on defects in the silicone edge region, the aforementioned edge band is expanded outward by 5 pixels to obtain the edge neighborhood set. The defect evaluation factor is defined as follows: in, As a defect evaluation factor, it is dimensionless and quantifies the severity of edge burrs; This is the set of pixels in the edge's neighborhood. Threshold. The calibration is performed using a statistical method: for a batch of qualified products that have been manually confirmed to be defect-free, their... mean of values and standard deviation ,set up: in, The threshold value is dimensionless. This threshold corresponds to the upper limit of the 99.7% confidence interval. When Exceed At that time, the location of burrs or cracks is marked, and the polar coordinates (angle, radial distance) of the defects are output for subsequent repair or sorting.

[0039] The interface bonding quality analysis approach is the core innovation of this invention, distinguishing it from existing technologies. Its physical principle lies in the following: When the material and silicone bond well, there are no air gaps between them, and the refractive index at the interface exhibits a continuous or gradual transition (silicone has a refractive index of approximately 1.4, while the metal material has a higher refractive index but is opaque; the interface primarily considers the optical contact between the silicone and the material surface). In this case, under transmitted or reflected illumination, the grayscale change near the interface is smooth, and the grayscale gradient modulus is small. When debonding or voids exist, an air layer (refractive index approximately 1.0) enters the interface, producing significant Fresnel reflection or total internal reflection effects, causing a drastic jump in grayscale at the interface and a significant increase in the grayscale gradient modulus. Therefore, the degree of bonding is negatively correlated with the interface gradient modulus: the worse the bonding, the larger the gradient.

[0040] This path first involves the three-dimensional interface curve. Projected onto a two-dimensional image plane. The projection process depends on the camera's intrinsic parameter matrix. (3×3 matrix) and extrinsic parameter matrix (a 3×4 matrix, where) It is a 3×3 rotation matrix. (These are 3×1 translation vectors), and these parameters are obtained in advance using the Zhang Zhengyou calibration method. For each discrete point on the three-dimensional curve... Its projection point Satisfies the projection equation: in, This is the intrinsic parameter matrix. It is a rotation matrix; It is a translation vector. is a non-zero scale factor (dimensionless), used as a normalization parameter when transforming three-dimensional world coordinates to two-dimensional homogeneous coordinates of the image plane; The image pixel coordinates of the projection point (unit: pixels); Here are the world coordinates (in mm) of the points in 3D space. Connect all projected points in their original order and generate a smooth projection curve using cubic spline interpolation. (here) (Represents the projection curve itself). Along The gradient vector of image gray levels is calculated point by point. The gradient calculation uses the Scharr operator (which has higher rotational symmetry than the Sobel operator). direction and The convolution kernels in the directions are as follows: in, They are respectively and Scharr operator convolution kernel in the direction.

[0041] The gradient magnitude is: in, For the image at points Gray-level gradient magnitude at a given location (unit: gray level / pixel); They are respectively and Partial derivative of the direction (unit: grayscale / pixel). Due to the projection curve It is continuous; in actual calculations, it is discretized into sampling points with equal arc length intervals (0.5 pixels apart), and numerical integration is performed using the compound Simpson's rule or the Gauss-Legend quadrature method to improve accuracy. The interface bonding index is defined as the average value of the gradient magnitude along the curve: in, Interface integration index (unit: grayscale / pixel); The total length of the projection curve (unit: pixels, which can be converted to millimeters through camera calibration). This is the differential of the curve arc length (unit: pixel); this index comprehensively reflects the degree of grayscale abrupt change across the entire interface area. To eliminate local noise interference caused by surface contamination or scratches, this invention also calculates the standard deviation of the gradient modulus. ,like If the value is too large (e.g., exceeding 50% of the average), it indicates a possible local anomaly, requiring manual review.

[0042] Critical threshold The calibration employed a combination of statistical and destructive testing methods. A batch of at least 30 product samples, confirmed to have good bonding through tensile testing or microscopic examination of sections, were selected, and the bonding strength of each sample was calculated under identical imaging conditions. Values, to obtain the sample set ,in Given the sample size (n≥30), calculate the mean of this sample set. and standard deviation : , Set threshold: in, The threshold for determining interface debonding (unit: grayscale / pixel) is used. This method is based on the normal distribution assumption, ensuring that the false positive rate for well-bonded samples is below 0.3%. When measured in actual tests... Greater than At that time, it was determined that there was a degumming or void defect.

[0043] Furthermore, according to The degree to which the delamination exceeds the threshold can be graded into levels of severity: mild delamination is... Moderate degumming is Severe degumming is This grading information can be used to guide the quantitative adjustment of process parameters.

[0044] To further enhance detection robustness, this invention also introduces a multi-view fusion strategy: [the strategy involves] collecting data from the same product... Each image is analyzed, and the values ​​in each image are calculated separately. The values ​​are then weighted and averaged to obtain the final interface integration index. in, The interface integration index after multi-view fusion (unit: grayscale / pixel). The weight for the j-th viewpoint is typically the projected length of the interface curve from that viewpoint. (Unit: pixels). This strategy can avoid misjudgments caused by occlusion or uneven lighting from a single viewpoint.

[0045] Step 4, Comprehensive Quality Scoring Stage: After analyzing the above three paths, the system obtained multiple detection indicators: radial deformation of the material. (Unit: mm) Texture contrast offset (Unit: %) Total area of ​​air bubbles and impurities (Unit: mm) 2 Only areas with areas within the lower limit of defect detection are included. Compared to the upper limit of direct invalidation The potential defects between them exceed Defects will result in immediate disqualification and will no longer be included in the scoring; total area of ​​the silicone region. (Unit: mm) 2 Burr Defect Evaluation Factor (Dimensionless), interface integration index (unit: grayscale / pixel), and preset upper limit parameters for each indicator. These indicators have different dimensions and physical meanings, making direct comparison or weighted summation impossible. Therefore, this invention constructs a multivariate nonlinear comprehensive quality evaluation function that maps all indicators to a unified score space. .

[0046] The evaluation function is designed according to the following principles: First, the scores of each dimension should monotonically decrease as the severity of the corresponding defect increases; second, when any indicator exceeds its tolerance limit, the score of that dimension should approach 0; third, multiple indicators within a dimension are coupled through a product to ensure that the score approaches 1 only when all sub-indicators are qualified; fourth, dimensions are fused using a weighted sum of squares and square root method to amplify the impact of severe defects. The complete form of the evaluation function is as follows: in, The overall quality score is dimensionless and ranges from [0,1]. The material quality weighting coefficient is dimensionless and satisfies 0 ≤ ≤1; The mass weighting coefficient for silicone is dimensionless and satisfies 0 ≤ ≤1; The interface incorporates a quality weight coefficient, which is dimensionless and satisfies 0 ≤ ≤1, and ; The measured radial deformation (unit: mm); This is the upper limit for radial deformation scoring (unit: mm). If the score exceeds this value, the score for the material will approach 0. The measured relative offset of contrast (unit: %) This is the upper limit for contrast offset scoring (unit: %). If the score of the material exceeds this value, the score will approach 0. is the total area of bubble impurities (unit: mm²); is the total area of the silica gel region (unit: mm²); is the maximum proportion of the allowable defect area to the total area of the silica gel, dimensionless; is the measured burr defect evaluation factor, dimensionless; is the upper limit of the burr evaluation factor score, dimensionless; is the measured interface bonding degree index (unit: gray level / pixel); is the interface delamination determination threshold (unit: gray level / pixel); is the slope parameter of the Sigmoid function, dimensionless, with a typical value of 0.2; is the minimum value function; is the natural exponential function.

[0047] The material score item and the silica gel score item both adopt the product form. Its physical meaning is that the two types of defects, deformation and surface damage, are independent of each other. Excessive defects of any type will render the whole material unusable. Similarly, the two types of defects, bubbles and burrs, are also independent. Even if there are no bubbles in the product but there are serious burrs, it should be judged as unqualified. The interface score item adopts the Sigmoid function , and its advantages are as follows: when is much smaller than , ; when is much larger than , ; the score changes continuously near the threshold, avoiding the boundary sensitivity caused by hard threshold determination, and at the same time retaining the quantization information of the delamination degree.

[0048] It should be noted especially that the direct rejection threshold (for example the direct rejection threshold is 0.05 mm, the rejection threshold is 0.2 mm²) and the upper limit parameter of the score ( = 0.20 mm, = 0.05) can be different: the former is used to quickly eliminate products with serious defects, and the latter is used for continuous score quantization. The two together constitute a two-level quality evaluation system.

[0049] Finally, compare the calculated comprehensive quality score with the preset qualified threshold ; is a dimensionless constant, The typical value range of is from 0.7 to 0.85. If , the product is judged as qualified; otherwise, it is judged as unqualified. This threshold can be adjusted dynamically according to the requirements of the production line for the missed inspection rate and the false inspection rate: when it is desired to reduce the missed inspection rate (even if it misjudges qualified products), reduce When the goal is to reduce the false positive rate (preferring to miss defective products), improve... In actual production, ROC curve analysis is usually used to determine the optimal threshold, which minimizes the weighted sum of the false negative rate and the false positive rate.

[0050] In addition to providing a binary pass / fail assessment, this invention also offers detailed reports on scores for each dimension, including material quality scores, silicone quality scores, interface quality scores, and specific numerical and location information for each defect indicator. This data can be uploaded to a Manufacturing Execution System (MES) in real time for process traceability and quality statistical analysis. For example, when the mean interface score continues to decline over a period of time, it suggests that the injection pressure or mold temperature may need adjustment; when the variance of the material deformation score increases, it suggests that the mold closing mechanism may be worn.

[0051] To adapt to process drift during production, this invention also introduces an online adaptive update mechanism: the system accumulates detections... After processing 500 products, recalculate the number of qualified products in the current batch. CON The statistical parameters of indicators are determined, and the thresholds are updated using the Exponentially Weighted Moving Average (EWMA) method. The formula is as follows: in, The updated threshold parameter (units depend on the specific metric); The threshold parameter currently in use; The threshold parameter is calculated based on the latest batch of qualified products; The forgetting factor is dimensionless (ranging from 0.7 to 0.9). This mechanism can automatically compensate for the effects of slow changes such as light source aging and lens contamination, maintaining detection consistency.

[0052] During calibration and testing, attention should be paid to The calibration and detection should use the same image spatial resolution (mm / pixel); otherwise, adjustments are required. A conversion is performed to ensure the consistency of the physical meaning of the threshold.

[0053] like Figure 2 As shown, this embodiment also provides a multi-dimensional quality inspection system for liquid silicone coating products, the system including: The geometric modeling unit is used to obtain the geometric topology model of the product and extract the 3D interface curves. The adaptive imaging unit includes a programmable multispectral light source, multiple independently adjustable cameras, and a central control unit. The central control unit is used to configure the light source parameters according to the material reflectivity difference between the material and the silicone part, and to control the shooting direction of each camera according to the interface curve. The image analysis unit is configured with a multi-threaded parallel architecture to perform material quality analysis, silicone part quality analysis, and interface integration quality analysis in parallel. The comprehensive scoring unit receives the output from the image analysis unit, calculates the comprehensive quality score, and outputs the judgment result.

[0054] Example 2 This embodiment uses a cylindrical product coated with liquid silicone as an example. The product consists of a 304 stainless steel cylinder as the inner core material, with a semi-transparent liquid silicone component coated on the outer periphery using an injection molding process. The hardness is 30 Shore A. The manufacturing process conditions are: injection temperature 130℃, injection pressure 65MPa, and holding time 8 seconds. Potential defects under these conditions include radial compression deformation of the material, bubbles in the silicone component, burrs, and interface delamination.

[0055] In the geometric modeling and feature extraction stage, a structured light 3D scanner (point cloud density 0.05mm) was used to scan the standard qualified products to obtain a geometric topological model. The centroid coordinates of the material were obtained through spatial analysis. =(0,0,0), average coating thickness of silicone parts =2.5mm, thickness standard deviation is 0.12mm. The three-dimensional interface curve is located at the radius Two independent circular curves at 25mm are located 2mm below the upper surface and 2mm above the lower surface of the product, respectively, with a circumference of 157.08. During the scanning process, point cloud data of 10 qualified products were collected simultaneously to establish a standard template library.

[0056] During the adaptive imaging configuration phase, the material reflectance was measured offline. =0.85 (measured using an integrating sphere spectrophotometer at a wavelength of 635 nm), silica gel reflectivity =0.12. According to the rules, a red light source is selected, with a center wavelength of... =635nm, bandwidth ±10nm. Base light intensity The calibration was set at 2000 lux, under the following conditions: camera ISO 400, exposure time 10ms, and a grayscale response of 200 on a standard white card (reflectivity 0.99). The contrast gain coefficient was then calculated. =1.35, substituting into the light intensity compensation function, we get In actual control, the LED array is driven by pulse width modulation (PWM) with a duty cycle set to 85%. The maximum grayscale value of the material area is monitored in real time and found to be 235 (unsaturated), therefore the current parameters are maintained. The camera layout parameters are: product outer diameter... =50mm, working distance =150mm, lens field of view =45°, calculate the field of view width ≈124.26mm. Take the image overlap rate. =0.20, substituting into the camera count function, we get To eliminate detection blind spots and improve system robustness, four cameras were actually installed, evenly distributed around the product's circumference, with one camera every 90°. The optical axis direction of each camera was adjusted via a motorized pan-tilt unit: for the camera at 0°, the normal vector of the interface curve in the 0° direction was calculated to be (1,0,0), and the pan-tilt unit was adjusted to make the optical axis parallel to this direction; the other cameras were adjusted in the same way.

[0057] During the image acquisition and multipath analysis phase, the system simultaneously triggers four cameras to acquire images, each with a resolution of 1280×1024 pixels and 256 gray levels. The following explanation uses the image processing results from the 0° camera as an example. The material quality analysis thread first uses Canny edge detection and ellipse fitting to measure the major axis radius of the material outline as 24.96mm, the minor axis radius as 24.88mm, and the equivalent radius... =24.92mm, standard radius =25.00mm, yielding the radial deformation. =0.08mm, exceeding the direct rejection threshold of 0.05mm, indicating a deformation defect. The standard deviation of the ellipse fitting residual is 0.008mm, not exceeding 0.02mm, indicating no local indentation. Gray-level co-occurrence matrix calculation ( =1, directional average, grayscale level 64) to obtain contrast. =85, Standard Template =72, relative offset ≈18.1%, which does not exceed the 20% threshold, therefore no heat-melting damage is determined; at the same time, it is calculated that... =0.023 (template 0.025), =0.87 (template 0.85), the change is not significant, further confirming no material damage. The silicone quality analysis thread uses adaptive threshold segmentation (window size 21×21 pixels). =2.5), two connected regions were detected within the silicone area, with areas of 152 pixels and 63 pixels respectively, which were converted to 0.12 mm after calibration. 2 and 0.05mm 2 Set a lower limit for defect detection. =0.05mm 2 The upper limit is directly invalidated. =0.2mm 2 Both defect areas are greater than the detection limit. Therefore, they are marked as potential defects; however, all are below the upper limit for direct rejection. Therefore, it is not directly deemed invalid, but its area is 0.17mm. 2 Added to the total defect area In the middle. Gaussian difference operator taking =0.8、 =1.6, calculate the defect evaluation factor after convolution. =1250. Threshold The calibration is based on 100 qualified products: =850, =83.3, ≈1100. If the threshold is exceeded, a burr defect is identified. The mark is located at the top edge of the product, at an angle of approximately 12°, and the burr width is approximately 0.15mm. The interface quality analysis thread projects the 3D interface curve onto the image plane to obtain the projected curve. (Approximately 520 pixels in length, corresponding to 157mm). Samples are taken along the curve at 0.5 pixel intervals. The gradient magnitude at each sampling point is calculated, and integrated using the compound Simpson's rule. =18.5 grayscale / pixel. Critical threshold calibration: Take 30 products with good bonding confirmed by slicing, and their... The values ​​were 8.2, 9.1, 7.8, ..., with a maximum value of 11.2 and a mean value of ... =9.5, standard deviation =1.2, take =13.1; simultaneously, multiplying the maximum value by the safety margin of 1.1 yields 12.32. Take the larger of the two values. =13.1. Actual measurement =18.5>13.1, indicating a delamination defect, and the proportion exceeding the threshold is 41%, classifying it as moderate delamination. (Four perspectives) The values ​​are obtained after weighted fusion. =17.9 grayscale / pixel.

[0058] In the comprehensive quality scoring stage, the above-mentioned testing indicators are substituted into the evaluation function. Parameter values: =0.20mm, =80%, =1200mm² (total area of ​​silicone area). =0.05, =3000 (take 3.5 times the average value of qualified products). =0.2, =13.1, weight =0.3、 =0.3、 =0.4, acceptable threshold =0.75. Calculate the material score: (1−0.08 / 0.20)=0.60, (1−18.1 / 80)=0.77375, product=0.464. Silicone score: (1−0.17 / (1200×0.05))=1−0.17 / 60=0.997, (1−1250 / 3000)=0.583, product=0.581. Interface score: Sigmoid function value=1 / (1+2.945)=0.253. Substituting into the function gives... =0.245. Therefore, the product was deemed unqualified. Subsequent destructive testing (microscopic observation after cutting) confirmed that the product did indeed have radial deformation of the material (measured 0.07mm), burrs on the upper edge (length 0.3mm, width 0.12mm), and continuous gaps at the interface (maximum width 0.03mm, arc length accounting for approximately 30%). Using traditional methods that only inspect silicone parts, this product was mistakenly judged as qualified because its silicone surface had no obvious bubbles or cracks. The multi-dimensional method of this invention successfully identified all three types of defects, preventing unqualified products from entering the next process.

[0059] To verify the statistical significance and process adaptability of this invention, 100 products with known defect distributions were selected for comparative testing. Among these, 30 products had only material defects (deformation or heat melting), 30 products had only silicone defects (bubbles or burrs), 30 products had only interface debonding, and 10 were defect-free qualified products. The overall accuracy of the traditional method (fixed light source, detecting only silicone) was 72%, with a 0% detection rate for material defects (completely ignored), an 86.7% detection rate for silicone defects, and a 0% detection rate for interface debonding. The overall accuracy of the method of this invention was 98%, with a 96.7% detection rate for material defects (29 / 30), a 96.7% detection rate for silicone defects (29 / 30), and a 96.7% detection rate for interface debonding (29 / 30). The false detection rate was 2% (one defect-free product was mistakenly identified as having an interface defect because oil contamination on the product surface caused an abnormal gradient). Analysis showed that this false detection could be further reduced by adding a preprocessing step (such as median filtering for noise reduction). The comparative data fully demonstrates the significant progress made by this invention in terms of detection dimensional completeness and detection accuracy.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for quality inspection of liquid coating products based on image recognition, characterized in that: Includes the following steps: Geometric modeling steps: Obtain the geometric topology model of the product to be inspected, and extract the 3D interface curves from the geometric topology model. Interface curve Defined as the contact boundary line between the outer surface of the product material and the inner surface of the silicone part; Adaptive imaging steps: Based on the difference in reflectivity between the substrate and the silicone part, the light source wavelength and intensity of the imaging system are dynamically configured, and the interface curve is used as a reference. Adjust the camera's shooting direction to acquire at least one composite feature image containing the material area, the silicone area, and the area where the two meet; Multi-path analysis steps: At least three analysis paths are executed in parallel on the composite feature image, including: Material quality analysis path, used to detect geometric deformation and surface damage of materials; The silicone part quality analysis path is used to detect bubbles and burr defects in silicone parts. The interface, combined with the quality analysis path, is used to display the interface curve. Projection curves are obtained by projecting onto the composite feature image. And calculate along the projection curve The image grayscale gradient magnitude is used to quantify the degree of interface bonding. Comprehensive scoring steps: Obtain the output indicators of the material quality analysis path, silicone part quality analysis path, and interface combination quality analysis path, and input them into a multivariate nonlinear comprehensive quality evaluation function to calculate a comprehensive quality score. Based on the comparison result of the comprehensive quality score and the preset qualified threshold, the product quality judgment result is output.

2. The method for quality inspection of liquid coating products based on image recognition according to claim 1, characterized in that: In the adaptive imaging step, dynamically configuring the light source wavelength and light intensity of the imaging system further includes: Based on the reflectivity of the material Reflectivity of silicone parts The ratio of the wavelengths is used to select a light source with the corresponding wavelength. The output light intensity of the light source is dynamically configured based on the light intensity compensation function, as shown in the following formula: in, This is the contrast gain coefficient. Based on the light intensity.

3. The method for quality inspection of liquid coating products based on image recognition according to claim 2, characterized in that: In the geometric modeling step, the 3D interface curves are extracted. Further includes: For products of rotation, radial sections with equal angular intervals are selected on the outer surface of the material. The intersection points of each section with the outer surface of the material and the inner surface of the silicone part are obtained. All intersection points are connected and fitted in angular order to obtain the interface curve. ;or, For non-rotational products, interface points are extracted on the outer surface of the material using a constant arc length sampling method, and then connected and fitted to obtain the interface curve. .

4. The method for quality inspection of liquid coating products based on image recognition according to claim 1, characterized in that: In the comprehensive scoring process, the multivariate nonlinear comprehensive quality evaluation function is: in, To calculate the overall quality score, These are the weighting coefficients for the quality of the material, silicone, and interface bonding, respectively. Sub-ratings for material quality, silicone quality, and interface quality. Through the Sigmoid function The calculation yielded that, The interface integration index. The threshold for determining interface debonding. This is the slope parameter.

5. The method for quality inspection of liquid coating products based on image recognition according to claim 4, characterized in that: In the interface integration quality analysis path, quantifying the degree of interface integration further includes: Interface integration index The calculation formula is: in, This represents the total length of the projected curve. For the image at points grayscale gradient magnitude at that location The differential of the arc length of the curve; Furthermore, by employing a multi-view fusion strategy, the interface integration index of images acquired from multiple viewpoints is calculated for each image. The results are then weighted and averaged to obtain the final interface bonding index. .

6. The method for quality inspection of liquid coating products based on image recognition according to claim 5, characterized in that: In the interface bonding quality analysis path, the threshold for determining interface debonding is... The method combines statistical analysis with destructive testing to determine: A batch of product samples confirmed to have good bonding through destructive testing were taken, and the interfacial bonding index of each sample was calculated under the same imaging conditions. This yields the sample set; Calculate the mean of the sample set. and standard deviation ; Set threshold .

7. The method for quality inspection of liquid coating products based on image recognition according to claim 1, characterized in that: In the adaptive imaging step, based on the interface curve Adjusting the camera's shooting direction further includes: The boundary interface curve Discretize the material into multiple equally spaced points, and calculate the unit outward normal vector of the outer surface of the material or the inner surface of the silicone at each point. ; Adjust the camera's orientation so that its optical axis aligns with the unit outward normal vector. parallel.

8. The method for quality inspection of liquid coating products based on image recognition according to claim 1, characterized in that: In the material quality analysis path, detecting the geometric deformation of the material includes: The edge contours of the material are extracted from the composite feature image, and ellipse fitting is performed on the edge contours to obtain the actual major axis radius of the material. and minor axis radius ; Calculate the equivalent radius based on the actual major and minor axis radii. ; Calculate radial deformation ,in The standard radius is extracted from the geometric topology model; When radial deformation If the value exceeds the preset threshold, it is determined to be a deformation defect.

9. The method for quality inspection of liquid coating products based on image recognition according to claim 1, characterized in that: In the quality analysis path for silicone parts, the detection of bubbles and burrs includes: For bubble defects, an adaptive threshold segmentation algorithm is used, which segments the silicone area based on the average grayscale value of the local window. and standard deviation Calculate the local threshold: in The empirical coefficients are used to determine the area of ​​the segmented connected components; For burr defects, the high-frequency components of the silicone edge are extracted using the difference of Gaussian operator, and a defect evaluation factor is defined: in It is the set of pixels in the neighborhood of the silicone edge. The response image is generated by convolving the image with the difference of Gaussian operator, and the defect evaluation factor is included. Thresholds obtained from statistical calibration Compare and determine whether there are burrs or cracks.

10. A quality inspection system for liquid coating products based on image recognition, characterized in that: The system for performing the method as described in any one of claims 1 to 9 comprises: The geometric modeling unit is used to obtain the geometric topology model of the product and extract the 3D interface curves. The adaptive imaging unit includes a programmable multispectral light source, multiple independently adjustable cameras, and a central control unit. The central control unit is used to configure the light source parameters according to the material reflectivity difference between the material and the silicone part, and to control the shooting direction of each camera according to the interface curve. The image analysis unit is configured with a multi-threaded parallel architecture to perform material quality analysis, silicone part quality analysis, and interface integration quality analysis in parallel. The comprehensive scoring unit receives the output from the image analysis unit, calculates the comprehensive quality score, and outputs the judgment result.

Images