A method for improving the silk printing precision of a display module by using an image enhancement algorithm

Abstract

The application discloses a method for improving the silk printing precision of a display module by using an image enhancement algorithm and relates to the technical field of image enhancement. The method first collects the visible light and near-infrared waveband images of a display module substrate, suppresses the background noise of the substrate and enhances the weak texture of the silk printing edge through multi-spectral differential operation, and generates a fusion enhancement image; then a pixel-level texture generation diffusion model is adopted to complete the silk printing edge texture in combination with a boundary constraint condition to generate a high-resolution silk printing alignment reference image; finally, the alignment deviation correction amount is obtained through pixel-level hierarchical feature matching, and the matching result is fed back to optimize the multi-spectral waveband weight. The application synchronously realizes the collaborative optimization of the silk printing geometric alignment precision and the photoelectric performance of the display module, solves the problems of low imaging contrast of the transparent substrate, difficulty in weak texture extraction and display defects caused by silk printing deviation, improves the silk printing precision, alignment robustness and system real-time performance, and is suitable for the silk printing production scene of a high-precision display module.

Description

Technical Field

[0001] This invention discloses a method for improving the screen printing accuracy of display modules using image enhancement algorithms, which relates to the field of image enhancement technology. Background Technology

[0002] In the manufacturing process of display modules, screen printing is a crucial step that determines the bezel width and light-shielding performance of the product. Existing technologies typically employ machine vision-based alignment systems. These systems use a single-band visible light camera to acquire images of the substrate, and then identify the positions of screen printing marks through edge detection or template matching algorithms. This allows for the control of the screen printing stencil's movement to achieve pattern alignment. This approach primarily relies on the geometric features of the image for positional deviation calculations, and the processing is limited to image information in a single spectral dimension.

[0003] Display module substrates are typically transparent or semi-transparent, and the edges of silkscreen patterns often exhibit weak texture areas with gradients or blurring. Single-band visible light imaging struggles to penetrate transparent substrates to suppress background stray light interference, resulting in insufficient image contrast and an inability to clearly extract the weak texture features of the silkscreen edges. Furthermore, existing image enhancement processes only correct geometric positions, neglecting the impact of silkscreen pattern boundary offsets on pixel unit light extraction efficiency and crosstalk, leading to silkscreen alignment that merely meets geometric dimensional requirements without considering the optical performance of the display module. This makes it difficult to achieve synergistic optimization of silkscreen geometric accuracy and display performance. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for improving the screen printing accuracy of a display module using an image enhancement algorithm includes the following steps: acquiring visible light and near-infrared band images of the display module substrate and performing multispectral difference operations to suppress substrate background noise and enhance weak textures at the screen printing edges, thereby generating a fused and enhanced image; Based on the extraction of pixel unit structure parameters from the fused and enhanced image, the influence of the screen printing pattern on the photoelectric performance is calculated through photoelectric property simulation, and the optimal boundary constraint parameters are output. A pixel-level texture generation diffusion model is used to process the fused and enhanced image and the optimal boundary constraint parameters to generate a screen printing alignment reference image. Perform pixel-level hierarchical feature matching between the silkscreen alignment reference map and the real-time enhanced image, output the deviation correction amount, and feed the matching result back to the band weight optimization process of multispectral difference operation.

[0006] Preferably, the multispectral difference operation includes: constructing a Laplacian pyramid decomposition layer for visible light and near-infrared band images, and calculating the difference image of the dual-band images in the high-frequency decomposition layer to extract high-frequency texture features of the screen printing edge; The texture richness of the difference image is evaluated using the information entropy metric, and the weighting coefficients of the dual-band image in the fused and enhanced image are dynamically adjusted.

[0007] Preferably, the photoelectric characteristic simulation calculation includes: constructing a microscopic optical simulation field of the pixel unit of the display module, and simulating the occlusion and reflection effects of the change in the boundary position of the screen printing pattern on the light path of the pixel based on the Monte Carlo ray tracing principle; Establish a mapping matrix between screen printing boundary offset and light extraction efficiency and crosstalk index, and solve for the boundary range that minimizes photoelectric performance loss as the optimal boundary constraint parameter.

[0008] Preferably, the generation of the diffusion model using pixel-level texture includes: in the inverse denoising stage of the diffusion model, encoding the optimal boundary constraint parameters into a condition vector and injecting it into the intermediate feature layer of the denoising network through a cross-attention mechanism; A denoising network is used to semantically complete the missing texture of the silkscreen edges in the fused and enhanced image, generating a high-resolution silkscreen alignment reference map that conforms to photoelectric property constraints.

[0009] Preferably, the pixel-level hierarchical feature matching includes: constructing a Gaussian pyramid hierarchical structure of the silkscreen alignment reference map and extracting accelerated and robust feature points of the real-time enhanced image; The feature point matching and geometric transformation matrix calculation are performed layer by layer from the top to the bottom. The deviation correction amounts of horizontal translation, rotation angle and scaling factor are decomposed and obtained to eliminate mismatched feature points.

[0010] Preferably, it also includes: introducing a multi-objective particle swarm optimization algorithm to construct a fitness function by combining the sharpness evaluation value of the fused enhanced image with the silkscreen edge gradient energy function; By iteratively optimizing the combination of weighting coefficients for the dual-band image at each decomposition layer, and adjusting the particle search step size using inertial weights, the background noise suppression ratio is maximized while preserving the weak texture details of the silkscreen.

[0011] Preferably, it also includes: collecting historical silkscreen defect data of the display module to construct a training set, and training a long short-term memory network temporal prediction model; The boundary offset calculated by real-time simulation is input into the prediction model to predict the risk probability of the screen printing deviation causing the display to malfunction in the next moment. Based on the risk probability, the optimal boundary constraint parameters are dynamically adjusted to achieve predictive compensation of photoelectric properties.

[0012] Preferably, before utilizing the denoising network, the method further includes: constructing a graph structure representation of the geometric topology of the silkscreen pattern, and using a graph convolutional neural network to extract the spatial adjacency relationship features of the silkscreen pattern; By fusing spatial adjacency features with the latent space features of the denoising network, the spatial continuity of silkscreen texture generation is constrained.

[0013] Preferably, the calculation of the geometric transformation matrix includes: using a random sampling consensus algorithm to remove outlier outliers from the feature points, and using the least squares method to fit the affine transformation matrix of the interior point set; Based on the residual distribution of the affine transformation matrix, the feature point search radius and matching threshold of the next level are adaptively adjusted to optimize the computational efficiency of hierarchical matching.

[0014] Preferably, it also includes: monitoring the real-time load status of the image processing unit and establishing a computing resource scheduling cost function; When the load exceeds the preset threshold, the feature point extraction density of hierarchical feature matching is reduced, and the pyramid downsampling rate in multispectral difference operation is increased simultaneously to optimize the allocation of system computing resources while ensuring the real-time alignment of silkscreen printing.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention acquires visible and near-infrared dual-band images and performs multispectral difference operations. It utilizes the difference features of the dual bands at the high-frequency decomposition layer to extract weak textures at the screen printing edges, suppressing background noise on transparent substrates and solving the problem of insufficient contrast in single-band imaging. By constructing a microscopic optical simulation field to calculate the impact of the screen printing pattern on photoelectric performance, it outputs optimal boundary constraint parameters, incorporating the photoelectric characteristics of the display module into the image enhancement process, thus achieving synergistic optimization of screen printing geometric alignment accuracy and display performance.

[0016] 2. This invention utilizes a pixel-level texture generation diffusion model combined with boundary constraint parameters to generate an alignment reference map, thus completing the missing edge texture details. Through hierarchical feature matching and feedback mechanisms, it achieves refined correction of alignment deviations and dynamic adjustment of algorithm parameters, improving the robustness and accuracy of screen printing alignment. Attached Figure Description

[0017] Figure 1 This is a general technical flow diagram of the present invention; Figure 2 This is a flowchart of the multispectral difference and weight optimization process of this invention; Figure 3 This is a flowchart of the photoelectric property simulation and boundary constraint process of the present invention; Figure 4 Flowchart for generating alignment reference diagrams for the diffusion model of this invention; Figure 5This is a flowchart of the pixel-level hierarchical feature matching process of the present invention; Figure 6 This is a flowchart of the adaptive scheduling process for computing resources in this invention. Detailed Implementation

[0018] This specific embodiment provides a method for improving the screen printing accuracy of display modules using image enhancement algorithms, applied to a screen printing alignment control system for display modules. The system includes a visible light image acquisition unit, a near-infrared image acquisition unit, an image processing unit, a photoelectric characteristic simulation unit, a diffusion model inference unit, an alignment matching unit, and a screen printing execution unit, all coaxially arranged. The optical axes of the visible light image acquisition unit and the near-infrared image acquisition unit coincide, and their field of view covers the area to be screen printed on the display module substrate. The image processing unit is coupled to the image acquisition unit, the photoelectric characteristic simulation unit, the diffusion model inference unit, and the alignment matching unit, respectively. The output of the alignment matching unit is coupled to the parameter optimization input of the screen printing execution unit and the image processing unit.

[0019] Example 1: Please refer to the appendix Figure 1 The image processing unit simultaneously acquires visible light and near-infrared images of the display module substrate at the same time and within the same field of view through a visible light image acquisition unit and a near-infrared image acquisition unit. The visible light band ranges from 400nm to 760nm, and the near-infrared band ranges from 780nm to 940nm. The image processing unit performs preprocessing on the acquired dual-band images, including distortion correction based on calibration parameters and pixel-level registration based on the homography matrix, to align the dual-band images to the same pixel coordinate system.

[0020] refer to Figure 2 The image processing unit performs multispectral difference operations on the registered dual-band image. First, it performs grayscale normalization on the dual-band image. The grayscale normalization calculation formula is as follows: ; in, These are the pixel grayscale values ​​of the original image. , These are the minimum and maximum gray values ​​of the corresponding image, respectively. This represents the normalized pixel grayscale value.

[0021] The image processing unit performs a difference operation based on the normalized dual-band image. The calculation formula for the difference operation is as follows: ; in, This is a normalized image for visible light. This is a near-infrared normalized image. , These are the initial band weighting coefficients, with initial values ​​set to 0.6 and 0.4 respectively. This is a differential image. The differential operation utilizes the consistent optical response of the transparent substrate in both wavelengths to cancel low-frequency background noise, while simultaneously enhancing weak edge texture based on the response differences of the silkscreen edges in both wavelengths. The image processing unit fuses the high-frequency texture component of the differential image with the low-frequency background component of the visible light image to generate a fused enhanced image. .

[0022] The image processing unit extracts pixel unit structural parameters based on the fused and enhanced image, performs threshold segmentation on the fused and enhanced image using the Otsu method to distinguish the luminous and non-luminous regions of the pixel unit, extracts the contour features of a single pixel unit through connected component analysis, calculates structural parameters such as the horizontal / vertical dimensions, pixel period, aperture ratio, and black matrix width of the pixel unit, and transmits the structural parameters to the photoelectric characteristic simulation unit.

[0023] The photoelectric characteristic simulation unit constructs a microscopic optical simulation field of the display module pixel unit based on the received pixel unit structure parameters, and sets the medium parameters of the simulation field, including the substrate refractive index. ITO layer refractive index Refractive index of screen printing ink layer Ink layer absorption coefficient The simulation area covers a single pixel unit and the boundary region of adjacent pixels, with the incident light rays set to a Lambertian distribution of 0°-30°. The photoelectric characteristic simulation unit simulates the occlusion and reflection effects of changes in the boundary position of the screen-printed pattern on the light-emitting path of the pixel, establishes a mapping relationship matrix between the screen-printed boundary offset and the light-emitting efficiency and crosstalk index, solves for the boundary range that minimizes photoelectric performance loss, and outputs the optimal boundary constraint parameters. The optimal boundary constraint parameters include the maximum allowable offset of the screen-printed boundary, the minimum line width, and the boundary tolerance band.

[0024] The diffusion model inference unit employs a denoising diffusion probability model with a U-Net structure. It receives the fused and enhanced image and optimal boundary constraint parameters, and generates a silkscreen alignment reference image through a pixel-level texture generation diffusion model. The calculation formula for the forward denoising process of the diffusion model is as follows: ; in, For the input fused and enhanced image, for The image after adding noise at different times. , for noise variance at time step This refers to random noise that follows a standard normal distribution.

[0025] In the inverse denoising stage of the diffusion model, the diffusion model inference unit normalizes the optimal boundary constraint parameters and encodes them into a fixed-dimensional condition vector through a fully connected layer. This condition vector is then injected into the intermediate feature layer of the denoising network via a cross-attention mechanism. The denoising network is then used to semantically complete the missing texture of the silkscreen edges in the fused and enhanced image, gradually restoring a high-resolution silkscreen alignment reference image that conforms to photoelectric property constraints. The resolution of the silkscreen alignment reference map is consistent with the original resolution of the acquired image.

[0026] The alignment matching unit receives the silkscreen alignment reference map and the real-time enhanced image generated by the image processing unit, and performs pixel-level hierarchical feature matching between the silkscreen alignment reference map and the real-time enhanced image. The alignment matching unit constructs a Gaussian pyramid hierarchical structure of the silkscreen alignment reference map, and performs feature point extraction, feature matching and geometric transformation matrix calculation layer by layer from the top to the bottom of the pyramid, decomposes and obtains the pixel-level deviation correction amount of horizontal translation, rotation angle and scaling factor, and removes mismatched feature points.

[0027] The alignment and matching unit outputs the deviation correction amount to the screen printing execution unit to control the screen printing plate to complete the position correction and alignment operation. At the same time, the matching results, including the matching accuracy, the number of effective matching feature points, and the average gradient of the screen printing edge, are fed back to the parameter optimization input of the image processing unit. The image processing unit dynamically adjusts the band weight coefficients in the multispectral difference operation according to the matching results to achieve synergistic optimization of screen printing accuracy and display performance.

[0028] In a preferred embodiment, the image processing unit constructs layers of [number] for the registered visible light image and near-infrared image, respectively. The Gaussian pyramid is used as a basis for constructing the corresponding Laplace pyramid decomposition layers. The Laplace pyramid's first... Layer Image The calculation formula is: ; in, For the first Layered Gaussian image, For upsampling operation, A 5×5 Gaussian convolution kernel; the top layer of the Laplace pyramid. The top of the Gauss Pyramid The remaining layers are high-frequency detail layers, containing the image's edge and texture features.

[0029] The image processing unit to The high-frequency Laplacian layer of the layer is used to calculate the difference image of the two-band images. The formula for calculating the difference image is as follows: ; in, For visible light images Laplace high-frequency layer, For near-infrared images Laplace high-frequency layer, For the first The difference image of the layers is used to extract high-frequency texture features of the silkscreen edges.

[0030] The image processing unit uses information entropy as a measure to evaluate the texture richness of the difference image. The formula for calculating information entropy is: ; in, For difference image medium grayscale value The probability of a pixel appearing. The information entropy of the difference image is positively correlated with texture richness. The image processing unit calculates the information entropy for each layer of the difference image. ,in accordance with The weighting coefficients of the dual-band Laplace layer are dynamically adjusted based on the comparison results with the preset entropy threshold. and ,satisfy ;when When the entropy exceeds the preset threshold, increase Weights are adjusted to preserve texture details when When the entropy is less than the preset threshold, increase The weights are assigned to suppress background noise. The image processing unit performs pyramid reconstruction on the weighted Laplacian image of each layer to obtain a fused and enhanced image.

[0031] Furthermore, the image processing unit introduces a multi-objective particle swarm optimization algorithm to optimize the optimal weighting coefficient combination of the dual-band image at each decomposition layer. The image processing unit constructs a fitness function by combining the sharpness evaluation value of the fused enhanced image with the silkscreen edge gradient energy function, where the sharpness evaluation value uses the image variance function, calculated as follows: ; in, , To fuse and enhance the width and height of the image, To fuse and enhance the average grayscale value of the image, The numerical value is positively correlated with image sharpness.

[0032] The gradient energy function for silkscreen edges uses the Sobel operator to calculate the gradient magnitude. The calculation formula is as follows: ; in, , The fused and enhanced images are respectively in direction and Sobel gradient magnitude in the direction, The numerical value is positively correlated with the clarity of the silkscreen edge texture.

[0033] The fitness function for the multi-objective optimization is: ; in, , Let be the weighting coefficient, satisfying The optimization objective is to minimize value.

[0034] The image processing unit performs iterative optimization using a multi-objective particle swarm optimization algorithm, where the particle dimension is... ,correspond Layer decomposition layer and The constraints are and The particle velocity update formula is: ; in, The inertial weight is initially set to 0.9, and decreases linearly to 0.4 with the number of iterations. , The learning factor is set to 2 for all of them; , A random number in the interval [0,1]. For the first The optimal position of each individual particle. The optimal position for the entire population; For the first Individual particles Location at any given moment The velocity corresponds to the given moment. The image processing unit calculates the optimal weighting coefficient combination for the dual-band image at each decomposition layer through iterative optimization, and adjusts the particle search step size through inertial weights to maximize the background noise suppression ratio while preserving the weak texture details of the silkscreen.

[0035] As a preferred embodiment, refer to Figure 3The photoelectric characteristic simulation unit constructs a three-dimensional microscopic optical simulation field of the display module pixel unit based on the pixel unit structural parameters. The simulation field covers the complete structure of the glass substrate, ITO conductive layer, pixel electrode, light-emitting layer, encapsulation layer and screen printing ink layer. Unstructured tetrahedral mesh is used for mesh generation, with a minimum mesh size of 10nm and a maximum mesh size of 100nm. The boundary conditions of the simulation field are set as follows: the incident surface is a Lambertian light source, the exit surface is a perfectly matched layer, and the side surface is a periodic boundary condition, which is used to simulate the optical characteristics of the array pixels.

[0036] The photoelectric characteristic simulation unit is based on the Monte Carlo ray tracing principle to simulate the occlusion and reflection effects of changes in the boundary position of the screen-printed pattern on the light path emitted by the pixels. The specific execution process is as follows: initialize photon packets, each photon packet carrying energy. The initial position is a random position of the emitting layer, and the initial propagation direction is a random direction conforming to the Lambertian distribution. The propagation path of the photon packet is tracked. When the photon packet reaches the interface between the two media, the reflectivity and transmittance are calculated according to the Fresnel formula, and a random number is used to determine whether the photon packet is reflected or transmitted. When the photon packet reaches the screen printing ink layer, the absorption probability is calculated based on the absorption coefficient of the ink layer. If it is absorbed, the tracking of the photon packet is terminated. If the photon packet exits from the exit surface, the exit energy, exit angle, and exit position of the photon packet are recorded. The above steps are repeated for at least [number missing] hours. Each photon packet is analyzed, and the total energy of the emitted photon packets is counted to calculate the light extraction efficiency of the pixel. ,in The total energy emitted. The total incident energy is given; the energy of photon packets emitted to adjacent pixel regions is counted, and the crosstalk rate is calculated. ,in This represents the energy emitted to adjacent pixels.

[0037] The photoelectric characteristic simulation unit establishes a mapping matrix between the screen printing boundary offset and the light extraction efficiency and crosstalk index, and sets the offset of the screen printing boundary. and The values ​​range for each value are [-5μm, 5μm], with a step size of 0.5μm. Negative offsets represent the silkscreen boundary shifting towards the black matrix region, while positive offsets represent the silkscreen boundary shifting towards the pixel opening region. For each group... Perform Monte Carlo ray tracing simulations to obtain the corresponding light extraction efficiency. and crosstalk rate Construct a mapping matrix The rows of the matrix correspond to The value of , the corresponding column The values ​​of are used to determine the output efficiency and crosstalk rate for the two channels, respectively.

[0038] The photoelectric characteristic simulation unit solves for the boundary range that minimizes photoelectric performance loss, which is used as the optimal boundary constraint parameter. The objective function constructed is: ; in, The baseline light output efficiency is the efficiency when there is no offset at the screen printing boundary. This is the maximum allowable value for crosstalk rate. , Let be the weighting coefficient, satisfying The photoelectric characteristic simulation unit solves the objective function. All The set of, where The maximum allowable threshold for photoelectric performance loss is the set corresponding to... , The range of values ​​for is the optimal boundary constraint parameter.

[0039] Furthermore, the photoelectric characteristic simulation unit is equipped with a long short-term memory network time-series prediction model to achieve predictive compensation of photoelectric characteristics. The training process of the model is as follows: A training set is constructed by collecting historical screen printing defect data from the display module. This historical screen printing defect data includes the screen printing boundary offset of historical batches, the corresponding light extraction efficiency, crosstalk rate, occurrence and severity level of Mura defects; the training set data is normalized to construct a time-series sequence, with each sequence having a length of [length missing]. , corresponding to continuous The detection data of each screen printing unit is labeled with the probability that the screen printing deviation will cause the display of Mura at the next moment.

[0040] The structure of the Long Short-Term Memory (LSTM) network temporal prediction model includes an input layer, two LSTM hidden layers, a fully connected layer, and a Sigmoid output layer; the dimension of the input layer is... The input feature is the silkscreen boundary offset. , Crosstalk rate The LSTM hidden layer has 64 neurons, and a Dropout layer is configured to prevent overfitting with a Dropout rate of 0.2. The fully connected layer maps the LSTM output features to a 1D output, and the Sigmoid function maps the output to the [0,1] interval, which represents the risk probability of the Mura defect. The model uses binary cross-entropy loss as its loss function, calculated as follows: ; in, The true label for the sample is 1 for the occurrence of Mura defect and 0 for the absence of Mura defect. This represents the risk probability predicted by the model. The model is trained using the Adam optimizer with a learning rate of 0.001 and 100 iterations until convergence.

[0041] The photoelectric characteristic simulation unit inputs the boundary offset calculated in real time into the trained prediction model to predict the probability of the screen printing deviation causing the display to malfunction in the next moment. The optimal boundary constraint parameters are dynamically adjusted based on the risk probability: when When this is the case, reduce the tolerance zone range of the optimal boundary constraint parameters and decrease the maximum permissible offset; when When, fine-tune the boundary constraint parameters and tighten the upper limit of the offset; when At that time, the original optimal boundary constraint parameters remain unchanged, where The high-risk threshold This is the low-risk threshold.

[0042] As a preferred embodiment, refer to Figure 4 The diffusion model inference unit adopts an improved denoised diffusion probability model. The model structure includes a conditional encoding module, a cross-attention mechanism module, a U-Net denoising network, and a graph convolutional feature fusion module.

[0043] In the inverse denoising stage of the diffusion model, the conditional coding module receives optimal boundary constraint parameters, which include the allowable offset of the silkscreen boundary. , Boundary tolerance zone Minimum line width There are a total of 4-dimensional parameters; the conditional encoding module normalizes the 4-dimensional constraint parameters, mapping them to the interval [-1, 1]. The first fully connected layer maps the 4-dimensional parameters into a 256-dimensional feature vector. After passing through the ReLU activation function and layer normalization, the second fully connected layer maps them into a 512-dimensional conditional vector. Conditional vector The dimension is consistent with the channel dimension of the intermediate feature layer of the denoising network.

[0044] The cross-attention mechanism module is embedded in the intermediate feature layer between the downsampling bottleneck layer and the upsampling layer of the U-Net denoising network, with a total of 6 cross-attention modules configured. The input of each cross-attention module is the intermediate feature map of the denoising network. With condition vector First, the feature map Flattened The sequence, used as the query vector ;Condition vector The key vectors are obtained through linear transformation. AND value vector Perform scaling dot product attention calculation, the formula is: Compare the output of the attention calculation with the original feature map. Residual connections are made, and after layer normalization, the output is sent to the next layer of the network. The photoelectric characteristic constraints of the optimal boundary constraint parameters are injected into the denoising process through the cross-attention mechanism.

[0045] The U-Net denoising network performs semantic completion on the missing texture of the silkscreen edges in the fused and enhanced image. The number of iterations in the reverse denoising process is set to... Step, noise variance The value increases linearly from 0.0001 to 0.02. Starting with random noise at time step, the noise is gradually predicted and removed until... A clear silkscreen alignment reference image is obtained at all times; during each step of the reverse denoising process, a conditional vector is injected through the cross-attention module. This ensures that the boundaries of the generated silkscreen pattern conform to photoelectric property constraints.

[0046] Furthermore, before executing the denoising network inference, the diffusion model inference unit first constructs a graph structure representation of the geometric topology of the silkscreen pattern, and uses a graph convolutional neural network to extract the spatial adjacency features of the silkscreen pattern. Specifically, the process involves: performing instance segmentation on the silkscreen pattern in the fused and enhanced image, extracting the contour point set of the silkscreen pattern, and constructing a graph structure representing the geometric topology of the silkscreen pattern. vertex set For the sampling points of the silkscreen pattern outline, the feature of each vertex is the coordinates of that point. Four-dimensional features including grayscale value, gradient magnitude, and edge set; To establish adjacency relationships between vertices, when the Euclidean distance between two vertices is less than a preset threshold, an edge is created between the two vertices. The adjacency matrix... It is a symmetric matrix. Represents vertices With vertex There are edges between them. This indicates that no edge exists.

[0047] The diffusion model inference unit extracts the spatial adjacency features of the silkscreen pattern through a two-layer graph convolutional neural network. The formula for calculating graph convolution is: ; in, For the first The vertex feature matrix of the layer, This is the initial vertex feature matrix; For degree matrix, ; For the first The trainable weight matrix of the layer; The ReLU activation function is used. After two layers of graph convolutional layers, high-dimensional features of each vertex are obtained. After global average pooling, the resulting feature has a dimension of... Spatial adjacency characteristics The dimension is consistent with the latent space features of the denoising network.

[0048] The diffusion model inference unit will incorporate spatial adjacency features. The residuals of the latent space features of the denoising network are added and fused together. The fused features are then input into the next layer of the denoising network to constrain the spatial continuity of the silkscreen texture generation and prevent artifacts from interfering with the alignment accuracy.

[0049] As a preferred embodiment, refer to Figure 5 The alignment matching unit constructs a Gaussian pyramid hierarchy structure for the silkscreen alignment reference map, wherein the number of pyramid layers is... Depending on the image resolution settings, for 4K resolution images, The Gaussian pyramid is constructed as follows: The fourth layer is the top layer, with a resolution of 1 / 16 of the original image; the 0th layer is the bottom layer, with the original resolution. The layer image is convolved and filtered using a 5×5 Gaussian kernel, and then downsampled by a factor of 2 to obtain the [then] layer. Layered image, standard deviation of Gaussian kernel The formula for calculating convolution is: ; in, The weights are for a 5×5 Gaussian kernel. For the first Layered Gaussian image, For the first Layered Gaussian image.

[0050] The alignment matching unit extracts accelerated and robust feature points from the real-time enhanced image. The extraction process is as follows: a Hessian matrix is ​​constructed, and for each pixel in the image, the determinant value of the Hessian matrix is ​​calculated. Points with determinant values ​​greater than a preset threshold are candidate feature points. Non-maximum suppression is performed in a 3×3×3 stereo neighborhood to select local extrema points as feature points. The Haar wavelet response in the neighborhood of the feature point is calculated, and the sum of responses within a 60° fan-shaped window is statistically analyzed. The direction corresponding to the maximum response is set as the main direction of the feature point. The neighborhood of the main direction of the feature point is divided into 4×4 sub-regions, and a 5-dimensional Haar wavelet response statistical value is calculated for each sub-region to generate a 64-dimensional feature descriptor.

[0051] The alignment matching unit performs feature point matching and geometric transformation matrix calculation layer by layer from the top to the bottom, decomposing and obtaining the deviation correction amounts for horizontal translation, rotation angle, and scaling factor, and eliminating mismatched feature points. Specifically, at the top of the pyramid, accelerated and robust feature points are extracted from the reference image and the real-time enhanced image, respectively. Euclidean distance is used for feature descriptor matching, with the matching criterion being that the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than 0.7, resulting in initial matching point pairs. Based on the initial matching point pairs, an affine transformation matrix is ​​calculated, and the initial deviation correction amount is obtained. The affine transformation matrix obtained from the upper layer is passed to the next layer to pre-correct the real-time enhanced image of the next layer, narrowing the search range of feature points. Feature points are extracted from the pre-corrected image and matched with the feature points of the corresponding layer of the reference image, updating the affine transformation matrix and refining the deviation correction amount. The above steps are repeated until the bottom of the pyramid to obtain the final pixel-level deviation correction amount, which is then output to the silkscreen execution unit.

[0052] The affine transformation matrix is ​​in the form of: ; in, The homogeneous coordinates of the feature points in the reference map. To enhance the homogeneous coordinates of image feature points in real time, For rotation and scaling parameters, The translation parameter is used; the horizontal translation amount is obtained through matrix decomposition. Vertical translation amount Rotation angle Scaling factor .

[0053] Furthermore, the alignment matching unit employs a random sampling consensus algorithm to remove outliers from the feature points, and uses the least squares method to fit the affine transformation matrix of the interior point set. Specifically, it randomly selects four non-collinear matching point pairs from the matched point pairs and calculates the initial affine transformation matrix. Using matrices Calculate the projection error for all matching point pairs. The formula for calculating the projection error is: ; in, For the reference figure, the first Homogeneous coordinates of feature points To enhance the homogeneous coordinates of corresponding feature points in an image in real time, For projection error; if If the error is less than a preset error threshold, the matched point pair is marked as an interior point; otherwise, it is marked as an exterior point. The number of interior points is counted. This process is repeated 100 times, and the matrix with the largest number of interior points is selected. As the optimal initial matrix, the corresponding set of interior points is the optimal set of interior points. The optimal set of interior points is then fitted using the least squares method to obtain the final affine transformation matrix. The objective function of the fitting is to minimize the sum of squared projection errors of all interior points, i.e.: ; The alignment matching unit adaptively adjusts the feature point search radius and matching threshold of the next level based on the residual distribution of the affine transformation matrix. Specifically, it calculates the mean projection error of the optimal interior point set. with standard deviation ,like If the residual value is greater than the preset residual threshold, the search radius of the feature points is reduced in the next level, the matching threshold of the feature descriptors is lowered, and the matching accuracy is improved; if If the residual value is less than the preset residual threshold, the search radius of the feature points is increased in the next level to improve the matching threshold and enhance matching efficiency.

[0054] Furthermore, the image processing unit monitors its own load status in real time, including CPU utilization, GPU computing power utilization, and total image processing latency, with a sampling period of 10ms, and establishes a computing resource scheduling cost function: ; in, CPU utilization For GPU computing power utilization, The total delay for image processing. For the weighting coefficients, satisfying .

[0055] When the system load exceeds the preset threshold At the same time, the image processing unit reduces the feature point extraction density of hierarchical feature matching and simultaneously increases the pyramid downsampling rate in multispectral difference operations, optimizing the allocation of system computing resources while ensuring the real-time alignment of screen printing. The specific adjustment strategy is as follows: when When the threshold of the Hessian matrix for feature point extraction is increased, the number of feature points extracted is reduced, and the number of Laplacian pyramid layers in the multispectral difference operation is reduced by one; when When the threshold of the Hessian matrix extracted from feature points is increased, feature matching is performed only from the top layer to the first layer of the pyramid, while the pyramid downsampling rate of the multispectral difference operation is increased; when During this period, unnecessary feature extraction and simulation calculations are paused, and only top-level feature matching and basic multispectral difference operations are performed. Once the system load decreases, the entire computation process is resumed. (Refer to...) Figure 6 .

[0056] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for improving the screen printing accuracy of a display module using an image enhancement algorithm, characterized in that, The steps include: acquiring visible light and near-infrared images of the display module substrate and performing multispectral difference operations to suppress substrate background noise and enhance weak textures at the silkscreen edges, generating a fused and enhanced image; Based on the extraction of pixel unit structure parameters from the fused and enhanced image, the influence of the screen printing pattern on the photoelectric performance is calculated through photoelectric property simulation, and the optimal boundary constraint parameters are output. A pixel-level texture generation diffusion model is used to process the fused and enhanced image and the optimal boundary constraint parameters to generate a screen printing alignment reference image. Perform pixel-level hierarchical feature matching between the silkscreen alignment reference map and the real-time enhanced image, output the deviation correction amount, and feed the matching result back to the band weight optimization process of multispectral difference operation.

2. The method according to claim 1, characterized in that, The multispectral difference operation includes: constructing a Laplacian pyramid decomposition layer for visible light and near-infrared band images, and calculating the difference image of the dual-band images in the high-frequency decomposition layer to extract high-frequency texture features of the silkscreen edge. The texture richness of the difference image is evaluated using the information entropy metric, and the weighting coefficients of the dual-band image in the fused and enhanced image are dynamically adjusted.

3. The method according to claim 1, characterized in that, The photoelectric characteristic simulation calculation includes: constructing a microscopic optical simulation field for the pixel unit of the display module, and simulating the occlusion and reflection effects of the change in the boundary position of the screen printing pattern on the light path of the pixel based on the Monte Carlo ray tracing principle; Establish a mapping matrix between screen printing boundary offset and light extraction efficiency and crosstalk index, and solve for the boundary range that minimizes photoelectric performance loss as the optimal boundary constraint parameter.

4. The method according to claim 1, characterized in that, The generation of the diffusion model using pixel-level texture includes: in the inverse denoising stage of the diffusion model, the optimal boundary constraint parameters are encoded into a condition vector and injected into the intermediate feature layer of the denoising network through a cross-attention mechanism. A denoising network is used to semantically complete the missing texture of the silkscreen edges in the fused and enhanced image, generating a high-resolution silkscreen alignment reference map that conforms to photoelectric property constraints.

5. The method according to claim 1, characterized in that, The pixel-level hierarchical feature matching includes: constructing a Gaussian pyramid hierarchical structure of the silkscreen alignment reference map and extracting accelerated and robust feature points of the real-time enhanced image; The feature point matching and geometric transformation matrix calculation are performed layer by layer from the top to the bottom. The deviation correction amounts of horizontal translation, rotation angle and scaling factor are decomposed and obtained to eliminate mismatched feature points.

6. The method according to claim 2, characterized in that, Also includes: A multi-objective particle swarm optimization algorithm is introduced, and the sharpness evaluation value of the fused enhanced image and the silkscreen edge gradient energy function are used to construct a fitness function; By iteratively optimizing the combination of weighting coefficients for the dual-band image at each decomposition layer, and adjusting the particle search step size using inertial weights, the background noise suppression ratio is maximized while preserving the weak texture details of the silkscreen.

7. The method according to claim 3, characterized in that, Also includes: Collect historical silkscreen defect data of the display module to build a training set, and train a long short-term memory network temporal prediction model. The boundary offset calculated by real-time simulation is input into the prediction model to predict the risk probability of the screen printing deviation causing the display to malfunction in the next moment. Based on the risk probability, the optimal boundary constraint parameters are dynamically adjusted to achieve predictive compensation of photoelectric properties.

8. The method according to claim 4, characterized in that, Before utilizing the denoising network, the method further includes: constructing a graph structure representation of the geometric topology of the silkscreen pattern, and using a graph convolutional neural network to extract the spatial adjacency relationship features of the silkscreen pattern; By fusing spatial adjacency features with the latent space features of the denoising network, the spatial continuity of silkscreen texture generation is constrained.

9. The method according to claim 5, characterized in that, The calculation of the geometric transformation matrix includes: using a random sampling consensus algorithm to remove outlier points from the feature points, and using the least squares method to fit the affine transformation matrix of the interior point set; Based on the residual distribution of the affine transformation matrix, the feature point search radius and matching threshold of the next level are adaptively adjusted to optimize the computational efficiency of hierarchical matching.

10. The method according to claim 9, characterized in that, Also includes: Monitor the real-time load status of the image processing unit and establish a computational resource scheduling cost function; When the load exceeds the preset threshold, the feature point extraction density of hierarchical feature matching is reduced, and the pyramid downsampling rate in multispectral difference operation is increased simultaneously.

Images