Wafer defect classification algorithm fusing extended receptive field and local weighting

Abstract

The application discloses a wafer defect classification algorithm fusing extended receptive field and local weighting, relates to image enhancement, and inputs a wafer image to be detected; feature extraction is performed on the wafer image through adaptive weighted down-sampling operation; multi-scale feature extraction and fusion are performed on the feature map through dilation convolution operation and reparameterization operation; defect information in the global feature map is optimized through local weighting and context information guiding operation; the optimized global feature map is input into a classification network, defects in the wafer image are classified through a full connection layer; and whether the wafer image exists defects is judged according to a classification result. Through comprehensive use of extended receptive field, local weighting and context guiding technology, the application effectively improves the performance of the module in multi-scale feature fusion and small target detection, greatly reduces the false positive rate and improves the detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to image enhancement, more particularly, it relates to a wafer defect classification algorithm fusing extended receptive field and local weighting. BACKGROUND

[0002] With the continuous development of integrated circuit technology and the increasing miniaturization of microelectronic products, the quality detection of wafers in the semiconductor manufacturing process has become a key link to ensure product performance and stability. Wafer defect detection is not only an important step in the semiconductor manufacturing process, but also directly related to production efficiency, yield and manufacturing cost. In the traditional defect detection process, there are certain limitations in using manual visual inspection and traditional image processing methods, which are difficult to meet the increasingly sophisticated detection needs.

[0003] In recent years, computer vision and deep learning technology have made significant progress in image recognition and defect detection. In particular, image classification and object detection methods based on convolutional neural networks (CNN) and deep learning have become the mainstream technology in semiconductor wafer defect detection. These methods use the powerful features of neural networks to extract deep features from a large amount of image data, thereby achieving automated and accurate defect detection.

[0004] In existing wafer defect detection systems, common image processing techniques include traditional machine learning-based methods and deep learning methods. Traditional image processing methods, such as edge detection, template matching and morphological operations, rely on the design of handcrafted features and cannot fully exploit the deep information of images. Moreover, when faced with complex defect types, the detection effect is often unsatisfactory. With the continuous development of deep learning technology, detection methods based on deep neural networks have begun to be applied to wafer defect detection and have made some progress.

[0005] Existing deep learning methods usually use convolutional neural networks (CNN) or their variants such as YOLO, ResNet, etc., to extract image features through multiple convolution operations and perform classification. However, these methods still have problems such as insufficient receptive field, neglect of local information, and ineffective use of contextual information, which limit the detection accuracy, especially in the detection of complex backgrounds or small-sized defects. The main manifestations are:

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

[0007] Most existing deep learning methods rely on traditional convolution operations, whose receptive field is limited and cannot fully cover the global information of the image. In wafer defect detection, defects often have great variability, and different types of defects may appear at different spatial locations. Due to the small receptive field of traditional convolution layers, the detection results may be unstable and inaccurate.

[0008] 2. Local information processing is insufficient:

[0009] In wafer defect detection, defects are often local and small, and traditional methods often overlook the importance of these local features. Most existing technologies use global feature extraction, which is difficult to handle local details on the wafer surface.

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

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

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

[0013] Existing technologies have low accuracy and high false positive rate when facing complex backgrounds and multiple defect types. Some methods perform poorly when detecting small-sized defects, which is closely related to their lack of multi-scale feature processing and insufficient ability to capture fine-grained defects.

[0014] 5. Complex structure, large computational overhead:

[0015] Some existing deep learning modules have complex structures and large computational overhead, making them difficult to apply in real-time detection. These modules require a large amount of computational resources and storage space, resulting in high cost and slow running speed. SUMMARY

[0016] The technical problem to be solved by the present application is to address the shortcomings of the prior art and provide a wafer defect classification algorithm that combines extended receptive field and local weighting. By optimizing the algorithm structure, the algorithm maintains high accuracy while reducing computational load and storage requirements, has higher operation efficiency and lower implementation cost, and effectively solves multiple problems in wafer defect detection in the prior art.

[0017] The wafer defect classification algorithm combining extended receptive field and local weighting comprises the following steps:

[0018] S1, input the wafer image to be detected;

[0019] S2, construct a DPLAWDS module, the DPLAWDS module extracts features from the wafer image through adaptive weighted down-sampling operation to obtain a feature map with key defect regions;

[0020] S3, constructing a BNDWR_DRB module, which performs multi-scale feature extraction and fusion on the feature map through an expansion convolution operation and a reparameterization operation to form a global feature map;

[0021] S4, constructing a LRContextGuidedBlock_Down module, which

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

[0023] S5, inputting the optimized global feature map into a classification network to classify defects in the wafer image through a fully connected layer;

[0024] S6, determining whether the wafer image has defects according to the classification result, and outputting a defect category if there are defects, or returning a defect-free state if there are no defects.

[0025] Preferably, the processing step of the DPLAWDS module on the wafer image is:

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

[0027] S22, using a 1x1 convolution to process the aggregated local feature information to extract deep features in the wafer image representing key defect regions, forming an attention feature map;

[0028] S23, unfolding 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 down-sampled feature map;

[0030] S25, unfolding the down-sampled feature map into a four-channel down-sampled map;

[0031] S26, aligning the down-sampled map with the attention map, and mapping the attention weight of the attention map to the down-sampled map by element-wise multiplication;

[0032] S27, summing the last dimension of the down-sampled map output by step S26 to obtain a weighted fusion feature map;

[0033] S28, applying a Dropout layer to the weighted fusion feature map and outputting a feature map with key defect regions.

[0034] Preferably, in step S26, the calculation method of the attention weight of the attention map is:

[0035] The weight of the attention map is calculated using a Softmax function, and the weight of each spatial position of the attention map is between [0, 1].

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

[0037] Preferably, the attention feature map is unfolded by a rearrange operation, and the down-sampling feature map is unfolded by a rearrange operation.

[0038] Preferably, in step S24, the wafer image is convoluted by a depth separable convolution self.ds_conv.

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

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

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

[0042] S33, using 1x1 convolution to fuse feature information of multiple scales to form a global feature map.

[0043] Preferably, in step S31, the preliminary features are extracted from the feature map by 3x3 convolution.

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

[0045] S41, performing down-sampling processing on the global feature map by 1x1 convolution to reduce the size of the global feature map;

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

[0047] S43, extracting feature information around the local image features, and extracting context information based on the local image features and the feature information around the local image features by dilated convolution.

[0048] Advantages

[0049] The present application has the advantages that:

[0050] 1. The DPLAWDS module is designed to dynamically adjust the size of the receptive field through adaptive weighted downsampling operation, so that the model can adjust the feature extraction weight in the region according to the distribution and importance of defects. This innovation not only expands the receptive field, but also focuses better on the defect area, 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 the 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 improve the model's detection ability for large-scale defects, ensuring the model's detection performance at multiple scales, avoiding the defect blind area in traditional methods.

[0053] 2. The BNDWR_DRB module designed in the present application extracts image features at different scales through dilation convolution and reparameterization technology, and fuses features at multiple scales. Dilation convolution can significantly expand the receptive field and enhance the recognition ability of defects at different scales. Feature fusion operation ensures the effective integration of multi-scale information, improving the detection accuracy of defects of different sizes.

[0054] In terms of improving multi-scale detection capability: through multi-scale feature extraction and fusion, the present application can capture both detailed features and large-scale defect information in the image, significantly improving the recognition ability of 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 positive rate and improving the reliability in practical applications.

[0056] 3. The LRContextGuidedBlock_Down module designed in the present application combines local convolution and dilation convolution to extract local details and global context information of the image, ensuring that the model can not only process local features, but also use global information to improve the accuracy of defect detection. The module refines features guided by context, so that the model can eliminate background noise while enhancing the expression ability of tiny defects.

[0057] In terms of enhancing fine-grained detection accuracy: through the context guidance of combining local features and global information, the present application can improve the recognition accuracy of tiny defects based on fine-grained feature extraction, especially suitable for the detection of small and complex defects.

[0058] In reducing false positives: By eliminating background noise and enhancing the expression of defect areas, the model can more accurately distinguish between defects and backgrounds, reducing false positives and improving the reliability of the detection system.

[0059] 4. Overall system performance improvement

[0060] Although the present invention enhances the complexity of the model by adding extended receptive field, local weighting and context guidance, etc. Through optimizing the model structure and integrating multi-level features, the present invention maintains high computational efficiency while ensuring high precision. For example, in batch processing, the model can process large-scale images with low latency, ensuring the feasibility of real-time detection. The advantages are:

[0061] Efficiency: Even with increased model complexity, the optimized algorithm can still maintain low latency and high FPS, ensuring real-time processing capability and suitability for automated detection in large-scale production environments.

[0062] Economy: By improving detection accuracy and reducing false positives, the present invention reduces the need for subsequent manual intervention, saving labor costs and improving production efficiency.

[0063] Accuracy: Among multiple models, the accuracy gradually improves from 0.982 to 0.987 with the enhancement of modules, proving the significant effect of adding additional modules on improving model performance.

[0064] Model size: Although the model size has increased, the increase in parameter quantity has also brought better accuracy. In particular, the large model with the addition of the LRContextGuidedBlock_Down module, despite the increase in size and computational load, the model accuracy continues to improve.

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

[0066] Adaptability: Training an artificial intelligence model using this algorithm does not receive restrictions from the training platform, and different hardware devices of the training platform have little impact on the training results of the model. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The wafer defect classification algorithm flowchart of the present invention;

[0068] Figure 2 The DPLAWDS module execution flowchart of the present invention;

[0069] Figure 3A BNDWR_DRB module flowchart of the present application;

[0070] Figure 4 An LRContextGuidedBlock_Down module execution flowchart of the present application. DETAILED DESCRIPTION

[0071] The present application is further described below in conjunction with embodiments, but does not constitute any limitation to the present application, and any limited number of modifications made by anyone within the scope of the claims of the present application is still within the scope of the claims of the present application.

[0072] A wafer defect classification algorithm fusing an extended receptive field and local weighting of the present application, which mainly fuses the technologies of extended receptive field, local weighting and context guidance, can improve the accuracy of wafer defect detection, especially in the detection of small targets and micro defects, and reduce the recognition errors caused by insufficient receptive field and incomplete local feature extraction. To realize the algorithm, it is necessary to first construct a DPLAWDS module, a BNDWR_DRB module and an LRContextGuidedBlock_Down module.

[0073] The DPLAWDS module extracts features from the wafer image through adaptive weighted down-sampling operation to obtain a feature map with key defect regions. The adaptive weighted down-sampling operation of the DPLAWDS module can enhance the sensitivity of the network to the key defect regions in the feature map. Its core principle is to focus on important defect regions in the image by dynamically adjusting the receptive field of each region. And the DPLAWDS module can adaptively identify micro defects in the image and improve the expression ability of fine-grained features by weighted sampling of different regions.

[0074] The BNDWR_DRB module extracts and fuses multi-scale features from the feature map through dilated convolution operation and reparameterization operation to form a global feature map, thereby expanding the receptive field range and improving the multi-scale feature extraction capability, which can capture defect information in the key defect region, especially suitable for complex background and multi-scale defect detection. Its main function is to extract image features at different scales through dilated convolution operation, and optimize the parameters of the convolution layer through reparameterization means, thereby improving the recognition accuracy of the model for defects of different sizes.

[0075] The LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through context information guided operation, so as to further improve the detection capability of micro defects. The core idea is to guide the learning of local features through context information, so that the model can capture local details while obtaining global feature support, thereby improving the recognition accuracy of micro and complex defects.

[0076] Referring to Figure 1 Based on the above module, the specific flow steps of the algorithm of the present application are as follows:

[0077] S1, input the wafer image to be detected. The image size is 52x52 pixels, and the gray image format.

[0078] S2, the DPLAWDS module extracts features from the wafer image through adaptive weighted down-sampling operation, focusing on extracting key defect areas in the wafer image.

[0079] S3, the BNDWR_DRB module performs multi-scale feature extraction and fusion on the feature map through dilated convolution operation and reparameterization operation to form a global feature map, so that the model can capture large-scale defect information in the image. Since most existing deep learning methods rely on traditional convolution operation, the receptive field is limited and cannot fully cover the global information of the image, therefore, the present application adopts dilated convolution operation and reparameterization operation technology, which can effectively expand the receptive field and improve the capture capability of global information, thereby improving the accuracy of defect detection.

[0080] S4, the LRContextGuidedBlock_Down module optimizes the defect information in the global feature map through local weighting and context information guided operation. The main purpose is to optimize the expression of micro defects in the global feature map, and ensure the effective combination of local details and global information. Since the traditional deep learning method often ignores the guidance of context information when classifying defects, in order to solve this problem, the present application introduces a context guidance mechanism, which integrates local and global information to improve the recognition ability of the module to complex background and diversified defects. In addition, the present application adopts the strategy of local weighting, which can more accurately focus on the local area in the image and improve the recognition ability of fine-grained defects, especially in the detection of micro defects.

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

[0082] S6, judging whether the wafer image exists defects according to the classification result, outputting a defect category if the wafer image exists defects, and returning a no-defect state if the wafer image does not exist defects.

[0083] The application proposes a high-precision wafer defect detection and classification algorithm fusing extended receptive field, local weighting and context guidance, aiming to overcome the problems in the prior art and improve the precision and efficiency of wafer defect detection. At the same time, the high-precision detection and classification in the field of wafer image dataset research for small targets are effectively made up, and by effectively combining local and global information, improving the receptive field and optimizing the model structure, the application can efficiently and accurately detect various types of wafer defects, especially micro and fine-grained defects, and significantly improve the performance of the detection system.

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

[0085] As shown in the DPLAWDS module, first introduce the core function part: Figure 2

[0086] Softmax activation function: the Softmax activation function is used to calculate the weight of each spatial position in the subsequent calculation, so as to weight the feature map.

[0087] Attention mechanism: an average pooling (AvgPool2d) and a convolution operation (Conv(ch, ch, k = 1)) are used to generate an attention feature map. In general, the average pooling operation extracts spatial information by averaging over a 3x3 window, while the convolution operation is used for feature conversion to facilitate the next step of weighted fusion.

[0088] Down-sampling convolution (ds_conv): the down-sampling convolution uses a 3x3 convolution kernel with a step size of 2, which reduces the resolution of the input feature map by half, and controls the receptive field and learned features of each convolution kernel by using group convolution (g = (ch / / group)). This operation will increase the number of channels of the feature map to 4 times the original.

[0089] Dropout layer (dropout): the dropout layer is used for network regularization, which prevents overfitting by randomly discarding a certain proportion of features. Here, dropout_rate is a hyperparameter that controls the dropout rate.

[0090] These steps together constitute the core function of the DPLAWDS module, i.e. the importance preservation of adaptive weight down-sampling, which can effectively extract key features in the image.

[0091] ​Based on the above concept, in step S2, the processing step of the wafer image by the DPLAWDS module is specifically as follows:

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

[0093] S22, using a 1x1 convolution to process the aggregated local feature information to extract deep features in the wafer image for representing key defect regions, and forming an attention feature map.

[0094] Specifically, the wafer image x has a size of (bs, ch, h, w). First, it is processed by the attention network. The attention network is composed of average pooling and 1x1 convolution, and generates an attention feature map with a size of (bs, ch, h, w).

[0095] S23, the attention feature map is unfolded into a four-channel attention map by the rearrange operation, i.e. with a shape of (bs, ch, h, w, 4). This step realizes the spatial rearrangement of features.

[0096] Then, the weight of the attention map is calculated using the Softmax function, and the weight of each spatial position of the attention map is between 0 and 1, and the sum of the weights of all spatial positions of the attention map is 1, so that the weighting can be effectively performed.

[0097] S24, the wafer image is processed by a depth separable convolution self.ds_conv, and its resolution is reduced by half to obtain a down-sampled feature map. The output map has a size of (bs, 4*ch, h / 2, w / 2).

[0098] S25, the down-sampled feature map is also unfolded into a four-channel down-sampled map by the rearrange operation, with a size of (bs, ch, h, w, 4).

[0099] S26, the down-sampled map is aligned with the attention map, and the attention weight of the attention map is mapped onto the down-sampled map by element-wise multiplication.

[0100] S27, summing the last dimension of the down-sampled map output by step S26 to obtain a weighted fusion feature map;

[0101] S28, applying a Dropout layer to the weighted fusion feature map, and outputting a feature map with key defect regions, which contains more representative, multi-scale information and is subjected to spatial weighting processing, enhancing the attention of important features. This step helps to improve the generalization ability of the model and reduce the risk of overfitting.

[0102] As Figure 3 shown in step S3, the BNDWR_DRB module performs the following steps for multi-scale feature extraction and fusion on the feature map:

[0103] S31, extract preliminary features from the feature map with the key defect area through 3x3 convolution.

[0104] S32, perform dilation convolution operation and reparameterization operation on the feature map with the extracted preliminary features to extract feature information of different scales. The purpose of dilation convolution is to expand the receptive field of the convolution kernel, which can extract large-scale context information without increasing the amount of calculation. Therefore, dilation convolution can not only extract detailed information, but also obtain large-scale context information.

[0105] S33, use 1x1 convolution to fuse feature information of multiple scales to form a global feature map. The 1x1 convolution here plays a role in channel fusion, which can fuse feature maps from different scales while compressing or expanding the number of channels for subsequent processing. The global feature map contains information from details to large-scale, which can effectively deal with diversified defects and improve the recognition ability of the model for multiple defects.

[0106] Fusing feature maps of different scales through 1x1 convolution is a common method in existing technology. This approach is mainly used to fuse feature information of different channels, and then compress or adjust the number of channels, and merge multi-scale features through weight learning of the convolution kernel. Specifically, the role of 1x1 convolution has the following points:

[0107] Channel fusion: the core role of 1x1 convolution is to fuse different scale feature maps in the channel, compress or expand the channel information of multiple different scale feature maps, so as to obtain a more compact and information-rich feature map. It is actually a pixel-by-pixel linear transformation that performs information fusion in the channel dimension.

[0108] Reducing the amount of calculation: using 1x1 convolution can reduce the complexity of convolution calculation, because it only focuses on channel information and does not involve convolution operation in spatial dimension.

[0109] Information integration: 1x1 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 detailed features and global context features.

[0110] As Figure 4 shown in step S4, the LRContextGuidedBlock_Down performs the following steps to optimize the defect information in the global feature map:

[0111] S41, performing down-sampling processing on the global feature map through 1x1 convolution to reduce the size of the global feature map.

[0112] S42, extracting local image features of the global feature map output by step S41 through a context module or attention mechanism. The purpose of context-guided feature refinement is to optimize local features, so that the network's recognition ability for defects is enhanced. Specifically, the specific technical content of the context-guided mechanism can 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 context information in the local area and ignore irrelevant background information. Common techniques include:

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

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

[0115] S43, extracting feature information around the local image features, and extracting context information based on the local image features and their surrounding feature information through dilated convolution, thereby refining the image. In this step, the convolutional network is used to further enhance and optimize the details, improving the recognition ability for small size defects.

[0116] Regarding the fully connected layer (Fully Connected Layer), its role is to convert input features through weighting and bias and generate outputs, which will become the basis for the final classification of the network. The specific implementation process of the fully connected layer in step S5 for classifying defects is as follows:

[0117] First fully connected layer: This layer inputs the flattened feature vector into a larger output space, usually with a ReLU activation function to add nonlinearity to enhance the network's expression ability.

[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 dimension, mapping information from high dimension to lower dimension to facilitate classification.

[0119] Last fully connected layer: This is the final output layer, which usually uses a Softmax activation function to map the output to a probability distribution of categories. 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, thus increasing the expressive power of the model. Commonly used activation functions include:

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

[0122] Softmax: usually used in the last layer, especially in multi-classification tasks, it will convert the output value of each class to a probability distribution, output a normalized probability value.

[0123] The Softmax activation function is used in this embodiment. For classification tasks, especially multi-class problems, the Softmax activation function will normalize the prediction value of each class and calculate the probability of each class. The class 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 classes is 1. For example, if the output of the model is [2.3, -0.7, 1.2], Softmax will convert them to [0.76, 0.01, 0.23], indicating that the probability of the first class is 0.76, the second is 0.01, and the third is 0.23. The final model will choose the first class as the prediction.

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

[0125] The above is only the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the structure of the present application, can make several variations and improvements, these will not affect the effect of the present application and the practicality of the patent.

Claims

1. A wafer defect classification algorithm fusing extended receptive field and local weighting, characterized in that, The method comprises the following steps: S1, inputting a wafer image to be detected; S2, constructing a DPLAWDS module, which extracts features of the wafer image through adaptive weighted down-sampling operation to obtain a feature map with key defect regions; S3, constructing a BNDWR_DRB module, which extracts and fuses multi-scale features of the feature map through dilated convolution operation and reparameterization operation to form a global feature map; S4, constructing a LRContextGuidedBlock_Down module, which optimizes defect information in the global feature map through local weighting and context information guided operation; S5, inputting the optimized global feature map into a classification network, and classifying defects in the wafer image through a full connection layer; S6, judging whether the wafer image has defects according to the classification result, and outputting a defect category if there are defects, or returning a defect-free state if there are no defects; The specific steps of the LRContextGuidedBlock_Down module for optimizing defect information in the global feature map are as follows: S41, down-sampling the global feature map through 1x1 convolution to reduce the size of the global feature map; S42, extracting local image features of the global feature map output by step S41 through a context module or attention mechanism; S43, extracting feature information around the local image features, and extracting context information based on the local image features and the surrounding feature information through dilated convolution.

2. The wafer defect classification algorithm of claim 1, wherein, The processing steps of the DPLAWDS module for the wafer image are as follows: S21, applying average pooling operation to aggregate local feature information of the wafer image; S22, using 1x1 convolution to process the aggregated local feature information to extract deep features in the wafer image for representing key defect regions, and forming 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 down-sampled feature map; S25, expanding the down-sampled feature map into a four-channel down-sampled map; S26, aligning the down-sampled map with the attention map, and mapping attention weights of the attention map to the down-sampled map through element-wise multiplication; S27, summing the last dimension of the down-sampled map output by step S26 to obtain a weighted fusion feature map; S28, applying a Dropout layer to the weighted fusion feature map, and outputting a feature map with key defect regions.

3. The wafer defect classification algorithm of claim 2, wherein, In step S26, the calculation method of the attention weights of the attention map is as follows: The weights of the attention map are calculated using a Softmax function, and the weights of each spatial position of the attention map are between 0 and 1.

4. The wafer defect classification algorithm of claim 3, wherein, The sum of the weights of all spatial positions of the attention map is 1.

5. The wafer defect classification algorithm of claim 2, wherein, The attention feature map is unfolded through a rearrange operation, and the down-sampling feature map is unfolded through a rearrange operation.

6. The wafer defect classification algorithm of claim 2, wherein, In step S24, the wafer image is processed through a deep separable convolution self.ds_conv.

7. The wafer defect classification algorithm of claim 1, wherein, The processing steps of the BNDWR_DRB module for multi-scale feature extraction and fusion of the feature map are as follows: S31, extracting preliminary features from the feature map with the key defect area; S32, performing dilated convolution operation and reparameterization operation on the feature map with the preliminary features extracted to extract feature information of different scales; S33, using 1*1 convolution to fuse the feature information of multiple scales to form a global feature map.

8. The wafer defect classification algorithm of claim 7, wherein, In step S31, preliminary features are extracted from the feature map through 3*3 convolution.

Images