Wafer imaging electrostatic compensation method, device, equipment, medium and program product

Abstract

The application relates to the technical field of electron microscopic imaging, in particular to a wafer imaging static compensation method, device, equipment, medium and program product. The method comprises the following steps: obtaining an initial scanning image of a target scanning position, and obtaining an image quality score based on a preset quality evaluation network; in response to the image quality score being lower than a preset threshold, searching for optimal hardware parameters in a preset hardware parameter value range according to a preset parameter adjustment strategy; collecting a compensation scanning image based on the optimal hardware parameters, evaluating the compensation scanning image through the quality evaluation network, and obtaining a compensated image quality score; if the compensated image quality score is lower than the preset threshold, updating the parameter adjustment strategy and repeatedly searching for and evaluating the optimal hardware parameters until the compensated image quality score reaches or exceeds the preset threshold. The method can improve the robustness and compensation accuracy of imaging compensation.

Description

Technical Field

[0001] This application relates to the field of electron microscopy imaging technology, and in particular to a wafer imaging electrostatic compensation method, apparatus, device, medium, and program product. Background Technology

[0002] An electron microscope is a precision analytical instrument that uses a focused electron beam to scan the surface of a sample and acquire signals to form a high-resolution image. Its core principle is as follows: a high-energy electron beam is emitted from an electron gun, focused into a nanometer-scale beam spot by an electromagnetic lens system, and scans the sample surface point by point under the control of deflection coils. After the electron beam interacts with the sample, it generates secondary electrons and backscattered electrons, which are received by a detector and converted into image grayscale, ultimately revealing information such as the morphology and composition of the sample surface. Compared to optical microscopes, electron microscopes have sub-nanometer resolution, making them particularly suitable for detecting the microstructure and analyzing defects on wafer surfaces in semiconductor manufacturing.

[0003] In electron microscopy imaging of wafers, the wafer is typically fixed on a vacuum stage, and an electron beam is incident on the wafer surface at a set energy. Because wafers undergo various processes during manufacturing (such as oxidation, etching, and chemical mechanical polishing), their surfaces are highly susceptible to static charge accumulation. This is especially true in areas covered by insulating layers (such as silicon dioxide or silicon nitride) or high-resistivity materials, where conductivity is poor and charge cannot dissipate quickly, leading to a rise in local potential. The static charge on the wafer surface exerts a Coulomb force on the incident electron beam, altering its trajectory and landing energy, causing image distortion, focus drift, and abnormal brightness. In severe cases, image jitter, blurring, or even the inability to form a stable image may occur. Furthermore, electrostatic accumulation can trigger breakdown discharge, damaging the delicate structures or electronic devices on the wafer, directly affecting the reliability and yield of the detection.

[0004] In related technologies, the problem of unstable imaging due to electrostatic charge on the wafer surface is usually addressed from both hardware and software perspectives. Hardware solutions typically employ low-vacuum scanning modes, add charge neutralizers, or optimize stage design, all of which essentially neutralize the charge on the wafer surface. Software solutions rely on optimizing scanning methods or image processing to mitigate the impact of the surface electric field on imaging.

[0005] However, current electrostatic compensation methods have the following technical problems:

[0006] Existing electrostatic compensation methods are complex, have poor robustness, and have limited effect on improving image quality, and need to be optimized. Summary of the Invention

[0007] Therefore, it is necessary to provide a wafer imaging electrostatic compensation method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the robustness and accuracy of imaging compensation, thereby addressing the aforementioned technical problems.

[0008] In a first aspect, this application provides a method for electrostatic compensation in wafer imaging. The method includes:

[0009] Acquire an initial scan image of the target scan location, process the initial scan image based on a preset quality assessment network, and obtain an image quality score;

[0010] In response to the image quality score being lower than a preset threshold, the optimal hardware parameters are searched within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy.

[0011] Based on the optimal hardware parameters, a compensated scan image is acquired, and the compensated scan image is evaluated through the quality assessment network to obtain the compensated image quality score;

[0012] If the compensated image quality score is lower than the preset threshold, the parameter adjustment strategy is updated and the optimal hardware parameters are searched and evaluated repeatedly until the compensated image quality score reaches or exceeds the preset threshold.

[0013] In one embodiment, the step of responding to the image quality score being lower than a preset threshold and searching for optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy includes:

[0014] Based on the preset first step long traversal of the hardware parameter value range, the compensated scan image corresponding to each group of hardware parameters and the corresponding image quality score are collected.

[0015] The hardware parameters and corresponding image quality scores of each group are subjected to curve fitting, and the hardware parameters corresponding to the peak points of the curve are taken as the optimal hardware parameters.

[0016] In one embodiment, the step of updating the parameter adjustment strategy and repeatedly searching for and evaluating the optimal hardware parameters if the compensated image quality score is lower than the preset threshold, until the compensated image quality score reaches or exceeds the preset threshold, includes:

[0017] The search step size is adjusted to a preset second step size. Taking the optimal hardware parameters as the center, the hardware parameters are traversed and adjusted in the neighborhood of the center according to the second step size. The compensated scan image is acquired and the corresponding image quality score is calculated. The second step size is smaller than the first step size.

[0018] Update the optimal hardware parameters to the hardware parameters corresponding to the highest image quality score, and repeat the search process until the search image quality score meets the standard or the step size is less than the preset minimum step size.

[0019] In one embodiment, obtaining an initial scan image of the target scan location and processing the initial scan image based on a preset quality assessment network to obtain an image quality score includes:

[0020] Multi-scale high-frequency features are extracted from the initial scan image and fused with the initial scan image to obtain an enhanced scan image;

[0021] The brightness distribution and local contrast distribution of the enhanced scanned image are statistically analyzed to generate an image quality prior vector;

[0022] The enhanced scanned image is input into the convolutional branch and attention branch of the quality assessment network to extract local and global features.

[0023] Dynamic fusion weights are generated based on the image quality prior vector, and the local features and global features are weighted and fused to map the image quality score.

[0024] In one embodiment, the step of inputting the enhanced scanned image into the convolutional branch and attention branch of the quality assessment network respectively to extract local and global features includes:

[0025] Dynamic convolution kernel weights are generated based on input features, and an adaptive dynamic convolution kernel is obtained by weighted combination of multiple static convolution kernels.

[0026] The adaptive dynamic convolution kernel with multiple scales is used to extract local features at multiple scales. A dynamic activation threshold is calculated based on the image quality prior vector. The local features at multiple scales are then gated and activated before being output.

[0027] In one embodiment, the step of inputting the enhanced scanned image into the convolutional branch and attention branch of the quality assessment network respectively to extract local and global features includes:

[0028] The input feature map is divided into multiple local windows. Attention scores are calculated independently within each local window, and the highest attention scores are retained and weighted summed.

[0029] A quality deviation term is calculated based on the image quality prior vector, and the quality deviation term is added to the attention score.

[0030] The sparse attention outputs of each of the local windows are then stitched together to restore them to their original size, thus obtaining the global features.

[0031] Secondly, this application also provides a wafer imaging electrostatic compensation device. The device includes:

[0032] The quality assessment module is used to acquire the initial scan image of the target scan position, process the initial scan image based on a preset quality assessment network, and obtain an image quality score.

[0033] The parameter search module is used to search for the optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy in response to the image quality score being lower than a preset threshold.

[0034] The compensation evaluation module is used to acquire a compensated scan image based on the optimal hardware parameters, evaluate the compensated scan image through the quality evaluation network, and obtain the compensated image quality score.

[0035] An iterative search module is used to update the parameter adjustment strategy and repeatedly search for and evaluate the optimal hardware parameters if the compensated image quality score is lower than the preset threshold, until the compensated image quality score reaches or exceeds the preset threshold.

[0036] In one embodiment, the parameter search module includes:

[0037] The first step long traversal module is used to collect the compensated scan image and the corresponding image quality score for each set of hardware parameters based on the preset range of hardware parameter values ​​in the first step long traversal.

[0038] The quality curve fitting module is used to perform curve fitting on each set of hardware parameters and the corresponding image quality scores, and to use the hardware parameters corresponding to the peak points of the curve as the optimal hardware parameters.

[0039] In one embodiment, the iterative search module includes:

[0040] The second step size search module is used to adjust the search step size to a preset second step size. Taking the optimal hardware parameters as the center, the module traverses and adjusts the hardware parameters in the neighborhood of the center according to the second step size, collects the compensated scan image and calculates the corresponding image quality score. The second step size is smaller than the first step size.

[0041] The repeated search module is used to update the optimal hardware parameters to the hardware parameters corresponding to the highest image quality score, and repeat the search process until the searched image quality score meets the standard or the step size is less than the preset minimum step size.

[0042] In one embodiment, the quality assessment module includes:

[0043] An enhanced scan image module is used to extract multi-scale high-frequency features from the initial scan image and fuse them with the initial scan image to obtain an enhanced scan image;

[0044] The quality prior vector module is used to statistically analyze the brightness distribution and local contrast distribution of the enhanced scan image and generate an image quality prior vector.

[0045] The branch processing module is used to input the enhanced scan image into the convolutional branch and attention branch of the quality assessment network respectively, and extract local features and global features;

[0046] The weighted fusion module is used to generate dynamic fusion weights based on the image quality prior vector, and to map the weighted fusion of the local features and the global features to the image quality score.

[0047] In one embodiment, the branch processing module includes:

[0048] The dynamic convolution kernel module is used to generate dynamic convolution kernel weights based on input features, and to obtain an adaptive dynamic convolution kernel by weighted combination of multiple static convolution kernels.

[0049] The dynamic activation threshold module is used to extract multi-scale local features using the multi-scale adaptive dynamic convolution kernel, calculate the dynamic activation threshold based on the image quality prior vector, and output the multi-scale local features after gating activation.

[0050] In one embodiment, the branch processing module includes:

[0051] The local window module is used to divide the input feature map into multiple local windows, independently calculate the attention score within each local window, and retain the part of the attention score with the highest score for weighted summation;

[0052] The quality deviation module is used to calculate a quality deviation term based on the image quality prior vector and to add the quality deviation term to the attention score.

[0053] The local window stitching module is used to output the sparse attention of each local window and stitch them together to restore them to their original size, thereby obtaining the global feature.

[0054] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a wafer imaging electrostatic compensation method as described in any embodiment of the first aspect.

[0055] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of a wafer imaging electrostatic compensation method as described in any embodiment of the first aspect.

[0056] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of a wafer imaging electrostatic compensation method as described in any embodiment of the first aspect.

[0057] The above-described wafer imaging electrostatic compensation method, apparatus, computer equipment, storage medium, and computer program product, derived from the technical features in the embodiments, can achieve the following beneficial effects to address the technical problems in the background art:

[0058] This application provides a wafer imaging electrostatic compensation method, comprising: acquiring an initial scan image of a target scanning position; processing the initial scan image based on a preset quality assessment network to obtain an image quality score; responding to the image quality score being lower than a preset threshold, searching for optimal hardware parameters within a preset hardware parameter value range according to a preset parameter adjustment strategy; acquiring a compensated scan image based on the optimal hardware parameters; evaluating the compensated scan image through the quality assessment network to obtain a compensated image quality score; if the compensated image quality score is lower than the preset threshold, updating the parameter adjustment strategy and repeatedly searching for and evaluating the optimal hardware parameters until the compensated image quality score reaches or exceeds the preset threshold. In implementation, the initial scan image is first rapidly diagnosed based on the quality assessment network, and the compensation process is triggered only when the image quality does not meet the preset threshold, which helps to avoid unnecessary parameter adjustments. When electrostatic compensation is required, the optimal parameters are automatically searched within the hardware parameter range according to the preset strategy, and the compensated image is acquired for secondary evaluation. If the compensation fails to meet the standard in one attempt, the search strategy is dynamically updated, and the parameters are continuously iterated and optimized with finer step sizes until the image quality meets the requirements. In this way, the entire process forms a closed-loop adaptive feedback mechanism. Without additional hardware, it can automatically match the optimal imaging parameters for wafers with different electrostatic distribution states, which helps to effectively solve problems such as image blurring and loss of detail caused by local electric field interference. The clear and stable images obtained directly improve the accuracy and reliability of subsequent visual positioning, critical dimension measurement, and defect detection. At the same time, the simplified iterative strategy controls the compensation time, balancing imaging quality and production efficiency. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of the first process of a wafer imaging electrostatic compensation method in one embodiment;

[0061] Figure 2 This is a schematic diagram of the network architecture of a quality assessment network in one embodiment;

[0062] Figure 3 This is a schematic diagram of the second process of a wafer imaging electrostatic compensation method in another embodiment;

[0063] Figure 4 This is a schematic diagram of the third process of a wafer imaging electrostatic compensation method in another embodiment;

[0064] Figure 5 This is a schematic diagram of the fourth process of a wafer imaging electrostatic compensation method in another embodiment;

[0065] Figure 6 This is a schematic diagram of the network architecture of the preprocessing section in a quality assessment network in one embodiment.

[0066] Figure 7 This is a schematic diagram of the fifth process of a wafer imaging electrostatic compensation method in another embodiment;

[0067] Figure 8 This is a schematic diagram of the architecture of the dynamic threshold convolution module in the quality assessment network in one embodiment;

[0068] Figure 9 This is a schematic diagram of the sixth process of a wafer imaging electrostatic compensation method in another embodiment;

[0069] Figure 10 This is a schematic diagram of the architecture of the sparse attention module in the quality assessment network in one embodiment;

[0070] Figure 11 This is a schematic diagram of the architecture of the dynamic focusing module in a quality assessment network in one embodiment;

[0071] Figure 12 This is a schematic diagram comparing images before and after electrostatic compensation in one embodiment;

[0072] Figure 13 This is a structural block diagram of a wafer imaging electrostatic compensation device in one embodiment;

[0073] Figure 14 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0075] In related technologies, the problem of unstable imaging due to electrostatic charge on the wafer surface is usually addressed from both hardware and software perspectives. Hardware solutions typically employ low-vacuum scanning modes, add charge neutralizers, or optimize stage design, all of which essentially neutralize the charge on the wafer surface. Software solutions rely on optimizing scanning methods or image processing to mitigate the impact of the surface electric field on imaging.

[0076] However, current electrostatic compensation methods have the following technical problems:

[0077] Existing electrostatic compensation methods are complex, have poor robustness, and have limited effect on improving image quality, and need to be optimized.

[0078] To address the aforementioned issues, this application provides a wafer imaging electrostatic compensation method, apparatus, device, medium, and program product.

[0079] In one embodiment, such as Figure 1 As shown, a wafer imaging electrostatic compensation method is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0080] Step 102: Obtain the initial scan image of the target scan position, process the initial scan image based on a preset quality assessment network, and obtain an image quality score.

[0081] Among them, the quality assessment network can refer to a hybrid neural network used to judge the quality of scanned images.

[0082] For example, it can be as follows Figure 2 As shown, the quality assessment network can include a quality-aware preprocessing module, a dynamic thresholding convolution module, a sparse-aware attention module, and an adaptive fusion module. Specifically, a quality-aware preprocessing module can be designed at the very beginning of the network after the input nodes to obtain prior knowledge of image quality at the network input.

[0083] Step 104: In response to the image quality score being lower than a preset threshold, search for the optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy.

[0084] The range of hardware parameters refers to the allowable variation of adjustable imaging parameters in a scanning electron microscope, such as electron beam current and accelerating voltage. Adjusting parameters within this range can affect image quality; parameters that are too low may result in insufficient signal-to-noise ratio, while parameters that are too high may introduce new aberrations or damage the wafer.

[0085] The parameter tuning strategy refers to the rules and sequence of steps followed when searching for optimal hardware parameters. The parameter tuning strategy determines how to traverse the range of hardware parameter values, how to select the next set of parameter values ​​to be tested, and how to update the search direction based on the acquired image quality scores.

[0086] The optimal hardware parameters refer to the combination of hardware parameters that achieves the best image quality at the current scanning position. In electrostatic compensation scenarios, due to the uneven distribution of electrostatic charge on the wafer surface, the optimal hardware parameters change dynamically with the scanning position.

[0087] For example, in response to the image quality score being lower than a preset threshold, the terminal may search for the optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy.

[0088] Step 106: Acquire a compensated scan image based on the optimal hardware parameters, evaluate the compensated scan image through the quality assessment network, and obtain the compensated image quality score.

[0089] Step 108: If the compensated image quality score is lower than the preset threshold, update the parameter adjustment strategy and repeatedly search for and evaluate the optimal hardware parameters until the compensated image quality score reaches or exceeds the preset threshold.

[0090] In the above-described wafer imaging electrostatic compensation method, a reasonable derivation based on the technical features in the embodiments achieves the beneficial effect of solving the technical problems raised in the background art:

[0091] This application provides a wafer imaging electrostatic compensation method, comprising: acquiring an initial scan image of a target scanning position; processing the initial scan image based on a preset quality assessment network to obtain an image quality score; responding to the image quality score being lower than a preset threshold, searching for optimal hardware parameters within a preset hardware parameter value range according to a preset parameter adjustment strategy; acquiring a compensated scan image based on the optimal hardware parameters; evaluating the compensated scan image through the quality assessment network to obtain a compensated image quality score; if the compensated image quality score is lower than the preset threshold, updating the parameter adjustment strategy and repeatedly searching for and evaluating the optimal hardware parameters until the compensated image quality score reaches or exceeds the preset threshold. In implementation, the initial scan image is first rapidly diagnosed based on the quality assessment network, and the compensation process is triggered only when the image quality does not meet the preset threshold, which helps to avoid unnecessary parameter adjustments. When electrostatic compensation is required, the optimal parameters are automatically searched within the hardware parameter range according to the preset strategy, and the compensated image is acquired for secondary evaluation. If the compensation fails to meet the standard in one attempt, the search strategy is dynamically updated, and the parameters are continuously iterated and optimized with finer step sizes until the image quality meets the requirements. In this way, the entire process forms a closed-loop adaptive feedback mechanism. Without additional hardware, it can automatically match the optimal imaging parameters for wafers with different electrostatic distributions, effectively solving problems such as image blurring and detail loss caused by local electric field interference. Ultimately, it can be achieved as follows: Figure 12 As shown, the clear and stable images obtained directly improve the accuracy and reliability of subsequent visual positioning, critical dimension measurement and defect detection. At the same time, the simplified iteration strategy controls the compensation time, taking into account both imaging quality and production efficiency.

[0092] In one embodiment, it can be as follows Figure 3 As shown, step 104 includes:

[0093] Step 302: Based on the preset first step long traversal of the hardware parameter value range, collect the compensated scan image corresponding to each group of hardware parameters and the corresponding image quality score.

[0094] The first step length can refer to the parameter adjustment interval used in the initial search phase. The first step length can be set to a larger value. A larger step length can quickly cover the entire parameter range with fewer sampling times, which helps to quickly and initially locate the interval where the optimal parameter is located.

[0095] Step 304: Perform curve fitting on the hardware parameters and the corresponding image quality scores of each group, and take the hardware parameters corresponding to the peak points of the curve as the optimal hardware parameters.

[0096] Curve fitting refers to establishing a continuous function model based on discrete sampling points, i.e., hardware parameter values ​​and their corresponding image quality scores, to describe the mapping relationship between hardware parameters and image quality.

[0097] For example, the terminal can use fitting methods such as polynomial fitting or least squares to perform curve fitting on each set of hardware parameters and the corresponding image quality scores. In this way, the estimated quality score corresponding to any parameter value can be calculated from the fitted curve, and then the peak point of the curve can be found as the estimated position of the optimal hardware parameters.

[0098] In this embodiment, by traversing the range of hardware parameters with a fixed step size and fitting the parameter quality curve, a global mapping relationship between hardware parameters and imaging quality can be quickly established, and the peak point of the curve can be directly locked as the optimal parameter. Compared with exhaustive search or random search, the curve fitting method significantly reduces the number of samplings, significantly improves compensation efficiency while ensuring search accuracy, and is suitable for scenarios with unknown initial states or uneven electrostatic distribution. It helps to stably and efficiently locate the best imaging parameters, providing a reliable starting benchmark for subsequent fine adjustments.

[0099] In one embodiment, it can be as follows Figure 4 As shown, step 108 includes:

[0100] Step 402: Adjust the search step size to a preset second step size. Using the optimal hardware parameters as the center, traverse and adjust the hardware parameters in the neighborhood of the center according to the second step size. Collect the compensated scan image and calculate the corresponding image quality score. The second step size is smaller than the first step size.

[0101] The second step size refers to the parameter adjustment interval used in the fine-tuning search phase, and its value is smaller than the first step size. The first step size is used to quickly cover the entire parameter range to locate the optimal interval, while the second step size is used for fine-tuning within the range located in the optimal interval. The neighborhood can refer to a local parameter interval near the current optimal hardware parameters, and the radius of the neighborhood can be related to the second step size. For example, the neighborhood can be formed by expanding outwards from the optimal parameters by several step sizes. The setting of the neighborhood limits the spatial range of the fine-tuning search, helping to avoid repeated traversal across the entire parameter range, thereby improving search efficiency.

[0102] Step 404: Update the optimal hardware parameters to the hardware parameters corresponding to the highest image quality score, and repeat the search process until the search image quality score meets the standard or the step size is less than the preset minimum step size.

[0103] In this embodiment, by reducing the step size within the neighborhood of the curve peak for a fine-grained local search, the precise location of the optimal hardware parameters can be gradually approximated, effectively overcoming the positioning deviation caused by the initial large step size fitting. The adaptive mechanism of decreasing step size with iteration balances search efficiency and accuracy, avoiding getting trapped in local suboptimal solutions while achieving rapid convergence, thus helping to ensure that the compensated image quality stably reaches the preset threshold.

[0104] In one embodiment, it can be as follows Figure 5 and Figure 6 As shown, step 102 includes:

[0105] Step 502: Extract multi-scale high-frequency features from the initial scan image and fuse them with the initial scan image to obtain an enhanced scan image.

[0106] For example, the quality-aware preprocessing module may include two branch units: a multi-scale Laplacian pyramid and a brightness-contrast statistics unit. The multi-scale Laplacian pyramid is used to extract multi-scale features of the input image in parallel. Each layer is obtained through Gaussian blurring and downsampling. The "residual" calculated layer by layer is the high-frequency component feature map of the image, reflecting the image's sharpness and texture details. Finally, the high-frequency feature map is obtained by stitching the results together. The high-frequency feature map is then compared with the original input image. By concatenating along the channel dimension, an enhanced input is obtained. As the input to the subsequent two-branch network, it can be expressed as follows:

[0107]

[0108] Step 504: Statistically analyze the brightness distribution and local contrast distribution of the enhanced scan image to generate an image quality prior vector.

[0109] For example, brightness contrast statistics involve calculating the grayscale histogram of the image and obtaining the local contrast of the image by calculating the standard deviation of the sliding window. The results are then stitched together and mapped to a prior image quality vector q of a preset dimension, which is used to modulate subsequent modules as their gating signal.

[0110] Step 506: Input the enhanced scanned image into the convolutional branch and attention branch of the quality assessment network respectively to extract local and global features.

[0111] For example, it can be as follows Figure 8As shown, following the quality perception module, this embodiment provides a dual-branch hybrid network structure, namely a CNN (Convolutional Neural Network) branch and a Transformer (Transformer Network) branch. The CNN branch is mainly used for local detail feature extraction, while the Transformer branch is mainly used for global context feature extraction. Considering the requirements of device operating efficiency, both can be implemented in a lightweight manner, complementing each other to improve overall accuracy and robustness.

[0112] For example, the CNN branch body provided in this embodiment can be a dynamic threshold convolution module (DTConv). DTConv can be based on depthwise separable convolution and includes a dynamic convolution kernel generation module, a multi-scale dynamic convolution module, and a threshold modulation module.

[0113] Step 508: Generate dynamic fusion weights based on the image quality prior vector, and map the weighted fusion of the local features and the global features to the image quality score.

[0114] In this embodiment, image details are enhanced through multi-scale high-frequency feature fusion. A quality prior is generated by combining brightness and contrast statistics, guiding the convolutional and attention branches to extract local texture and global contextual features, respectively. The dynamic weighted fusion mechanism enables the network to adaptively adjust the contribution ratio of local and global features, helping to improve the robustness and accuracy of evaluating images of different qualities and providing a reliable basis for subsequent compensation decisions.

[0115] In one embodiment, it can be as follows Figure 7 and Figure 8 As shown, step 506 includes:

[0116] Step 702: Generate dynamic convolution kernel weights based on input features, and obtain an adaptive dynamic convolution kernel by weighting and combining multiple static convolution kernels.

[0117] For example, the dynamic convolution kernel generation module in this embodiment can be a lightweight sub-network used to adaptively adjust the weights of the convolution kernels based on local information of the input feature F, thereby achieving input adaptation and dynamic weights. The formula for calculating can be shown in the following equation:

[0118]

[0119] Here, FC represents a fully connected layer, and GAP represents global average pooling. This represents N depthwise separable static convolutional kernels. In this embodiment, the FC+GAP design is very lightweight, but it allows the model to dynamically adjust the kernel weights according to different input features, paying more attention to local high-frequency features such as edges and textures, thus enhancing robustness to images of different quality.

[0120] Step 704: Extract multi-scale local features using the multi-scale adaptive dynamic convolution kernel, calculate the dynamic activation threshold based on the image quality prior vector, and output the multi-scale local features after gating activation.

[0121] For example, in this embodiment, the multi-scale dynamic convolution module can use the dynamic weights generated above. The convolutional kernels are used to form 3×3 and 5×5 depthwise separable convolutions. The outputs of the two convolutions are added element-wise to obtain multi-scale local features. The threshold modulation module can set an adaptive activation function, dynamically adjusting the threshold of the activation function according to the quality of different input images. It can be represented by the following formula:

[0122]

[0123] Where q is the quality prior vector calculated by the quality perception preprocessing module. It is a learnable projection vector that maps the quality prior vector to the threshold domain. It is a learnable generalization threshold.

[0124] For example, the terminal can be based on the aforementioned dynamic threshold. The ReLU activation function can be modified and used as the activation function for multi-scale features, as shown in the following equation:

[0125]

[0126] Thus, as can be seen from the formula, activation only occurs when the input features exceed the dynamic threshold; otherwise, they are suppressed. From an image-level perspective, for noisy images, more quality-perceived evaluation is needed, resulting in a larger quality prior vector and a higher dynamic threshold. This suppresses weak feature responses such as noise, preserving more strong features. Conversely, for clean and clear images, more weaker features are retained for subsequent calculations. The overall output of the dynamic threshold convolutional block can then be expressed as follows:

[0127]

[0128] In this embodiment, by adaptively generating dynamic convolutional kernel weights, the network can flexibly adjust its filtering response according to the content features of different images, enhancing its ability to capture key textures and edges. Multi-scale convolution combined with dynamic activation thresholds helps to effectively suppress weak feature interference such as noise, while preserving the fine structure of clear images, thus improving the robustness and discriminative power of local feature representation.

[0129] In one embodiment, it can be as follows Figure 9 As shown, step 504 includes:

[0130] Step 902: Divide the input feature map into multiple local windows, calculate the attention score independently in each local window, and retain the part of the attention score with the highest score for weighted summation.

[0131] For example, it can be as follows Figure 10 As shown, the dual-branch quality assessment network provided in this embodiment introduces windowed sparse attention and a quality bias term. Specifically, windowed sparse attention divides the input feature map into multiple small windows, calculates the attention score within each window, and retains the top-K highest-scoring tokens for weighted summation, thus focusing only on the most effective information and reducing overall computational complexity. The quality bias term is calculated based on the quality prior vector q, which affects the degree to which the model pays attention to different regions in the feature map, highlighting potentially degraded regions such as blurred or noisy areas. The calculation formula is shown below:

[0132]

[0133] Step 904: Calculate the quality deviation term based on the image quality prior vector, and add the quality deviation term to the attention score.

[0134] For example, the terminal can use the quality prior vector q to calculate the bias values ​​at query position i and key position j respectively using the corresponding lightweight multilayer perceptron and then sum them. The final sparse attention calculation can be shown in the following equation:

[0135]

[0136] Where Q, K, and V are the query, key, and value matrices in the standard self-attention calculation formula, respectively. B is the dimension of the key, and B is the quality deviation matrix composed of quality deviation terms.

[0137] Step 906: Output the sparse attention of each of the local windows and stitch them together to restore them to their original size to obtain the global feature.

[0138] For example, it can be as follows Figure 11 As shown, after the dual-branch hybrid network, this embodiment can set a dual-branch adaptive fusion dynamic focusing module to organically fuse the features output by the two branches according to the quality of the input image, so as to achieve synergistic complementarity between local and global features of the image.

[0139] Specifically, in the dual-branch hybrid network, the local detail features output by the CNN branch (convolutional branch) are denoted as feature L, and the local detail features output by the Transformer branch (attention branch) are denoted as feature G. Feature L primarily includes local detail information of the image, such as edges, textures, and corners—microscopic structures captured by the convolutional kernel—used to perceive subtle quality degradation in the image, such as blurred edges or noise particles. Feature G primarily includes global contextual information of the image, such as overall brightness distribution, large-scale structural layout, and long-distance dependencies between regions, responsible for grasping the overall image quality and determining whether there are large areas of contrast imbalance or shadows. Considering the different spatial dimensions of the output features L and G from the CNN and Transformer branches, to facilitate subsequent feature fusion, the outputs of the two branches are respectively passed through an adaptive average pooling layer to unify them to the same size, obtaining spatially aligned feature vectors Ls and Gs, as shown in the following equation:

[0140]

[0141] Secondly, in order to adaptively focus on the effective region of the input image and achieve the fusion of local and global features, this embodiment sets up a lightweight multilayer perceptron, which can generate dynamic fusion weights β based on the quality prior vector q, as shown in the following formula:

[0142]

[0143] In this way, the dynamic fusion weight β depends entirely on the prior quality of the overall image and is not affected by the output features of the CNN branch and the Transformer branch, thus ensuring the global consistency of the fusion strategy.

[0144] For example, finally, based on the aligned feature vectors and dynamic fusion weights described above, this embodiment sets up a dynamic star connection for the features, as shown in the following formula:

[0145]

[0146] in, The first term represents the element-wise multiplication of two feature vectors, which means capturing the collaborative features of local and global features, emphasizing the areas that both need to focus on; the second term emphasizes the individual contributions of the two features, that is, based on the quality prior, it determines whether the network focuses more on local features or global features.

[0147] In this embodiment, a windowed sparse attention mechanism reduces the computational complexity of global self-attention to a level linearly related to image size, significantly improving processing efficiency. The introduction of a quality bias term enables the model to adaptively enhance its focus on degraded regions such as blur and noise based on prior image information, guiding attention weights towards key locations. This achieves more discriminative global contextual features while maintaining computational lightweightness. The adaptive fusion method proposed in this embodiment achieves the fusion of local and global features while maintaining computational lightweightness, highlighting their respective contributions and resolving the feature confusion that may result from fusing images of different quality using fixed weights.

[0148] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0149] Based on the same inventive concept, this application also provides a wafer imaging electrostatic compensation device for implementing the wafer imaging electrostatic compensation method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the wafer imaging electrostatic compensation device provided below can be found in the limitations of the wafer imaging electrostatic compensation method described above, and will not be repeated here.

[0150] In one embodiment, such as Figure 13 As shown, a wafer imaging electrostatic compensation device is provided, comprising: a quality assessment module, a parameter search module, a compensation assessment module, and an iterative search module, wherein:

[0151] The quality assessment module is used to acquire the initial scan image of the target scan position, process the initial scan image based on a preset quality assessment network, and obtain an image quality score.

[0152] The parameter search module is used to search for the optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy in response to the image quality score being lower than a preset threshold.

[0153] The compensation evaluation module is used to acquire a compensated scan image based on the optimal hardware parameters, evaluate the compensated scan image through the quality evaluation network, and obtain the compensated image quality score.

[0154] An iterative search module is used to update the parameter adjustment strategy and repeatedly search for and evaluate the optimal hardware parameters if the compensated image quality score is lower than the preset threshold, until the compensated image quality score reaches or exceeds the preset threshold.

[0155] In one embodiment, the parameter search module includes:

[0156] The first step long traversal module is used to collect the compensated scan image and the corresponding image quality score for each set of hardware parameters based on the preset range of hardware parameter values ​​in the first step long traversal.

[0157] The quality curve fitting module is used to perform curve fitting on each set of hardware parameters and the corresponding image quality scores, and to use the hardware parameters corresponding to the peak points of the curve as the optimal hardware parameters.

[0158] In one embodiment, the iterative search module includes:

[0159] The second step size search module is used to adjust the search step size to a preset second step size. Taking the optimal hardware parameters as the center, the module traverses and adjusts the hardware parameters in the neighborhood of the center according to the second step size, collects the compensated scan image and calculates the corresponding image quality score. The second step size is smaller than the first step size.

[0160] The repeated search module is used to update the optimal hardware parameters to the hardware parameters corresponding to the highest image quality score, and repeat the search process until the searched image quality score meets the standard or the step size is less than the preset minimum step size.

[0161] In one embodiment, the quality assessment module includes:

[0162] An enhanced scan image module is used to extract multi-scale high-frequency features from the initial scan image and fuse them with the initial scan image to obtain an enhanced scan image;

[0163] The quality prior vector module is used to statistically analyze the brightness distribution and local contrast distribution of the enhanced scan image and generate an image quality prior vector.

[0164] The branch processing module is used to input the enhanced scan image into the convolutional branch and attention branch of the quality assessment network respectively, and extract local features and global features;

[0165] The weighted fusion module is used to generate dynamic fusion weights based on the image quality prior vector, and to map the weighted fusion of the local features and the global features to the image quality score.

[0166] In one embodiment, the branch processing module includes:

[0167] The dynamic convolution kernel module is used to generate dynamic convolution kernel weights based on input features, and to obtain an adaptive dynamic convolution kernel by weighted combination of multiple static convolution kernels.

[0168] The dynamic activation threshold module is used to extract multi-scale local features using the multi-scale adaptive dynamic convolution kernel, calculate the dynamic activation threshold based on the image quality prior vector, and output the multi-scale local features after gating activation.

[0169] In one embodiment, the branch processing module includes:

[0170] The local window module is used to divide the input feature map into multiple local windows, independently calculate the attention score within each local window, and retain the part of the attention score with the highest score for weighted summation;

[0171] The quality deviation module is used to calculate a quality deviation term based on the image quality prior vector and to add the quality deviation term to the attention score.

[0172] The local window stitching module is used to output the sparse attention of each local window and stitch them together to restore them to their original size, thereby obtaining the global feature.

[0173] The various modules in the aforementioned wafer imaging electrostatic compensation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0174] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 14As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a wafer imaging electrostatic compensation method. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0175] Those skilled in the art will understand that Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0176] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0177] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0178] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0179] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0180] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0181] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0182] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for electrostatic compensation in wafer imaging, characterized in that, The method includes: Acquire an initial scan image of the target scan location, process the initial scan image based on a preset quality assessment network, and obtain an image quality score; In response to the image quality score being lower than a preset threshold, the optimal hardware parameters are searched within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy. Based on the optimal hardware parameters, a compensated scan image is acquired, and the compensated scan image is evaluated through the quality assessment network to obtain the compensated image quality score; If the compensated image quality score is lower than the preset threshold, the parameter adjustment strategy is updated and the optimal hardware parameters are searched and evaluated repeatedly until the compensated image quality score reaches or exceeds the preset threshold. The initial scan image of the target scan location is obtained, and the initial scan image is processed based on a preset quality assessment network to obtain an image quality score, including: Multi-scale high-frequency features are extracted from the initial scan image and fused with the initial scan image to obtain an enhanced scan image; The brightness distribution and local contrast distribution of the enhanced scanned image are statistically analyzed to generate an image quality prior vector; The enhanced scanned image is input into the convolutional branch and attention branch of the quality assessment network to extract local and global features. Dynamic fusion weights are generated based on the image quality prior vector, and the local features and global features are weighted and fused to map the image quality score. The step of inputting the enhanced scanned image into the convolutional branch and attention branch of the quality assessment network to extract local and global features includes: Dynamic convolution kernel weights are generated based on input features, and an adaptive dynamic convolution kernel is obtained by weighted combination of multiple static convolution kernels. The multi-scale adaptive dynamic convolution kernel is used to extract the multi-scale local features, and a dynamic activation threshold is calculated based on the image quality prior vector. The multi-scale local features are then gated and activated before being output. The step of inputting the enhanced scanned image into the convolutional branch and attention branch of the quality assessment network to extract local and global features further includes: The input feature map is divided into multiple local windows. Attention scores are calculated independently within each local window, and the highest attention scores are retained and weighted summed. A quality deviation term is calculated based on the image quality prior vector, and the quality deviation term is added to the attention score. The sparse attention outputs of each of the local windows are then stitched together to restore them to their original size, thus obtaining the global features.

2. The method according to claim 1, characterized in that, The step of responding to an image quality score below a preset threshold by searching for optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy includes: Based on the preset first step long traversal of the hardware parameter value range, the compensated scan image corresponding to each group of hardware parameters and the corresponding image quality score are collected. The hardware parameters and corresponding image quality scores of each group are subjected to curve fitting, and the hardware parameters corresponding to the peak points of the curve are taken as the optimal hardware parameters.

3. The method according to claim 2, characterized in that, If the compensated image quality score is lower than the preset threshold, then updating the parameter adjustment strategy and repeatedly searching for and evaluating the optimal hardware parameters until the compensated image quality score reaches or exceeds the preset threshold includes: The search step size is adjusted to a preset second step size. Taking the optimal hardware parameters as the center, the hardware parameters are traversed and adjusted in the neighborhood of the center according to the second step size. The compensated scan image is acquired and the corresponding image quality score is calculated. The second step size is smaller than the first step size. Update the optimal hardware parameters to the hardware parameters corresponding to the highest image quality score, and repeat the search process until the search image quality score meets the standard or the step size is less than the preset minimum step size.

4. A wafer imaging electrostatic compensation device, characterized in that, The device includes: The quality assessment module is used to acquire the initial scan image of the target scan position, process the initial scan image based on a preset quality assessment network, and obtain an image quality score. The parameter search module is used to search for the optimal hardware parameters within a preset range of hardware parameter values ​​according to a preset parameter adjustment strategy in response to the image quality score being lower than a preset threshold. The compensation evaluation module is used to acquire a compensated scan image based on the optimal hardware parameters, evaluate the compensated scan image through the quality evaluation network, and obtain the compensated image quality score. An iterative search module is used to update the parameter adjustment strategy and repeatedly search for and evaluate the optimal hardware parameters if the compensated image quality score is lower than the preset threshold, until the compensated image quality score reaches or exceeds the preset threshold. The quality assessment module includes: An enhanced scan image module is used to extract multi-scale high-frequency features from the initial scan image and fuse them with the initial scan image to obtain an enhanced scan image; The quality prior vector module is used to statistically analyze the brightness distribution and local contrast distribution of the enhanced scan image and generate an image quality prior vector. The branch processing module is used to input the enhanced scan image into the convolutional branch and attention branch of the quality assessment network respectively, and extract local features and global features; The weighted fusion module is used to generate dynamic fusion weights based on the image quality prior vector, and to map the weighted fusion of the local features and the global features to the image quality score. The branch processing module includes: The dynamic convolution kernel module is used to generate dynamic convolution kernel weights based on input features, and to obtain an adaptive dynamic convolution kernel by weighted combination of multiple static convolution kernels. The dynamic activation threshold module is used to extract multi-scale local features using the multi-scale adaptive dynamic convolution kernel, calculate the dynamic activation threshold based on the image quality prior vector, and output the multi-scale local features after gating activation. The branch processing module includes: The local window module is used to divide the input feature map into multiple local windows, independently calculate the attention score within each local window, and retain the part of the attention score with the highest score for weighted summation; The quality deviation module is used to calculate a quality deviation term based on the image quality prior vector and to add the quality deviation term to the attention score. The local window stitching module is used to output the sparse attention of each local window and stitch them together to restore them to their original size, thereby obtaining the global feature.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images