A training method and a classification method, system and device for real-time classification of wafer defects based on privileged learning

Abstract

The application provides a wafer defect real-time classification training method, a classification method, a system and equipment based on privileged learning, comprising the following steps: S1, obtaining low-resolution images and high-resolution images corresponding to the same wafer defect, and labeling the defect category; S2, inputting the high-resolution images into a privileged branch and inputting the low-resolution images into a main branch; S3, performing privileged encoding and main encoding on the high-resolution images and the low-resolution images respectively to obtain high-resolution feature maps and low-resolution feature maps; S4, performing feature alignment on the two kinds of feature maps to obtain aligned low-resolution feature maps; S5, performing privileged decoding and main decoding respectively to obtain privileged decoding features and main decoding features; S6, performing image reconstruction and defect classification according to the decoding features to obtain a reconstructed image and an image category; and S7, updating model parameters according to the reconstructed image, the image category, the high-resolution image and the feature alignment result.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection and image classification technology, and more specifically, to a training method, classification method, system and device for real-time classification of wafer defects based on privileged learning. Background Technology

[0002] Semiconductor manufacturing is the core of the modern electronics industry, and its yield rate directly affects production costs and efficiency. Wafers are the fundamental material for manufacturing semiconductor chips. During the complex manufacturing process, wafer surfaces are prone to various minute defects, such as particles, scratches, and pattern losses. These defects can severely impact chip performance and reliability. Therefore, rapid and accurate defect detection of wafers during the manufacturing process is crucial.

[0003] Defect detection comprises two main steps: defect localization and defect classification. Defect localization outputs the coordinates of all suspected defects, while defect classification determines whether the current defect is a nuisance or a killer defect. For killer defects, further subdivision is required. Nuisance defects typically include minute dust particles on the wafer surface, trace amounts of photoresist remaining after photolithography, and tiny particles deposited in non-critical areas. Although these defects are abnormal, they do not affect the chip's electrical performance. Therefore, the final defect detection results should minimize the reporting of nuisance defects to reduce the time engineers spend reviewing meaningless defects and avoid incorrect adjustments to process parameters due to false alarms.

[0004] To keep up with production capacity, defect detection typically utilizes low-resolution line scan images (above 100nm), which are fast but have low accuracy. However, image resolution significantly impacts classification accuracy during defect classification. For example, scanning electron microscope (SEM) images, considered the gold standard for defect classification, typically have a resolution of 1nm or even lower, but SEM is very slow and can only be performed offline. A current compromise is to use a repeat camera to acquire images (30nm) and perform classification after the line scan. However, this still slows down the overall defect detection process. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and proposes a training method, classification method, system, and device for real-time classification of wafer defects based on privileged learning. It trains a bi-branch defect classification model using low-resolution and high-resolution images corresponding to the same wafer defect. The bi-branch defect classification model includes a privileged branch and a main branch. The method includes the following steps:

[0006] S1. Obtain low-resolution and high-resolution images corresponding to the same wafer defect, and label the wafer defect with category tags according to its type; S2. Input the high-resolution image into the privileged branch, and input the low-resolution image into the main branch; S3. The privileged branch performs privileged encoding on the high-resolution image to obtain a high-resolution feature map; The main branch performs main encoding on the low-resolution image to obtain a low-resolution feature map; S4. Align the low-resolution feature map and the high-resolution feature map to obtain the aligned low-resolution feature map. S5. The privileged branch performs privileged decoding based on the high-resolution feature map to obtain privileged decoding features; The main branch performs main decoding based on the aligned low-resolution feature map to obtain the main decoded features; S6. The privileged branch performs image reconstruction and defect classification based on the privileged decoding features to obtain a privileged reconstructed image and a privileged image category; The main branch performs image reconstruction and defect classification based on the main decoding features to obtain the main reconstructed image and the main image category; S7. Update the parameters of the dual-branch defect classification model based on the privileged reconstructed image, the privileged image category, the main reconstructed image, the main image category, the high-resolution image, the high-resolution feature map, and the aligned low-resolution feature map.

[0007] In wafer defect detection scenarios, to keep up with production line inspection speeds, low-resolution images are typically acquired using line scan cameras. While these images are fast and suitable for online detection, they are insufficient for clearly representing the edges and textures of minute defects such as particles, scratches, and missing patterns. Scanning electron microscope (SEM) images or repeater camera images offer higher resolution and can reveal the detailed morphology of defects, but their slow acquisition speed makes them difficult to use directly for real-time online classification. Therefore, this solution uses both low-resolution and high-resolution images corresponding to the same wafer defect during the training phase. The low-resolution image corresponds to the input during actual deployment, while the high-resolution image is used as privileged information during training. Both types of images correspond to the same wafer defect and share the same category label, enabling the model to establish a correspondence between the blurred defect representation in the low-resolution image and the clear defect representation in the high-resolution image during training.

[0008] Through the training method described above, high-resolution images are only used for model parameter updates during the training phase, and no high-resolution images are required for subsequent classification. After training, the main branch is used for real-time classification of low-resolution images. In this way, defect details in scanning electron microscope images or repeat camera images are utilized during model training, while classification is still performed solely using line scan camera images during actual inspection, eliminating the need for additional high-resolution image acquisition. This approach is well-suited for the rapid scanning and online classification inspection processes in wafer fabrication lines.

[0009] Furthermore, step S7, which updates the parameters of the dual-branch defect classification model, specifically involves: The privileged reconstruction loss is calculated based on the privileged reconstructed image and the high-resolution image, and the main reconstruction loss is calculated based on the main reconstructed image and the high-resolution image; Calculate the privileged classification loss based on the privileged image category and the category label, and calculate the main classification loss based on the main image category and the category label; Calculate the feature alignment loss based on the high-resolution feature map and the aligned low-resolution feature map; The parameters of the bi-branch defect classification model are updated based on the aforementioned loss.

[0010] The privileged branch takes a high-resolution image as input and uses it as the reconstruction target. Since the privileged reconstruction loss is calculated from the difference between the privileged reconstructed image and the original high-resolution image, the difference between the privileged reconstructed image and the original high-resolution image increases when defect edges, texture, or morphological information in the high-resolution image is lost during privileged encoding or decoding. When the model updates its parameters in reverse, it adjusts the privileged encoding and decoding processes based on this difference, making the privileged reconstructed image closer to the input high-resolution image in terms of defect details.

[0011] The main branch also aims to reconstruct high-resolution images. Since the main reconstruction loss is calculated from the difference between the main reconstructed image and the high-resolution image, the main branch cannot only restore the blurred contours in the low-resolution image; it also needs to learn the detailed features of the corresponding high-resolution image during training.

[0012] In this way, after training is completed, when only low-resolution images are input during the deployment phase, the classification results obtained by inputting the low-resolution images to be detected into the main branch are closer to the true defect categories.

[0013] Furthermore, the privileged classification loss and the main classification loss are calculated using cross-entropy loss; The privileged reconstruction loss and the primary reconstruction loss are calculated using L1 loss. The feature alignment loss is calculated using multi-granularity high-frequency sensing loss.

[0014] Furthermore, both the main branch and the privileged branch adopt a reversible neural network structure, which includes a reversible downsampling operator and a reversible upsampling operator.

[0015] Furthermore, after the dual-branch defect classification model is trained, the network structure used for main encoding, main decoding and defect classification in the main branch is deployed as the final defect classification model.

[0016] Furthermore, the final defect classification model does not execute the privileged encoding, privileged decoding, image reconstruction, and defect classification processes corresponding to the privileged branch, nor does it execute the image reconstruction process of the main branch.

[0017] Furthermore, the high-resolution image is a scanning electron microscope image or a repeater camera image, and the low-resolution image is a line scan camera image.

[0018] This invention also provides a real-time wafer defect classification method based on privileged learning, comprising the following steps: Obtain the low-resolution image to be classified; The low-resolution image to be classified is input into the defect classification model, which performs main encoding, main decoding and defect classification on the low-resolution image to be classified, and outputs the defect category corresponding to the low-resolution image to be classified. The defect classification model is the final defect classification model described above.

[0019] This invention also provides a real-time wafer defect classification system based on privileged learning, comprising: The image acquisition module is used to acquire low-resolution and high-resolution images corresponding to the same wafer defect, and to label the wafer defect with category tags according to the type of the wafer defect; The input module is used to input the high-resolution image into the privileged branch of the two-branch defect classification model and input the low-resolution image into the main branch of the two-branch defect classification model. The feature encoding module is used to perform privileged encoding on the high-resolution image through the privileged branch to obtain a high-resolution feature map, and to perform main encoding on the low-resolution image through the main branch to obtain a low-resolution feature map; The feature alignment module is used to align the low-resolution feature map and the high-resolution feature map to obtain an aligned low-resolution feature map. The feature decoding module is used to perform privileged decoding based on the high-resolution feature map through the privileged branch to obtain privileged decoded features, and to perform main decoding based on the aligned low-resolution feature map through the main branch to obtain main decoded features; The reconstruction head and classification head modules are used to perform image reconstruction and defect classification based on the privileged decoding features to obtain a privileged reconstructed image and a privileged image category, and to perform image reconstruction and defect classification based on the master decoding features to obtain a master reconstructed image and a master image category. The parameter update module is used to update the parameters of the dual-branch defect classification model based on the privileged reconstructed image, the privileged image category, the main reconstructed image, the main image category, the high-resolution feature map, and the aligned low-resolution feature map. The present invention also includes a computer device comprising a memory and a processor, the memory storing a computer program, the processor executing the above-described training method for real-time classification of wafer defects based on privileged learning.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a privileged learning method during training, which enables the classification model to learn complex features on high-resolution images through indirectly transmitted high-resolution image information, thereby significantly improving the accuracy of defect classification on low-resolution images.

[0021] 2. The main branch and privileged branches adopt the I-NUET structure. The core innovation of I-NUET lies in the introduction of learnable and strictly reversible downsampling and upsampling operators, rather than traditional irreversible operations (such as pooling, strided convolution, or bilinear interpolation). Traditional U-Net encoders irreversibly lose information during downsampling, regardless of whether max pooling or strided convolution is used—the lost details are often the key high-frequency information needed to distinguish nanoscale defects. I-NUET achieves zero information loss during information compression through a mathematically rigorous bijection design, providing the richest available information for subsequent decoding tasks. MGHF Loss is a novel perceptual loss function based on reversible neural networks, specifically designed to address the problem of high-frequency information loss during feature extraction in traditional CNN perceptual losses. Unlike the perceptual losses commonly used in traditional CNN networks such as VGG, MGHF Loss uses INNs as feature extractors, theoretically achieving information fidelity and better recovering sharp edges and realistic textures. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the training method for real-time wafer defect classification based on privileged learning according to the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of the dual-branch defect classification model of the present invention. Detailed Implementation

[0024] Example 1 like Figure 1 As shown, this embodiment provides a training method for real-time wafer defect classification based on privileged learning. A bi-branch defect classification model is trained using low-resolution and high-resolution images corresponding to the same wafer defect. The bi-branch defect classification model includes a privileged branch and a main branch. The method includes the following steps: S1. Obtain low-resolution and high-resolution images corresponding to the same wafer defect, and label the wafer defect with category tags according to its type; S2. Input the high-resolution image into the privileged branch, and input the low-resolution image into the main branch; S3. The privileged branch performs privileged encoding on the high-resolution image to obtain a high-resolution feature map; The main branch performs main encoding on the low-resolution image to obtain a low-resolution feature map; S4. Align the low-resolution feature map and the high-resolution feature map to obtain the aligned low-resolution feature map. S5. The privileged branch performs privileged decoding based on the high-resolution feature map to obtain privileged decoding features; The main branch performs main decoding based on the aligned low-resolution feature map to obtain the main decoded features; S6. The privileged branch performs image reconstruction and defect classification based on the privileged decoding features to obtain a privileged reconstructed image and a privileged image category; The main branch performs image reconstruction and defect classification based on the main decoding features to obtain the main reconstructed image and the main image category; S7. Update the parameters of the dual-branch defect classification model based on the privileged reconstructed image, the privileged image category, the main reconstructed image, the main image category, the high-resolution image, the high-resolution feature map, and the aligned low-resolution feature map.

[0025] like Figure 2 As shown, the dual-branch defect classification model includes a privileged branch for processing high-resolution images and a main branch for processing low-resolution images. Let the high-resolution image corresponding to the same wafer defect be... Low-resolution images are The category labels are one-hot vectors of category y.

[0026] During training, high-resolution images Input privileged branch, which extracts high-resolution feature maps by privileged encoder. Low-resolution images The input to the main branch is used by the main encoder to extract low-resolution feature maps. Subsequently, the high-resolution feature maps are processed by the feature alignment module. By introducing a low-resolution feature map, an aligned low-resolution feature map is obtained. .

[0027] Then, the privileged branch is based on the high-resolution feature map. Privileged decoding is performed, and the main branch uses the aligned low-resolution feature map. Perform main decoding. The decoded features are then input into the corresponding reconstruction and classification processes to obtain the reconstructed high-resolution image output by the main branch. and category probability vector and the reconstructed high-resolution image output by the privileged branch. and category probability vector .

[0028] In wafer defect detection scenarios, to keep up with production line inspection speeds, low-resolution images are typically acquired using line scan cameras. While these images are fast and suitable for online detection, they are insufficient for clearly representing the edges and textures of minute defects such as particles, scratches, and missing patterns. Scanning electron microscope (SEM) images or repeater camera images offer higher resolution and can reveal the detailed morphology of defects, but their slow acquisition speed makes them difficult to use directly for real-time online classification. Therefore, this solution uses both low-resolution and high-resolution images corresponding to the same wafer defect during the training phase. The low-resolution image corresponds to the input during actual deployment, while the high-resolution image is used as privileged information during training. Both types of images correspond to the same wafer defect and share the same category label, enabling the model to establish a correspondence between the blurred defect representation in the low-resolution image and the clear defect representation in the high-resolution image during training.

[0029] Through the training method described above, high-resolution images are only used for model parameter updates during the training phase, and no high-resolution images are required for subsequent classification. After training, the main branch is used for real-time classification of low-resolution images. In this way, defect details in scanning electron microscope images or repeat camera images are utilized during model training, while classification is still performed solely using line scan camera images during actual inspection, eliminating the need for additional high-resolution image acquisition. This approach is well-suited for the rapid scanning and online classification inspection processes in wafer fabrication lines.

[0030] Preferably, step S7 updates the parameters of the dual-branch defect classification model as follows: The privileged reconstruction loss is calculated based on the privileged reconstructed image and the high-resolution image, and the main reconstruction loss is calculated based on the main reconstructed image and the high-resolution image; Calculate the privileged classification loss based on the privileged image category and the category label, and calculate the main classification loss based on the main image category and the category label; Calculate the feature alignment loss based on the high-resolution feature map and the aligned low-resolution feature map; The parameters of the bi-branch defect classification model are updated based on the aforementioned loss.

[0031] During training, the total loss Ltotal of the dual-branch defect classification model includes the main classification loss, privileged classification loss, main reconstruction loss, privileged reconstruction loss, and feature alignment loss, specifically:

[0032] in, Indicates the primary classification loss. This represents the weight corresponding to the main classification loss; Indicates privileged classification loss, This represents the weight corresponding to the privileged classification loss; Indicates the main reconstruction loss. This represents the weight corresponding to the primary reconstruction loss; Indicating the loss of privileges during the reconstruction, This represents the weight corresponding to the privilege reconstruction loss; This represents the feature alignment loss between the high-resolution feature map and the aligned low-resolution feature map. This represents the weights corresponding to the feature alignment loss.

[0033] Each weight satisfies:

[0034] The classification results of the main branch, the classification results of the privileged branch, the reconstruction results of the two branches, and the feature alignment results, based on the total loss mentioned above, are all used to update the model parameters.

[0035] The privileged branch takes a high-resolution image as input and uses it as the reconstruction target. Since the privileged reconstruction loss is calculated from the difference between the privileged reconstructed image and the original high-resolution image, the difference between the privileged reconstructed image and the original high-resolution image increases when defect edges, texture, or morphological information in the high-resolution image is lost during privileged encoding or decoding. When the model updates its parameters in reverse, it adjusts the privileged encoding and decoding processes based on this difference, making the privileged reconstructed image closer to the input high-resolution image in terms of defect details.

[0036] The main branch also aims to reconstruct high-resolution images. Since the main reconstruction loss is calculated from the difference between the main reconstructed image and the high-resolution image, the main branch cannot only restore the blurred contours in the low-resolution image; it also needs to learn the detailed features of the corresponding high-resolution image during training.

[0037] In this way, after training is completed, when only low-resolution images are input during the deployment phase, the classification result obtained by inputting the low-resolution image to be detected into the main branch is closer to the actual classification result.

[0038] Preferably, the privileged classification loss and the primary classification loss are calculated using cross-entropy loss; The privileged reconstruction loss and the primary reconstruction loss are calculated using L1 loss. The feature alignment loss is calculated using multi-granularity high-frequency sensing loss.

[0039] Both the privileged classification loss and the main classification loss are calculated using cross-entropy loss, specifically:

[0040]

[0041] in, This represents the class probability vector output by the privileged branch. This represents the class probability vector output by the main branch, where y represents the class label.

[0042] Furthermore, both the privileged reconstruction loss and the primary reconstruction loss are calculated using L1 loss, specifically as follows:

[0043]

[0044] in, This represents a high-resolution image corresponding to a defect on the same wafer. This represents the main reconstructed image output by the main branch. This represents the reconstructed image output by the privileged branch. The privileged branch takes a high-resolution image as input and uses that high-resolution image as the reconstruction target; the main branch takes a low-resolution image as input, but also uses a high-resolution image as the reconstruction target. Thus, during training, the main branch does not reproduce the low-resolution image itself, but rather learns the edge, texture, and morphological features of the same defect in the high-resolution image based on the low-resolution input.

[0045] Furthermore, the feature alignment loss is calculated using multi-granularity high-frequency sensing loss, specifically as follows:

[0046] in, This represents the main reconstructed image output by the main branch. This represents the privileged reconstructed image output by the privileged branch.

[0047] Preferably, both the main branch and the privileged branch adopt a reversible neural network structure, which includes a reversible downsampling operator and a reversible upsampling operator.

[0048] The reversible neural network structure is an I-NUET structure. During the encoding process, the I-NUET structure uses a reversible downsampling operator to downsample the input image or intermediate features, and during the decoding process, it uses a corresponding reversible upsampling operator to upsample the input image.

[0049] For wafer defect images, defects such as particles, scratches, and missing patterns typically manifest as local edge variations, abrupt texture changes, or subtle morphological differences. If irreversible downsampling methods such as pooling and stride convolution are used during the encoding process, some high-frequency details will be weakened during the downsampling stage, making them difficult to recover during subsequent decoding. The I-NUET structure transmits features through reversible downsampling and upsampling operators, enabling the features obtained in the encoding stage to retain more detailed information in the decoding stage. This is beneficial for preserving the edge, texture, and morphological features of wafer defects in both the main reconstructed image and the privileged reconstructed image.

[0050] Preferably, after the dual-branch defect classification model is trained, the network structure used for main encoding, main decoding and defect classification in the main branch is deployed as the final defect classification model.

[0051] Preferably, the final defect classification model does not execute the privileged encoding, privileged decoding, image reconstruction, and defect classification processes corresponding to the privileged branch, nor does it execute the image reconstruction process of the main branch.

[0052] Preferably, the high-resolution image is a scanning electron microscope image or a repeater camera image, and the low-resolution image is a line scan camera image.

[0053] In this embodiment, the low-resolution image has a resolution of 100nm or higher, and the high-resolution image has a resolution of 30nm or lower.

Claims

1. A training method for real-time wafer defect classification based on privileged learning, wherein a two-branch defect classification model is trained using low-resolution and high-resolution images corresponding to the same wafer defect, the two-branch defect classification model comprising a privileged branch and a main branch, characterized in that, Includes the following steps: S1. Obtain low-resolution and high-resolution images corresponding to the same wafer defect, and label the wafer defect with category tags according to its type; S2. Input the high-resolution image into the privileged branch, and input the low-resolution image into the main branch; S3. The privileged branch performs privileged encoding on the high-resolution image to obtain a high-resolution feature map; The main branch performs main encoding on the low-resolution image to obtain a low-resolution feature map; S4. Align the low-resolution feature map and the high-resolution feature map to obtain the aligned low-resolution feature map. S5. The privileged branch performs privileged decoding based on the high-resolution feature map to obtain privileged decoding features; The main branch performs main decoding based on the aligned low-resolution feature map to obtain the main decoded features; S6. The privileged branch performs image reconstruction and defect classification based on the privileged decoding features to obtain a privileged reconstructed image and a privileged image category; The main branch performs image reconstruction and defect classification based on the main decoding features to obtain the main reconstructed image and the main image category; S7. Update the parameters of the dual-branch defect classification model based on the privileged reconstructed image, the privileged image category, the main reconstructed image, the main image category, the high-resolution image, the high-resolution feature map, and the aligned low-resolution feature map.

2. The training method for real-time wafer defect classification based on privileged learning according to claim 1, characterized in that, The specific steps in step S7 for updating the parameters of the bi-branch defect classification model are as follows: The privileged reconstruction loss is calculated based on the privileged reconstructed image and the high-resolution image, and the main reconstruction loss is calculated based on the main reconstructed image and the high-resolution image; Calculate the privileged classification loss based on the privileged image category and the category label, and calculate the main classification loss based on the main image category and the category label; Calculate the feature alignment loss based on the high-resolution feature map and the aligned low-resolution feature map; The parameters of the bi-branch defect classification model are updated based on the aforementioned loss.

3. The training method for real-time wafer defect classification based on privileged learning according to claim 2, characterized in that, The privileged classification loss and the primary classification loss are calculated using cross-entropy loss. The privileged reconstruction loss and the primary reconstruction loss are calculated using L1 loss. The feature alignment loss is calculated using multi-granularity high-frequency sensing loss.

4. The training method for real-time wafer defect classification based on privileged learning according to claim 1, characterized in that, Both the main branch and the privileged branch adopt a reversible neural network structure, which includes a reversible downsampling operator and a reversible upsampling operator.

5. The training method for real-time wafer defect classification based on privileged learning according to claim 1, characterized in that, After the dual-branch defect classification model is trained, the network structure used for main encoding, main decoding and defect classification in the main branch is deployed as the final defect classification model.

6. The training method for real-time wafer defect classification based on privileged learning according to claim 5, characterized in that, The final defect classification model does not execute the privileged encoding, privileged decoding, image reconstruction, and defect classification processes corresponding to the privileged branch, nor does it execute the image reconstruction process of the main branch.

7. A training method for real-time wafer defect classification based on privileged learning according to any one of claims 1 to 6, characterized in that, The high-resolution image is a scanning electron microscope image or a repeater camera image, and the low-resolution image is a line scan camera image.

8. A real-time wafer defect classification method based on privileged learning, characterized in that, Includes the following steps: Obtain the low-resolution image to be classified; The low-resolution image to be classified is input into the defect classification model, which performs main encoding, main decoding and defect classification on the low-resolution image to be classified, and outputs the defect category corresponding to the low-resolution image to be classified. The defect classification model is the final defect classification model as described in claim 5 or 6.

9. A real-time wafer defect classification system based on privileged learning, characterized in that, include: The image acquisition module is used to acquire low-resolution and high-resolution images corresponding to the same wafer defect, and to label the wafer defect with category tags according to the type of the wafer defect; The input module is used to input the high-resolution image into the privileged branch of the two-branch defect classification model and input the low-resolution image into the main branch of the two-branch defect classification model. The feature encoding module is used to perform privileged encoding on the high-resolution image through the privileged branch to obtain a high-resolution feature map, and to perform main encoding on the low-resolution image through the main branch to obtain a low-resolution feature map; The feature alignment module is used to align the low-resolution feature map and the high-resolution feature map to obtain an aligned low-resolution feature map. The feature decoding module is used to perform privileged decoding based on the high-resolution feature map through the privileged branch to obtain privileged decoded features, and to perform main decoding based on the aligned low-resolution feature map through the main branch to obtain main decoded features; The reconstruction head and classification head modules are used to perform image reconstruction and defect classification based on the privileged decoding features to obtain a privileged reconstructed image and a privileged image category, and to perform image reconstruction and defect classification based on the master decoding features to obtain a master reconstructed image and a master image category. The parameter update module is used to update the parameters of the dual-branch defect classification model based on the privileged reconstructed image, the privileged image category, the main reconstructed image, the main image category, the high-resolution image, the high-resolution feature map, and the aligned low-resolution feature map.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the training method for real-time classification of wafer defects based on privileged learning as described in any one of claims 1 to 7.

Images