Microfluidic heat dissipation channel alignment positioning method suitable for optoelectronic co-packaging

Abstract

The present application relates to the technical field of image processing, and more particularly to a microfluidic heat dissipation channel alignment positioning method suitable for optoelectronic co-packaging, comprising: acquiring a first image of an optoelectronic co-packaging chip and a second image of a microfluidic heat dissipation channel, calculating a Hessian matrix of the first image to screen strong isotropic regions as anchor point blocks, searching for an initial displacement vector on a phase consistency graph of the first image and the second image with normalized mutual information as a measure to obtain an initial displacement vector; constructing a feature vector set from the initial displacement vector, performing unsupervised clustering based on a density-based non-parametric Bayesian model, removing outlier vectors to obtain a main matching cluster and retaining membership posterior probability; determining a normalized weight based on spatial proximity, normalized mutual information score and membership posterior probability, constructing a local weighted affine transformation model to solve local affine transformation parameters of each pixel, generating a dense displacement field, and improving the alignment accuracy and stability of the microfluidic heat dissipation channel and the optoelectronic co-packaging chip.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to a microfluidic heat dissipation channel alignment and positioning method suitable for optoelectronic co-packaging. Background Technology

[0002] Optoelectronic co-packaging technology integrates optical engines and switching chips on the same substrate with high density, shortening electrical interconnect paths and reducing signal attenuation and transmission power consumption. It represents a significant development direction for high-speed optical communication packaging. As the integration density of optoelectronic devices continues to increase, the local heat flux density during chip operation also increases. Traditional air cooling or conventional water cooling is insufficient to meet the heat dissipation requirements of high heat flux areas, thus microfluidic heat dissipation channels have been introduced into the packaging structure. During manufacturing and packaging, the microfluidic heat dissipation channels must be accurately aligned with the hot spots inside the optoelectronic co-packaged chip. The alignment quality directly affects heat dissipation efficiency, device stability, and long-term reliability.

[0003] Currently, conventional machine vision alignment methods mostly rely on grayscale similarity metrics or global geometric transformations to register two images. In the prior art, Chinese patent application CN101567051A discloses a feature point-based image registration method. This method defines a ring-shaped edge-corner code model based on the K-nearest neighbor structure of feature points and uses it for point pattern matching. It then optimizes the matching results through local mapping clustering to perform optimal mapping estimation.

[0004] However, the aforementioned methods have limitations in the context of optoelectronic co-packaging microscopic alignment. Optoelectronic chips and microfluidic structures differ in material properties, surface reflectivity, and processing textures, leading to inconsistent contrast, blurred structural edges, and interference from complex backgrounds in the acquired images. Matching based on grayscale similarity is prone to mismatches at areas with repetitive textures and blurred structures. Furthermore, microfluidic structures may undergo minor local deformations during processing, bonding, and packaging, resulting in a non-rigid and non-uniform spatial correspondence between the two, making it difficult for global geometric transformations to stably adapt to local deformations. Therefore, extracting stable features from microscopic images to accurately characterize local deformations while eliminating abnormal matching interference, enabling high-precision and high-stability alignment and positioning of the microfluidic heat dissipation channel and the optoelectronic co-packaged chip, has become a critical technical problem to be solved in advanced packaging manufacturing. Summary of the Invention

[0005] To address the technical problems of insufficient alignment accuracy and stability caused by grayscale differences, noise, and local deformation interference in existing alignment methods, this invention provides a microfluidic heat dissipation channel alignment and positioning method suitable for optoelectronic co-packaging, including: S1. Acquire the first image of the optoelectronic co-packaged chip and the second image of the microfluidic heat dissipation channel. Calculate the Hessian matrix of the first image to select strongly isotropic regions as anchor blocks. Generate phase consistency maps of the first image and the second image respectively. Use normalized mutual information as a metric to search and match the anchor blocks between the phase consistency maps of the first image and the second image to obtain a set of initial displacement vectors. S2, construct a feature vector set from each initial displacement vector, perform unsupervised clustering analysis using a density-based nonparametric Bayesian model, remove outlier vectors with probability densities below a preset threshold to obtain the main matching cluster, and retain the posterior probability of membership of each member of the main matching cluster. S3. Based on the main matching cluster, a local weighted affine transformation model is constructed. For each pixel position in the first image, the normalized weight of each member of the main matching cluster is determined based on spatial proximity, normalized mutual information score and membership posterior probability. The local affine transformation parameters of each pixel position are obtained by weighted solution based on the normalized weights. A dense displacement field is generated by combining the parameters, and the alignment relationship between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip is determined.

[0006] This invention first uses the eigenvalue analysis of the Hessian matrix of the first image to select highly isotropic regions as anchor blocks. On the phase consistency map, normalized mutual information is used as a metric for search and matching, mitigating the impact of image grayscale differences and noise interference on the matching results. Next, a feature vector set is constructed from the initial displacement vector. Outlier vectors are eliminated using unsupervised clustering based on a density-based nonparametric Bayesian model, resulting in a main matching cluster with high geometric consistency while retaining the posterior probability of membership. Then, based on spatial proximity, normalized mutual information score, and posterior probability of membership, normalized weights are determined, a local weighted affine transformation model is constructed, and a dense displacement field is generated, enabling pixel-by-pixel characterization of local deformation. This determines the alignment relationship between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip, improving the accuracy and stability of the alignment.

[0007] Preferably, the step of selecting strongly isotropic regions as anchor blocks includes: constructing a two-dimensional Hessian matrix by calculating the second-order partial derivatives of the horizontal, vertical, and mixing directions for each pixel of the first image; solving for a first eigenvalue with a larger absolute value and a second eigenvalue with a smaller absolute value; using the absolute value of the second eigenvalue as the isotropic response value; marking the pixel as a candidate anchor point when the isotropic response value is greater than a preset isotropic threshold and the ratio of the absolute values ​​of the two eigenvalues ​​is less than a preset proportional threshold; suppressing non-maximum values ​​of the candidate anchor points within a preset local neighborhood; taking the pixel with the maximum local response value as the anchor point center; and extracting an image block of a preset size as the anchor block.

[0008] By using the dual determination of the absolute value of the two eigenvalues ​​of the Hessian matrix and their ratio, combined with non-maximum suppression to select the anchor point center, the selected anchor point blocks are concentrated in the microstructure region with obvious curvature response in both orthogonal directions, reducing the interference of weak response noise points and unidirectional edges.

[0009] Preferably, the step of searching for and matching anchor blocks between the phase consistency maps of the first image and the second image includes: extracting the amplitude and phase information of the first image and the second image at different scales and directions using a two-dimensional Log-Gabor filter bank, generating phase consistency maps of the first image and the second image respectively, and cropping the phase consistency block of the first image corresponding to the anchor block; sliding the matching box within a preset search neighborhood of the phase consistency map of the second image, calculating the normalized mutual information score between the matching box and the phase consistency block of the first image at each position, and taking the position with the maximum score as the candidate matching coordinates; taking the center pixel coordinates of the anchor block as the center coordinates of the anchor block, and calculating the initial displacement vector of the anchor block based on the difference between the center coordinates of the anchor block and the candidate matching coordinates.

[0010] Preferably, the unsupervised clustering analysis using a density-based nonparametric Bayesian model includes: separating the spatial horizontal and vertical coordinates and horizontal and vertical displacement components of each initial displacement vector in the first image; performing Z-score zero-mean standardization on the spatial horizontal and vertical coordinates and displacement components respectively; concatenating the standardized spatial horizontal and vertical coordinates and displacement components to construct a four-dimensional feature vector with a uniform scale and forming a feature vector set; using a Dirichlet process mixture model as a nonparametric Bayesian model for iterative inference; calculating the probability of each feature vector belonging to each existing data cluster and generating a new data cluster in each iteration; updating the cluster affiliation of each feature vector; recalculating the mean vector and covariance matrix of each data cluster based on the updated cluster affiliation; determining the data cluster to which each feature vector belongs after iterative convergence and outputting the posterior probability of each feature vector belonging to the corresponding data cluster.

[0011] The Dirichlet process mixture model is used to adaptively estimate the number of clusters according to the data distribution. It can distinguish between the main matching distribution and the outlier distribution without pre-specifying the number of clusters, and outputs the posterior probability of membership of each feature vector for subsequent weighting.

[0012] Preferably, the process of removing outlier vectors with probability densities lower than a preset threshold to obtain the main matching cluster includes: counting the number of feature vectors contained in each data cluster in the clustering results, and marking the data cluster containing the most feature vectors as the main candidate cluster; calculating the multidimensional Gaussian probability density distribution value corresponding to each feature vector in the main candidate cluster based on the mean vector and covariance matrix of the main candidate cluster; identifying feature vectors with multidimensional Gaussian probability density distribution values ​​lower than the preset threshold as outliers and removing them; and retaining the initial displacement vectors corresponding to the feature vectors that were not removed, and combining them to form a main matching cluster with geometric consistency.

[0013] Outlier vectors within the main candidate cluster are removed by combining the multidimensional Gaussian probability density distribution value with a preset threshold, so that the retained initial displacement vectors have high geometric consistency and reduce the impact of mismatched points on the affine solution.

[0014] Preferably, when calculating the multidimensional Gaussian probability density distribution value, a preset small regularization term is added to the main diagonal of the covariance matrix before inverting the covariance matrix to ensure the positive definiteness and inversion stability of the covariance matrix.

[0015] Preferably, the step of determining the normalized weights of each member of the main matching cluster based on spatial proximity, normalized mutual information score, and membership posterior probability includes: for any target pixel position in the first image, taking the center coordinates of the anchor block corresponding to each member of the main matching cluster as the starting spatial coordinates of each member, calculating the Euclidean distance between the target pixel position and the starting spatial coordinates of each member, and substituting the Euclidean distance into the Gaussian kernel function to obtain the spatial proximity weight; extracting the normalized mutual information score and membership posterior probability corresponding to each member of the main matching cluster, multiplying the spatial proximity weight, normalized mutual information score, and membership posterior probability to obtain the comprehensive weight value of each member; and normalizing the comprehensive weight value of each member so that the sum of the normalized weights of each member is 1.

[0016] Multiply and normalize the spatial proximity weight, normalized mutual information score, and membership posterior probability to give members with high spatial proximity, reliable matching, and high membership probability a larger share of constraints in the solution, thereby constraining the solution of local affine parameters.

[0017] Preferably, the step of obtaining the local affine transformation parameters of each pixel position by weighted calculation based on normalized weights and combining them to generate a dense displacement field includes: for any target pixel position in the first image, constructing a cost function based on the normalized weights of the target pixel position, with the goal of minimizing the weighted square error between the predicted coordinates of the initial spatial coordinates of each member of the main matching cluster after local affine transformation and the corresponding candidate matching coordinates of each member as the actual matching coordinates; setting the derivative of the cost function with respect to the parameters to 0, constructing a system of linear equations containing the observation matrix and the weight matrix, and solving for the local affine transformation parameters of the target pixel position using the weighted least squares method; traversing all pixel positions, calculating the deformation displacement according to the corresponding local affine transformation parameters, and combining them to generate a dense displacement field covering the entire image.

[0018] We solve the local affine transformation parameters pixel by pixel using weighted least squares and combine them to generate a dense displacement field, so that the local deformation at each position in the whole image can be independently represented, thus enhancing the ability to express non-uniform deformation.

[0019] Preferably, determining the alignment relationship between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip includes: applying a smoothing filter to the dense displacement field to filter out high-frequency local abnormal fluctuations in the displacement field; extracting the average horizontal displacement and average vertical displacement of the filtered dense displacement field as the overall translational deviation between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip; calculating the average displacement gradient matrix of the filtered dense displacement field, adding it to the identity matrix to construct the overall deformation gradient matrix, extracting the rotational component from the singular value decomposition of the overall deformation gradient matrix as the overall rotational deviation; and combining the overall translational deviation and the overall rotational deviation to generate an alignment compensation command for the microfluidic heat dissipation channel relative to the optoelectronic co-packaged chip to complete the alignment.

[0020] After smoothing and filtering the dense displacement field, the overall translational deviation and the overall rotational deviation obtained by singular value decomposition are extracted respectively. Based on this, an alignment compensation command is generated so that both translational and rotational alignment deviations are compensated.

[0021] Preferably, acquiring the first image of the optoelectronic co-packaged chip and the second image of the microfluidic heat dissipation channel includes: acquiring a surface image of the optoelectronic co-packaged chip using a charge-coupled device camera as the first image, and acquiring an internal structure image of the microfluidic heat dissipation channel using an infrared transmission microscopy imaging system as the second image; based on the calibration parameters of the microscopy imaging system, performing scale normalization, rotation correction, and coordinate origin alignment on the first and second images, so that the first and second images are mapped to a unified pixel coordinate system.

[0022] The technical solution of the present invention has the following beneficial technical effects: This invention combines Hessian matrix isotropic screening with phase consistency normalized mutual information matching to obtain stable anchor point matching under complex backgrounds and gray-scale differences. It introduces nonparametric Bayesian clustering to estimate the number of clusters according to the data distribution and remove outlier vectors, which reduces the interference of mismatched points on subsequent solutions and obtains main matching clusters with high geometric consistency.

[0023] Furthermore, this invention comprehensively determines the normalization weight by considering spatial proximity, normalized mutual information score, and membership posterior probability, constructs a locally weighted affine transformation model, and solves the affine parameters of each pixel in a weighted manner to generate a dense displacement field with local deformation expression capability. Based on this, the overall translational deviation and overall rotational deviation are extracted and alignment compensation instructions are generated, so that the non-rigid and non-uniform correspondence between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip can be characterized, and the alignment accuracy and stability are improved. Attached Figure Description

[0024] Figure 1 A flowchart of a microfluidic heat dissipation channel alignment and positioning method suitable for optoelectronic co-packaging; Figure 2A schematic diagram of the distribution of the master matching cluster and the outlier displacement vector; Figure 3 This is a schematic diagram comparing alignment errors. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a microfluidic heat dissipation channel alignment and positioning method applicable to optoelectronic co-packaging, referring to... Figure 1 This includes steps S1 to S3: S1. Acquire images and match anchor point blocks to generate initial displacement vectors.

[0027] A surface image of the optoelectronic co-packaged chip is acquired using a charge-coupled device (CCD) camera as the first image, and an internal structure image of the microfluidic heat dissipation channel is acquired using an infrared transmission microscopy system as the second image. After acquiring both images, scale normalization, rotation correction, and coordinate origin alignment are performed on the first and second images based on the calibration parameters of the microscopy system, mapping both images to a unified pixel coordinate system, providing a prerequisite for feature matching under the same coordinate reference. Taking a grayscale image with a resolution of 1024×1024 as an example, to suppress the interference of high-frequency noise on the second derivative, a Gaussian smoothing filter is first used to preprocess the first image, with a standard deviation of... It is 1.2.

[0028] After obtaining the preprocessed first image, the second-order partial derivatives in the horizontal direction, vertical direction, and mixed direction are calculated pixel by pixel using the Sobel operator or central difference template. A 2×2 Hessian matrix is ​​constructed for each pixel, and the two eigenvalues ​​of this matrix are solved. and Agreement The absolute values ​​of the larger eigenvalues ​​correspond to the maximum principal curvature response, while the absolute values ​​of the smaller eigenvalues ​​correspond to the minimum principal curvature response. The absolute values ​​of the smaller eigenvalues ​​are extracted. As an isotropic response value, it is used to characterize the degree of second-order grayscale variation of microstructures in two orthogonal directions. A strongly isotropic region is a microstructure region that exhibits significant local curvature response in both orthogonal directions; when the isotropic response value of a pixel is greater than the isotropic threshold, and the ratio of the absolute values ​​of the two feature values ​​is... When the value is less than the proportional threshold, the pixel is marked as a candidate anchor point. In this embodiment, the isotropic threshold is set to 0.05 and the proportional threshold is set to 5: a threshold that is too small will introduce noise points with weak responses, while a threshold that is too large will result in too sparse anchor points; the proportional threshold is used to constrain the proximity of two feature values, and a threshold that is too large will mix in edge structures extending along a single direction. After changing the chip surface texture or imaging device, the response value distribution of several typical samples can be re-statistically calibrated, and the above two thresholds can be adjusted accordingly.

[0029] After the candidate anchor points are determined, non-maximum suppression is performed within a circular neighborhood of 5 pixels centered on the candidate anchor point. Only the pixel with the largest isotropic response value in the neighborhood is retained as the anchor point center, and local redundant responses are eliminated. A 31×31 pixel square image block is cropped from the anchor point center as the anchor point block, which serves as the primitive for subsequent matching.

[0030] A phase consistency algorithm based on a two-dimensional Log-Gabor filter is used to process the first and second images respectively. A two-dimensional Log-Gabor filter bank covering multiple scales and directions is constructed, with a scale parameter of 4, a wavelength increment factor of 2.1, and a direction parameter of 6, i.e., dividing 180° into semicircles at 30° intervals, with a center frequency range of 0.1 to 0.4 and an angle standard deviation of 1.5. The first and second images are transformed into the frequency domain by Fourier transform, multiplied with the filter bank, and then mapped back to the spatial domain by inverse Fourier transform to obtain the even-symmetric and odd-symmetric filter responses of each pixel at different scales and directions. The amplitude response is calculated for the even-symmetric and odd-symmetric responses at the same scale and direction; the even-symmetric responses and odd-symmetric responses at different scales in the same direction are accumulated, and the local energy in that direction is calculated based on the two accumulation results; then the local energy in each direction is accumulated, and after deducting the noise compensation term, it is compared with the sum of the amplitude responses at all scales and directions to obtain the phase consistency value of the pixel. The noise compensation term is used to suppress low-energy noise response. A preset small constant is added to the sum of amplitude responses to avoid the denominator being 0. The phase consistency value is limited to between 0 and 1, and phase consistency maps of the first and second images are generated respectively.

[0031] A 31×31 pixel phase consistency block, identical in size to the anchor block, is extracted from the phase consistency map of the first image. On the phase consistency map of the second image, a 101×101 pixel search neighborhood is defined based on the anchor center coordinates. A 31×31 pixel matching box is slid within this neighborhood with a step size of 1 pixel. The floating-point phase consistency values ​​between the matching box and the phase consistency block of the first image are discretized into 64 gray levels position by position, and a two-dimensional joint histogram is constructed. The edge entropy and joint entropy of both are calculated, and the normalized mutual information score is obtained, typically between 1 and 2. The sliding position with the largest normalized mutual information score is taken as the candidate matching coordinate. The original coordinates of the anchor center are subtracted from the candidate matching coordinates to obtain the result including the horizontal offset. With vertical offset The initial displacement vector is obtained by traversing all anchor blocks and combining the initial displacement vectors of each anchor block to form a set of initial displacement vectors. Taking a certain anchor block as an example, its normalized mutual information score has a peak value of 1.68 in the search neighborhood, and the corresponding position is the candidate matching coordinate of the anchor block.

[0032] S2. Clustering removes outlier vectors to obtain the main matching cluster.

[0033] After obtaining a set of initial displacement vectors, the spatial position coordinates of each vector in the first image are extracted. and the displacement calculated from the search matching The position coordinates and displacement offsets are respectively standardized using Z-score zero-mean, and then concatenated into a four-dimensional feature vector with a uniform scale. The eigenvector set is composed of the four-dimensional eigenvectors corresponding to all initial displacement vectors. Spatial position coordinates Displacement offset is taken from the center coordinates of each anchor block. The initial displacement vector is obtained from the anchor block search and matching.

[0034] A Dirichlet process mixture model is employed as a density-based nonparametric Bayesian model for unsupervised clustering analysis of the feature vector set. This model can adaptively estimate the number of clusters according to the data distribution. A concentration hyperparameter for the Dirichlet process is defined. Initial value is 1, basic probability distribution The inverse normal Wissaud distribution is used as the conjugate prior. Its mean prior is the global four-dimensional mean of the entire eigenvector set, and its covariance prior is the global 4×4 covariance matrix. The inverse normal Wissaud distribution is used to describe the uncertainty of the mean vector and covariance matrix in the Gaussian distribution at the same time.

[0035] In the unsupervised clustering inference stage, Gibbs sampling iteration is used, with a maximum of 500 iterations. Before the first iteration, all feature vectors are initialized to the same data cluster as the initial class labels for Gibbs sampling. In the next iteration, based on the Chinese restaurant process, for the first... The eigenvector belongs to the existing eigenvector. The probability of each data cluster is proportional to In the formula For clusters Current number of samples, Let Gaussian probability density function be used. and For clusters The four-dimensional mean vector and 4×4 covariance matrix; the probability of generating a new data cluster is proportional to... The class labels of each feature vector are updated according to the two types of probabilities mentioned above. After one round of allocation, the posterior four-dimensional mean vector of each data cluster is recalculated. With 4×4 covariance matrix When the change in the number of samples in each cluster is less than 1% within 50 consecutive iterations, the iteration is considered to have converged. After convergence, based on the sampling records of the convergence phase, the frequency of each feature vector being assigned to each data cluster is counted, and it is assigned to the data cluster with the highest frequency. The highest frequency value is used as the posterior probability of its belonging to the corresponding data cluster.

[0036] Taking a clustering result as an example: after iterative convergence, three data clusters were obtained, with sample sizes of 265, 35, and 12 respectively. The data cluster containing 265 feature vectors was determined as the primary candidate cluster. Generally, after iterative convergence, all cluster assignment results are traversed, and the number of four-dimensional feature vectors contained in each data cluster is counted. The data cluster with the largest number of samples is marked as the primary candidate cluster, and the remaining data clusters are considered outliers. The four-dimensional posterior mean vector and 4×4 covariance matrix of the primary candidate cluster are extracted. For each four-dimensional feature vector in the primary candidate cluster, the multivariate normal distribution probability density function is called to obtain the multidimensional Gaussian probability density distribution value of the vector under the corresponding Gaussian distribution of the primary candidate cluster. Optionally, a preset small regularization term is added to the main diagonal before inverting the covariance matrix to ensure the positive definiteness and inversion stability of the covariance matrix.

[0037] The preset threshold is the fifth percentile of the multidimensional Gaussian probability density distribution values ​​of each member within the primary candidate cluster. The probability density distribution values ​​of each member are compared one by one; feature vectors below the preset threshold are identified as outliers and removed. The initial displacement vectors corresponding to the remaining feature vectors are retained, and these are combined to form a geometrically consistent primary matching cluster. Continuing the previous example, after removing outliers, the primary candidate cluster retains 248 feature vectors, and the corresponding 248 initial displacement vectors constitute the primary matching cluster. The distribution of the primary matching cluster and the outlier displacement vectors is as follows: Figure 2 As shown.

[0038] After clustering, the posterior responsibility matrix output by the Dirichlet process mixture model is extracted. Each row of this matrix corresponds to an eigenvector, and each column corresponds to a Gaussian distribution component. Matrix elements represent the probability that the corresponding eigenvector belongs to the corresponding Gaussian distribution component. After determining the Gaussian distribution component corresponding to the principal candidate cluster, the posterior responsibility of each member in the principal matching cluster on that component is read and saved as the posterior membership probability of each member for subsequent weight calculation in the local weighted affine transformation. Taking a member in the principal matching cluster as an example, its posterior membership probability is 0.94.

[0039] S3. Weighted affine transformation generates a displacement field to complete alignment.

[0040] In the two-dimensional coordinate system of the first image, a locally weighted affine transformation model is established for each pixel. For any target pixel position to be solved... traversal of contents Extracting the primary matching cluster of the nth member, the nth member... The initial spatial coordinates of each member when no displacement has occurred. Calculate the Euclidean distance between the target pixel and the member. Because members closer to the target pixel exert a stronger constraint on the local deformation at that location, Substituting into the Gaussian kernel function yields the spatial proximity weights. In the formula This is a bandwidth adjustment parameter. Target pixel position. Pixel grid taken from the first image, member starting space coordinates The coordinates are taken from the anchor point center and retained after outlier removal through clustering. Bandwidth Ideally, use 30 to 80 pixels: If the value is too small, only the nearest members participate in the interpolation, while the weights of distant members approach zero, making them susceptible to local noise. If the value is too large, distant members will participate excessively and smooth out the true local deformation. In this embodiment, 50 pixels is used. When changing the main matching cluster density or target interpolation accuracy requirements, it can be adjusted within the above range according to the average spacing of members.

[0041] After obtaining the spatial proximity weight, the normalized mutual information score and the membership posterior probability assigned by the clustering step are extracted from the member in the search matching. The spatial proximity weight, normalized mutual information score, and membership posterior probability are multiplied to obtain the comprehensive weight value of the member for the target pixel. The normalized mutual information score is taken from the anchor block search matching result, and the membership posterior probability is taken from the posterior responsibility matrix output by clustering. The comprehensive weight values ​​of all members in the main matching cluster are normalized so that the sum of the normalized weights of all members is 1. Let the sum of the normalized weights of all members be denoted as the first... The normalized weights of the members are This allows members with high spatial proximity, high normalized mutual information scores, and high posterior probability of membership to occupy a larger share of constraints in subsequent solutions.

[0042] Let the local affine transformation parameter matrix be... Extract the first matching cluster from the main matching cluster. The original homogeneous coordinates of each member And the matching target coordinates determined by the member's initial spatial coordinates and the corresponding initial displacement vector. Construct a cost function that minimizes the weighted prediction residuals. Summation permeates all clusters. Each member has a weighted average. The cost function with respect to the matrix Find the derivative and set it to zero to construct a system of normal linear equations. In the formula, the observation matrix From the original homogeneous coordinates of each member Concatenate line by line For diagonal elements that are all normalized weights Diagonal matrix, constant matrix Matching target coordinates of each member Arranged by row. Solving using weighted least squares, we obtain an analytical solution. ;when When the rank is less than 1, use singular value decomposition or pseudo-inverse to solve. This yields the six-parameter local affine transformation matrix of the current target pixel. The normalized weights are given in the equation. The larger the member, the better. The stronger the constraint, the more likely it is that spatially adjacent, reliably matched, and highly probable members will dominate the solution of the pixel's affine parameters, thus characterizing the non-uniform deformation at that location with locally dominant members. The homogeneous coordinates of the target pixel are then used. and Multiplying the original coordinates yields the projected coordinates, which are then subtracted from the original coordinates to obtain the predicted horizontal displacement of the current pixel. Vertical predicted displacement The displacement components of each pixel in the first image are solved independently, and then combined to generate a dense displacement field covering the entire image.

[0043] After obtaining the dense displacement field, the horizontal and vertical displacement component matrices are extracted and smoothed using a mask with a size of 15×15 and a standard deviation of [missing information]. A Gaussian smoothing filter with a value of 3 is used to filter out high-frequency local abnormal fluctuations in the displacement field; If the value is too small, high-frequency local abnormal fluctuations will not be adequately filtered out; if it is too large, the true local deformation will be overly smoothed. In this embodiment, a value of 3 is used based on the displacement field noise level. This is based on the pixel calibration coefficient of the microscopic imaging system. Convert pixel displacement to physical displacement. The unit is μm / pixel, which is obtained by calibration of the microscopic imaging system.

[0044] The verification was performed using a set of measured displacement field data: A region of interest (ROI) template, pre-defined in a unified pixel coordinate system, was used to mark the effective alignment evaluation area. This ROI template was a binary mask that marked whether each pixel in the filtered displacement field participated in the statistics, retaining only the pixels covering the hotspot alignment area of ​​the optoelectronic co-package chip. Based on the ROI template, the average displacement of the retained pixels in the filtered displacement field was calculated. The average vertical displacement is 1.58 μm. It is -2.34 μm, which constitutes the overall translational deviation vector characterizing the overall translation. Before calculating the displacement gradient, the coordinate grid is calibrated according to pixel calibration coefficients. The displacement was converted to physical coordinates to ensure that the displacement and coordinate quantities were in the same unit. Then, a 3×3 Sobel operator was used to differentiate the filtered horizontal and vertical displacement fields in both directions. The Jacobian partial derivatives were extracted pixel-by-pixel to construct a dimensionless local gradient matrix. The average of this matrix over the entire image yielded a 2×2 average displacement gradient matrix. .Will With the second-order identity matrix Add them together to construct the overall deformation gradient matrix. ,right Perform singular value decomposition In the formula and For orthogonal matrices Given a singular value diagonal matrix, the rotation components are reconstructed from the orthogonal matrix. Then, the overall rotational deviation angle is calculated from the rotational component elements. In the formula This is the element in the 2nd row and 1st column of the overall rotation matrix. This is the element in the first row and first column. The average displacement gradient matrix is ​​obtained from this set of data. Adding them together gives After singular value decomposition, the reconstructed result is It is 0.035. The value is 0.999. Substituting this into the equation, we get... It is 0.035 radians. This is combined with the overall translational deviation. Overall rotational deviation The alignment compensation command is generated, and the lateral offset, longitudinal offset and rotation angle deviation are converted into control command pulses for the piezoelectric displacement stage. This controls the multi-degree-of-freedom alignment platform to perform relative motion compensation, thereby completing the alignment and positioning of the microfluidic heat dissipation channel relative to the optoelectronic co-packaged chip.

[0045] To verify the effectiveness of the above schemes, tests were conducted on a dataset of microscopic images of optoelectronic co-packaged chips containing complex illumination variations and local noise interference. The test platform was a vision alignment system configured with a six-DOF high-precision displacement stage, and four control schemes were set up. Scheme A uses Harris corner detection, gray-level normalized cross-correlation, random sampling consistency elimination algorithm, and global affine transformation model; Scheme B, based on Scheme A, introduces Hessian matrix isotropic screening and phase consistency search matching; Scheme C, based on Scheme B, introduces nonparametric Bayesian clustering and multidimensional Gaussian probability density outlier elimination; Scheme D is the complete scheme, which, based on Scheme C, introduces a local affine transformation dense displacement field solution based on spatial proximity, normalized mutual information score, and membership posterior probability weighted comprehensively. Using the compensated average translational alignment error and average rotational alignment error as evaluation indicators, Scheme A has an average translational error of 4.52 μm and an average rotational error of 2.35 milliradians; Scheme B reduces these to 3.18 μm and 1.62 milliradians, Scheme C reduces them to 2.05 μm and 0.94 milliradians, and Scheme D further reduces them to 0.86 μm and 0.31 milliradians. The alignment errors are compared... Figure 3 As shown.

[0046] The results show that with the gradual introduction of isotropic anchor point matching, nonparametric Bayesian outlier removal, and local weighted affine transformation, the translation and rotation compensation errors decrease group by group, and the alignment accuracy of the complete scheme D is better than that of the first three control schemes.

[0047] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for aligning and positioning microfluidic heat dissipation channels suitable for optoelectronic co-packaging, characterized in that, include: S1. Acquire a first image of the optoelectronic co-packaged chip and a second image of the microfluidic heat dissipation channel. Calculate the Hessian matrix of the first image to select strongly isotropic regions as anchor blocks. Generate phase consistency maps of the first and second images respectively. Use normalized mutual information as a metric to search and match the anchor blocks between the phase consistency maps of the first and second images to obtain a set of initial displacement vectors. The isotropic regions are microstructure regions with significant local curvature responses in both orthogonal directions. S2. Construct a feature vector set from the initial displacement vectors and perform unsupervised analysis using a density-based nonparametric Bayesian model. Cluster analysis is performed to remove outlier vectors with probability densities below a preset threshold, resulting in the main matching cluster. The posterior probability of membership for each member of the main matching cluster is retained. The non-parametric Bayesian model is a Dirichlet process mixture model. S3: Based on the main matching cluster, a locally weighted affine transformation model is constructed. For each pixel position in the first image, the normalized weights of each member of the main matching cluster are determined based on spatial proximity, normalized mutual information score, and posterior probability of membership. This includes: for any target pixel position in the first image, using the center coordinates of the anchor blocks corresponding to each member of the main matching cluster as the starting spatial coordinates of each member, calculating the relationship between the target pixel position and each member's position. The Euclidean distance between the initial spatial coordinates of the members is used to obtain the spatial proximity weights by substituting the Euclidean distance into the Gaussian kernel function. The normalized mutual information score and membership posterior probability corresponding to each member of the main matching cluster are extracted. The spatial proximity weights, normalized mutual information scores, and membership posterior probabilities are multiplied to obtain the comprehensive weight value of each member. The comprehensive weight value of each member is normalized so that the sum of the normalized weights of each member is 1. The local affine transformation parameters of each pixel position are obtained by weighted calculation based on the normalized weights. A dense displacement field is generated by combining these parameters, including: for any target pixel position in the first image, based on the normalized values ​​of the target pixel positions... We weight the data and construct a cost function with the objective of minimizing the weighted square error between the predicted coordinates of each member's initial spatial coordinates after local affine transformation and the corresponding candidate matching coordinates of each member as the actual matching coordinates. We set the derivative of the cost function with respect to the parameters to 0 and construct a system of linear equations containing the observation matrix and the weight matrix. We then use the weighted least squares method to solve for the local affine transformation parameters of the target pixel position. We iterate through all pixel positions and calculate the deformation displacement according to the corresponding local affine transformation parameters. We combine these to generate a dense displacement field covering the entire image and determine the alignment relationship between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip.

2. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 1, characterized in that, The process of selecting strongly isotropic regions as anchor blocks includes: constructing a two-dimensional Hessian matrix by calculating the second-order partial derivatives of the horizontal, vertical, and mixing directions for each pixel in the first image; solving for a first eigenvalue with a larger absolute value and a second eigenvalue with a smaller absolute value; using the absolute value of the second eigenvalue as the isotropic response value; marking the pixel as a candidate anchor point when the isotropic response value is greater than a preset isotropic threshold and the ratio of the absolute values ​​of the two eigenvalues ​​is less than a preset proportional threshold; suppressing non-maximum values ​​of the candidate anchor points within a preset local neighborhood; taking the pixel with the maximum local response value as the anchor point center; and extracting an image block of a preset size as the anchor block.

3. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 1, characterized in that, The step of searching for and matching anchor blocks between the phase consistency maps of the first and second images includes: extracting the amplitude and phase information of the first and second images at different scales and directions using a two-dimensional Log-Gabor filter bank, generating phase consistency maps of the first and second images respectively, and cropping the phase consistency block of the first image corresponding to the anchor block; sliding the matching box within a preset search neighborhood of the phase consistency map of the second image, calculating the normalized mutual information score between the matching box and the phase consistency block of the first image at each position, and taking the position with the maximum score as the candidate matching coordinate; taking the center pixel coordinate of the anchor block as the center coordinate of the anchor block, and calculating the initial displacement vector of the anchor block based on the difference between the center coordinate of the anchor block and the candidate matching coordinate.

4. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 1, characterized in that, The unsupervised clustering analysis using a density-based nonparametric Bayesian model includes: separating the spatial horizontal and vertical coordinates and horizontal and vertical displacement components of each initial displacement vector in the first image; performing Z-score zero-mean standardization on the spatial horizontal and vertical coordinates and displacement components respectively; concatenating the standardized spatial horizontal and vertical coordinates and displacement components to construct a four-dimensional feature vector with a uniform scale and forming a feature vector set; using a nonparametric Bayesian model for iterative inference, calculating the probability of each feature vector belonging to each existing data cluster and generating a new data cluster in each iteration, and updating the cluster affiliation of each feature vector; recalculating the mean vector and covariance matrix of each data cluster based on the updated cluster affiliation; and determining the data cluster to which each feature vector belongs after iterative convergence and outputting the posterior probability of each feature vector belonging to the corresponding data cluster.

5. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 4, characterized in that, The process of removing outlier vectors with probability densities below a preset threshold to obtain the main matching cluster includes: counting the number of feature vectors contained in each data cluster in the clustering results, and marking the data cluster containing the most feature vectors as the main candidate cluster; calculating the multidimensional Gaussian probability density distribution value corresponding to each feature vector in the main candidate cluster based on the mean vector and covariance matrix of the main candidate cluster; identifying feature vectors with multidimensional Gaussian probability density distribution values ​​below a preset threshold as outliers and removing them; and retaining the initial displacement vectors corresponding to the feature vectors that were not removed, and combining them to form a main matching cluster with geometric consistency.

6. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 5, characterized in that, When calculating the multidimensional Gaussian probability density distribution value, a preset small regularization term is added to the main diagonal of the covariance matrix before inverting the covariance matrix to ensure the positive definiteness and inversion stability of the covariance matrix.

7. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 1, characterized in that, Determining the alignment relationship between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip includes: applying a smoothing filter to the dense displacement field to filter out high-frequency local abnormal fluctuations in the displacement field; extracting the average horizontal displacement and average vertical displacement of the filtered dense displacement field as the overall translational deviation between the microfluidic heat dissipation channel and the optoelectronic co-packaged chip; calculating the average displacement gradient matrix of the filtered dense displacement field, adding it to the identity matrix to construct the overall deformation gradient matrix, extracting the rotational component from the singular value decomposition of the overall deformation gradient matrix as the overall rotational deviation; and combining the overall translational deviation and the overall rotational deviation to generate an alignment compensation command for the microfluidic heat dissipation channel relative to the optoelectronic co-packaged chip to complete the alignment.

8. The microfluidic heat dissipation channel alignment and positioning method for optoelectronic co-packaging according to claim 1, characterized in that, The acquisition of the first image of the optoelectronic co-packaged chip and the second image of the microfluidic heat dissipation channel includes: using a charge-coupled device camera to acquire a surface image of the optoelectronic co-packaged chip as the first image, and using an infrared transmission microscopy system to acquire an internal structure image of the microfluidic heat dissipation channel as the second image; based on the calibration parameters of the microscopy system, performing scale normalization, rotation correction, and coordinate origin alignment on the first and second images, so that the first and second images are mapped to a unified pixel coordinate system.

Images