Wafer particle defect detection method, device, equipment, medium and program product

Abstract

The application relates to a wafer particle defect detection method, device, equipment, medium and program product, comprising: acquiring an image of a wafer to be detected; performing adaptive residual filtering preprocessing on the image to obtain a preprocessed image; wherein the preprocessed image comprises a particle candidate region sub-image subjected to first filtering processing and residual enhancement processing and a substrate region sub-image subjected to second filtering processing and residual suppression processing; a local attention enhanced feature extraction network is used to extract a candidate defect region from the preprocessed image; and authenticity verification is performed on the candidate defect region to obtain a target defect region; wherein the target defect region is a candidate defect region with a similarity greater than a similarity threshold to a reference template defect. Through adaptive residual filtering preprocessing and local attention enhancement, the application effectively highlights the feature details of small particle defects, and effectively improves the accuracy and reliability of small particle defect detection by combining authenticity verification on the candidate defect region.

Description

Technical Field

[0001] This application relates to the field of wafer defect detection technology, and in particular to a wafer particle defect detection method, apparatus, equipment, medium, and program product. Background Technology

[0002] In advanced integrated circuit manufacturing processes, small particle defects on the wafer surface are a key factor leading to decreased chip yield. When the signal-to-noise ratio (SNR) of the inspection scenario is high, such as greater than or equal to 20dB, the strong reflected signal from the substrate material is easily superimposed on the signal from small particle defects, resulting in signal redundancy interference.

[0003] In relevant defect detection schemes, filtering algorithms such as Gaussian filtering and median filtering are commonly used to obtain defect information on the wafer surface. However, Gaussian filtering and median filtering methods tend to over-smooth the edge features of small particles under high signal-to-noise ratio conditions, leading to missed defects. For the unfiltered raw signal, strong signals on the substrate are easily misidentified as defects. Summary of the Invention

[0004] Therefore, it is necessary to provide a wafer particle defect detection method, device, equipment, medium, and program product that can achieve accurate defect identification and effectively avoid false detection and missed detection in high SNR scenarios to address the above-mentioned technical problems.

[0005] In a first aspect, this application provides a method for detecting wafer particle defects, including:

[0006] Acquire an image of the wafer under test;

[0007] The image is preprocessed by adaptive residual filtering to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after first filtering and residual enhancement processing and a base region sub-image after second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering processing are both smaller than the filter kernel and smoothing intensity parameter of the second filtering processing;

[0008] Candidate defect regions are extracted from the preprocessed image using a local attention-enhanced feature extraction network; wherein, the local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect.

[0009] Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0010] In one embodiment, the adaptive residual filtering preprocessing of the image to obtain a preprocessed image includes:

[0011] The image is divided into multiple pixel blocks, and the local variance of each pixel block is calculated;

[0012] The pixel block is partitioned according to the local variance and variance threshold to obtain a particle candidate region sub-image and a base region sub-image;

[0013] The particle candidate region sub-image is subjected to a first filtering process and a residual enhancement process;

[0014] The base region sub-image is subjected to a second filtering process and residual suppression process;

[0015] The preprocessed image is obtained by fusing the candidate particle region sub-image and the base region sub-image.

[0016] In one embodiment, the step of partitioning the pixel block according to the local variance and variance threshold to obtain a particle candidate region sub-image and a base region sub-image includes:

[0017] If the local variance is greater than or equal to the variance threshold, the pixel block is determined as a candidate sub-image for particles;

[0018] If the local variance is less than the variance threshold, the pixel block is determined as a base region sub-image.

[0019] In one embodiment, the variance threshold is calculated using the following formula:

[0020]

[0021] in, Let V be the local variance of each pixel block. The maximum local variance of each pixel block. The variance threshold, The weighting coefficients are determined based on the variance of the particle signal from Monte Carlo simulation.

[0022] In one embodiment, the step of extracting candidate defect regions from the preprocessed image using a local attention-enhanced feature extraction network includes:

[0023] In the encoding stage, the preprocessed image is normalized to obtain a normalized image; a preset number of convolutional blocks are used to encode the normalized image to obtain a feature map; the target receptive field is used to traverse the feature map, extract the defective local pixels, and amplify the attention weight of the defective local pixels to obtain an attention-enhanced feature map.

[0024] During the decoding stage, upsampling and feature fusion processing are performed sequentially based on the attention-enhanced feature map to output a defect probability map; wherein, the feature fusion processing is used to fuse high-resolution shallow features and semantic deep features; threshold segmentation processing is performed on the defect probability map to obtain the candidate defect region.

[0025] In one embodiment, the pixel size of the target receptive field is ,in, are positive integers and The step of traversing the feature map using the target receptive field to extract local pixels of defects includes:

[0026] Calculate the signal difference entropy of pixels within the target's receptive field;

[0027] If the signal difference entropy is greater than or equal to a preset entropy threshold, the pixel within the target receptive field is determined to be the defective local pixel.

[0028] In one embodiment, the method further includes:

[0029] If the signal difference entropy is less than the preset entropy threshold, the pixel in the target receptive field is determined to be the base pixel;

[0030] The attention weights of the base pixels are reduced to obtain the attention-enhanced feature map.

[0031] In one embodiment, calculating the similarity between the candidate defect region and a pre-generated defect template using Monte Carlo simulation, and selecting candidate defect regions with a similarity greater than a similarity threshold as target defect regions, includes:

[0032] Obtain the physical feature parameters of the candidate defect region; wherein, the physical feature parameters include signal intensity distribution, mean edge gradient, and grayscale standard deviation features;

[0033] Calculate the similarity between the physical feature parameters and the template feature parameters, wherein the template feature parameters are the physical feature parameters of the defect template pre-generated using Monte Carlo simulation;

[0034] If the similarity is greater than or equal to a preset similarity threshold, the candidate defect region is determined to be the target defect region.

[0035] In one embodiment, the method further includes:

[0036] Monte Carlo simulation is performed based on the scene parameters of the detection scenario to pre-generate defect templates corresponding to multiple types of small particle defects; wherein, the scene parameters include electron beam parameters, substrate parameters and particle parameters.

[0037] Secondly, this application also provides a wafer particle defect detection device, comprising:

[0038] The acquisition module is used to acquire images of the wafer under test;

[0039] A preprocessing module is used to perform adaptive residual filtering preprocessing on the image to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after a first filtering process and residual enhancement processing and a base region sub-image after a second filtering process and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering process are both smaller than the filter kernel and smoothing intensity parameter of the second filtering process;

[0040] An extraction module is used to extract candidate defect regions from the preprocessed image using a local attention enhancement feature extraction network; wherein the local attention enhancement feature extraction network is used to amplify the feature intensity of local pixels of the defect.

[0041] The verification module is used to calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and to take the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the wafer particle defect detection method described in the first aspect.

[0043] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wafer particle defect detection method described in the first aspect.

[0044] Fifthly, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the wafer particle defect detection method described in the first aspect.

[0045] In summary, this application proposes a wafer particle defect detection method, apparatus, device, medium, and program product, comprising: acquiring an image of the wafer to be tested; performing adaptive residual filtering preprocessing on the image to obtain a preprocessed image; wherein the preprocessed image includes a particle candidate region sub-image after a first filtering and residual enhancement processing and a base region sub-image after a second filtering and residual suppression processing; extracting candidate defect regions from the preprocessed image using a local attention-enhanced feature extraction network; and performing authenticity verification on the candidate defect regions to obtain a target defect region; wherein the target defect region is a candidate defect region whose similarity to a reference template defect is greater than a similarity threshold. This application effectively highlights the feature details of small particle defects through adaptive residual filtering preprocessing and local attention enhancement, and combined with authenticity verification of candidate defect regions, effectively improves the accuracy and reliability of small particle defect detection. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the application environment of a wafer particle defect detection method in one embodiment.

[0047] Figure 2 This is a flowchart illustrating a wafer particle defect detection method in one embodiment;

[0048] Figure 3 This is a schematic diagram of the steps for obtaining a preprocessed image in one embodiment;

[0049] Figure 4 This is a flowchart illustrating the steps for determining a target defect region in one embodiment;

[0050] Figure 5 This is a structural block diagram of a wafer particle defect detection device in one embodiment;

[0051] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0052] To facilitate understanding of this disclosure, a more complete description will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the present disclosure. However, this disclosure can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure.

[0054] When using the terms "comprising," "having," and "including" as described herein, another component may be added unless explicitly qualifying terms such as "only," "consisting of," etc. are used. Unless otherwise stated, singular terms may include plural forms and should not be construed as having a quantity of one. Those skilled in the art will understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0055] In advanced integrated circuit manufacturing processes, tiny particle defects on the wafer surface are a key factor leading to decreased chip yield. In high signal-to-noise ratio (SNR) inspection scenarios, the strong reflected signals from the substrate material are easily superimposed on the signals from small particle defects, resulting in severe signal redundancy interference.

[0056] Specifically, in this embodiment, advanced integrated circuit manufacturing processes refer to processes of 7 nanometers (nm) and below. Microparticle defects can be small particle defects with size parameters smaller than a preset size parameter threshold. For example, in this embodiment, microparticle defects can be defects with a diameter less than or equal to 50 nm, such as small particle defects with a diameter of 20 nm, 30 nm, or 40 nm. It should be noted that the actual values ​​of the size parameters and the preset size parameter threshold can be determined according to the needs of the actual application scenario. A high signal-to-noise ratio (SNR) detection scenario refers to a detection scenario where the SNR of the acquired optical image signal or electron beam image signal from the wafer surface is greater than or equal to 20 dB.

[0057] In specific application scenarios, small particle defects on the wafer surface ranging from 20nm to 50nm include various types such as process residue particles, metallic particles, environmental and foreign particles, film-related particles, and native crystal particles. Examples include photoresist residue, metal debris, dielectric powder, silicon chips, organic contaminants, film peelings, and native crystal microparticles. In one embodiment, the wafer particle defect detection method provided in this embodiment is used to detect small particle defects ranging from 20nm to 50nm on the wafer surface.

[0058] Traditional wafer defect detection methods generally employ unified image filtering algorithms such as Gaussian filtering and median filtering. These methods tend to over-smooth the edge features of tiny particles under high signal-to-noise ratio conditions, leading to a large number of missed small particle defects. Without filtering, strong substrate signals are easily misidentified as defects, resulting in a high false detection rate. Furthermore, traditional feature extraction networks, which use large, fixed receptive fields and global attention mechanisms, cannot focus on weak signals from small particles, and defect features are easily suppressed by substrate signals, making accurate identification difficult.

[0059] More significantly, traditional inspection processes can only roughly locate defects and cannot verify the physical authenticity of suspected defects. When it is necessary to confirm the type and cause of defects, wafers need to be broken up and sampled for offline analysis using a TEM (Transmission Electron Microscope). This method not only results in the direct scrapping of wafers and high material costs, but also has a total cycle of more than seven days for sample preparation, delivery, and analysis, requiring a large amount of manpower for tracking and verification, and cannot achieve real-time inspection and rapid feedback. By the time the defect analysis results are available, a large number of wafers from the same batch have already flowed into subsequent processes, and the defect problem continues to spread, seriously delaying the opportunity to improve yield and failing to meet the needs of advanced processes for efficient, non-destructive, and high-precision online inspection.

[0060] The wafer particle defect detection method provided in this embodiment can not only accurately identify the pixels where small particle defect regions of 20nm-50nm are located on the wafer surface, but also compare the pre-generated defect template based on Monte Carlo simulation with the candidate defect region to complete the authenticity verification of suspected defects. There is no need to break the wafer for sampling and send it to TEM for offline analysis. The type and cause of small particle defects can be determined online through template comparison, which effectively improves the efficiency of wafer particle defect detection and greatly reduces the material and labor costs of wafer particle defect detection solutions.

[0061] The wafer particle defect detection method provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown is as follows. Terminal 102 communicates with server 104 via a network. Terminal 102 is used to acquire images of the wafer under test, issue inspection commands, and display defect detection results. Server 104 is used to perform core calculations such as adaptive residual filtering, local attention enhancement feature extraction, and defect template matching. The data storage system is used to store the original wafer images, Monte Carlo simulation defect templates, inspection parameters, and defect analysis reports. It can be integrated on server 104 or deployed in the cloud or on a dedicated storage server.

[0062] The terminal 102 includes, but is not limited to, specialized equipment for semiconductor inspection scenarios such as wafer optical inspection machines, electron beam inspection equipment, industrial computers, and inspection station host computers. The server 104 can be a standalone server, a computing cluster composed of multiple servers, or a cloud-based elastic computing node to meet the computing power requirements for high-speed wafer image processing and online real-time inspection.

[0063] In one embodiment, such as Figure 2 As shown, a method for detecting wafer grain defects is provided, which is applied to... Figure 1 Taking the server in the example, the following steps are included:

[0064] Step 201: Obtain an image of the wafer to be tested.

[0065] In this embodiment, the wafer under test refers to a semiconductor wafer that has entered the defect detection process and whose surface particle defects need to be identified. The image of the wafer under test is an image of the wafer surface acquired by an optical imaging device or an electron beam imaging device. It should be noted that the process technology of the wafer under test can be determined according to the needs of the actual application scenario. For example, the process technology of the wafer under test can be an advanced process technology of 7nm or below.

[0066] Specifically, this step involves acquiring a raw image of the wafer surface to be inspected, providing a data foundation for subsequent defect detection. The image can be acquired by optical inspection equipment or electron beam inspection equipment in a semiconductor production line, covering the entire wafer surface. The image resolution must meet the imaging requirements for micro-particle defects, ensuring that defect signals can be clearly captured. For example, the image resolution of the wafer under test can be 1024×1024 or higher. It should be noted that the size parameter range of the micro-particle defects can be determined according to the needs of the actual application scenario; for example, the size parameter range of micro-particle defects may be 20nm-50nm.

[0067] It should be noted that in practical application scenarios, real-time wafer images can be directly obtained from optical inspection equipment or electron beam inspection equipment in semiconductor production lines, or wafer images of the wafer under test can be retrieved from storage systems or the cloud. In this embodiment, the specific method of obtaining images of the wafer under test can be determined according to the actual application requirements.

[0068] Step 202: Perform adaptive residual filtering preprocessing on the image to obtain a preprocessed image.

[0069] In this embodiment, the preprocessed image includes a particle candidate region sub-image that has undergone a first filtering process and residual enhancement processing, and a base region sub-image that has undergone a second filtering process and residual suppression processing. The filter kernel and smoothing intensity parameter of the first filtering process are both smaller than those of the second filtering process.

[0070] In this embodiment, adaptive residual filtering refers to a filtering method that automatically selects the filtering intensity based on the local signal characteristics of the image to enhance or suppress the residual signal. Specifically, the particle candidate region refers to the area in the image where the signal fluctuates greatly and small particle defects may exist. The substrate region refers to the flat area of ​​the wafer in the image where the signal is uniform and defect-free.

[0071] In this embodiment, the first filtering process can also be called weak filtering. Using a smaller filter kernel and a smaller smoothing intensity parameter to filter the particle candidate region sub-image can effectively protect defect details. The second filtering process can also be called strong filtering. Using a larger filter kernel and a larger smoothing intensity parameter to filter the basal region sub-image can effectively suppress strong basal interference. For example, taking Gaussian filtering as an example, a 3×3 Gaussian filter kernel and a Gaussian function standard deviation of 0.5 can be used. Weak filtering was applied to the particle candidate region sub-image using a 5×5 Gaussian filter kernel and a Gaussian function with a standard deviation of 1.2. Strong filtering is applied to the basal region sub-image. It should be noted that the filter kernel and smoothing intensity parameters can be adaptively replaced according to the specific type of filtering method in the actual application scenario.

[0072] In this embodiment, after filtering the particle candidate region sub-image and the base region sub-image, differential residual signal reconstruction processing is performed on the particle candidate region sub-image and the base region sub-image respectively. Residual enhancement processing is used to amplify the defect-related residual signal to highlight the particle edges and details, and residual suppression processing is used to weaken the base noise residual and reduce interference.

[0073] Based on the above steps, this embodiment obtains a preprocessed image that reduces noise while preserving defect details by performing partitioned differential filtering and residual processing on the acquired original wafer image. This preprocessed image consists of two parts: a particle candidate region sub-image and a substrate region sub-image. This embodiment avoids the excessive smoothing of small particle edges caused by traditional one-size-fits-all filtering. It performs weak filtering on the particle candidate region sub-image and strong filtering on the substrate region sub-image, achieving a balance between noise reduction and detail preservation. Differential residual processing further enhances the defect signal, suppresses substrate interference, and improves the reliability of feature extraction.

[0074] Step 203: A local attention-enhanced feature extraction network is used to extract candidate defect regions from the preprocessed image. The local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect area.

[0075] In this embodiment, the feature extraction network refers to a deep network used to learn features such as defect texture, edges, and gradients from images. The local attention-enhanced feature extraction network is a network model obtained by combining a deep network with local attention enhancement techniques. Specifically, the local attention-enhanced feature extraction network provided in this embodiment can be improved based on the U-Net network. In practical application scenarios, the local attention-enhanced feature extraction network provided in this embodiment includes an encoder module, a decoder module, and a skip connection module. The encoder module gradually extracts deep semantic features of the image through multiple convolution and pooling operations, while compressing the spatial resolution. The decoder module is symmetrical to the encoder module and gradually restores the spatial size of the image through upsampling operations, mapping high-level semantics to pixel-level details. The skip connection module directly passes the high-resolution feature maps of the corresponding layers in the encoder to the decoder for concatenation. In this embodiment, the local attention-enhanced feature extraction network sets a corresponding local attention module after each convolutional layer of the encoder module. This module is used to enhance the feature intensity of local defect pixels when the receptive field identifies pixels in a local defect region, thereby strengthening the feature intensity of pixels corresponding to small-particle defects.

[0076] In this embodiment, the preprocessed image is input into a Local Attention Enhancement Feature Extraction Network (LIFRIN), which automatically locates and outputs candidate defect regions that may contain defects. LIFRIN focuses on local pixels of defects, amplifying their feature intensity while suppressing interference signals from uniform substrate regions. This step accurately identifies pixels corresponding to small particles, amplifies defect signals, and significantly improves the detection capability for small defects. Structured candidate regions are output, reducing the computational load for subsequent matching.

[0077] In this embodiment, candidate defect regions are image regions that the network initially determines may be defective, including real defect regions and pseudo-defect regions. This embodiment compares the similarity between candidate defect regions and pre-generated defect templates using Monte Carlo simulation, eliminates pseudo-defect regions, and retains real defect regions. This allows for verification of defect authenticity from a physical principle perspective, effectively excluding scratches, noise, bumps, etc., as defects.

[0078] Step 204: Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0079] In this embodiment, Monte Carlo simulation is a method based on the physical laws of electron scattering to simulate real particle signals. The defect template refers to a simulated standard small particle defect template of 20nm-50nm. Similarity refers to the degree of closeness between the candidate region features and the template features. The similarity threshold is the critical value for determining whether it is a real defect. The target defect region is the finally confirmed real small particle defect.

[0080] In this embodiment, the candidate defect regions obtained in step 203 are compared with the standard defect templates generated in advance through Monte Carlo physical simulation. Candidate regions with similarity greater than a set threshold are determined as real target defect regions, while those with similarity less than the threshold are determined as false defects and eliminated.

[0081] Based on the above steps, the authenticity of defects is verified from a physical principle perspective, completely eliminating false defects such as scratches, noise, and bumps. It does not rely on human experience or purely visual features, resulting in more stable and accurate judgments. It can achieve precise detection with low false positives and low false negatives in high SNR scenarios.

[0082] Based on the above steps, this embodiment provides a wafer particle defect detection method. It achieves differentiated noise reduction without losing defect details through adaptive residual filtering, accurately focuses on small particles and enhances defect features through a local attention network, and eliminates false defects through Monte Carlo template matching physical verification, ultimately outputting the true defect. In practical applications, applying the wafer particle defect detection method provided in this embodiment to a small particle defect detection system on a wafer production line can detect real small particle defects on the wafer surface online. Furthermore, through feature enhancement and template comparison, the type and cause of small particle defects can be quickly determined without requiring wafer breakage and inspection. This significantly improves the detection efficiency of small particle defects on the wafer surface while greatly reducing the detection cost.

[0083] In one embodiment, such as Figure 3 As shown, adaptive residual filtering preprocessing is performed on the image to obtain a preprocessed image, including:

[0084] Step 301: Divide the image into multiple pixel blocks and calculate the local variance of each pixel block.

[0085] Step 302: Divide the pixel block into sub-images based on local variance and variance threshold to obtain particle candidate region sub-images and base region sub-images.

[0086] Step 303: Perform first filtering and residual enhancement processing on the particle candidate region sub-image.

[0087] Step 304: Perform a second filtering process and residual suppression process on the base region sub-image.

[0088] Step 305: Fuse the candidate particle region sub-image and the base region sub-image to obtain a preprocessed image.

[0089] In this embodiment, the image of the wafer under test obtained in step 201 can be divided into blocks according to a preset pixel block size to obtain multiple pixel blocks. The preset pixel block size is adapted to the pixel range of small particle defects, avoiding misjudgment due to single-pixel noise. For example, a 1024×1024 resolution reflected electronic signal image, divided into 8×8 pixel blocks, yields 16384 independent 8×8 pixel blocks. A 2024×2024 resolution image, divided into 8×8 pixel blocks, yields 65536 independent 8×8 pixel blocks.

[0090] After image segmentation, the local variance of each pixel block is calculated. Specifically, for pixel blocks In pixels Centered on, pixel blocks The local variance can be calculated using the following formula:

[0091]

[0092]

[0093] in, For pixel blocks The local mean of all pixels within the range. For pixel blocks The value of a single pixel within. Let W be the pixel coordinates of a single pixel within pixel block W. For pixel blocks The local variance.

[0094] In this embodiment, the pixel block is partitioned according to the local variance and the variance threshold to obtain a particle candidate region sub-image and a base region sub-image, including: if the local variance is greater than or equal to the variance threshold, the pixel block is determined as a particle candidate region sub-image; if the local variance is less than the variance threshold, the pixel block is determined as a base region sub-image.

[0095] Specifically, if the local variance is greater than or equal to the variance threshold, it indicates that the region corresponding to the pixel block has obvious structural changes or boundaries, suggesting that the region contains the edges, texture details, or significant gray-level transitions of the target particles. Therefore, the region corresponding to the pixel block can be identified as a candidate particle sub-image. If the local variance is less than the variance threshold, it indicates that the pixel value distribution within the pixel block is relatively uniform and smooth, lacking significant high-frequency features. Therefore, the region corresponding to the pixel block can be identified as a relatively flat basal region sub-image.

[0096] In one embodiment, the formula for calculating the variance threshold is:

[0097]

[0098] in, Let V be the local variance of each pixel block. The maximum local variance of each pixel block. The variance threshold, The weighting coefficients are determined based on the variance of the particle signal from Monte Carlo simulation.

[0099] In this embodiment, the variance threshold is determined based on the particle signal variance obtained from Monte Carlo simulation. By employing a dynamic variance threshold, the threshold more closely matches the physical signal, enabling more accurate partitioning and thus improving the accuracy of defect feature extraction.

[0100] In this embodiment, the specific implementation methods of the first and second filtering processes can refer to the specific implementation methods in the foregoing embodiments, and will not be repeated here. For the residual reconstruction processing, the residual signal of the corresponding pixel block can be calculated based on the pixels of the original pixel block and the pixels of the filtered pixel block, respectively. , the residual signal This includes critical details such as defect edges and minute particles. Residual signal. The calculation formula is as follows:

[0101]

[0102] in, The pixels of the original pixel block, These are the pixels of the filtered pixel block.

[0103] For the particle candidate region sub-image, residual enhancement processing is performed, i.e. This amplifies the difference in defect signals. Residual suppression processing is then applied to the sub-image of the base region, i.e. This is to reduce residual noise in the base region. It should be noted that the coefficients of the residual enhancement treatment... A value greater than 1 indicates the coefficient of residual suppression treatment. Less than 1. For example, It can be set to any value in the range of 1.2-1.8, such as 1.5. It can be set to any value within the range of 0.2-0.5, such as 0.3. It should be noted that the coefficients for residual enhancement and residual suppression can be configured according to the needs of the actual application scenario.

[0104] After performing residual enhancement processing on the particle candidate region sub-image and residual suppression processing on the basal region sub-image respectively, the reconstructed preprocessed image is shown as follows:

[0105]

[0106] in, This is the preprocessed signal obtained from the reconstruction. In this embodiment, the preprocessed image is the image obtained by fusing the particle candidate region sub-image and the base region sub-image. Through differential filtering and residual processing, noise reduction can be achieved without losing details, and noise suppression can be achieved without masking defects.

[0107] Based on the above steps, this embodiment partitions regions according to local variance, which can automatically and accurately distinguish candidate regions from base regions without the need for manual threshold setting. The use of dynamic variance thresholding makes the threshold more closely match the physical signal, resulting in more accurate partitioning. Block-based computation and fusion are highly efficient, with natural edge transitions and no block artifacts.

[0108] In one embodiment, a local attention-enhanced feature extraction network is used to extract candidate defect regions from a preprocessed image, including:

[0109] In the encoding stage, the preprocessed image is normalized to obtain a normalized image; a preset number of convolutional blocks are used to encode the normalized image to obtain a feature map; the target receptive field is used to traverse the feature map, extract local pixels of defects, and amplify the attention weights of local pixels of defects to obtain an attention-enhanced feature map.

[0110] During the decoding stage, upsampling and feature fusion are performed sequentially based on the attention-enhanced feature map to output a defect probability map. The feature fusion process is used to fuse high-resolution shallow features and semantic deep features. Threshold segmentation is performed on the defect probability map to obtain candidate defect regions.

[0111] The pixel size of the target receptive field is... ,in, are positive integers and The target receptive field traverses the feature map to extract local pixels of defects, including: calculating the signal difference entropy of pixels within the target receptive field; if the signal difference entropy is greater than or equal to a preset entropy threshold, the pixels within the target receptive field are determined to be local pixels of defects.

[0112] In one embodiment, the wafer particle defect detection method further includes: if the signal difference entropy is less than a preset entropy threshold, determining the pixel in the target receptive field as the base pixel; reducing the attention weight of the base pixel to obtain an attention-enhanced feature map.

[0113] In this embodiment, the local attention enhancement feature extraction network is an improvement on the U-net network, and the decoder module normalizes the pixel values ​​of the preprocessed image to... A normalized image is obtained from the interval, eliminating computational interference caused by differences in image brightness. Four sets of cascaded convolutional blocks are selected to encode the normalized image layer by layer, with the convolutional channels progressively expanded from 64 to 128, 256, and 512. Batch Normalization (BN) and Rectified Linear Unit (ReLU) activation functions are used to extract image features of different dimensions layer by layer, generating the original feature map. A target receptive field of size 3×3-5×5 is selected and slides through the entire feature map window by window. The signal difference entropy of pixels within each window is calculated, and a preset entropy threshold is used as the judgment criterion: if the window signal difference entropy is greater than or equal to the threshold, the corresponding pixel is judged as a defective local pixel, and the attention weight of this part of the pixel is significantly increased; if the window entropy value is lower than the threshold, it is assigned to the base pixel, and its attention weight is compressed. After weight mapping and multiplication, an attention-enhanced feature map is generated.

[0114] Based on the above steps, the encoder module uses a 3×3-5×5 receptive field size to match the pixel coverage of imaging for small particles ranging from 20nm to 50nm, effectively avoiding the problem of large-area substrate signals being mixed in with large receptive fields. By judging defect pixels based on signal difference entropy, the defect location can be automatically identified. Furthermore, by directionally enhancing weak features of small particles and suppressing strong signal interference from the substrate, the problem of small target features being submerged in traditional large receptive field networks can be effectively solved.

[0115] In a specific embodiment, when assigning attention weights, pixels containing small particles can be given high weights to amplify weak signals, while base pixels can be given low weights to suppress strong signal interference. For example, if a pixel signal abruptly changes within a 3×3 region, matching the characteristics of small particles, the attention weights will focus on the abruptly changing pixel, amplifying its feature intensity by 2-3 times, while the weights of the surrounding uniform base pixels are compressed to below 0.5. It should be noted that the degree of attention weight focusing on the abruptly changing pixel and the degree of attention weight compression on the base pixels can be configured according to the actual application requirements.

[0116] The decoder module uses the attention-enhanced feature map output by the encoder module as a basis, and performs progressive upsampling through transposed convolution. Simultaneously, it performs cross-layer feature fusion, splicing together high-resolution shallow features that retain detailed information from the encoder with deep features rich in defect semantic information from the decoder, thus compensating for the loss of detail in small-granular features during the upsampling process. After multi-layer feature fusion, a pixel-level defect probability map is output, and the probability is adjusted according to a preset probability threshold. Figure 2 Value-based segmentation is used, and the highlighted areas obtained after segmentation are the candidate defect areas.

[0117] Based on the above steps, the decoder module, in conjunction with the encoder module, can achieve deep and shallow feature fusion that balances positional accuracy and semantic features, accurately delineate all suspected defects through threshold segmentation, and output complete candidate defects. The local attention-enhanced feature extraction network provided in this embodiment balances detection rate and positioning accuracy.

[0118] In one embodiment, such as Figure 4 As shown, the similarity between candidate defect regions and pre-generated defect templates using Monte Carlo simulation is calculated. Candidate defect regions with similarity greater than a similarity threshold are selected as target defect regions, including:

[0119] Step 401: Obtain the physical feature parameters of the candidate defect region. These physical feature parameters include signal intensity distribution, mean edge gradient, and grayscale standard deviation.

[0120] Step 402: Calculate the similarity between the physical feature parameters and the template feature parameters. The template feature parameters are the physical feature parameters of the defect template pre-generated using Monte Carlo simulation.

[0121] Step 403: If the similarity is greater than or equal to the preset similarity threshold, the candidate defect region is determined as the target defect region.

[0122] In this embodiment, key physical features are unified for each candidate defect block obtained after feature extraction, including the signal intensity distribution within the region, the mean gradient of the defect edge, and the standard deviation of the block gray level. These three features are derived based on the physical laws of electron beam imaging and can objectively characterize the scattering imaging characteristics of small particle defects in the 20nm-50nm range, distinguishing them from the imaging features of pseudo-defects such as scratches, substrate protrusions, and random noise.

[0123] A cosine similarity algorithm is used to compare the measured physical characteristics of candidate defect regions with the standard feature parameters of pre-stored defect templates. The feature parameters of the defect templates are generated in advance based on the physical mechanism of electron scattering through Monte Carlo simulation. The defect template library includes standard features corresponding to small particles of different materials and particle sizes.

[0124] In practical applications, a similarity threshold is pre-configured. When the similarity calculated between the candidate region features and the template features is greater than or equal to this threshold, it indicates that the imaging pattern of the region conforms to the electron scattering characteristics of real particles, and it is determined to be a real target defect region. If the similarity is lower than the threshold, it is determined to be a false defect and directly rejected, thus eliminating false detections at the physical level. It should be noted that the similarity threshold can be customized according to the needs of the actual application scenario, for example, set to 0.85.

[0125] In one embodiment, the wafer particle defect detection method further includes:

[0126] Monte Carlo simulation is performed based on the scene parameters of the detection scenario to pre-generate defect templates corresponding to multiple types of small particle defects; among which, the scene parameters include electron beam parameters, substrate parameters and particle parameters.

[0127] In this embodiment, defect templates are pre-built offline in batches. Monte Carlo simulation modeling is performed based on the actual scene parameters of the on-site inspection. These scene parameters specifically include electron beam operating parameters, wafer substrate material parameters, and the physical property parameters of the particles to be tested. By transforming the various scene parameters, standard defect templates corresponding to multiple types of small particle defects are generated through simulation and archived into a database. The electron beam operating parameters include accelerating voltage and incident beam current, while the wafer substrate material parameters include silicon substrate or metal wiring substrate. The physical property parameters of the particles to be tested include particle material and particle size.

[0128] In practical applications, the corresponding working condition template is directly retrieved from the standard defect module library during actual online detection, eliminating the need for real-time simulation and balancing detection accuracy and operational efficiency.

[0129] Based on the above steps, this embodiment abandons the traditional visual judgment method that relies solely on image grayscale and contours. Instead, it uses the physical laws of electron scattering as constraints to eliminate detected false defects, significantly reducing the false detection rate in high SNR small particle defect detection conditions. Furthermore, this embodiment employs three-dimensional physical feature joint matching, which is more stable than single-feature comparison and adaptable to the detection of tiny particles from different processes and materials. Templates are pre-simulated and prefabricated, and online detection allows for rapid retrieval, ensuring the algorithm meets the real-time detection speed requirements of the production line. The entire process achieves non-destructive wafer inspection, eliminating the need for fragmented wafers for TEM analysis. The inspection cycle is shortened from the traditional 7 days to less than 1 hour, saving wafer scrapping costs and manual inspection costs.

[0130] In summary, this embodiment provides a wafer particle defect detection method. Through the coordinated operation of a three-layer architecture—adaptive residual filtering preprocessing, local attention-enhanced feature extraction, and Monte Carlo template similarity verification—the following technical effects are achieved:

[0131] First, for high signal-to-noise ratio detection environments with SNR≥20dB, image partitioning is achieved based on the local variance of pixel blocks. Different filtering intensities and residual processing strategies are adopted for the particle candidate region and the substrate region, respectively. This avoids the problem of missed detection caused by excessive smoothing of the edges of 20nm-50nm micro particles due to traditional unified filtering, and can effectively suppress the redundant interference caused by the strong reflection signal of the substrate.

[0132] Second, a 3×3-5×5 small target receptive field is used in combination with signal difference entropy to determine defect pixels. By leveraging the local attention mechanism, the feature weights of defect pixels are dynamically amplified and the features of base pixels are suppressed. This solves the problem that large receptive fields in conventional networks easily bury weak features of small particles, and effectively improves the ability to capture fine particle defects.

[0133] Third, Monte Carlo simulations are completed based on electron beam parameters, substrate parameters, and particle parameters, and multiple defect templates are prefabricated. Similarity comparison is completed based on multiple physical features such as signal intensity distribution, mean edge gradient, and gray scale standard deviation. False defects such as scratches, substrate protrusions, and random noise are eliminated from the perspective of physical scattering laws. The false defect elimination rate can reach more than 99%, which significantly reduces the probability of false detection.

[0134] Fourth, the entire algorithm can complete online non-destructive testing at the optical or electron beam inspection equipment, abandoning the traditional testing mode that relies on offline TEM analysis after fragment sampling. The testing time is shortened from the traditional seven days to less than one hour, avoiding the scrapping and loss of wafer fragments, saving material and labor costs, and facilitating the production line to quickly locate the source of defects and carry out timely process rectification.

[0135] Fifth, the algorithm architecture is adaptable to the detection of multiple substrates and small particles of different materials. The defect template can be updated simply by adjusting the Monte Carlo simulation parameters. At the same time, the preprocessing blocks can be computed in parallel, and the local attention module has a smaller number of parameters, which can meet the real-time detection requirements of 1024×1024 resolution images at more than 20fps, taking into account both versatility and detection efficiency.

[0136] 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.

[0137] Based on the same inventive concept, this application also provides a wafer particle defect detection device for implementing the wafer particle defect detection method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more wafer particle defect detection device embodiments provided below can be found in the limitations of the wafer particle defect detection method described above, and will not be repeated here.

[0138] In one embodiment, such as Figure 5 As shown, a wafer particle defect detection device 500 is provided, including: an acquisition module 510, a preprocessing module 520, an extraction module 530, and a verification module 540, wherein:

[0139] The acquisition module 510 is used to acquire an image of the wafer under test;

[0140] The preprocessing module 520 is used to perform adaptive residual filtering preprocessing on the image to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after the first filtering and residual enhancement processing and a base region sub-image after the second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering are both smaller than the filter kernel and smoothing intensity parameter of the second filtering.

[0141] The extraction module 530 is used to extract candidate defect regions from the preprocessed image using a local attention enhancement feature extraction network; wherein, the local attention enhancement feature extraction network is used to amplify the feature intensity of local pixels of the defect.

[0142] The verification module 540 is used to calculate the similarity between candidate defect regions and defect templates pre-generated using Monte Carlo simulation, and to select candidate defect regions with similarity greater than a similarity threshold as target defect regions.

[0143] In one embodiment, the preprocessing module 520 is further configured to divide the image into multiple pixel blocks and calculate the local variance of each pixel block; partition the pixel blocks according to the local variance and a variance threshold to obtain a particle candidate region sub-image and a base region sub-image; perform a first filtering process and a residual enhancement process on the particle candidate region sub-image; perform a second filtering process and a residual suppression process on the base region sub-image; and fuse the particle candidate region sub-image and the base region sub-image to obtain a preprocessed image.

[0144] In one embodiment, the preprocessing module 520 is further configured to determine the pixel block as a candidate particle sub-image if the local variance is greater than or equal to a variance threshold, and to determine the pixel block as a base region sub-image if the local variance is less than the variance threshold.

[0145] In one embodiment, the extraction module 530 is further configured to, during the encoding stage, normalize the preprocessed image to obtain a normalized image; encode the normalized image using a preset number of convolutional blocks to obtain a feature map; and extract local pixels of defects by traversing the feature map using the target receptive field, and amplify the attention weights of the local pixels of defects to obtain an attention-enhanced feature map.

[0146] During the decoding stage, upsampling and feature fusion are performed sequentially based on the attention-enhanced feature map to output a defect probability map. The feature fusion process is used to fuse high-resolution shallow features and semantic deep features. Threshold segmentation is performed on the defect probability map to obtain candidate defect regions.

[0147] In one embodiment, the extraction module 530 is further configured to calculate the signal difference entropy of pixels within the target receptive field; if the signal difference entropy is greater than or equal to a preset entropy threshold, the pixels within the target receptive field are determined to be defective local pixels.

[0148] In one embodiment, the preprocessing module 520 is further configured to determine the pixels within the target receptive field as base pixels if the signal difference entropy is less than a preset entropy threshold; reduce the attention weight of the base pixels to obtain an attention-enhanced feature map.

[0149] In one embodiment, the verification module 540 is further configured to acquire physical feature parameters of the candidate defect region; wherein the physical feature parameters include signal intensity distribution, mean edge gradient, and grayscale standard deviation features; calculate the similarity between the physical feature parameters and the template feature parameters, wherein the template feature parameters are the physical feature parameters of the defect template pre-generated using Monte Carlo simulation; if the similarity is greater than or equal to a preset similarity threshold, the candidate defect region is determined to be the target defect region.

[0150] In one embodiment, the verification module 540 is further configured to perform Monte Carlo simulation based on the scene parameters of the detection scene to pre-generate defect templates corresponding to multiple types of small particle defects; wherein the scene parameters include electron beam parameters, substrate parameters and particle parameters.

[0151] In summary, this embodiment provides a wafer particle defect detection device that achieves differentiated noise reduction without losing defect details through adaptive residual filtering, accurately focuses on small particles and enhances defect features through local attention networks, and eliminates false defects through Monte Carlo template matching physical verification, ultimately outputting the true defects.

[0152] Each module in the aforementioned wafer particle defect detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0153] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6As 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 the computer program is executed by the processor, it implements a wafer particle defect detection 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.

[0154] Those skilled in the art will understand that Figure 6 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.

[0155] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0156] Acquire an image of the wafer under test;

[0157] The image is preprocessed by adaptive residual filtering to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after first filtering and residual enhancement processing and a base region sub-image after second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering processing are both smaller than the filter kernel and smoothing intensity parameter of the second filtering processing;

[0158] Candidate defect regions are extracted from the preprocessed image using a local attention-enhanced feature extraction network; wherein, the local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect.

[0159] Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0161] Acquire an image of the wafer under test;

[0162] The image is preprocessed by adaptive residual filtering to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after first filtering and residual enhancement processing and a base region sub-image after second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering processing are both smaller than the filter kernel and smoothing intensity parameter of the second filtering processing;

[0163] Candidate defect regions are extracted from the preprocessed image using a local attention-enhanced feature extraction network; wherein, the local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect.

[0164] Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0165] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0166] Acquire an image of the wafer under test;

[0167] The image is preprocessed by adaptive residual filtering to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after first filtering and residual enhancement processing and a base region sub-image after second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering processing are both smaller than the filter kernel and smoothing intensity parameter of the second filtering processing;

[0168] Candidate defect regions are extracted from the preprocessed image using a local attention-enhanced feature extraction network; wherein, the local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect.

[0169] Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

[0170] 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.

[0171] 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.

[0172] 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 detecting wafer particle defects, characterized in that, The method includes: Acquire an image of the wafer under test; The image is preprocessed by adaptive residual filtering to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after first filtering and residual enhancement processing, and a base region sub-image after second filtering and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering process are both smaller than the filter kernel and smoothing intensity parameter of the second filtering process; Candidate defect regions are extracted from the preprocessed image using a local attention-enhanced feature extraction network; wherein, the local attention-enhanced feature extraction network is used to amplify the feature intensity of local pixels in the defect. Calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and select the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

2. The method according to claim 1, characterized in that, The step of performing adaptive residual filtering preprocessing on the image to obtain a preprocessed image includes: The image is divided into multiple pixel blocks, and the local variance of each pixel block is calculated; The pixel block is partitioned according to the local variance and variance threshold to obtain a particle candidate region sub-image and a base region sub-image; The particle candidate region sub-image is subjected to a first filtering process and a residual enhancement process; The base region sub-image is subjected to a second filtering process and residual suppression process; The preprocessed image is obtained by fusing the candidate particle region sub-image and the base region sub-image.

3. The method according to claim 2, characterized in that, The step of partitioning the pixel block according to the local variance and variance threshold to obtain a particle candidate region sub-image and a base region sub-image includes: If the local variance is greater than or equal to the variance threshold, the pixel block is determined as a candidate sub-image for particles; If the local variance is less than the variance threshold, the pixel block is determined as a base region sub-image.

4. The method according to claim 2, characterized in that, The formula for calculating the variance threshold is: in, Let V be the local variance of each pixel block. The maximum local variance of each pixel block. The variance threshold, The weighting coefficients are determined based on the variance of the particle signal from Monte Carlo simulation.

5. The method according to claim 1, characterized in that, The step of extracting candidate defect regions from the preprocessed image using a local attention-enhanced feature extraction network includes: In the encoding stage, the preprocessed image is normalized to obtain a normalized image; a preset number of convolutional blocks are used to encode the normalized image to obtain a feature map; the target receptive field is used to traverse the feature map, extract the defective local pixels, and amplify the attention weight of the defective local pixels to obtain an attention-enhanced feature map. During the decoding stage, upsampling and feature fusion processing are performed sequentially based on the attention-enhanced feature map to output a defect probability map; wherein, the feature fusion processing is used to fuse high-resolution shallow features and semantic deep features; threshold segmentation processing is performed on the defect probability map to obtain the candidate defect region.

6. The method according to claim 5, characterized in that, The pixel size of the target receptive field is ,in, are positive integers and The step of traversing the feature map using the target receptive field to extract local pixels of defects includes: Calculate the signal difference entropy of pixels within the target's receptive field; If the signal difference entropy is greater than or equal to a preset entropy threshold, the pixel within the target receptive field is determined to be the defective local pixel.

7. The method according to claim 6, characterized in that, The method further includes: If the signal difference entropy is less than the preset entropy threshold, the pixel in the target receptive field is determined to be the base pixel; The attention weights of the base pixels are reduced to obtain the attention-enhanced feature map.

8. The method according to claim 1, characterized in that, The calculation of the similarity between the candidate defect region and the pre-generated defect template using Monte Carlo simulation, and the selection of candidate defect regions with a similarity greater than a similarity threshold as target defect regions, includes: Obtain the physical feature parameters of the candidate defect region; wherein, the physical feature parameters include signal intensity distribution, mean edge gradient, and grayscale standard deviation features; Calculate the similarity between the physical feature parameters and the template feature parameters, wherein the template feature parameters are the physical feature parameters of the defect template pre-generated using Monte Carlo simulation; If the similarity is greater than or equal to a preset similarity threshold, the candidate defect region is determined to be the target defect region.

9. The method according to claim 8, characterized in that, The method further includes: Monte Carlo simulation is performed based on the scene parameters of the detection scenario to pre-generate defect templates corresponding to multiple types of small particle defects; wherein, the scene parameters include electron beam parameters, substrate parameters and particle parameters.

10. A wafer particle defect detection device, characterized in that, The device includes: The acquisition module is used to acquire images of the wafer under test; A preprocessing module is used to perform adaptive residual filtering preprocessing on the image to obtain a preprocessed image; wherein, the preprocessed image includes a particle candidate region sub-image after a first filtering process and residual enhancement processing and a base region sub-image after a second filtering process and residual suppression processing; the filter kernel and smoothing intensity parameter of the first filtering process are both smaller than the filter kernel and smoothing intensity parameter of the second filtering process; An extraction module is used to extract candidate defect regions from the preprocessed image using a local attention enhancement feature extraction network; wherein the local attention enhancement feature extraction network is used to amplify the feature intensity of local pixels of the defect. The verification module is used to calculate the similarity between the candidate defect region and the defect template pre-generated using Monte Carlo simulation, and to take the candidate defect region with a similarity greater than the similarity threshold as the target defect region.

11. 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 9.

12. 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 9.

13. 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 9.

Images