Wafer defect classification algorithm fusing extended receptive field and local weighting

Abstract

The invention discloses a wafer defect classification algorithm fusing extended receptive field and local weighting, and relates to image enhancement, which comprises the following steps: inputting a wafer image to be detected; performing feature extraction on the wafer image through adaptive weighted downsampling operation; performing multi-scale feature extraction and fusion on the feature map through expansion convolution operation and re-parameterization operation; optimizing defect information in the global feature map through local weighting and context information guide operation; inputting the optimized global feature map into a classification network, and classifying defects in the wafer image through a full connection layer; and judging whether the wafer image has defects or not according to a classification result. According to the method, by comprehensively applying an extended receptive field, local weighting and context guidance technology, the performance of the module in multi-scale feature fusion and small target detection is effectively improved, the false alarm rate is greatly reduced, and the detection precision is improved.

Description

Technical Field

[0001] The present invention relates to image enhancement, and more particularly, to a wafer defect classification algorithm integrating extended receptive field and local weighting. Background Art

[0002] With the continuous advancement of integrated circuit technology and the increasing sophistication of microelectronic products, wafer quality inspection during semiconductor manufacturing has become a critical step in ensuring product performance and stability. Wafer defect detection is not only a vital step in semiconductor manufacturing but also directly impacts production efficiency, yield rate, and manufacturing costs. Traditional defect detection methods, such as manual visual inspection and traditional image processing, have limitations and are unable to meet the demands of increasingly sophisticated inspections.

[0003] In recent years, computer vision and deep learning technologies have made significant progress in image recognition and defect detection. In particular, image classification and object detection methods based on convolutional neural networks (CNNs) and deep learning have become mainstream technologies for semiconductor wafer defect inspection. These methods leverage the powerful properties of neural networks to extract deep features from large amounts of image data, enabling automated and accurate defect detection.

[0004] In existing wafer defect inspection systems, common image processing technologies include traditional machine learning-based methods and deep learning-based approaches. Traditional image processing methods, such as edge detection, template matching, and morphological operations, rely on handcrafted feature design and fail to fully exploit the deep-level information in images. Furthermore, detection results are often suboptimal when faced with complex defect types. However, with the continuous advancement of deep learning technology, detection methods based on deep neural networks have begun to be applied to wafer defect inspection and have achieved considerable progress.

[0005] Existing deep learning methods typically use convolutional neural networks (CNNs) or their variants, such as YOLO and ResNet, to extract image features and perform classification through multi-layer convolution operations. However, these methods still suffer from problems such as insufficient receptive field, neglect of local information, and ineffective use of contextual information, which limits detection accuracy. They are particularly poor in detecting complex backgrounds or small defects. This is mainly manifested in:

[0006] 1. Insufficient receptive field, unable to capture global information:

[0007] Most existing deep learning methods rely on traditional convolutional operations, which have limited receptive fields and fail to fully capture the global information of an image. In wafer defect detection, defects often exhibit significant variability, with different types of defects appearing in different spatial locations. The small receptive field of traditional convolutional layers can easily lead to unstable and inaccurate detection results.

[0008] 2. Insufficient local information processing:

[0009] In wafer defect detection, defects are often localized and tiny, and traditional methods often overlook the importance of these local features. Existing technologies mostly use global feature extraction, which has difficulty processing local details on the wafer surface.

[0010] 3. Contextual information is not fully utilized:

[0011] Traditional deep learning methods often ignore the guidance of contextual information when classifying defects. In actual inspections, the nature of defects can be affected by the surrounding environment and background, and relying solely on local features cannot accurately determine the type and nature of defects.

[0012] 4. Low detection accuracy and high false alarm rate:

[0013] Existing technologies suffer from low accuracy and high false positive rates when faced with complex backgrounds and multiple defect types. Some methods perform particularly poorly when detecting small defects, which is closely related to their lack of processing multi-scale features and insufficient ability to capture fine-grained defects.

[0014] 5. Complex structure and high computational overhead:

[0015] Some existing deep learning modules are complex in structure and computationally expensive, making them difficult to apply in real-time detection. These modules require a large amount of computing resources and storage space, resulting in high costs and slow operation. Summary of the Invention

[0016] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a wafer defect classification algorithm that integrates extended receptive field and local weighting. By optimizing the algorithm structure, it maintains high precision while reducing the amount of calculation and storage requirements, has higher computing efficiency and lower implementation cost, and effectively solves multiple problems of the existing technology in wafer defect detection.

[0017] The wafer defect classification algorithm of the present invention that integrates extended receptive field and local weighting includes the following steps:

[0018] S1. Input the wafer image to be inspected;

[0019] S2. Constructing a DPLAWDS module, wherein the DPLAWDS module performs feature extraction on the wafer image through an adaptive weighted downsampling operation to obtain a feature map having a key defect area;

[0020] S3. Construct a BNDWR_DRB module, which performs multi-scale feature extraction and fusion on the feature map through dilated convolution operations and reparameterization operations to form a global feature map;

[0021] S4, build LRContextGuidedBlock_Down module,

[0022] The LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through local weighting and context information guidance operations;

[0023] S5. Inputting the optimized global feature map into the classification network, and classifying the defects in the wafer image through a fully connected layer;

[0024] S6. Determine whether the wafer image has defects based on the classification result. If defects are present, output the defect category. If no defects are present, return to a no-defect state.

[0025] Preferably, the DPLAWDS module processes wafer images in the following steps:

[0026] S21, applying an average pooling operation to aggregate local feature information of the wafer image;

[0027] S22. Performing feature processing on the aggregated local feature information using a 1×1 convolution to extract deep features for characterizing key defect areas in the wafer image and form an attention feature map;

[0028] S23, expanding the attention feature map into a four-channel attention map;

[0029] S24, performing convolution processing on the wafer image and reducing its resolution by half to obtain a downsampled feature map;

[0030] S25, expanding the downsampled feature map into a four-channel downsampled map;

[0031] S26. Align the downsampled image with the attention image, and map the attention weight of the attention image to the downsampled image by element-by-element multiplication;

[0032] S27, summing the last dimension of the downsampled image output in step S26 to obtain a weighted fusion feature map;

[0033] S28. Apply a Dropout layer to the weighted fusion feature map, and output a feature map with key defect areas.

[0034] Preferably, in step S26, the attention weight of the attention map is calculated as follows:

[0035] The Softmax function is used to calculate the weight of the attention map, and the weight of each spatial position of the attention map is set between [0, 1].

[0036] Preferably, the sum of the weights of all spatial positions in the attention map is 1.

[0037] Preferably, the attention feature map is expanded by a rearrange operation; and the downsampled feature map is expanded by a rearrange operation.

[0038] Preferably, in step S24, the wafer image is convolved by depthwise separable convolution self.ds_conv.

[0039] Preferably, the BNDWR_DRB module performs the following processing steps for multi-scale feature extraction and fusion on the feature map:

[0040] S31, extracting preliminary features from the feature map having the key defect area;

[0041] S32, performing dilated convolution operations and reparameterization operations on the feature map from which the preliminary features are extracted, to extract feature information of different scales;

[0042] S33. Use 1×1 convolution to fuse feature information of multiple scales to form a global feature map.

[0043] Preferably, in step S31, preliminary features are extracted from the feature map through 3×3 convolution.

[0044] Preferably, the specific steps of optimizing the defect information in the global feature map by the LRContextGuidedBlock_Down are as follows:

[0045] S41, performing downscaling processing on the global feature map through 1×1 convolution to reduce the size of the global feature map;

[0046] S42, extracting local image features of the global feature map output in step S41 through a context module or an attention mechanism;

[0047] S43: Extract feature information around the local image feature, and extract context information based on the local image feature and its surrounding feature information through dilated convolution.

[0048] Beneficial effects

[0049] The advantages of the present invention are:

[0050] 1. This paper designs the DPLAWDS module, which dynamically adjusts the receptive field size through adaptive weighted downsampling. This allows the model to adjust feature extraction weights within a region based on the distribution and importance of defects. This innovation not only expands the receptive field but also better focuses on defect areas, effectively enhancing the detection accuracy of fine-grained defects.

[0051] In terms of performance improvement: Through adaptive weighted sampling, the model can more accurately capture tiny defects in the image, improving the overall detection accuracy, especially when dealing with small targets and complex backgrounds.

[0052] In terms of improving detection accuracy: Expanding the receptive field can effectively enhance the model's ability to detect a wide range of defects, ensuring the model's detection performance at multiple scales and avoiding the blind spots of defects in traditional methods.

[0053] 2. The BNDWR_DRB module designed in this invention uses dilated convolution and reparameterization techniques to extract image features at different scales and fuse features at multiple scales. Dilated convolution significantly expands the receptive field, enhancing the ability to recognize defects of different scales. Feature fusion ensures the effective integration of multi-scale information, improving the detection accuracy of defects of varying sizes.

[0054] In terms of improving multi-scale detection capabilities: through multi-scale feature extraction and fusion, the present invention can simultaneously capture detailed features and a wide range of defect information in the image, significantly improving the ability to identify different types of defects.

[0055] In terms of robustness enhancement: For complex backgrounds or defects of different scales, the model shows stronger adaptability and robustness, reducing the false alarm rate and improving reliability in practical applications.

[0056] 3. The LRContextGuidedBlock_Down module designed by this invention combines local convolution with dilated convolution to extract local image details and global contextual information, ensuring that the model not only processes local features but also leverages global information to improve defect detection accuracy. This module's context-guided feature refinement enables the model to eliminate background noise while enhancing the representation of subtle defects.

[0057] In terms of enhancing fine-grained detection accuracy: by combining local features with contextual guidance of global information, the present invention can improve the recognition accuracy of tiny defects based on fine-grained feature extraction, and is particularly suitable for the detection of small and complex defects.

[0058] In terms of reducing false alarm rates: By eliminating background noise and enhancing the expression of defect areas, the model can more accurately distinguish defects from background, reduce the occurrence of false alarms, and improve the reliability of the detection system.

[0059] 4. Overall system performance improvement

[0060] Although the present invention enhances model complexity by incorporating multiple technical modules such as extended receptive field, local weighting, and context guidance, it maintains high computational efficiency while ensuring high accuracy by optimizing the model structure and integrating multi-level features. For example, during batch processing, the model can process large-scale images with low latency, ensuring the feasibility of real-time detection. The advantages it brings include:

[0061] Efficiency: Even with increasing model complexity, the optimized algorithm maintains low latency and high FPS, ensuring real-time processing capabilities suitable for automated testing in large-scale production environments.

[0062] Economical: By improving detection accuracy and reducing false alarm rates, the present invention can reduce the need for subsequent manual intervention, save labor costs, and improve production efficiency.

[0063] Accuracy: Across multiple models, as modules were added, accuracy gradually increased from 0.982 to 0.987, demonstrating the significant impact of adding additional modules on improving model performance.

[0064] Model size: Although the model size has increased, the increase in parameters also brings better accuracy. In particular, large models with the addition of the LRContextGuidedBlock_Down module continue to improve model accuracy despite the increase in size and computational effort.

[0065] Computing speed: Different models have varying processing speeds, but compared to the official model, the model with the optimization module can still maintain a high FPS, and the latency during batch processing is relatively small, demonstrating good real-time performance.

[0066] Adaptability: Using this algorithm to train artificial intelligence models will not be restricted by the training platform, and different hardware devices on the training platform will not have a significant impact on the model training results. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flow chart of the wafer defect classification algorithm of the present invention;

[0068] Figure 2 This is a flowchart of the DPLAWDS module execution of the present invention;

[0069] Figure 3This is a flow chart of the BNDWR_DRB module of the present invention;

[0070] Figure 4 This is a flowchart of the execution of the LRContextGuidedBlock_Down module of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described below in conjunction with the embodiments, but this does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0072] This paper presents a wafer defect classification algorithm that integrates an extended receptive field and local weighting. This algorithm primarily integrates the extended receptive field, local weighting, and context-guided techniques to improve the accuracy of wafer defect detection, particularly in the detection of small objects and minor defects, by reducing recognition errors caused by insufficient receptive field and incomplete local feature extraction. To implement this algorithm, you first need to build the DPLAWDS module, the BNDWR_DRB module, and the LRContextGuidedBlock_Down module.

[0073] The DPLAWDS module extracts features from the wafer image through adaptive weighted downsampling to obtain a feature map of key defect areas. This adaptive weighted downsampling enhances the network's sensitivity to key defect areas in the feature map. Its core principle is to dynamically adjust the receptive field of each region, focusing on important defect areas in the image. Furthermore, by weighted sampling of different regions, the DPLAWDS module enables the network to adaptively identify small defects in the image, improving the ability to express fine-grained features.

[0074] The BNDWR_DRB module extracts and fuses multi-scale features from the feature map through dilated convolution and reparameterization operations to form a global feature map. This expands the receptive field and improves the ability to extract multi-scale features. It can capture defect information in key defect areas and is particularly suitable for defect detection in complex backgrounds and at multiple scales. Its main function is to extract image features at different scales through dilated convolution operations and optimize the parameters of the convolution layer through reparameterization, thereby improving the model's recognition accuracy for defects of different sizes.

[0075] The LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through context-guided operations, further improving the detection of minor defects. Its core idea is to use context-guided learning of local features, enabling the model to capture local details while gaining support from global features, thereby improving the recognition accuracy of minor and complex defects.

[0076] See Figure 1 Based on the above modules, the specific steps of the algorithm of the present invention are as follows:

[0077] S1. Input the wafer image to be inspected. The image size is 52x52 pixels and the image format is grayscale.

[0078] S2, DPLAWDS module performs feature extraction on the wafer image through adaptive weighted downsampling operation, focusing on extracting key defect areas in the wafer image.

[0079] The S3 and BNDWR_DRB modules perform multi-scale feature extraction and fusion on the feature map through dilated convolution and reparameterization operations to form a global feature map, enabling the model to capture a wide range of defect information in the image. Because most existing deep learning methods rely on traditional convolution operations, their receptive fields are limited and cannot fully cover the global information of the image. Therefore, this invention uses dilated convolution and reparameterization operations. By integrating these techniques to expand the receptive field, it can effectively expand the receptive field and improve the ability to capture global information, thereby improving the accuracy of defect detection.

[0080] S4, LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through local weighting and context information guidance operations. It mainly optimizes the expression of small defects in the global feature map to ensure the effective combination of local details and global information. Since traditional deep learning methods often ignore the guidance of context information when classifying defects, in order to solve this problem, the present invention introduces a context guidance mechanism to enhance the module's recognition ability for complex backgrounds and diverse defects by integrating local and global information. In addition, the present invention adopts a local weighting strategy, which can focus on local areas in the image more accurately and improve the recognition ability of fine-grained defects, especially in the detection of small defects.

[0081] S5. Input the optimized global feature map into the classification network, and classify the defects in the wafer image through the fully connected layer and the Softmax activation function.

[0082] S6. Determine whether the wafer image has defects based on the classification result. If defects are present, output the defect category. If no defects are present, return to a no-defect state.

[0083] This paper proposes a high-precision wafer defect detection and classification algorithm that integrates an extended receptive field, local weighting, and context guidance. This algorithm aims to overcome the challenges of existing technologies and improve the accuracy and efficiency of wafer defect detection. It also effectively addresses the challenges of high-precision detection and classification in wafer image datasets for small objects. By effectively combining local and global information, enhancing the receptive field, and optimizing the model structure, this paper can efficiently and accurately detect various types of wafer defects, especially tiny and fine-grained defects, significantly improving the performance of the detection system.

[0084] The specific process of the corresponding steps in the above algorithm will be described in detail below.

[0085] like Figure 2 As shown in the DPLAWDS module, the core functional parts are introduced first:

[0086] Softmax activation function: The Softmax activation function is used to calculate the weight of each spatial position proportionally in subsequent calculations in order to weight the feature map.

[0087] Attention mechanism: Average pooling (AvgPool2d) and a convolution operation (Conv(ch,ch,k=1)) are used to generate the attention feature map. Generally speaking, the average pooling operation extracts spatial information by averaging over a 3×3 window, while the convolution operation transforms the features to facilitate weighted fusion in the next step.

[0088] Downsampling convolution (ds_conv): Downsampling convolution uses a 3×3 convolution kernel with a stride of 2 to reduce the resolution of the input feature map by half, and controls the receptive field of each convolution kernel and the learned features by using grouped convolution (g=(ch / / group)). This operation increases the number of channels of the feature map by 4 times.

[0089] Dropout layer (dropout): The Dropout layer is used to regularize the network and prevent overfitting by randomly dropping a certain proportion of features. Here, dropout_rate is a hyperparameter that controls the dropout ratio.

[0090] These steps together constitute the core functionality of the DPLAWDS module, namely the importance preservation of adaptive weighted downsampling, which can effectively extract key features in images.

[0091] Based on the above concepts, in step S2, the DPLAWDS module processes the wafer image in the following steps:

[0092] S21. Apply an average pooling operation to aggregate local feature information of the wafer image.

[0093] S22. Use 1×1 convolution to perform feature processing on the aggregated local feature information to extract deep features in the wafer image for characterizing key defect areas and form an attention feature map.

[0094] Specifically, a wafer image x of size (bs, ch, h, w) is first processed by an attention network. The attention network consists of average pooling and 1×1 convolution to generate an attention feature map of size (bs, ch, h, w).

[0095] S23. Expand the attention feature map into a four-channel attention map (bs, ch, h, w, 4) through a rearrange operation. This step achieves spatial feature rearrangement.

[0096] The Softmax function is then used to calculate the weight of the attention map, and the weight of each spatial position of the attention map is made between [0, 1], and the sum of the weights of all spatial positions of the attention map is 1, so that weighting can be performed effectively.

[0097] S24. Convolve the wafer image using a depthwise separable convolution (self.ds_conv) and reduce its resolution by half to obtain a downsampled feature map. The output map size is (bs, 4*ch, h / 2, w / 2).

[0098] S25. Similarly, the downsampled feature map is expanded into a four-channel downsampled map with a size of (bs, ch, h, w, 4) through a rearrange operation.

[0099] S26. Align the downsampled image with the attention image, and map the attention weight of the attention image to the downsampled image by element-by-element multiplication.

[0100] S27, summing the last dimension of the downsampled image output in step S26 to obtain a weighted fusion feature map;

[0101] S28. Apply a Dropout layer to the weighted fusion feature map and output a feature map of key defect areas. This feature map contains more representative, multi-scale information and is spatially weighted to enhance the focus on important features. This step helps improve the generalization ability of the model and reduce the risk of overfitting.

[0102] like Figure 3 As shown, in step S3, the BNDWR_DRB module performs multi-scale feature extraction and fusion processing on the feature map as follows:

[0103] S31. Extract preliminary features from the feature map having the key defect area through 3×3 convolution.

[0104] S32. Performing a dilated convolution operation and a reparameterization operation on the feature map from which the preliminary features were extracted to extract feature information at different scales. The purpose of the dilated convolution is to expand the receptive field of the convolution kernel, thereby extracting a wide range of contextual information without increasing the amount of computation. Therefore, the dilated convolution can not only extract detailed information, but also obtain a wide range of contextual information.

[0105] S33 uses 1×1 convolution to fuse feature information from multiple scales to form a global feature map. The 1×1 convolution performs channel fusion, fusing feature maps from different scales while compressing or expanding the number of channels for subsequent processing. The resulting global feature map encompasses information from both detailed and large-scale areas, effectively addressing a wide range of defects and improving the model's ability to identify multiple defects.

[0106] A common technique in existing technologies is to fuse feature maps of different scales through 1×1 convolution. This approach is primarily used to fuse feature information from different channels, thereby compressing or adjusting the number of channels, and combining multi-scale features through convolution kernel weight learning. Specifically, 1×1 convolution has the following effects:

[0107] Channel fusion: The core function of 1×1 convolution is to fuse feature maps of different scales on the channel level. This compresses or expands the channel information of multiple feature maps of different scales to obtain a more compact and information-rich feature map. It is actually a pixel-by-pixel linear transformation that fuses information on the channel dimension.

[0108] Reduce the amount of computation: Using 1×1 convolution can reduce the complexity of convolution calculation because it only focuses on channel information and does not involve convolution operations in the spatial dimension.

[0109] Information integration: 1×1 convolution can help the network better integrate multi-scale information and improve the fusion ability of multi-scale features, so that the network can retain both detail features and global context features.

[0110] like Figure 4 As shown, in step S4, the specific steps of LRContextGuidedBlock_Down optimizing the defect information in the global feature map are:

[0111] S41. Downsampling the global feature map through 1×1 convolution to reduce the size of the global feature map.

[0112] S42. Extract local image features from the global feature map output in step S41 using a context module or attention mechanism. The purpose of context-guided feature refinement is to optimize local features, thereby enhancing the network's ability to identify defects. For more information on the specific technical aspects of the context-guided mechanism, please refer to the "context module" or "attention mechanism" in the prior art. The core idea of this mechanism is to guide the network to focus on important contextual information in a local area while ignoring irrelevant background information. Commonly used techniques include:

[0113] Self-Attention: Automatically selects and enhances useful information by calculating the relationship between different regions in the image;

[0114] Spatial attention: focusing on important spatial regions in the image to enhance attention to defects.

[0115] S43: Extract feature information surrounding the local image feature and extract contextual information based on the local image feature and its surrounding features through dilated convolution, thereby refining the image. In this step, the convolutional network is used to further enhance and optimize details, improving the ability to recognize small defects.

[0116] The fully connected layer transforms input features and generates outputs through weighting and biasing. These outputs become the basis for the network's final classification. The specific implementation process for defect classification in the fully connected layer in step S5 is as follows:

[0117] The first fully connected layer: This layer inputs the flattened feature vector into a larger output space, and usually adds nonlinearity through the ReLU activation function to enhance the expressive power of the network.

[0118] Second fully connected layer (optional): Depending on the complexity of the network, multiple fully connected layers can be used. The second layer further compresses the feature dimensions, mapping information from high dimensions to lower dimensions for easier classification.

[0119] The last fully connected layer: This is the final output layer, which usually uses the Softmax activation function to map the output to the probability distribution of the category. In the wafer defect classification task, the size of the output layer should be consistent with the number of defect categories.

[0120] After each fully connected layer, an activation function is usually used to introduce nonlinearity to increase the expressive power of the model. Common activation functions include:

[0121] ReLU (Rectified Linear Unit): used in hidden layers, which can effectively alleviate the gradient disappearance problem.

[0122] Softmax: Usually used in the last layer, especially in multi-classification tasks, it converts the output value of each category into a probability distribution and outputs a normalized probability value.

[0123] This embodiment uses the Softmax activation function. For classification tasks, especially multi-category problems, the Softmax activation function normalizes the predicted values of each category and calculates the probability of each class. The category corresponding to the maximum output value is the prediction result of the model. The output of Softmax is a probability distribution, and the sum of the probabilities of all categories is 1. For example, if the output of the model is [2.3, -0.7, 1.2], Softmax will convert them into [0.76, 0.01, 0.23], indicating that the probability of the first category is 0.76, the second is 0.01, and the third is 0.23. The final model will select the first category as the prediction.

[0124] Therefore, for the classification result output of this algorithm, based on the probability output by the Softmax activation function, the category with the largest probability value is selected as the final classification result.

[0125] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention. These modifications and improvements will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A wafer defect classification algorithm that integrates extended receptive field and local weighting, characterized in that: The following steps are involved: S1. Input the wafer image to be inspected; S2. Constructing a DPLAWDS module, wherein the DPLAWDS module performs feature extraction on the wafer image through an adaptive weighted downsampling operation to obtain a feature map having a key defect area; S3. Construct a BNDWR_DRB module, which performs multi-scale feature extraction and fusion on the feature map through dilated convolution operations and reparameterization operations to form a global feature map; S4, build LRContextGuidedBlock_Down module, The LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through local weighting and context information guidance operations; S5. Inputting the optimized global feature map into the classification network, and classifying the defects in the wafer image through a fully connected layer; S6. Determine whether the wafer image has defects based on the classification result. If defects are present, output the defect category. If no defects are present, return to a no-defect state.

2. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 1 is characterized in that: The DPLAWDS module processes wafer images in the following steps: S21, applying an average pooling operation to aggregate local feature information of the wafer image; S22. Performing feature processing on the aggregated local feature information using a 1×1 convolution to extract deep features for characterizing key defect areas in the wafer image and form an attention feature map; S23, expanding the attention feature map into a four-channel attention map; S24, performing convolution processing on the wafer image and reducing its resolution by half to obtain a downsampled feature map; S25, expanding the downsampled feature map into a four-channel downsampled map; S26. Align the downsampled image with the attention image, and map the attention weight of the attention image to the downsampled image by element-by-element multiplication; S27, summing the last dimension of the downsampled image output in step S26 to obtain a weighted fusion feature map; S28. Apply a Dropout layer to the weighted fusion feature map, and output a feature map with key defect areas.

3. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 2 is characterized in that: In step S26, the attention weight of the attention map is calculated as follows: The Softmax function is used to calculate the weight of the attention map, and the weight of each spatial position of the attention map is set between [0, 1].

4. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 3 is characterized in that: The sum of the weights of all spatial positions in the attention map is 1.

5. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 2 is characterized in that: The attention feature map is expanded through a rearrange operation; the down-sampled feature map is expanded through a rearrange operation.

6. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 2, characterized in that: In step S24, the wafer image is convolved using a depthwise separable convolution self.ds_conv.

7. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 1 is characterized in that: The BNDWR_DRB module performs multi-scale feature extraction and fusion processing on the feature map as follows: S31, extracting preliminary features from the feature map having the key defect area; S32, performing dilated convolution operations and reparameterization operations on the feature map from which the preliminary features are extracted, to extract feature information of different scales; S33. Use 1×1 convolution to fuse feature information of multiple scales to form a global feature map.

8. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 7 is characterized in that: In step S31, preliminary features are extracted from the feature map through 3×3 convolution.

9. The wafer defect classification algorithm integrating extended receptive field and local weighting according to claim 1, characterized in that: The specific steps of the LRContextGuidedBlock_Down to optimize the defect information in the global feature map are: S41, performing downscaling processing on the global feature map through 1×1 convolution to reduce the size of the global feature map; S42, extracting local image features of the global feature map output in step S41 through a context module or an attention mechanism; S43: Extract feature information around the local image feature, and extract context information based on the local image feature and its surrounding feature information through dilated convolution.

Images