Touch module automatic laminating and deviation correcting method and system

Abstract

This invention provides an automatic bonding and correction method and system for touch modules. The method includes acquiring reference marker images of a first module and a second module; extracting corner points in the core region of the markers and unifying them to a visual measurement coordinate system; matching corresponding corner points to generate initial displacement vectors; calculating position confidence weights by combining the gradient direction distribution of the corner point neighborhood, local second derivative eigenvalues, and structural entropy; and solving the best-fit affine transformation model using an iterative reweighted least squares method based on the weights; determining the pose dispersion based on the maximum eigenvalue of the residual covariance matrix; decomposing the affine model to obtain translation, rotation, and non-rigid deformation information; and adaptively adjusting the dispersion threshold according to the confidence weight distribution; and calculating the correction amount and performing bonding when the pose dispersion and deformation degree meet the judgment conditions.

Description

Technical Field

[0001] This application belongs to the field of visual feedback, and in particular relates to a method and system for automatic alignment and correction of touch modules. Background Technology

[0002] In the production of touch modules, the automated bonding process is a critical step affecting product yield and optical display quality. This process typically requires precise alignment of two modules to be bonded, followed by pressing after the positional requirements are met. Current industrial production often employs machine vision-based automated bonding and correction systems. These systems use industrial cameras to capture images of reference marks on each module, extract image features to identify relative positional deviations between modules, and calculate translational and rotational compensation amounts. This information then drives a motion control platform to perform fine-tuning and correction, thereby completing the bonding of the touch modules.

[0003] However, in actual production environments, the acquired reference marker images are prone to distortion and noise due to factors such as pre-processing errors, differences in material physical properties, changes in ambient lighting, and minor defects or stains on the module surface. This leads to insufficient stability in position feature extraction, matching errors, and outliers, which in turn interfere with subsequent pose calculations. Furthermore, flexible or ultra-thin touch modules are prone to slight local non-rigid deformation during handling and clamping. Traditional alignment correction methods typically approximate the module as an ideal rigid body, lacking quantitative assessment of the reliability of each position point and failing to distinguish the impact of local deformation on pose calculation. During the overall alignment error consistency assessment, if adaptive judgment cannot be made based on the current feature extraction quality, the output correction amount may be significantly deviated under conditions of interference or deformation, resulting in misalignment, air bubbles, or even product scrap. Therefore, how to suppress the influence of feature noise, identify local module deformation, and achieve reliable adaptive bonding pose calculation under production conditions has become a pressing technical problem in the touch module manufacturing field. Summary of the Invention

[0004] To address the problems of existing technologies, such as feature extraction being susceptible to noise interference, abnormal matching points affecting pose calculation, difficulty in identifying local non-rigid deformations, and lack of adaptive adjustment of correction thresholds, this invention provides an automatic bonding and correction method and system for touch modules.

[0005] In a first aspect, the present invention provides an automatic bonding and correction method for a touch module, comprising:

[0006] Images of multiple reference marks on the first and second modules to be bonded are acquired. Corner feature sets are extracted within the preset core area of ​​each reference mark. The position of each corner point in the coordinate system of its respective module is determined. The positions of each corner point are unified to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship. The corresponding corner points of the first and second modules are matched to generate multiple sets of initial displacement vectors.

[0007] The initial position confidence of the corner point is calculated based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point. The structural entropy of the neighborhood image of the corner point is calculated and the initial position confidence is modulated using the structural entropy to generate position confidence weights. Multiple initial displacement vectors are weighted using the position confidence weights, and the best-fit affine transformation model is solved by iterative reweighted least squares method.

[0008] Calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector, and define the maximum eigenvalue of the covariance matrix as the dispersion of the overall pose. Decompose the best-fit affine transformation model into translation components, rotation matrix, and non-rigid deformation matrix. Adaptively adjust the preset dispersion benchmark reference value according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, calculate the correction amount based on the translation components and rotation matrix to complete the fitting.

[0009] On the other hand, the present invention also provides an automatic bonding and correction system for touch modules, comprising the following modules:

[0010] The generation module is used to acquire images of multiple reference marks on the first module and the second module to be bonded, extract corner feature sets in the preset core area of ​​each reference mark, determine the position of each corner point in the coordinate system of its respective module, unify the positions of each corner point to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship, and match the corresponding corner points of the first module and the second module to generate multiple sets of initial displacement vectors.

[0011] The solution module is used to calculate the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point, calculate the structural entropy of the neighborhood image of the corner point and use the structural entropy to modulate the initial position confidence to generate position confidence weights, use the position confidence weights to weight multiple sets of initial displacement vectors, and solve the best-fit affine transformation model by iterative reweighted least squares method.

[0012] The calculation module is used to calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector, and define the largest eigenvalue of the covariance matrix as the dispersion of the overall pose. The best-fit affine transformation model is decomposed into translation components, rotation matrix, and non-rigid deformation matrix. The preset dispersion benchmark reference value is adaptively adjusted according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, the correction amount is calculated based on the translation components and rotation matrix to complete the fitting.

[0013] This invention extracts and matches corner features in the core region of a reference marker to generate an initial displacement vector in a unified visual measurement coordinate system. Combining the gradient direction distribution of the corner's neighborhood, the eigenvalues ​​of the local second-order derivative matrix, and structural entropy, it calculates the position confidence weight of each corner. An iterative reweighted least squares method is then used to solve for the best-fit affine transformation model, reducing the impact of image noise and abnormal matching points on pose calculation. Furthermore, the overall pose dispersion is quantified by the maximum eigenvalue of the residual vector covariance matrix. The affine model is decomposed to obtain translation, rotation, and non-rigid deformation information. The dispersion judgment threshold is adaptively adjusted based on the confidence weight distribution, thereby improving the stability of the correction calculation and the reliability of the bonding process, ultimately enhancing the bonding accuracy and product yield of the touch module. Attached Figure Description

[0014] Figure 1 A flowchart of the first embodiment;

[0015] Figure 2 A top-view diagram showing the layout of reference marks for the first and second modules;

[0016] Figure 3 This is a schematic diagram illustrating the calculation of the structural entropy and confidence weight of the neighborhood of a corner point. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the first embodiment, the present invention proposes an automatic bonding and correction method for touch modules, such as... Figure 1 As shown, it includes:

[0019] S1. Acquire images of multiple reference marks on the first module and the second module to be bonded. Extract corner feature sets within the preset core area of ​​each reference mark. Determine the position of each corner point in its respective module coordinate system. Unify the positions of each corner point to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship. Match the corresponding corner points of the first module and the second module to generate multiple sets of initial displacement vectors.

[0020] High-resolution industrial cameras combined with coaxial light sources are used to acquire high-definition image data of the first and second modules. A normalized cross-correlation (NCC) template matching algorithm is employed to locate multiple reference markers in the acquired images. Reference markers are placed at the four corners of the first and second modules. Each module establishes its own independent coordinate system. A top-view diagram of the reference marker placement for the first and second modules is shown below. Figure 2 As shown.

[0021] A specified pixel region extending outward from the center of each reference mark is selected as the preset core region. Within this region, a corner detection algorithm based on structural tensor response is run to extract the corner feature set.

[0022] The camera intrinsic and extrinsic parameter matrices obtained using the Zhang Zhengyou calibration method are used to correct image distortion. Based on a pre-defined transformation relationship from pixel coordinates to the visual measurement coordinate system, the two-dimensional coordinates of corner points in the image pixel coordinate system are converted into planar position coordinates in a unified visual measurement coordinate system, using micrometers as the uniform unit of length. Specifically, during the equipment debugging phase, the correspondence between image pixel coordinates and actual planar coordinates in the visual measurement coordinate system is obtained through a calibration board. The two-dimensional homography transformation matrix or affine transformation matrix is ​​solved and stored. Subsequently, the corner point pixel coordinates are substituted into this transformation matrix to obtain the planar position coordinates in the unified visual measurement coordinate system.

[0023] The Scale Invariant Feature Transform (SIFT) algorithm is used to extract feature descriptors for each corner point. Then, the matching algorithm from the FLANN fast nearest neighbor search library is used to test the ratio of the nearest neighbor to the second nearest neighbor distance between the corner feature descriptors of the first and second modules based on Euclidean distance, thus completing corner point pairing. Under the same visual measurement coordinate system, the corresponding first module corner point coordinate vector is subtracted from the paired second module corner point coordinate vector to calculate multiple sets of initial displacement vectors.

[0024] In some embodiments, the step of extracting a set of corner feature points within a preset core region of each reference marker and determining the position of each corner point in its respective module coordinate system includes:

[0025] Extract the grayscale image of the preset core region;

[0026] Calculate the first gray-level gradient in the horizontal direction and the second gray-level gradient in the vertical direction for each pixel in the image;

[0027] Calculate the square values ​​of the first gray-level gradient and the second gray-level gradient, as well as their product. Based on the calculation results of all adjacent pixels, obtain the Gaussian weighted cumulative value of the local window and construct the structure tensor matrix corresponding to each pixel.

[0028] Extract the principal and secondary eigenvalues ​​of the structure tensor matrix, calculate their product, and then calculate the product of the square of the sum of the two and the constant weight. Subtract the weighted square term from the product to obtain the corner response value of each pixel.

[0029] Select pixels whose corner response values ​​are greater than a preset value requirement as initial feature points;

[0030] Within a defined grid range, the single initial feature point with the largest corner response value is retrieved as the target corner point, and the Cartesian coordinates of the target corner point within the corresponding module are output as the desired position.

[0031] Extract a 64×64 pixel core region grayscale image, with pixel grayscale levels ranging from 0 to 255. Perform spatial convolution on the image using a 3×3 differential convolution kernel to obtain the first grayscale gradient in the horizontal direction and the second grayscale gradient in the vertical direction for each pixel.

[0032] Calculate the squares of the two gradients and their product, and set a local window of size 5×5. Using a Gaussian filter kernel with a standard deviation between 1.0 and 2.0 and an example value of 1.2, perform Gaussian weighted summation on the squares of the first gray-level gradient, the squares of the second gray-level gradient, and their product within the local window to obtain the corresponding elements in the structure tensor matrix.

[0033] The principal and secondary eigenvalues ​​of the structural tensor matrix are extracted using a matrix factorization algorithm. Their product is calculated, and the square of this sum is multiplied by a constant weight. This weighted square term is then subtracted from the product to obtain the corner response value for each pixel. The constant weight is used to adjust the sensitivity of corner detection, ranging from 0.04 to 0.06; in this example, it is set to 0.04.

[0034] When the corner response value of a certain pixel exceeds a preset value requirement, for example, setting the response threshold to 2% of the theoretical maximum response peak value within the region, i.e., 10... 5When the initial feature point is selected, it is marked as the initial feature point. A non-maximum suppression mechanism is applied to remove false corner points. Using the initial feature point as the center, a traversal comparison is performed within a defined grid range, such as a 3×3 pixel or 5×5 pixel neighborhood array. The single initial feature point with the largest corner response value within the neighborhood is retained as the target corner point. Using the corner response values ​​of each pixel within a preset neighborhood around the target corner point as fitting samples, a quadratic surface function with respect to the pixel's horizontal and vertical coordinates is established. The extreme point position of the quadratic surface function is then determined, and this extreme point position is used as the sub-pixel-level planar coordinates of the target corner point, for example, the sub-pixel position with coordinates (32.54, 41.27), for use in the matching calculation.

[0035] S2, calculate the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point, calculate the structural entropy of the neighborhood image of the corner point and use the structural entropy to modulate the initial position confidence to generate position confidence weights, use the position confidence weights to weight multiple sets of initial displacement vectors, and solve the best-fit affine transformation model by iterative reweighted least squares method.

[0036] For each corner point, a neighborhood image patch of fixed pixel size centered on that point is extracted. The Sobel operator is used to calculate the gradient magnitude of the neighborhood image in the horizontal and vertical directions, and the gradient direction of each pixel is solved by the arctangent function.

[0037] Multiple pixel-level Hessian matrices are constructed based on the second-order partial derivatives of the neighborhood image. These matrices are then weighted and averaged according to Gaussian weights within the neighborhood to obtain the local second-order derivative matrix for each corner point. The two eigenvalues ​​of this local second-order derivative matrix are solved, and the ratio of the smaller to the larger absolute value is calculated. This ratio, combined with the two-dimensional structural features of the gradient direction distribution, is used to obtain the initial position confidence of the corner point.

[0038] In the neighborhood image of the corner point, the gray-level distribution information and frequency domain amplitude distribution information are statistically analyzed. Based on this, the structural entropy value of the local image is calculated. The structural entropy is normalized using the maximum theoretical entropy value. The structural stability coefficient is obtained by subtracting the normalized structural entropy from 1. The structural stability coefficient is then multiplied by the initial position confidence to generate the position confidence weight.

[0039] Using the position confidence weights, a weighted displacement residual equation is constructed, and a weighted error objective function is built. Iterative reweighted least squares (IRLS) is performed, adjusting the penalty weights of each corner data point in each iteration based on the fitting residuals of the previous round to suppress the interference of outliers, until the weighted squared error function converges to a minimum value, thus solving for the best-fit affine transformation model that describes the coordinate mapping relationship from the first module to the second module.

[0040] In some embodiments, calculating the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point includes:

[0041] Calculate the directional gradient value of each pixel in the neighborhood image of the corner point, construct a directional gradient histogram based on the angle interval distribution, and extract the main peak direction and the secondary main peak direction in the directional gradient histogram;

[0042] The second-order partial derivatives of the horizontal and vertical axes and the mixed partial derivatives of the pixels in the neighborhood image are calculated respectively to establish the pixel-level Hessian matrix. The weighted average of multiple pixel-level Hessian matrices in the neighborhood of the corner point is then calculated to obtain the local second-order derivative matrix of the corresponding corner point.

[0043] Extract the minimum and maximum absolute eigenvalues ​​of the local second derivative matrix, and divide the minimum absolute eigenvalue by the maximum absolute eigenvalue to obtain the matrix deformation ratio parameter;

[0044] The matrix deformation ratio parameter is normalized using the positive exponential mapping operator to obtain the first evaluation parameter;

[0045] The second evaluation parameter is calculated based on the angle between the main peak direction and the secondary main peak direction, as well as the degree of balance between their frequencies.

[0046] The product of the first evaluation parameter and the second evaluation parameter is used as the confidence level of the initial position of the corner point.

[0047] A 16×16 pixel neighborhood around the corner point is selected as the analysis benchmark, and the gradient magnitude and gradient direction of each pixel in this region are calculated. The angular space from 0° to 360° is divided into multiple preset angular intervals, with a total of 36 intervals and a span of 10° for each interval. The frequency of the pixel gradient is accumulated in the corresponding angular interval as a weight to construct an directional gradient histogram.

[0048] The interval with the highest cumulative frequency in the histogram is selected as the primary peak direction interval. The interval with the highest cumulative frequency in the directional gradient histogram (excluding the primary peak direction interval) is selected as the candidate secondary primary peak direction interval. The angle between the primary peak direction and the candidate secondary primary peak direction is calculated. When the angle is within a preset two-dimensional corner structure angle range, the candidate secondary primary peak direction interval is determined as the secondary primary peak direction interval, and the ratio of the secondary primary peak frequency to the primary peak frequency is calculated as a second evaluation parameter. When the angle is not within the preset two-dimensional corner structure angle range, the second evaluation parameter is set to a preset low value, such as 0.1. The preset two-dimensional corner structure angle range is preferably 60° to 120°. This avoids misjudging straight line edges with an excessively high proportion of a single direction as high-confidence corner points.

[0049] The second-order partial derivatives of pixels in the neighborhood image are calculated in the X and Y directions using a 3×3 second-order difference operator, and the mixed second-order partial derivatives in the X and Y directions are calculated using a mixed difference operator to obtain the pixel-level Hessian matrix for each pixel. Multiple pixel-level Hessian matrices within a 16×16 neighborhood are then weighted and averaged using Gaussian weights to generate the local second-order derivative matrix for the corresponding corner point. The Gaussian weights are calculated based on the spatial distance between each pixel in the neighborhood and the target corner point, centered on the sub-pixel position of the target corner point. Pixels closer to the target corner point have higher weights, and pixels farther away have lower weights. The standard deviation of the Gaussian kernel can be set to 2 to 4 pixels, preferably 3 pixels.

[0050] The local second derivative matrix is ​​solved to extract its minimum and maximum absolute eigenvalues, and the two are divided to obtain the matrix deformation ratio parameter. The value of this deformation ratio parameter is between 0 and 1. If the minimum eigenvalue is 15 and the maximum eigenvalue is 50, then the ratio is 0.3.

[0051] The deformation ratio parameter ρ is normalized using a positive exponential mapping operator with a smoothing coefficient, where the smoothing coefficient α ranges from 0.1 to 0.5; the example value is 0.3. Calculate the first evaluation parameter; when ρ is 0.3 and α is 0.3, the first evaluation parameter is approximately 0.656.

[0052] If the frequency of the primary peak is 80 and the frequency of the secondary primary peak is 70, and the angle between the directions of the primary and secondary primary peaks is within the preset two-dimensional corner point structure angle range, then the second evaluation parameter is 70 / 80, or 0.875. Multiplying the first evaluation parameter 0.656 by the second evaluation parameter 0.875 yields approximately 0.574, and this product value is used as the confidence score output for the initial position of the corner point.

[0053] In some embodiments, calculating the structural entropy of the corner neighborhood image and using the structural entropy to modulate the initial position confidence to generate position confidence weights includes:

[0054] The proportion of pixels at all discrete gray levels in the neighborhood image of the corner point is statistically analyzed. The proportion of pixels at each gray level is multiplied by the corresponding logarithm with the natural constant as the base, the sum is taken and the negative is obtained to obtain the one-dimensional information distribution parameters.

[0055] Two-dimensional discrete Fourier transform is applied to obtain the spectral features of the neighborhood image, and the regions are divided according to the amplitude level and the probability of the number of frequency points in the corresponding regions is counted.

[0056] Multiply the probability of the number of each frequency point by the corresponding logarithm with the natural constant as the base, sum them and take the negative to obtain the frequency domain fluctuation distribution parameters;

[0057] The overall structural entropy is obtained by adding the one-dimensional information distribution parameters and the frequency domain fluctuation distribution parameters and taking the arithmetic mean.

[0058] The normalized structural entropy is obtained by normalizing the overall structural entropy result using the maximum theoretical entropy value of the neighborhood image.

[0059] The product of the structural stability coefficient obtained by subtracting the normalized structural entropy from 1 and the initial position confidence is defined as the position confidence weight.

[0060] For a given 21×21 pixel corner neighborhood image, scan the discrete gray levels within the region. For an 8-bit grayscale image, there are a total of 256 gray levels. Calculate the proportion of pixels of each gray level to the total number of pixels. The one-dimensional information distribution parameter is obtained by calculating using the Shannon entropy formula; the example calculation result is 3.25.

[0061] The fast two-dimensional discrete Fourier transform is applied to convert the neighborhood image from the spatial domain to the frequency domain, obtaining a spectral feature map in complex matrix form, and the amplitude of the frequency domain components is extracted. Within the frequency domain, the amplitude interval is divided into equal-width amplitude level regions on a logarithmic scale, with the number of regions ranging from 10 to 20 (16 levels are used in this example). The number of frequency points falling into each amplitude level region is counted, and the probability of it occupying all the total frequency points is calculated.

[0062] The frequency domain fluctuation distribution parameter is calculated by multiplying the probability of the number of frequency points at each amplitude level by the natural logarithm of the corresponding probability, taking the negative value, and summing the results. An example result is 2.15. The one-dimensional information distribution parameter calculated in the spatial domain is added to the frequency domain fluctuation distribution parameter, and the arithmetic mean is calculated, for example, 2.70, to obtain the overall structure entropy result. A schematic diagram illustrating the calculation of corner point neighborhood structure entropy and confidence weights is shown below. Figure 3 As shown.

[0063] Using the maximum theoretical entropy value of this neighborhood under an ideal uniform distribution, for example, by averaging the maximum entropy of the 256-level spatial domain and the 16-level frequency domain to obtain a constant of 4.16, the overall structural entropy result is normalized by division, resulting in a normalized structural entropy of 0.649. Subtracting the normalized structural entropy of 0.649 from 1 yields a structural stability coefficient of 0.351; multiplying the initial position confidence score of 0.574 by the structural stability coefficient of 0.351 generates a position confidence weight of approximately 0.201.

[0064] In some embodiments, the step of weighting multiple initial displacement vectors using position confidence weights and solving for the best-fit affine transformation model using an iterative reweighted least squares method includes:

[0065] A two-dimensional homogeneous initial prediction transformation matrix is ​​established, the position coordinates of each corner point in a unified visual measurement coordinate system are extracted, and the position coordinates are substituted into the two-dimensional homogeneous initial prediction transformation matrix to perform forward mapping to obtain the estimated prediction position points.

[0066] Calculate the difference between the estimated predicted location point and the corresponding actual target location reference point to obtain the corresponding spatial geometric residual.

[0067] Divide the position confidence weight corresponding to each diagonal point by the sum of the absolute value of the spatial geometric residual and the preset positive bias constant to obtain the combined weight factor, and construct a diagonal weight matrix based on the combined weight factor. Introduce the diagonal weight matrix into the least squares error function to form a weighted nonlinear error cost function.

[0068] The weighted nonlinear error cost function is solved iteratively by calling the weighted iterative solution method to obtain the pose parameter update amount of the current iteration step, and the global root mean square difference between the spatial geometric residuals corresponding to the current iteration step and the previous iteration step is calculated.

[0069] When the global root mean square difference is less than the preset tolerance constant, the update loop calculation is determined to end, and the transformation matrix corresponding to the current round is taken as the best-fit affine transformation model.

[0070] The two-dimensional initial prediction transformation matrix representing the pose transformation is initialized as a 3×3 unit diagonal matrix. The planar position coordinates of each corner point extracted from the first module in the unified visual measurement coordinate system are transformed into homogeneous coordinate vectors containing 1, and substituted into the initial transformation matrix to perform linear forward mapping, projecting them onto the corresponding positions of the second module in the same visual measurement coordinate system to obtain the estimated predicted position points of each feature point.

[0071] Calculate the Euclidean distance between each estimated predicted location point and the actual matched target reference point in the second module to obtain the spatial geometric residual for each diagonal point pair. Extract the final position confidence weight for each diagonal point. Use a positive bias constant, with a constant range of 10. -5 Up to 10 -3 Between, the example is set to 10 -4 The combined weighting factor is obtained by dividing the position confidence weight by the sum of the absolute value of the spatial geometric residual and the bias constant. The spatial geometric residual and the positive bias constant use the same unit of length in a unified visual measurement coordinate system.

[0072] A diagonal weight matrix is ​​constructed using the combined weight factors, and this diagonal weight matrix is ​​introduced into the least squares error function to form a weighted nonlinear error cost function. The Gauss-Newton iterative algorithm is called to locally linearize the weighted nonlinear error cost function, construct the Jacobian matrix, and solve for the pose parameter update amount in the current iteration step. The pose parameter update amount is accumulated into the current transformation matrix parameters to complete one parameter update.

[0073] The difference between the spatial geometric residual output in the current step and the residual from the previous iteration is calculated. The global root mean square difference of this residual difference is then compared with a preset tolerance constant. The tolerance constant ranges from 0.001 to 0.01 micrometers, with an example value of 0.005 micrometers. When the global root mean square difference is less than the tolerance constant, the iterative operation is considered converged, and the loop ends. The retained 3×3 transformation matrix is ​​extracted as the best-fit affine transformation model.

[0074] S3, calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector and define the largest eigenvalue of the covariance matrix as the dispersion of the overall pose, decompose the best-fit affine transformation model into translation components, rotation matrix and non-rigid deformation matrix, and adaptively adjust the preset dispersion benchmark reference value according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, calculate the correction amount based on the translation components and rotation matrix to complete the fitting.

[0075] Substitute the position coordinates of the first module corner point in the unified visual measurement coordinate system into the aforementioned best-fit affine transformation model to calculate the predicted position coordinates of the second module corner point. Subtract the corresponding position coordinates of the first module corner point from the predicted position coordinates to obtain the displacement vector predicted by the model. Subtract the displacement vector predicted by the model from the initial displacement vector to obtain the residual vector of each paired point.

[0076] Concatenate the residual vectors of all matched corner points column-wise to construct a residual matrix. Calculate the mean vector of each residual vector, and subtract the mean vector from each residual vector to obtain a centered residual matrix. Multiply the centered residual matrix by its transpose and divide by the total number of corner point samples minus 1 to obtain the covariance matrix of the residual vectors. Calculate the eigenvalue sequence of this covariance matrix, extract the largest eigenvalue, and define it as the dispersion of the overall pose fluctuation state.

[0077] The linear transformation part of the best-fit affine transformation model is decomposed into the product of an orthogonal matrix and a symmetric positive definite matrix using an extreme decomposition algorithm. The orthogonal matrix represents the rotation matrix, and the symmetric positive definite matrix represents the non-rigid deformation components. The translation vector in the affine transformation model is extracted as the translation component.

[0078] Calculate the mean, standard deviation, and median of the confidence weights for all matching points. Generate an adaptive adjustment coefficient greater than zero based on the mean, standard deviation, and median. Use the adaptive adjustment coefficient to scale the preset dispersion benchmark reference value of the system to generate a first threshold.

[0079] The Frobenius norm of the difference between the non-rigid deformation matrix and the identity matrix of the same order is calculated to obtain a deformation degree value used to characterize the degree of non-rigid deformation. The second threshold is predetermined based on the statistical distribution of the deformation degree values ​​in qualified fitting samples. Specifically, multiple sets of qualified fitting historical samples are extracted, and the mean and standard deviation of their deformation degree values ​​are calculated. The sum of the mean and a preset multiple of the standard deviation is used as the second threshold. The second threshold is a dimensionless threshold, preferably ranging from 0.03 to 0.08, with an example value of 0.05.

[0080] The system determines whether the overall pose dispersion is less than a first threshold and whether the deformation degree is less than a second threshold. When both conditions are met, it indicates that the pose data has the reliability to continue the correction calculation, and the non-rigid deformation is within the allowable range of the process. At this point, based on the decomposition results of the affine transformation model, the yaw angle correction is extracted from the rotation matrix using inverse trigonometric functions, and the horizontal and vertical translation distance corrections are extracted from the translation components. The controller sends the yaw angle correction and the translation distance correction to the programmable logic controller (PLC) via an industrial communication protocol. The PLC generates corresponding pulse control signals based on the received correction values, driving a multi-axis servo motor to move the bonding robot arm to adjust the planar pose of the first module and complete the touch module bonding process. When the overall pose dispersion is not less than the first threshold, or the deformation degree is not less than the second threshold, bonding stops and an alignment error message is output.

[0081] In some embodiments, the decomposition of the best-fit affine transformation model into translation components, rotation matrices, and non-rigid deformation matrices includes:

[0082] From the three-row, three-column standard parameter matrix corresponding to the best-fit affine transformation model, extract the first two data points of the third column to construct translation components used to control the horizontal and vertical offsets.

[0083] Extract the data from the intersection of the first two rows and the first two columns of the standard parameter matrix to form a core linear transformation matrix with two rows and two columns;

[0084] Singular value decomposition is performed on the core linear transformation matrix to obtain the left singular value matrix, the diagonal singular value matrix, and the right singular value matrix.

[0085] Multiply the left singular value matrix by the transpose of the right singular value matrix to generate a rotation matrix representing a pure rotation action.

[0086] Multiply the right singular value matrix, the diagonal singular value matrix, and the transpose of the right singular value matrix in sequence to generate a non-rigid deformation matrix representing the shear deformation and non-proportional expansion effect.

[0087] For the converged 3×3 best-fit affine transformation standard parameter matrix, locate the third column of the parameter matrix and extract the two scalar data corresponding to the first two rows, for example, extract the... =15.3μm and =-8.7μm. Reorganize these two values ​​into a two-dimensional column vector format, and output the translation components. Extract the intersection of the first two rows and the first two columns of four elements, for example, a subarray with element values ​​of 0.99, -0.03, 0.04, and 1.01, and construct a 2×2 order core linear transformation matrix.

[0088] Initiate a singular value decomposition process on the 2×2 core linear transformation matrix, splitting it into a product sequence of three matrices: a left singular value matrix, a diagonal singular value matrix, and a right singular value matrix.

[0089] Multiplying the left singular value matrix by the transpose of the right singular value matrix generates a 2×2 rotation matrix representing a pure rotation effect, for example, solving for a rotation angle of 2.1°. The diagonal singular value matrix reflecting the deformation is retained, and the right singular value matrix is ​​multiplied by its transpose at both ends. This transformation generates a non-rigid deformation matrix, providing a basis for control judgment.

[0090] In some embodiments, the step of adaptively adjusting a preset dispersion benchmark reference value based on the statistical distribution of location confidence weights to generate a first threshold includes:

[0091] Extract and record the position confidence weights calculated from all corner point matching, and arrange them to construct a one-dimensional numerical array;

[0092] Calculate the expected mean and standard deviation of the confidence weights for all locations in a one-dimensional numerical array;

[0093] The first difference correction term is calculated by adding the product of the preset dimensionless basic adjustment constant and the standard deviation and the preset deviation adjustment coefficient.

[0094] Extract the median of the one-dimensional numerical array, calculate the root mean square of the median and the expected mean, and subtract the root mean square from 1 to obtain the second difference correction term;

[0095] The first difference correction term and the second difference correction term are added together, and the amplitude is limited by a preset minimum positive value to obtain an adaptive adjustment coefficient greater than zero. The dispersion benchmark reference value is divided by the adaptive adjustment coefficient, and the new value obtained by division is used as the first threshold.

[0096] For the corner points that have successfully completed the association mapping in the current detection batch, such as a total of 64 pairs of matching features, the position confidence weights are extracted, arranged according to the numerical value, and a one-dimensional numerical array is generated.

[0097] Iterate through all data elements in the one-dimensional numerical array, calculate their expected average (e.g., 0.78), and calculate the standard deviation parameter representing the dispersion of the weights (e.g., 0.12). Using a given dimensionless basic adjustment constant (ranging from 1.5 to 3.0, with a fixed value of 2.0 in this example) and a preset deviation adjustment factor (set between 1.0 and 2.5, with 1.5 used in this example), multiply the standard deviation by the deviation adjustment factor to obtain 0.18, and add this to the dimensionless basic adjustment constant of 2.0 to calculate the first difference correction term of 2.18.

[0098] Extract the median value from the one-dimensional numerical array, such as 0.81. Calculate the square root of half the sum of the squares of the median and the expected mean, yielding a root mean square (RMS) of approximately 0.795. Subtract the RMS from 1 to obtain the second difference correction term, which in the example is approximately 0.205. Add the first difference correction term (2.18) to the second difference correction term (0.205) to output the adaptive adjustment coefficient, which in the example is approximately 2.385.

[0099] Extract a baseline reference value for the dispersion, for example, 10.0 square micrometers. Divide the baseline reference value by the generated adaptive adjustment coefficient, and the resulting numerical index is used as the first threshold, for example, the new limit is approximately 4.19 square micrometers. This first threshold serves as the reference limit for evaluating the overall pose dispersion.

[0100] In a second embodiment, the present invention also proposes an automatic bonding and correction system for a touch module, comprising:

[0101] The generation module is used to acquire images of multiple reference marks on the first module and the second module to be bonded, extract corner feature sets in the preset core area of ​​each reference mark, determine the position of each corner point in the coordinate system of its respective module, unify the positions of each corner point to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship, and match the corresponding corner points of the first module and the second module to generate multiple sets of initial displacement vectors.

[0102] The solution module is used to calculate the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point, calculate the structural entropy of the neighborhood image of the corner point and use the structural entropy to modulate the initial position confidence to generate position confidence weights, use the position confidence weights to weight multiple sets of initial displacement vectors, and solve the best-fit affine transformation model by iterative reweighted least squares method.

[0103] The calculation module is used to calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector, and define the largest eigenvalue of the covariance matrix as the dispersion of the overall pose. The best-fit affine transformation model is decomposed into translation components, rotation matrix, and non-rigid deformation matrix. The preset dispersion benchmark reference value is adaptively adjusted according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, the correction amount is calculated based on the translation components and rotation matrix to complete the fitting.

[0104] In some embodiments, the step of extracting a set of corner feature points within a preset core region of each reference marker and determining the position of each corner point in its respective module coordinate system includes:

[0105] Extract the grayscale image of the preset core region;

[0106] Calculate the first gray-level gradient in the horizontal direction and the second gray-level gradient in the vertical direction for each pixel in the image;

[0107] Calculate the square values ​​of the first gray-level gradient and the second gray-level gradient, as well as their product. Based on the calculation results of all adjacent pixels, obtain the Gaussian weighted cumulative value of the local window and construct the structure tensor matrix corresponding to each pixel.

[0108] Extract the principal and secondary eigenvalues ​​of the structure tensor matrix, calculate their product, and then calculate the product of the square of the sum of the two and the constant weight. Subtract the weighted square term from the product to obtain the corner response value of each pixel.

[0109] Select pixels whose corner response values ​​are greater than a preset value requirement as initial feature points;

[0110] Within a defined grid range, the single initial feature point with the largest corner response value is retrieved as the target corner point, and the Cartesian coordinates of the target corner point within the corresponding module are output as the desired position.

[0111] In some embodiments, calculating the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point includes:

[0112] Calculate the directional gradient value of each pixel in the neighborhood image of the corner point, construct a directional gradient histogram based on the angle interval distribution, and extract the main peak direction and the secondary main peak direction in the directional gradient histogram;

[0113] The second-order partial derivatives of the horizontal and vertical axes and the mixed partial derivatives of the pixels in the neighborhood image are calculated respectively to establish the pixel-level Hessian matrix. The weighted average of multiple pixel-level Hessian matrices in the neighborhood of the corner point is then calculated to obtain the local second-order derivative matrix of the corresponding corner point.

[0114] Extract the minimum and maximum absolute eigenvalues ​​of the local second derivative matrix, and divide the minimum absolute eigenvalue by the maximum absolute eigenvalue to obtain the matrix deformation ratio parameter;

[0115] The matrix deformation ratio parameter is normalized using the positive exponential mapping operator to obtain the first evaluation parameter;

[0116] The second evaluation parameter is calculated based on the angle between the main peak direction and the secondary main peak direction, as well as the degree of balance between their frequencies.

[0117] The product of the first evaluation parameter and the second evaluation parameter is used as the confidence level of the initial position of the corner point.

[0118] In some embodiments, calculating the structural entropy of the corner neighborhood image and using the structural entropy to modulate the initial position confidence to generate position confidence weights includes:

[0119] The proportion of pixels at all discrete gray levels in the neighborhood image of the corner point is statistically analyzed. The proportion of pixels at each gray level is multiplied by the corresponding logarithm with the natural constant as the base, the sum is taken and the negative is obtained to obtain the one-dimensional information distribution parameters.

[0120] Two-dimensional discrete Fourier transform is applied to obtain the spectral features of the neighborhood image, and the regions are divided according to the amplitude level and the probability of the number of frequency points in the corresponding regions is counted.

[0121] Multiply the probability of the number of each frequency point by the corresponding logarithm with the natural constant as the base, sum them and take the negative to obtain the frequency domain fluctuation distribution parameters;

[0122] The overall structural entropy is obtained by adding the one-dimensional information distribution parameters and the frequency domain fluctuation distribution parameters and taking the arithmetic mean.

[0123] The normalized structural entropy is obtained by normalizing the overall structural entropy result using the maximum theoretical entropy value of the neighborhood image.

[0124] The product of the structural stability coefficient obtained by subtracting the normalized structural entropy from 1 and the initial position confidence is defined as the position confidence weight.

[0125] In some embodiments, the step of weighting multiple initial displacement vectors using position confidence weights and solving for the best-fit affine transformation model using an iterative reweighted least squares method includes:

[0126] A two-dimensional homogeneous initial prediction transformation matrix is ​​established, the position coordinates of each corner point in a unified visual measurement coordinate system are extracted, and the position coordinates are substituted into the two-dimensional homogeneous initial prediction transformation matrix to perform forward mapping to obtain the estimated prediction position points.

[0127] Calculate the difference between the estimated predicted location point and the corresponding actual target location reference point to obtain the corresponding spatial geometric residual.

[0128] Divide the position confidence weight corresponding to each diagonal point by the sum of the absolute value of the spatial geometric residual and the preset positive bias constant to obtain the combined weight factor, and construct a diagonal weight matrix based on the combined weight factor. Introduce the diagonal weight matrix into the least squares error function to form a weighted nonlinear error cost function.

[0129] The weighted nonlinear error cost function is solved iteratively by calling the weighted iterative solution method to obtain the pose parameter update amount of the current iteration step, and the global root mean square difference between the spatial geometric residuals corresponding to the current iteration step and the previous iteration step is calculated.

[0130] When the global root mean square difference is less than the preset tolerance constant, the update loop calculation is determined to end, and the transformation matrix corresponding to the current round is taken as the best-fit affine transformation model.

[0131] In some embodiments, the decomposition of the best-fit affine transformation model into translation components, rotation matrices, and non-rigid deformation matrices includes:

[0132] From the three-row, three-column standard parameter matrix corresponding to the best-fit affine transformation model, extract the first two data points of the third column to construct translation components used to control the horizontal and vertical offsets.

[0133] Extract the data from the intersection of the first two rows and the first two columns of the standard parameter matrix to form a core linear transformation matrix with two rows and two columns;

[0134] Singular value decomposition is performed on the core linear transformation matrix to obtain the left singular value matrix, the diagonal singular value matrix, and the right singular value matrix.

[0135] Multiply the left singular value matrix by the transpose of the right singular value matrix to generate a rotation matrix representing a pure rotation action.

[0136] Multiply the right singular value matrix, the diagonal singular value matrix, and the transpose of the right singular value matrix in sequence to generate a non-rigid deformation matrix representing the shear deformation and non-proportional expansion effect.

[0137] In some embodiments, the step of adaptively adjusting a preset dispersion benchmark reference value based on the statistical distribution of location confidence weights to generate a first threshold includes:

[0138] Extract and record the position confidence weights calculated from all corner point matching, and arrange them to construct a one-dimensional numerical array;

[0139] Calculate the expected mean and standard deviation of the confidence weights for all locations in a one-dimensional numerical array;

[0140] The first difference correction term is calculated by adding the product of the preset dimensionless basic adjustment constant and the standard deviation and the preset deviation adjustment coefficient.

[0141] Extract the median of the one-dimensional numerical array, calculate the root mean square of the median and the expected mean, and subtract the root mean square from 1 to obtain the second difference correction term;

[0142] The first difference correction term and the second difference correction term are added together, and the amplitude is limited by a preset minimum positive value to obtain an adaptive adjustment coefficient greater than zero. The dispersion benchmark reference value is divided by the adaptive adjustment coefficient, and the new value obtained by division is used as the first threshold.

[0143] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for automatic bonding and correction of a touch module, characterized in that, include: Images of multiple reference marks on the first and second modules to be bonded are acquired. Corner feature sets are extracted within the preset core area of ​​each reference mark. The position of each corner point in the coordinate system of its respective module is determined. The positions of each corner point are unified to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship. The corresponding corner points of the first and second modules are matched to generate multiple sets of initial displacement vectors. The initial position confidence of the corner point is calculated based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point. The structural entropy of the neighborhood image of the corner point is calculated and the initial position confidence is modulated using the structural entropy to generate position confidence weights. Multiple initial displacement vectors are weighted using the position confidence weights, and the best-fit affine transformation model is solved by iterative reweighted least squares method. Calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector, and define the maximum eigenvalue of the covariance matrix as the dispersion of the overall pose. Decompose the best-fit affine transformation model into translation components, rotation matrix, and non-rigid deformation matrix. Adaptively adjust the preset dispersion benchmark reference value according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, calculate the correction amount based on the translation components and rotation matrix to complete the fitting.

2. The method according to claim 1, characterized in that, The step of extracting a set of corner features within a preset core area of ​​each reference mark and determining the position of each corner point in its respective module coordinate system includes: Extract the grayscale image of the preset core region; Calculate the first gray-level gradient in the horizontal direction and the second gray-level gradient in the vertical direction for each pixel in the image; Calculate the square values ​​of the first gray-level gradient and the second gray-level gradient, as well as their product. Based on the calculation results of all adjacent pixels, obtain the Gaussian weighted cumulative value of the local window and construct the structure tensor matrix corresponding to each pixel. Extract the principal and secondary eigenvalues ​​of the structure tensor matrix, calculate their product, and then calculate the product of the square of the sum of the two and the constant weight. Subtract the weighted square term from the product to obtain the corner response value of each pixel. Select pixels whose corner response values ​​are greater than a preset value requirement as initial feature points; Within a defined grid range, the single initial feature point with the largest corner response value is retrieved as the target corner point, and the Cartesian coordinates of the target corner point within the corresponding module are output as the desired position.

3. The method according to claim 1, characterized in that, The calculation of the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point includes: Calculate the directional gradient value of each pixel in the neighborhood image of the corner point, construct a directional gradient histogram based on the angle interval distribution, and extract the main peak direction and the secondary main peak direction in the directional gradient histogram; The second-order partial derivatives of the horizontal and vertical axes and the mixed partial derivatives of the pixels in the neighborhood image are calculated respectively to establish the pixel-level Hessian matrix. The weighted average of multiple pixel-level Hessian matrices in the neighborhood of the corner point is then calculated to obtain the local second-order derivative matrix of the corresponding corner point. Extract the minimum and maximum absolute eigenvalues ​​of the local second derivative matrix, and divide the minimum absolute eigenvalue by the maximum absolute eigenvalue to obtain the matrix deformation ratio parameter; The matrix deformation ratio parameter is normalized using the positive exponential mapping operator to obtain the first evaluation parameter; The second evaluation parameter is calculated based on the angle between the main peak direction and the secondary main peak direction, as well as the degree of balance between their frequencies. The product of the first evaluation parameter and the second evaluation parameter is used as the confidence level of the initial position of the corner point.

4. The method according to claim 1, characterized in that, The calculation of the structural entropy of the corner neighborhood image and the modulation of the initial position confidence using the structural entropy to generate position confidence weights include: The proportion of pixels at all discrete gray levels in the neighborhood image of the corner point is statistically analyzed. The proportion of pixels at each gray level is multiplied by the corresponding logarithm with the natural constant as the base, the sum is taken and the negative is obtained to obtain the one-dimensional information distribution parameters. Two-dimensional discrete Fourier transform is applied to obtain the spectral features of the neighborhood image, and the regions are divided according to the amplitude level and the probability of the number of frequency points in the corresponding regions is counted. Multiply the probability of the number of each frequency point by the corresponding logarithm with the natural constant as the base, sum them and take the negative to obtain the frequency domain fluctuation distribution parameters; The overall structural entropy is obtained by adding the one-dimensional information distribution parameters and the frequency domain fluctuation distribution parameters and taking the arithmetic mean. The normalized structural entropy is obtained by normalizing the overall structural entropy result using the maximum theoretical entropy value of the neighborhood image. The product of the structural stability coefficient obtained by subtracting the normalized structural entropy from 1 and the initial position confidence is defined as the position confidence weight.

5. The method according to claim 4, characterized in that, The process of weighting multiple initial displacement vectors using position confidence weights and then solving for the best-fit affine transformation model using an iterative reweighted least squares method includes: A two-dimensional homogeneous initial prediction transformation matrix is ​​established, the position coordinates of each corner point in a unified visual measurement coordinate system are extracted, and the position coordinates are substituted into the two-dimensional homogeneous initial prediction transformation matrix to perform forward mapping to obtain the estimated prediction position points. Calculate the difference between the estimated predicted location point and the corresponding actual target location reference point to obtain the corresponding spatial geometric residual. Divide the position confidence weight corresponding to each diagonal point by the sum of the absolute value of the spatial geometric residual and the preset positive bias constant to obtain the combined weight factor, and construct a diagonal weight matrix based on the combined weight factor. Introduce the diagonal weight matrix into the least squares error function to form a weighted nonlinear error cost function. The weighted nonlinear error cost function is solved iteratively by calling the weighted iterative solution method to obtain the pose parameter update amount of the current iteration step, and the global root mean square difference between the spatial geometric residuals corresponding to the current iteration step and the previous iteration step is calculated. When the global root mean square difference is less than the preset tolerance constant, the update loop calculation is determined to end, and the transformation matrix corresponding to the current round is taken as the best-fit affine transformation model.

6. The method according to claim 1, characterized in that, The process of decomposing the best-fit affine transformation model into translation components, rotation matrices, and non-rigid deformation matrices includes: From the three-row, three-column standard parameter matrix corresponding to the best-fit affine transformation model, extract the first two data points of the third column to construct translation components used to control the horizontal and vertical offsets. Extract the data from the intersection of the first two rows and the first two columns of the standard parameter matrix to form a core linear transformation matrix with two rows and two columns; Singular value decomposition is performed on the core linear transformation matrix to obtain the left singular value matrix, the diagonal singular value matrix, and the right singular value matrix. Multiply the left singular value matrix by the transpose of the right singular value matrix to generate a rotation matrix representing a pure rotation action. Multiply the right singular value matrix, the diagonal singular value matrix, and the transpose of the right singular value matrix in sequence to generate a non-rigid deformation matrix representing the shear deformation and non-proportional expansion effect.

7. The method according to claim 1, characterized in that, The step of adaptively adjusting the preset dispersion benchmark reference value based on the statistical distribution of the location confidence weights to generate the first threshold includes: Extract and record the position confidence weights calculated from all corner point matching, and arrange them to construct a one-dimensional numerical array; Calculate the expected mean and standard deviation of the confidence weights for all locations in a one-dimensional numerical array; The first difference correction term is calculated by adding the product of the preset dimensionless basic adjustment constant and the standard deviation and the preset deviation adjustment coefficient. Extract the median of the one-dimensional numerical array, calculate the root mean square of the median and the expected mean, and subtract the root mean square from 1 to obtain the second difference correction term; The first difference correction term and the second difference correction term are added together, and the amplitude is limited by a preset minimum positive value to obtain an adaptive adjustment coefficient greater than zero. The dispersion benchmark reference value is divided by the adaptive adjustment coefficient, and the new value obtained by division is used as the first threshold.

8. An automatic bonding and correction system for a touch module, characterized in that, include: The generation module is used to acquire images of multiple reference marks on the first module and the second module to be bonded, extract corner feature sets in the preset core area of ​​each reference mark, determine the position of each corner point in the coordinate system of its respective module, unify the positions of each corner point to the same visual measurement coordinate system according to the pre-calibrated coordinate transformation relationship, and match the corresponding corner points of the first module and the second module to generate multiple sets of initial displacement vectors. The solution module is used to calculate the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point, calculate the structural entropy of the neighborhood image of the corner point and use the structural entropy to modulate the initial position confidence to generate position confidence weights, use the position confidence weights to weight multiple sets of initial displacement vectors, and solve the best-fit affine transformation model by iterative reweighted least squares method. The calculation module is used to calculate the residual vector between each initial displacement vector and the displacement vector predicted by the best-fit affine transformation model, construct the covariance matrix of the residual vector, and define the largest eigenvalue of the covariance matrix as the dispersion of the overall pose. The best-fit affine transformation model is decomposed into translation components, rotation matrix, and non-rigid deformation matrix. The preset dispersion benchmark reference value is adaptively adjusted according to the statistical distribution of the position confidence weight to generate a first threshold. When the dispersion of the overall pose is less than the first threshold and the deformation degree represented by the non-rigid deformation matrix is ​​less than the second threshold, the correction amount is calculated based on the translation components and rotation matrix to complete the fitting.

9. The system according to claim 8, characterized in that, The step of extracting a set of corner features within a preset core area of ​​each reference mark and determining the position of each corner point in its respective module coordinate system includes: Extract the grayscale image of the preset core region; Calculate the first gray-level gradient in the horizontal direction and the second gray-level gradient in the vertical direction for each pixel in the image; Calculate the square values ​​of the first gray-level gradient and the second gray-level gradient, as well as their product. Based on the calculation results of all adjacent pixels, obtain the Gaussian weighted cumulative value of the local window and construct the structure tensor matrix corresponding to each pixel. Extract the principal and secondary eigenvalues ​​of the structure tensor matrix, calculate their product, and then calculate the product of the square of the sum of the two and the constant weight. Subtract the weighted square term from the product to obtain the corner response value of each pixel. Select pixels whose corner response values ​​are greater than a preset value requirement as initial feature points; Within a defined grid range, the single initial feature point with the largest corner response value is retrieved as the target corner point, and the Cartesian coordinates of the target corner point within the corresponding module are output as the desired position.

10. The system according to claim 8, characterized in that, The calculation of the initial position confidence of the corner point based on the gradient principal direction and the eigenvalue distribution of the local second derivative matrix of the neighborhood image of each corner point includes: Calculate the directional gradient value of each pixel in the neighborhood image of the corner point, construct a directional gradient histogram based on the angle interval distribution, and extract the main peak direction and the secondary main peak direction in the directional gradient histogram; The second-order partial derivatives of the horizontal and vertical axes and the mixed partial derivatives of the pixels in the neighborhood image are calculated respectively to establish the pixel-level Hessian matrix. The weighted average of multiple pixel-level Hessian matrices in the neighborhood of the corner point is then calculated to obtain the local second-order derivative matrix of the corresponding corner point. Extract the minimum and maximum absolute eigenvalues ​​of the local second derivative matrix, and divide the minimum absolute eigenvalue by the maximum absolute eigenvalue to obtain the matrix deformation ratio parameter; The matrix deformation ratio parameter is normalized using the positive exponential mapping operator to obtain the first evaluation parameter; The second evaluation parameter is calculated based on the angle between the main peak direction and the secondary main peak direction, as well as the degree of balance between their frequencies. The product of the first evaluation parameter and the second evaluation parameter is used as the confidence level of the initial position of the corner point.

Images