A method and system for detecting overlay misregistration based on image processing

Abstract

The present application belongs to the technical field of ceramic product stack detection, and relates to a stack alignment offset detection method and system based on image processing. The method comprises: obtaining a smooth image of a multi-layer ceramic component, which comprises a top layer of green ceramic tape and a plurality of bottom layers of green ceramic tape stacked in sequence, and has a semi-transparent medium layer between the top layer of green ceramic tape and the bottom layers of green ceramic tape; extracting a plurality of candidate edge points of the bottom layer of green ceramic tape alignment mark in the smooth image; calculating the edge reliability score of each of the plurality of candidate edge points, taking the edge reliability score as a dynamic weighting coefficient, fitting the candidate edge points in the candidate edge point set to obtain the center coordinates of the bottom layer of green ceramic tape alignment mark, realizing the detection of the stack alignment offset, and improving the accuracy of interlayer correction.

Description

Technical Field

[0001] This invention belongs to the field of ceramic product stacking inspection technology, and relates to a stacking alignment offset detection method and system based on image processing. It is used for offset detection during the pressing of multilayer ceramic components and can improve the accuracy of ceramic component offset detection. Background Technology

[0002] In the high-end intelligent equipment manufacturing sector, represented by chip manufacturing and packaging, 5G communications, and aerospace, the manufacturing processes for multilayer electronic components such as LTCC (Low Temperature Co-fired Ceramic Capacitor), HTCC (High Temperature Co-fired Ceramic Capacitor), and MLCC (Multilayer Ceramic Capacitor) are rapidly evolving towards higher precision and integration. In a typical production process for these multilayer ceramic components, the first step involves uniformly coating and drying ceramic slurries such as alumina and aluminum nitride using a casting machine to prepare a translucent green ceramic tape of a specific thickness. Subsequently, conductive patterns and alignment marks are printed onto the surface of the green ceramic tape using processes such as screen printing. Finally, the most crucial lamination and pressing process begins: the multiple layers of green ceramic tapes with printed circuitry are physically aligned and sequentially laminated before being co-fired in a kiln to achieve the final shape.

[0003] In the lamination process, the visual alignment system must penetrate the semi-transparent dielectric layer between the two components—the upper ceramic green ceramic tape and its surface adhesive medium—to accurately image and extract the coordinates of the alignment marks on the lower layer, thereby guiding the automated equipment to complete inter-layer correction. Existing technologies for this type of penetrating visual alignment often rely directly on conventional image gradient operators to extract edges. However, due to the uneven thickness of the green ceramic tape caused by minute fluctuations in the casting process, the semi-transparent dielectric layer exhibits significant spatial differences in transmittance. This global dielectric turbidity significantly reduces the contrast of the lower layer features. Simultaneously, under the pressure of the equipment during lamination, microscopic overflows of adhesive or conductive paste can occur in localized areas. These complex high-frequency random textures induce optical scattering and refraction distortion, resulting in high-gradient textured pseudo-edges on the image. Traditional equal-weighted edge fitting algorithms cannot identify and eliminate these optical artifacts caused by physical process defects, easily misjudging high-intensity artifacts as real physical contours, leading to highly unstable alignment accuracy and severely restricting the yield of high-end LTCC / MLCC products.

[0004] Chinese patent document CN108074263B discloses a visual positioning method and system. The visual positioning method includes the following steps: acquiring an image at a standard positioning location; extracting edge points from the image; fitting the edge points to obtain a fitted curve; and calculating the coordinates of the standard positioning location in the world coordinate system based on the fitted curve. This scheme does not design any mechanism for identifying and removing textured pseudo-edges. When the gradient magnitude of a pseudo-edge is higher than that of the real edge after attenuation by the medium, the method directly uses the artifact as a real edge in the fitting process, causing the calculated center coordinates to deviate significantly from the true value.

[0005] Chinese patent document CN114723813B discloses a fiber optic end-face positioning method and its application. This method, by acquiring an image of the fiber optic end-face and adjusting the lighting method accordingly, can achieve efficient and precise splicing of polarization-maintaining fiber end-faces, thereby reducing fusion splicing loss and maintaining a stable extinction ratio. However, this method cannot identify or eliminate textured pseudo-edges caused by material overflow. Regardless of the lighting method used, the diffuse reflection characteristics of the overflow area may simultaneously enhance both real and pseudo-edges, even making the pseudo-edges more prominent.

[0006] Chinese patent document CN119152034B discloses an image localization method, apparatus, and related equipment, relating to the field of image localization. The image localization method includes the following steps: acquiring a test image of the localization markers; acquiring test feature points; matching the template image and the test image of the localization markers based on template feature points and test feature points; when matching fails, using the center point of the circumscribed rectangle of the first connected component as the localization position; and controlling the visual system's field of view movement based on the deviation between the localization position and the visual system's field of view center position. However, this method cannot identify or remove pseudo-connected components generated by overflow textures, nor can it handle situations where real markers are segmented by textures.

[0007] In summary, existing technologies cannot identify and eliminate optical artifacts caused by physical process defects in the through-vision alignment inspection of multilayer ceramic component lamination processes, and cannot achieve accurate positioning of alignment marks. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and solve the problem of poor alignment accuracy caused by medium turbidity and local glue texture interference in the lamination process, and to provide a lamination alignment offset detection method and system based on image processing.

[0009] To achieve the above-mentioned objectives, this invention provides a method for detecting stacked alignment offset based on image processing, comprising the following steps: acquiring a smooth image of a multilayer ceramic component, wherein the multilayer ceramic component comprises a top layer of green ceramic strips and multiple bottom layers of green ceramic strips stacked sequentially, with a semi-transparent dielectric layer between the top layer of green ceramic strips and the bottom layers of green ceramic strips; extracting multiple candidate edge points of the alignment marks of the bottom layer of green ceramic strips in the smooth image, wherein the multiple candidate edge points constitute a candidate edge point set; calculating the edge reliability score of each of the multiple candidate edge points, using the edge reliability score as a dynamic weighting coefficient, fitting the candidate edge points in the candidate edge point set to obtain the center coordinates of the alignment marks of the bottom layer of green ceramic strips, thereby realizing the detection of stacked alignment offset; The calculation method for edge reliability score includes: calculating the global turbidity of the medium in the smoothed image; calculating the local gray-level dispersion and radial deformation deviation rate of candidate edge points; calculating the optical scattering penalty index based on the global turbidity and local gray-level dispersion; calculating the geometric constraint confidence of candidate edge points, where the geometric constraint confidence is inversely proportional to the optical scattering penalty index; and calculating the refractive distortion penalty index based on the global turbidity and radial deformation deviation rate, where the edge reliability score is positively correlated with the geometric constraint confidence and negatively correlated with the refractive distortion penalty index.

[0010] In the lamination process of multilayer ceramic components, this invention calculates the degree of influence of optical stray light pollution and refractive distortion on edge points by comprehensively evaluating the global turbidity rate, local grayscale dispersion, and radial deformation deviation rate of the medium. Then, it dynamically assigns a reliability score to each candidate edge point and performs weighted fitting, thereby effectively eliminating the interference of high-intensity artifacts on the real physical contour. This significantly improves the accuracy and stability of bottom-layer alignment mark extraction through the semi-transparent dielectric layer, thereby improving the production yield of high-end ceramic electronic components.

[0011] The method for calculating the global turbidity rate of the medium described in this invention is as follows: extract all pixels in the background area of ​​the underlying green ceramic band in the smoothed image, remove abnormal pixels, and calculate its gray-level variance; use the ratio of the gray-level variance to the preset historical reference variance as the global turbidity rate of the medium.

[0012] Compared to existing technologies that cannot distinguish the fluctuations in the light transmission and shading effects of multi-layer components, this invention, in industrial casting coating scenarios, can accurately quantify the macroscopic light and dark interlacing and transmission attenuation caused by uneven thickness of the green ceramic strip by comparing the pixel grayscale variance of the current underlying background area with the historical maximum variance. This provides a reliable global environmental benchmark parameter for subsequent evaluation of light path scattering and refraction distortion, and enhances the algorithm's adaptability to changes in the light transmittance of different batches of materials.

[0013] The method for calculating the local grayscale dispersion described in this invention is as follows: a local square neighborhood window of a preset size is constructed with the target candidate edge point as the center; the grayscale values ​​of all pixels within the local square neighborhood window are extracted, and the variance of the grayscale values ​​is used as the local grayscale dispersion.

[0014] Compared to traditional edge detection, which only focuses on the gradient magnitude of a single pixel and cannot distinguish between regular abrupt changes in real contours and disordered fluctuations in rough textures, this invention utilizes a local statistical window to extract local grayscale dispersion in the real process environment of lamination and pressing accompanied by local stress. This effectively identifies and quantifies the high-frequency random rough texture interface caused by the micro-overflow of adhesive or ceramic slurry, accurately determines whether candidate points are in the high diffuse reflection zone of the lamination overflow, and avoids misjudging messy textures as real edges.

[0015] The method for calculating the radial deformation deviation rate described in this invention includes: obtaining the theoretical center coordinates and theoretical radius of the bottom alignment mark in the design drawing; calculating the actual Euclidean distance from the target candidate edge point to the theoretical center coordinates; extracting the difference between the actual Euclidean distance and the theoretical radius; and using the ratio of the difference to the theoretical radius as the radial deformation deviation rate.

[0016] Compared to existing technologies that struggle to address the microscopic refraction deviation of the transmission light path caused by uneven thickness in semi-transparent media, this invention combines theoretical parameters from engineering design drawings. By calculating the degree to which the actual geometric distance deviates from the theoretical standard radius, it precisely quantifies the degree of local geometric distortion caused by severe distortion of the light path. This effectively identifies disordered offset points affected by abrupt changes in the refractive index of the medium or deformation due to compressive stress, ensuring that the extracted contour points closely match the actual physical arc trajectory.

[0017] The method for calculating the optical scattering penalty index according to the present invention includes: calculating the sum of the global turbidity of the medium and a preset constant, and multiplying the sum by the local gray-level dispersion to obtain the optical scattering penalty index.

[0018] Compared to conventional methods that isolate and analyze local or global optical interference, this invention innovatively integrates the global medium turbidity characteristics with the optical scattering effect caused by local glue overflow in complex, multi-layered optical imaging environments. This scientifically reflects the comprehensive intensity of optical stray light pollution at specific points, enabling the system to accurately isolate artifacts severely affected by astigmatism under harsh conditions of multiple superimposed optical interferences.

[0019] The method for calculating the geometric constraint confidence level described in this invention is as follows: calculate the normalized basic gradient of the candidate edge point and use the normalized basic gradient as the numerator; calculate the product of the optical scattering penalty index of the candidate edge point and the preset noise reduction sensitivity coefficient, and the sum of the product and the preset constant, and use the sum as the denominator; use the ratio of the numerator to the denominator as the geometric constraint confidence level.

[0020] Compared to existing filtering mechanisms based on equal-weighted gradient thresholds, which are prone to missing weak edges or retaining high-intensity artifacts, this invention uses the calculated optical scattering penalty index to target the initial gradient response value for noise reduction during complex visual alignment processes. This significantly reduces the weight of high-contrast artifacts caused by messy textures or strong reflections, thus achieving the goal of completely eliminating high-intensity reflection artifacts while preserving real weak physical edges.

[0021] The method for calculating the edge reliability score described in this invention is as follows: the ratio of the geometric constraint confidence level to the refractive distortion penalty index is used as the edge reliability score.

[0022] Before extracting multiple candidate edge points as described in this invention, the method further includes the steps of: obtaining the region contour of the top layer of green ceramic band alignment mark, and performing mask processing on the pixel region corresponding to the region contour in the smooth image to eliminate top layer feature interference.

[0023] Compared to the problem that traditional coaxial light imaging can easily cause the superposition and confusion of upper and lower layer features in the same two-dimensional space, this invention performs masking processing on the obtained top layer alignment mark contour area before extracting the weak edges of the bottom layer, and forces the top layer pattern pixels to be assigned background values. This achieves absolute spatial decoupling of the high contrast patterns and weak features of the upper and lower layers from both physical and algorithmic levels, eliminating the pixel-level interference of the top layer pattern on the extraction of bottom layer features.

[0024] The method for obtaining a smooth image according to the present invention is as follows: acquiring the original image of the current multilayer ceramic component, then performing grayscale processing on the original image to obtain a grayscale image, and using Gaussian filtering to perform smoothing and noise reduction processing on the grayscale image to obtain a smooth image.

[0025] By sequentially performing grayscale conversion and Gaussian filtering to smooth and denoise the acquired original images of multilayer ceramic components, conventional sensor random noise and subtle environmental noise are effectively filtered out, providing a clean, stable, and reliable basic data source for subsequent complex micro-texture variance calculation and edge gradient analysis.

[0026] The present invention also provides an image processing-based overlay alignment offset detection system, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the overlay alignment offset detection method based on image processing described above is implemented.

[0027] Compared with existing technologies, this invention has at least the following beneficial effects: First, it solves the problem of poor alignment accuracy caused by uneven thickness of the semi-transparent green ceramic tape and local glue overflow during the lamination process of multilayer ceramic components, resulting in medium turbidity, optical diffuse reflection and refraction distortion. This is achieved by extracting parameters such as the global turbidity rate of the medium, the local glue texture variance, and the radial deformation deviation rate. Second, it constructs an optical scattering and refraction penalty model to calculate the edge reliability score. This score is used as a dynamic weight for weighted fitting, which removes high-intensity optical artifact interference and achieves high-precision coordinate extraction of the bottom alignment mark under poor light transmission imaging environment, thereby improving the accuracy of interlayer correction. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the image processing-based method for detecting layer alignment offsets that relates to the present invention.

[0029] Figure 2 This is a schematic diagram of the physical feature response curve of the candidate edge point involved in the present invention.

[0030] Figure 3 This is a schematic diagram of the edge reliability score distribution curve involved in the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings.

[0032] Example 1: This embodiment discloses a method for detecting layer alignment offset based on image processing, referring to... Figure 1 This includes steps S1-S6: Step S1: Image acquisition and top-level benchmark extraction.

[0033] A coaxial industrial camera is vertically fixed above a high-precision vacuum laminating machine, ensuring that the camera's optical axis is parallel to the surface normal of the multilayer ceramic component being tested. During the lamination process, the system acquires the original image of the multilayer ceramic component based on a pre-lamination readiness trigger signal issued by the industrial control computer. The original image is then converted to grayscale to obtain a grayscale image. Subsequently, a Gaussian filter is used to smooth and denoise the grayscale image, resulting in a smoothed image. Finally, the Otsu thresholding method is used to segment the top layer region without media occlusion in the smoothed image to obtain the foreground region. Morphological opening operations are then used to perform area filtering on the foreground region to remove isolated noise points. Finally, the connected component centroid algorithm is used to extract the center coordinates of the top layer alignment markers in the foreground region. .

[0034] It should be noted that in the lamination process, ceramic components are made by laminating multiple layers of transparent green ceramic strips. There is a semi-transparent dielectric layer between adjacent green ceramic strips, which serves as an adhesive. The topmost green ceramic strip is defined as the top layer, and all green ceramic strips below the top layer are defined as the bottom layer. The surface of each green ceramic strip is printed with circular alignment marks. In the design state, the centers of the circular alignment marks of all green ceramic strips coincide, and their radii decrease from bottom to top.

[0035] Step S2: Obtain the bottom candidate edge point set and calculate the global turbidity of the medium.

[0036] Since coaxial imaging superimposes the features of the top and bottom layers into the same two-dimensional image space, to eliminate pixel-level interference from the high-contrast pattern of the top layer on the extraction of weak edges of the bottom layer, the system first generates a corresponding digital mask on the current smooth image based on the top layer alignment mark coordinates and connected component contours obtained in step S1. If the top layer mark extraction fails, coordinate mapping is performed based on the theoretical position and size of the top layer mark in the design drawing, combined with the identifiable substrate edge features in the current image, to generate an approximate mask area. After the mask processing is completed, a prompt message is sent to the operator for manual review. Through this mask operation, the pixel area where the top layer pattern is located is forcibly assigned a background grayscale value, thereby achieving absolute spatial decoupling of the upper and lower layer features at both physical and algorithmic levels.

[0037] After removing top-level interference, the system delineates the region of interest (ROI) in the image. Because the ROI is optically occluded by the translucent ceramic green porcelain band, directly extracting it using a single global threshold would result in the loss of subtle edges. Therefore, the system constructs a sliding computation window that traverses the entire ROI, using the Sobel operator to calculate the gradient magnitude pixel-by-pixel.

[0038] Pixels whose gradients are greater than a preset gradient threshold Th are selected to form a candidate edge point set. The preset gradient threshold Th has a range of values. The gradient threshold Th can be adaptively adjusted according to the overall grayscale dynamic range of the image. Specifically, the gradient magnitude distribution of all pixels in the underlying region of interest is calculated, and 1.2 times the global mean of the gradient magnitude is used as the initial threshold. Then, an empirical range is used for constraint to ensure that while filtering out the background undulations of conventional ceramic grains, as many potential physical edge points that are severely weakened by the dielectric layer as possible are retained. For example, in this embodiment, the value is 45.

[0039] Meanwhile, to quantify the macroscopic light transmission and shading effect of current multilayer ceramic components, all pixels in the bottom layer region are obtained as data samples. To prevent local extrema from interfering with global statistics, the system uses a truncated statistical method to remove the 5% brightest and 5% darkest abnormal pixels in the gray-level distribution, extracts the gray-level distribution of the remaining bottom background region pixels, and calculates its gray-level variance. The ratio of the gray-level variance to the maximum variance in historical smoothed images is used as the global turbidity rate of the medium in this specific scene. Among them, the historical maximum variance is the statistical maximum value of the gray-level variance of the bottom background region of several smoothed images pre-collected under normal process conditions for the same batch of products. In this embodiment, the method for obtaining the historical maximum variance is as follows: during the offline calibration stage of the equipment, 50 sets of normal good product samples are pre-collected to extract corresponding features, and the maximum boundary value of the distribution is obtained by Gaussian fitting as the historical maximum variance.

[0040] When the overall turbidity of the medium When the turbidity is high, it reflects extremely uneven thickness of the translucent ceramic band, resulting in strong light and dark contrasts and transmission attenuation at the bottom layer; when the overall turbidity of the medium is high... When the value is relatively small, it reflects that the raw ceramic belt is evenly coated and has stable light transmittance.

[0041] Step S3: Calculate the local grayscale dispersion and radial deformation deviation rate.

[0042] Extracting candidate edge point sets Then, for any candidate edge point... To address this, the degree of physical interference in the surrounding environment must be quantified. During the lamination and pressing of multilayer ceramic components, local compressive stress can force the interlayer adhesive or ceramic slurry to overflow microscopically. These overflowed solidified materials form rough, high-frequency random textures on the originally smooth green ceramic interface, resulting in strong diffuse reflection of incident light. Traditional edge detection only focuses on the gradient amplitude of a single pixel and cannot distinguish between the regular abrupt changes of real edges and the disordered fluctuations of rough textures.

[0043] Therefore, this step utilizes the physical characteristic that the texture of the real edge region is smooth while the texture of the overflow region is messy. By constructing a local statistical window, the gray-level variance is calculated, and the gray-level variance is used as the local gray-level dispersion.

[0044] Specifically, using any candidate edge point coordinates Construct a space with a size of centered on . The local square neighborhood window is used to calculate the local gray-level dispersion using the following formula. :

[0045] in, Indicates the candidate edge point Local gray-level dispersion in the neighborhood; This represents the pixel size of the side of the local square neighborhood window. The range of values ​​is The window size should be adapted to the imaging size of the alignment mark, typically taken as 1 / 10 to 1 / 15 of the alignment mark diameter as the window side length (in pixels), and limited to a certain range. For product types with large variations in mark size, the window size can be automatically adjusted according to the mark's preset radius to fully encompass the microscopic overflow texture features while avoiding the window crossing and over-smoothing the real physical edges. For example, in this embodiment, the value is [value missing]. ; Indicates the candidate edge point Within the local square neighborhood window centered on the first The actual grayscale value of each pixel; Indicates the candidate edge point The average grayscale value of all pixels within a local square neighborhood window centered on the center.

[0046] When the actual gray value of a local pixel Deviation from the average When the amplitude is large, it reflects the presence of severe semi-cured slurry overflow or random microbubbles in the area surrounding the candidate edge point, forming a rough and uneven optical scattering interface, resulting in local grayscale dispersion. It tends to increase sharply; when the actual gray value near Physically, this reflects that the surface of the ceramic green porcelain zone around that point is smooth and has uniform light transmission, leading to a decrease in local grayscale dispersion.

[0047] In summary, local gray-level dispersion This reflects the roughness of the microscopic physical interface and the degree of optical diffuse reflection. When the value is large, it is determined that the point is likely located in the compression overflow region of the high-frequency texture; when... When the value is small, the point is determined to have the optical smoothness condition to become a real physical edge.

[0048] On the other hand, the uneven thickness of the translucent green ceramic strip not only induces diffuse reflection but also causes microscopic refraction shifts in the transmitted light path. The physical contour of the underlying true standard alignment ring has a fixed and continuous geometric curvature, while the refractive distortion of the medium causes the pseudo-edge pixels after imaging to randomly deviate from this theoretical geometric trajectory. Therefore, this step utilizes the geometric law that the interference artifacts exhibit disordered spatial deviation while the true edges closely adhere to the theoretical trajectory to calculate the radial deformation deviation rate to quantify the degree of light path refraction.

[0049] Specifically, based on the engineering design drawings of this multilayer ceramic component, the theoretical center coordinates and theoretical radius of the bottom alignment mark are obtained, and the radial deformation deviation rate is calculated using the following formula:

[0050] in, Indicates candidate edge points Radial deformation deviation rate; Indicates candidate edge points The horizontal coordinate in the current two-dimensional digital image coordinate system; Indicates candidate edge points The ordinate in the current two-dimensional digital image coordinate system; This indicates the initial center x-coordinate of the bottom alignment mark determined by the design drawings; This indicates the initial center ordinate of the bottom alignment mark as determined by the design drawings; This represents the initial standard radius of the underlying alignment mark contour; this theoretical standard radius... The value depends on the specific design specifications of the multilayer ceramic component; for example, its value is 150 pixels.

[0051] When the actual geometric distance from the candidate edge point to the theoretical center Deviation from theoretical standard radius A large absolute difference indicates that the candidate edge point has undergone severe distortion and shift in the optical path due to a sudden change in the refractive index of the upper green ceramic band or deformation caused by compressive stress. This means it no longer conforms to the expected physical arc trajectory, leading to a different result value. It tends to increase significantly; when the actual distance approaches This indicates that the transmission imaging optical path at that point is stable, with no significant refractive distortion, leading to the result value. It tends to approach 0.

[0052] In summary, radial deformation deviation rate Essentially, it reflects the degree of local geometric distortion caused by the offset of the transmitted light path. When the value is large, the point is determined to be in a region of severe refractive distortion; when... When the value is small, the point is determined to be a true contour point that is not significantly affected by refraction shift.

[0053] Step S4: Calculate the geometric constraint confidence of the candidate edge points.

[0054] First, an optical diffuse reflection feature fusion calculation model is constructed to calculate the optical scattering penalty index. The method is as follows: the local gray-level dispersion is normalized using a linear normalization algorithm to obtain the normalized local gray-level dispersion; the expression for the optical scattering penalty index is:

[0055] in, Indicates the candidate edge point The optical scattering penalty index integrates global dielectric attenuation and local grayscale fluctuations; Indicates the candidate edge point The local gray-level dispersion after normalization in the neighborhood; This indicates the overall turbidity rate of the medium under the current multilayer green ceramic belt.

[0056] When the local grayscale dispersion increases, it reflects the presence of severe overflow of adhesive material or a rough green ceramic interface around the pixel, resulting in strong local diffuse reflection and an increase in the optical scattering penalty index. When the global turbidity of the medium increases simultaneously, the turbidity physical properties of the global medium will further amplify this scattering, causing the optical scattering penalty index to increase further. When the local grayscale dispersion decreases and the global turbidity of the medium decreases, it physically reflects local smoothness and good global light transmission, causing the optical scattering penalty index to tend to decrease.

[0057] In summary, the optical scattering penalty index reflects the overall intensity of optical stray light pollution at a location. When the optical scattering penalty index is large, the location is considered to be severely affected by optical scattering interference; when the optical scattering penalty index is small, the location is considered to have a relatively pure optical physical environment.

[0058] Based on the optical scattering penalty exponent, the initial gradient response value is denoised and stripped, and the geometric constraint confidence is calculated. The gradient value of each pixel in the candidate edge point set is normalized using a linear normalization algorithm. The normalized candidate edge points are then... The magnitude of the gradient is denoted as the normalized fundamental gradient. The expression for the confidence level of geometric constraints is:

[0059] in, Indicates the candidate edge point The confidence level of geometric constraints after removing optical diffuse reflection interference; Indicates candidate edge points The normalized basic gradient; This represents the noise reduction sensitivity coefficient used to control the weighting intensity. The range of values ​​is usually 100. , The specific value can be determined as follows: Collect a batch of sample images containing typical optical artifacts, and, given the true center coordinates, determine the optimal value using a grid search method with the goal of minimizing the fitting error. Value. For example, in this embodiment, the value is 2.5; Indicates the candidate edge point The optical scattering penalty index integrates global medium attenuation and local grayscale fluctuations; the constant 1 prevents the denominator from being zero under extreme purity conditions.

[0060] when When the value increases, it reflects that the pixel region has a stronger grayscale abrupt change attribute, and the geometric constraint confidence level... Increase; when As the value increases, it reflects that this high contrast is most likely an artifact caused purely by messy textures deceiving the reader. The strong penalty term causes the denominator to increase dramatically, leading to a decrease in the geometric constraint confidence. It was suppressed.

[0061] In summary, the confidence level of geometric constraints This reflects the degree of pure physical structural response after eliminating reflective deception, when When the value is large, the probability of determining that the point is a real physical edge point is higher; when... When the value is small, the point is more likely to be an optical deception artifact.

[0062] Step S5: Calculate the edge reliability score of the candidate edge points.

[0063] In addition to texture scattering, the optical refraction of the translucent green ceramic strip also causes geometric distortion at the edges. Therefore, it is necessary to calculate the final edge reliability score, expressed as:

[0064]

[0065] in, Indicates the candidate edge point The refractive distortion penalty index at the location; Indicates the candidate edge point The radial deformation deviation rate of the characterization point deviating from the geometric error of the theoretical circular arc; This indicates the overall turbidity of the medium under the current multilayer green ceramic belt; Indicates candidate edge points Edge reliability score; Indicates the candidate edge point The geometric constraint confidence level after removing optical diffuse reflection interference.

[0066] when When the value increases, it reflects that the spatial displacement at that point no longer conforms to the predetermined physical circular arc trajectory, leading to... Increase, at this time candidate edge points The edge reliability score decreases; when the refractive distortion penalty index decreases... Decrease and A higher value indicates that the texture at that point is pure and the geometric curvature is regular, and the edge reliability score tends to remain high. Therefore, the edge reliability score reflects the absolute probability that a specific pixel belongs to the underlying non-deformed true edge in the harsh imaging physical environment of multi-layered green ceramic tapes.

[0067] Step S6: Fit the bottom layer alignment mark center and calculate the interlayer offset.

[0068] Extract all candidate edge point sets The edge reliability score of the candidate edge points obtained in step S5 is used. As dynamic weighting coefficients, a weighted least-squares circle fitting model is constructed for the candidate edge point set. Specifically, with the goal of minimizing the geometric distance error, the following objective function is constructed:

[0069] in, Represent the objective function for the weighted geometric distance error; The total number of pixels in the candidate edge point set; For the first Candidate edge points Pixel coordinates; Score the edge reliability of this point; and These are the x and y coordinates of the underlying true center to be solved; denoted as the true fitting radius of the underlying layer to be solved.

[0070] The optimal solution that minimizes the objective function yields the true center coordinates of the bottom layer after resisting diffuse reflection and refraction interference. For example, this embodiment employs a weighted least squares circle fitting method based on algebraic distance to transform geometric errors into an algebraic error matrix, which is then solved using singular value decomposition (SVD). Specifically, the circle equation is expanded into a linear form, and edge reliability scores are introduced as weighting coefficients to construct a weighted objective function. A closed-form solution for the center coordinates and radius is obtained by solving the system of linear equations.

[0071] Finally, based on the center coordinates of the top-level alignment markers extracted in step S1 Calculate the two-dimensional spatial offset vector of the top green ceramic strip relative to the bottom green ceramic strip. It will then be sent as a feedback signal to the motion control system of the high-precision pressing machine to guide the physical position compensation in the subsequent lamination process.

[0072] The technical effects of this invention can also be illustrated in conjunction with the accompanying drawings. Figure 2 The figure shows the physical feature response curves of candidate edge points, illustrating the physical features of different candidate edge points in different dimensions calculated in step S3. The curves show that the base gradient value remains at a high level in both the real physical edge region and the cluttered region with artifacts caused by compressive stress. However, the overflow texture variance and radial deformation deviation rate introduced in this invention show an increasing trend in the artifact region, accurately reflecting the physical interference and geometric distortion caused by the micro-overflow and light refraction of the ceramic slurry.

[0073] Figure 3 The figure shows the edge reliability score distribution curve. As can be seen from the figure, the geometric constraint confidence and edge reliability scores maintain a very high and stable score rate in the real edge region. However, once they cross into the artifact region, their scores are suppressed and approach zero. This scoring mechanism achieves the mathematical decoupling between the real edge and the messy optical artifacts.

[0074] This invention also discloses an image processing-based overlay alignment offset detection system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an image processing-based overlay alignment offset detection method according to the present invention.

[0075] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. An image processing-based overlay shift detection method, characterized by, The process includes the following steps: acquiring a smoothed image of a multilayer ceramic component, which comprises a top layer of green ceramic strips and multiple bottom layers of green ceramic strips stacked sequentially, with a semi-transparent dielectric layer between the top and bottom layers of green ceramic strips; extracting multiple candidate edge points from the alignment markers of the bottom layer green ceramic strips in the smoothed image, forming a candidate edge point set; calculating the edge reliability score of each candidate edge point, using the edge reliability score as a dynamic weighting coefficient, and fitting the candidate edge points in the candidate edge point set to obtain the center coordinates of the alignment markers of the bottom layer green ceramic strips, thereby detecting the stack alignment shift. The calculation method for edge reliability score includes: calculating the global turbidity of the medium in the smoothed image; calculating the local gray-level dispersion and radial deformation deviation rate of candidate edge points; calculating the optical scattering penalty index based on the global turbidity and local gray-level dispersion; calculating the geometric constraint confidence of candidate edge points, where the geometric constraint confidence is inversely proportional to the optical scattering penalty index; and calculating the refractive distortion penalty index based on the global turbidity and radial deformation deviation rate, where the edge reliability score is positively correlated with the geometric constraint confidence and negatively correlated with the refractive distortion penalty index.

2. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The method for calculating the global turbidity rate of the medium is as follows: extract all pixels in the background area of ​​the underlying green ceramic band in the smoothed image, remove abnormal pixels, and calculate its gray-level variance; use the ratio of the gray-level variance to the preset historical reference variance as the global turbidity rate of the medium.

3. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The method for calculating local grayscale dispersion is as follows: construct a local square neighborhood window of a preset size with the target candidate edge point as the center; extract the grayscale values ​​of all pixels within the local square neighborhood window, and use the variance of the grayscale values ​​as the local grayscale dispersion.

4. The image processing-based method for detecting layer alignment offset according to claim 1, characterized in that, The method for calculating the radial deformation deviation rate includes: obtaining the theoretical center coordinates and theoretical radius of the bottom alignment mark in the design drawing; calculating the actual Euclidean distance from the target candidate edge point to the theoretical center coordinates; extracting the difference between the actual Euclidean distance and the theoretical radius; and using the ratio of the difference to the theoretical radius as the radial deformation deviation rate.

5. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The calculation method for the optical scattering penalty index includes: calculating the sum of the global turbidity of the medium and a preset constant, and multiplying the sum by the local gray-level dispersion to obtain the optical scattering penalty index.

6. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The method for calculating the geometric constraint confidence is as follows: calculate the normalized basic gradient of the candidate edge points and use the normalized basic gradient as the numerator. Calculate the product of the optical scattering penalty index of the candidate edge point and the preset noise reduction sensitivity coefficient, and the sum of the product and the preset constant. Use the sum as the denominator. Use the ratio of the numerator to the denominator as the geometric constraint confidence level.

7. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The marginal reliability score is calculated by taking the ratio of the geometric constraint confidence level to the refractive distortion penalty index as the marginal reliability score.

8. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, Before extracting multiple candidate edge points, the process includes the following steps: obtaining the region contour of the top-layer green ceramic band alignment marker, and performing masking on the pixel region corresponding to the region contour in the smoothed image to eliminate top-layer feature interference.

9. The method for detecting layer alignment offset based on image processing according to claim 1, characterized in that, The method for obtaining a smooth image is as follows: acquire the original image of the current multilayer ceramic component, then perform grayscale processing on the original image to obtain a grayscale image, and use the Gaussian filtering method to perform smoothing and noise reduction processing on the grayscale image to obtain a smooth image.

10. A layer alignment offset detection system based on image processing, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a layer alignment offset detection method based on image processing according to any one of claims 1-9.

Images