Wafer defect multi-label detection method based on rotation invariant convolutional neural network

Abstract

The invention provides a wafer defect multi-label detection method based on a rotation-invariant convolutional neural network, and the method comprises the steps: receiving a wafer detection task and a plurality of wafer images corresponding to the wafer detection task, encoding the images into tensors, inputting the tensor corresponding to each wafer image into a trained rotation-invariant convolutional neural network, and obtaining a rotation-invariant convolutional neural network; the detection process of the trained rotation invariant convolutional neural network comprises the following steps: performing feature extraction on each wafer image to obtain an initial feature; and then rotation convolution, Gaussian smoothing filtering, direction fusion, global average pooling and classification are sequentially carried out on the initial features in each channel to obtain a detection result about whether a wafer in each wafer image has defects or not and a detection result about defect types. According to the method, rotation invariant convolution is introduced, so that the stability and robustness of the network under the geometric transformation condition are improved, independent and parallel judgment can be carried out on various defect types possibly existing in a single wafer graph at the same time, and accurate recognition in a multi-defect coexistence scene is realized.

Description

Technical Field

[0001] This application belongs to the field of industrial defect detection technology, specifically relating to a multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks. Background Technology

[0002] As integrated circuit manufacturing processes become increasingly miniaturized, the morphologies of defects generated during wafer fabrication are becoming more complex, subtle, and diverse. Wafer defect detection, a crucial step in ensuring chip yield and reliability, directly impacts the quality control level of the semiconductor manufacturing process due to its accuracy and robustness. Traditional inspection systems often employ rule-based image processing algorithms or single-label classification models, which struggle to meet the identification needs of multiple coexisting defects (such as scratches, rings, and edge cracks) in complex production scenarios. Furthermore, wafer defects can appear at arbitrary locations and orientations in actual images, exhibiting a degree of rotational and translational uncertainty.

[0003] Currently, in the field of wafer defect detection, researchers generally use deep learning-based image classification models to identify defects in wafer images. Common methods are mostly based on 2D convolutional neural networks (CNNs), such as ResNet, DenseNet, and EfficientNet, combined with attention mechanisms, multi-scale extraction modules, or lightweight structures to improve recognition accuracy and deployment efficiency. The patent "A Wafer Defect Detection Method and System Based on a Lightweight Multi-Scale Wafer Defect Detection Model" (Chi Ronghua, Ren Jie, Wu Teng, et al. A Wafer Defect Detection Method and System Based on a Lightweight Multi-Scale Wafer Defect Detection Model: CN202411886913.8[P].CN119810063A[2025-11-12].) proposes a lightweight multi-scale CNN model for wafer defect detection, and integrates an SE attention module to adaptively adjust the weights of each feature channel, enhancing the model's response to defects of different sizes and distributions. The patent "A Lightweight Wafer Defect Detection Method" invented by Wang Quan et al. (Wang Quan, Wang Mengnan, Sun Jiadong. A Lightweight Wafer Defect Detection Method: CN202411896705.6[P]. CN119359702A[2025-11-12].) improves the YOLOv8 network structure to adapt to the deployment requirements of edge devices. This method introduces a C2f-IDR feature extraction module and an HSPAN bidirectional feature fusion module, combined with the Focaler-Shape IoU loss function, effectively improving the ability to identify small and complex defects.

[0004] In scenarios with small samples or incomplete labels, some works employ generative models and knowledge distillation to enhance model robustness. The patent "Robust Wafer Defect Detection Method Based on Distillation Learning" by Ding Guiguang et al. (Ding Guiguang, Chen Kai. Robust Wafer Defect Detection Method and System Based on Distillation Learning: 202510269636 [P] [2025-11-12]) proposes a teacher-student network structure. By introducing a distillation mechanism during training, the student model learns stable and universal feature representations from the teacher model, effectively mitigating the impact of noisy labels on the accuracy of wafer defect identification and improving the model's robustness and generalization ability under complex data distributions. The patent "A Wafer Defect Detection Method Based on Dual Siamese Networks" by Ming Ming et al. (Ming Ming, Gu Gang. A Wafer Defect Detection Method Based on Dual Siamese Networks: 202510504150 [P] [2025-11-12]) proposes a wafer defect detection framework combining generative adversarial networks (GANs) and a dual siamese network structure. By generating pseudo-defect enhancement training samples using GANs, and extracting discriminative and reconstructed features using two Siamese networks, defect localization and visualization are finally achieved based on feature differences, improving the detection capability for minute defects. The paper "Multi-scale guidance diffusion network for wafer map defect recognition" (Long Z, Yan J, Piao M. Multi-scale guidance diffusion network for wafer map defect recognition[J]. Expert Systems with Applications, 2025, 267(000).DOI:10.1016 / j.eswa.2024.126134.) proposes a wafer defect recognition network WMDiff based on a denoising diffusion probability model (DDPM). It employs a multi-scale information guidance strategy (MSG) to regulate the diffusion denoising process, gradually guiding the model to remove image noise, and combines a multi-scale network structure to extract defect features at different scales, thereby improving the classification accuracy of complex defect patterns. These methods improve the model's recognition performance for some defect types to varying degrees.

[0005] Traditional convolutional neural networks (CNNs) are designed to be equivariant only to image translation, lacking the ability to model equivariant transformations such as rotation and mirroring. This makes them ill-suited to effectively adapting to equivalent representations of wafer defects under different orientations or angles. Consequently, when faced with images of the same type of defect but with different orientations, the model's output fluctuates, exhibiting insufficient robustness and severely impacting detection accuracy. Furthermore, current multi-label detection methods often treat different defect types as mutually exclusive categories or employ simple parallel multi-label modeling, ignoring potential co-occurrence relationships and semantic coupling between defect types. In actual production, a single wafer image often contains multiple defect types simultaneously, and these combinations may not be fully covered in the training set. If the model lacks the ability to model these combination patterns, weak labels are easily difficult to identify accurately, leading to missed detections or identification imbalances. Some emerging methods attempt to introduce knowledge distillation or generative adversarial networks to improve learning capabilities under conditions of few samples and weak labels. However, these methods have high requirements for the stability of the teacher model, the quality of pseudo-samples, and the sensitivity to training parameter tuning. In practical deployments, they still suffer from problems such as training instability and limited generalization ability, making it difficult to apply them on a large scale in industrial inspection scenarios. More importantly, existing model improvements are mostly focused on local strategies such as network lightweighting, attention mechanisms, or single-scale optimization. They lack a general framework that systematically integrates rotational and other variable modeling and multi-label imbalanced learning mechanisms at the structural level, making it difficult to fully meet the multiple requirements of complex wafer defect detection tasks for directional robustness, mixed defect perception capabilities, and generalization performance. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides a multi-label detection method for wafer defects based on a rotation-invariant convolutional neural network. The technical problem to be solved by this application is achieved through the following technical solution: In a first aspect, this application provides a multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks, including: S100, receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; S200, the multiple wafer images are encoded to obtain a tensor that retains the semantic information of the pixels; S300, the tensor corresponding to each wafer image is input into a trained rotation-invariant convolutional neural network to perform defect detection on each input wafer image, obtaining the detection results of whether defects exist in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, feature extraction is performed to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain the detection results of whether there are defects in the wafer and the type of defects in each wafer image.

[0007] Secondly, this application provides a wafer defect multi-label detection device based on rotation-invariant convolutional neural networks, comprising: The receiving module is configured to receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; The encoding module is configured to encode the multiple wafer images to obtain a tensor that retains the semantic information of each pixel; The detection module is configured to input the tensor corresponding to each wafer image into a trained rotation-invariant convolutional neural network, so that it performs defect detection on each input wafer image and obtains the detection results of whether there are defects in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, feature extraction is performed to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain the detection results of whether there are defects in the wafer and the type of defects in each wafer image.

[0008] Beneficial effects: (1) This application can achieve accurate identification in scenarios where multiple defects coexist, significantly improving the applicability of the model. Traditional wafer defect detection methods generally adopt a single-label multi-classification framework, which assumes that each wafer image contains only one type of defect, and therefore cannot handle the composite defects commonly encountered in actual production. This application uses a multi-label learning mechanism to model eight typical defects separately and output independent probabilities, enabling the model to identify multiple defect categories simultaneously, thereby significantly improving the accuracy and reliability of the detection system under complex working conditions.

[0009] (2) This application improves the stability and robustness of rotation-invariant convolutional neural networks under geometric transformation conditions by introducing rotation-invariant convolutional layers. This application uses rotation-invariant convolution (F-Conv) to construct the defect feature extraction network, enabling the model to maintain consistent feature representation under rotation, flipping, and orientation changes, overcoming the shortcomings of traditional convolutional neural networks that are sensitive to orientation changes and have unstable recognition results. This advantage is particularly prominent in the recognition of rotationally symmetric defects such as Donut and Edge-Ring, as well as orientation-sensitive defects such as Scratch, ensuring the model's universality and detection stability in actual manufacturing environments.

[0010] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a multi-label wafer defect detection method based on a rotation-invariant convolutional neural network provided in this application. Figure 2 This is a schematic diagram of the model training process provided in this application; Figure 3 These are eight multi-defect wafer images provided in this application, along with a schematic diagram of the actual defect categories; Figure 4 This is a schematic diagram of the rotation-invariant convolutional neural network provided in this application. Detailed Implementation

[0012] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0013] This application proposes a multi-label wafer defect detection method based on rotation-invariant convolutional neural networks (CNNs). Addressing the practical challenges of diverse defect types, random orientations, and frequent co-occurrences in wafer images, this method designs a CNN structure with rotational equivariance, anti-aliasing properties, and multi-label processing capabilities. By introducing group equivariant convolution (F-Conv), an anti-aliasing downsampling module, and global average pooling, a model representation robust to defect image orientation and location with strong generalization ability is constructed. Simultaneously, a multi-label imbalance optimization strategy effectively improves the recognition capability of weak labels and the perception accuracy of mixed defect combinations. This method features a novel structure and stable performance, making it particularly suitable for wafer defect detection tasks with significant rotational perturbations and the coexistence of multiple defect types, significantly improving detection accuracy and practicality for industrial deployment.

[0014] like Figure 1 As shown, this application provides a multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks, including: S100, receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; S200, the multiple wafer images are encoded to obtain a tensor that retains the semantic information of the pixels; S300, the tensor corresponding to each wafer image is input into a trained rotation-invariant convolutional neural network to perform defect detection on each input wafer image, obtaining the detection results of whether defects exist in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, initial features are extracted. Then, these initial features are sequentially processed through rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain detection results for the presence and type of defects in each wafer image. The rotation-invariant convolutional neural network used is WFConvNet.

[0015] As one specific implementation of this application, reference is made to... Figure 2As shown, the training process of the rotation-invariant convolutional neural network includes: a) Load the npz file containing the original wafer data; wherein the npz file includes wafer image tensor data and a multi-label Boolean matrix; wherein the Boolean matrix records the actual defect category labels corresponding to the wafer image; The original wafer data npz file contains two parts: Arr_0 and Arr_1. Arr_0 is the wafer image tensor data, and Arr_1 is a multi-label Boolean matrix.

[0016] The shape of the wafer image corresponding to the wafer image tensor data in the npz file containing the original wafer data in this application is ( N Arr_1 is a multi-label Boolean matrix with the following shape: (52, 52), integer values ​​in the range {0, 1, 2, 3}, representing the pixel category (background, good, defective, etc.) at different locations on the wafer. N ,8), N This represents the number of wafer images in the dataset, indicating the temporal relationship between the eight actual defect categories. Figure 3 Eight images of multi-defect wafers and their corresponding categories are presented.

[0017] b. Perform one-hot encoding on the wafer image represented by the wafer image tensor data to obtain data samples that retain the category semantic information of each pixel, and randomly divide all data samples into training set, validation set and test set; One-hot encoding is performed on each (52, 52) wafer image, converting it into a (4, 52, 52) floating-point tensor, preserving the category semantic information of each pixel, in the form:

[0018] In the formula, Represents the pixel position in the wafer image The pixel value at that location.

[0019] The data samples were divided into a training set (80%), a validation set (10%), and a test set (10%) using a fixed random seed. The training set labels were then... And calculate the weighting factor for each category in the training set. The formula for calculating the weighting factor is expressed as:

[0020] In the formula, The total number of data samples in the training set. Indicates the first The number of data samples that appear for each real defect category label. This represents the total number of true defect category labels.

[0021] c. In the current iteration, a batch of data samples are selected from the training set and input into a preset rotation-invariant convolutional neural network for iterative training to obtain the predicted defect category label for the current iteration. During actual training, this application constructs a DataLoader suitable for the PyTorch framework, supporting the following three settings: using or not using WeightedRandomSampler, optionally enabling data augmentation of the training set, and supporting multi-threaded loading and CUDA memory acceleration. A batch of data samples can specifically take 256 samples.

[0022] The AdamW optimizer and CosineAnnealingLR learning rate scheduling strategy are used to stabilize training. During the training phase, a mixed precision and gradient pruning mechanism is enabled to improve convergence speed and training stability. To improve recall performance for each type of defect and avoid performance bottlenecks caused by fixed thresholds, a validation set-driven threshold adaptation mechanism is adopted: a grid search (0.1 to 0.9) is performed on the validation set at regular intervals to obtain the optimal decision threshold for each label.

[0023] This application employs the following data augmentation strategy for semantically faithful one-hot images of wafers: augmentation methods that do not disrupt the semantic label distribution, including 90° rotation, horizontal flipping, and center cropping; all augmentation operations use nearest-neighbor interpolation to ensure semantic consistency across one-hot channels; this augmentation strategy is applied only to the training set. Optimal hyperparameter settings are selected by monitoring the F1 score on the validation set during training, and the optimal model checkpoint is saved for testing.

[0024] d. Calculate the loss function based on the predicted defect category label, the actual defect category label, and the weighting factor of the actual defect category label in the current iteration; This application addresses the problems of class imbalance, complex label co-occurrence, and difficulty in identifying minute defects in multi-label defect detection of wafer images. It designs a targeted training strategy and loss function system during the training process, employing binary cross-entropy loss with logits as the main loss function for multi-label classification scenarios. Let the sample label be... The model output is a logits vector. The loss function is a weighted binary cross-entropy loss, expressed as:

[0025] In the formula, The sigmoid activation function is used. Labels for real defects j The weighting factor, For data samples Belongs to the Predicted defect category labels for different defect categories. For data samples Belongs to the The actual defect category label for the defect category.

[0026] This application significantly improves the learning power of niche categories by pre-calculating the weighting factor of each category on the training set and dynamically inputting it, effectively alleviating the strong label dominance effect in the early stage of training.

[0027] e. Enable the mixed precision and gradient pruning strategy, and adjust the weights of the rotation-invariant convolutional neural network in the direction of loss function descent; f, repeat c to e until the total number of training iterations is reached or the loss function no longer decreases after multiple consecutive iterations, to obtain a trained rotation-invariant convolutional neural network.

[0028] In actual training, this application trained the model on the GPU version of PyTorch 2.7.1 framework, using an NVIDIA 3090 GPU. Training employed the AdamW optimizer with an initial learning rate of 0.001, coupled with the CosineAnnealing learning rate scheduler and a weight decay of 0.0001. The batch size was set to 256, and the total training epochs were 300. A model checkpoint was saved every 20 epochs, and training stopped when the validation set loss did not decrease for 30 consecutive epochs. To prevent overfitting and gradient explosion, gradient clipping (maximum norm of 1.0) and mixed precision training (AMP) were introduced.

[0029] As one specific implementation of this application, reference is made to... Figure 4 The rotation-invariant convolutional neural network includes an input layer, a backbone network, a feature compression module, and a classification head connected in sequence. The input layer is used to input the tensor corresponding to the wafer image. The backbone network is used to extract features from the tensor corresponding to the input wafer image to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution and Gaussian smoothing filtering in each channel to obtain feature maps that retain rotational equivariance and spatial consistency at the scale. The feature compression module is used to fuse the feature maps in the channel dimension and then perform average pooling in the orientation dimension to obtain rotation- or mirror-invariant representations. The classification head is used to perform convolution, global average pooling, and classification on the representations to obtain detection results of whether there are defects in the wafer in the wafer image and the type of defects. The backbone network comprises a first PCA-based turnable convolutional layer, two stacked first residual blocks, a first anti-aliasing average pooling layer, and two stacked second residual blocks and a second anti-aliasing average pooling layer, connected in sequence. The feature compression module comprises a decomposed convolutional layer and a grouped average pooling layer, connected in sequence. The classification head comprises a regular convolutional layer, a global average pooling layer, and a fully connected layer, connected in sequence. Both the first and second residual blocks comprise a second PCA-based turnable convolutional layer connected in sequence, with the second turnable convolutional layer introducing the residual structure using a skip connection. Batch normalization and ReLU activation functions are embedded after both the first and second turnable convolutional layers.

[0030] refer to Figure 4 The first steerable convolutional layer uses a 5×5 convolutional kernel for initial feature extraction, while retaining M directional channels. The second steerable convolutional layer uses a 3×3 convolutional kernel to perform rotational convolution on the initial features in each channel. The first anti-aliasing average pooling layer and the second anti-aliasing average pooling layer use Gaussian filtering to smooth each channel, obtaining a feature map that retains rotational equivariance and spatial consistency. The decomposed convolutional layer uses a 1×1 convolutional kernel to fuse the feature maps in the channel dimension, and the grouped average pooling layer averages and converges the fused feature maps in the directional dimension to obtain a rotation- or mirror-invariant representation. The regular convolutional layer uses a 3×3 convolutional kernel to convolve the representation, and the global average pooling layer performs global average pooling on the output of the regular convolutional layer to obtain a 32-dimensional global feature. The fully connected layer obtains an M-dimensional defect label prediction vector through linear mapping. The defect label prediction vector represents the detection result of whether there are defects in the wafer in the input wafer image and the type of defect.

[0031] Figure 4 This is the WFConvNet model, a rotation-invariant convolutional neural network used for wafer defect detection. The backbone network employs PCA-based steerable convolutional layers. Figure 4 In this context, a convolutional layer denoted as Fconv_PCA is essentially a set of steerable convolutional kernels built upon rotation groups (such as C8 or D4). Given a two-dimensional convolutional kernel... Its transformations under translation and rotation groups satisfy isovariability:

[0032] in, These represent the horizontal and vertical coordinate indices of the convolution kernel, respectively. , k This indicates the size of the convolution kernel. This represents the input to the convolution operation, i.e., the input image or feature map. Represents group elements Image transformation under the action, This corresponds to the rotated convolution kernel. The implementation of Fconv_PCA involves constructing the basic rotated kernel using principal component interpolation. This generates convolutional responses with consistent orientation:

[0033] in For direction index, K Input the number of channels. For direction t Rotating nucleus, This represents the spatial offset of the convolution kernel, i.e., at the center position. The location of surrounding pixels / feature points, This represents the input feature of the k-th channel.

[0034] The network structure includes an initial 5×5 Fconv_PCA convolutional layer to extract low-level features, followed by two stacked residual blocks (ResBlocks). Each residual block contains two 3×3 Fconv_PCA convolutions, an F_BN layer, and a ReLU activation function layer. To reduce aliasing, an anti-aliasing average pooling layer is added. Figure 4 The term is represented as AntialiasedAvgPool, which applies a Gaussian filter to each channel. Achieve smoothing:

[0035] In the formula, This represents the standard deviation of the Gaussian filter kernel, used to control the rate at which the weights decay with distance, thereby adjusting the smoothing intensity.

[0036] Then, directional fusion is achieved using a 1×1 Fconv_1x1 convolution, followed by a grouped average pooling layer. Figure 4 This is represented as GroupMeanPooling, where the grouped average pooling layer averages along the directional dimension t.

[0037] In the formula, Indicates after the first After a group transformation (such as rotation), in the channel Location The characteristic response value at that location.

[0038] The final output feature map is processed by a 3×3 regular convolutional layer, ReLU, and a global average pooling layer to obtain a global vector. Mapped to through a fully connected layer Output the defect probability vector.

[0039] After training, this application can use a test set to test the trained model, and the sample output... The optimal threshold was obtained by grid search on the validation set. The label was determined, and the result was:

[0040] To comprehensively evaluate the model's recognition performance on each defect category, metrics such as Precision, Recall, F1-score, ROC-AUC, and mAP were used, and the meanings of the symbols are shown in Table 1.

[0041] Table 1

[0042] The formulas for metrics such as Precision, Recall, F1-score, ROC-AUC, and mAP are as follows:

[0043] In the above formula, express The first curve n One Recall value, Indicates the corresponding first n Each Precision value is calculated from a different confidence threshold.

[0044] This application systematically tested the performance of the proposed rotation-invariant convolutional network model WFConvNet in the wafer multi-label defect detection task, and obtained the classification index results shown in Table 2.

[0045] Table 2

[0046] As shown in Tables 1 and 2, the rotation-invariant convolutional network model of this application achieved extremely high recognition accuracy for eight typical wafer defect patterns. Among them, the defect categories with regular structures and obvious spatial symmetry, such as Center, Donut, Edge-Ring, and Scratch, all had Precision, Recall, and F1-score close to or reached 1.0, and the corresponding ROC-AUC and mAP were also close to 1.0. This indicates that the model can stably recognize these patterns under different rotation angles and spatial scales, demonstrating the significant advantages of rotation-invariant convolutional structures in handling ring-shaped and linear defects.

[0047] For defect categories such as Edge-Loc, Loc, and Random, which are characterized by strong locality, large morphological differences, and high recognition difficulty, the model still achieved high Precision (up to 0.9966), Recall (up to 0.9922), and F1-score (all greater than 0.98), with ROC-AUC exceeding 0.99 and mAP exceeding 0.995. This indicates that the model possesses strong feature generalization and local sensitivity capabilities, accurately capturing anomalous patterns even in noisy and blurred defect boundaries. Near-full defects, which are large-area global anomalies, achieved a Recall of 1.0, a Precision of 0.9375, and an F1-score of 0.9677, while both ROC-AUC and mAP reached the maximum score of 1.0. This demonstrates that the model maintains extremely high stability and robustness in handling global anomalous patterns, exhibits high sensitivity to large-scale changes in brightness, and almost eliminates false negatives.

[0048] Overall, the model achieved F1 scores exceeding 0.96 across all categories, with five defect classes achieving F1 scores exceeding 0.98, and all metrics (Precision, Recall, F1, AUC, mAP) for four defect classes approaching 1.0. These results demonstrate that the proposed rotation-invariant convolution-based multi-label defect detection model can simultaneously achieve global pattern recognition, local detail extraction, and robustness to geometric transformations such as rotation and flipping, exhibiting significant technical effectiveness and practical application value in complex wafer defect scenarios. Therefore, the model structure, training method, and loss design in this embodiment effectively improve the accuracy, stability, and generalization ability of wafer defect detection, significantly outperforming traditional convolutional network methods and providing a high-precision intelligent discrimination solution for industrial wafer quality inspection.

[0049] As a specific implementation of this application, S300 includes: The tensor corresponding to each wafer image is input into the input layer of a trained rotation-invariant convolutional neural network. The input layer then feeds this tensor into the backbone network, which extracts features from the tensor corresponding to the input wafer image to obtain initial features. These initial features are then subjected to rotational convolution and Gaussian smoothing filtering in each channel to obtain feature maps that preserve rotational equivariance and spatial consistency across scales. The feature maps are then fused along the channel dimension using a feature compression module and averaged along the orientation dimension to obtain rotation- or mirror-invariant representations. Finally, the classification head performs convolution, global average pooling, and classification on these representations to obtain detection results for the presence and type of defects in the wafer image.

[0050] Secondly, this application provides a wafer defect multi-label detection device based on rotation-invariant convolutional neural networks, comprising: The receiving module is configured to receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; The encoding module is configured to encode the multiple wafer images to obtain a tensor that retains the semantic information of each pixel; The detection module is configured to input the tensor corresponding to each wafer image into a trained rotation-invariant convolutional neural network, so that it performs defect detection on each input wafer image and obtains the detection results of whether there are defects in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, feature extraction is performed to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain the detection results of whether there are defects in the wafer and the type of defects in each wafer image.

[0051] This application provides a method and apparatus for multi-label wafer defect detection based on rotation-invariant convolutional neural networks (RONs). The method includes: receiving a wafer inspection task and multiple wafer images corresponding to the task; encoding the images into tensors; and inputting the tensors corresponding to each wafer image into a trained RON. The detection process of the trained RON involves: extracting initial features from each wafer image; and then sequentially performing rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification on each channel to obtain detection results for the presence and type of defects in each wafer image. By introducing rotation-invariant convolution, this application improves the stability and robustness of the RON under geometric transformation conditions. It enables independent and parallel determination of multiple defect categories that may coexist in a single wafer image, effectively overcoming the technical limitation of traditional methods that only support single-label multi-classification. This achieves accurate identification in scenarios with multiple coexisting defects and offers greater adaptability.

[0052] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0053] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks, characterized in that, include: S100, receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; S200, the multiple wafer images are encoded to obtain a tensor that retains the semantic information of the pixels; S300, the tensor corresponding to each wafer image is input into a trained rotation-invariant convolutional neural network to perform defect detection on each input wafer image, obtaining the detection results of whether defects exist in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, feature extraction is performed to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain the detection results of whether there are defects in the wafer and the type of defects in each wafer image.

2. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 1, characterized in that, The training process of the rotation-invariant convolutional neural network includes: a) Load the npz file containing the original wafer data; wherein the npz file includes wafer image tensor data and a multi-label Boolean matrix; wherein the Boolean matrix records the actual defect category labels corresponding to the wafer image; b. Perform one-hot encoding on the wafer image represented by the wafer image tensor data to obtain data samples that retain the category semantic information of each pixel, and randomly divide all data samples into training set, validation set and test set; c. In the current iteration, a batch of data samples are selected from the training set and input into a preset rotation-invariant convolutional neural network for iterative training to obtain the predicted defect category label for the current iteration. d. Calculate the loss function based on the predicted defect category label, the actual defect category label, and the weighting factor of the actual defect category label in the current iteration; e. Enable the mixed precision and gradient pruning strategy, and adjust the weights of the rotation-invariant convolutional neural network in the direction of loss function descent; f, repeat c to e until the total number of training iterations is reached or the loss function no longer decreases after multiple consecutive iterations, to obtain a trained rotation-invariant convolutional neural network.

3. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 2, characterized in that, The formula for calculating the weighting factor is as follows: In the formula, The total number of data samples in the training set. Indicates the first The number of data samples that appear for each real defect category label. This represents the total number of true defect category labels.

4. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 3, characterized in that, The loss function is a weighted binary cross-entropy loss, expressed as: In the formula, The sigmoid activation function is used. Labels for real defects j The weighting factor, For data samples Belongs to the Predicted defect category labels for different defect categories. For data samples Belongs to the The actual defect category label for the defect category.

5. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to any one of claims 1-4, characterized in that, The rotation-invariant convolutional neural network includes an input layer, a backbone network, a feature compression module, and a classification head connected in sequence. The input layer is used to input the tensor corresponding to the wafer image. The backbone network is used to extract features from the tensor corresponding to the input wafer image to obtain initial features. Then, the initial features are sequentially rotated and Gaussian smoothed and filtered in each channel to obtain a feature map that retains rotational equivariance and spatial consistency at the scale. The feature compression module is used to fuse the feature map in the channel dimension and then perform average convergence in the orientation dimension to obtain a rotation- or mirror-invariant representation; the classification head is used to perform convolution, global average pooling and classification on the representation to obtain the detection results of whether there are defects in the wafer in the wafer image and the type of defects.

6. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 5, characterized in that, The backbone network comprises a first PCA-based turnable convolutional layer, two stacked first residual blocks, a first anti-aliasing average pooling layer, and two stacked second residual blocks and a second anti-aliasing average pooling layer, connected in sequence. The feature compression module comprises a decomposed convolutional layer and a grouped average pooling layer, connected in sequence. The classification head comprises a regular convolutional layer, a global average pooling layer, and a fully connected layer, connected in sequence. Both the first and second residual blocks comprise a second PCA-based turnable convolutional layer connected in sequence, with the second turnable convolutional layer introducing the residual structure using a skip connection. Batch normalization and ReLU activation functions are embedded after both the first and second turnable convolutional layers.

7. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 6, characterized in that, The first steerable convolutional layer uses a 5×5 convolutional kernel for initial feature extraction, while retaining M directional channels. The second steerable convolutional layer uses a 3×3 convolutional kernel to perform rotational convolution on the initial features in each channel. The first anti-aliasing average pooling layer and the second anti-aliasing average pooling layer use Gaussian filtering to smooth each channel, obtaining a feature map that retains rotational equivariance and spatial consistency. The decomposed convolutional layer uses a 1×1 convolutional kernel to fuse the feature maps in the channel dimension, and the grouped average pooling layer averages and converges the fused feature maps in the directional dimension to obtain a rotation- or mirror-invariant representation. The regular convolutional layer uses a 3×3 convolutional kernel to convolve the representation, and the global average pooling layer performs global average pooling on the output of the regular convolutional layer to obtain a 32-dimensional global feature. The fully connected layer obtains an M-dimensional defect label prediction vector through linear mapping. The defect label prediction vector represents the detection result of whether there are defects in the wafer in the input wafer image and the type of defect.

8. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 1, characterized in that, The rotation-invariant convolutional neural network used is WFConvNet.

9. The multi-label detection method for wafer defects based on rotation-invariant convolutional neural networks according to claim 2, characterized in that, The S300 includes: The tensor corresponding to each wafer image is input into the input layer of a trained rotation-invariant convolutional neural network. The input layer then feeds this tensor into the backbone network, which extracts features from the tensor corresponding to the input wafer image to obtain initial features. These initial features are then subjected to rotational convolution and Gaussian smoothing filtering in each channel to obtain feature maps that preserve rotational equivariance and spatial consistency across scales. The feature maps are then fused along the channel dimension using a feature compression module and averaged along the orientation dimension to obtain rotation- or mirror-invariant representations. Finally, the classification head performs convolution, global average pooling, and classification on these representations to obtain detection results for the presence and type of defects in the wafer image.

10. A multi-label wafer defect detection device based on rotation-invariant convolutional neural networks, characterized in that, include: The receiving module is configured to receive a wafer inspection task and multiple wafer images corresponding to the wafer inspection task; The encoding module is configured to encode the multiple wafer images to obtain a tensor that retains the semantic information of each pixel; The detection module is configured to input the tensor corresponding to each wafer image into a trained rotation-invariant convolutional neural network, so that it performs defect detection on each input wafer image and obtains the detection results of whether there are defects in the wafer and the type of defects in each wafer image; the detection process of the trained rotation-invariant convolutional neural network is as follows: For each wafer image, feature extraction is performed to obtain initial features. Then, the initial features are sequentially subjected to rotational convolution, Gaussian smoothing filtering, orientation fusion, global average pooling, and classification in each channel to obtain the detection results of whether there are defects in the wafer and the type of defects in each wafer image.

Images