A wafer defect detection method based on RefineDet model

Abstract

The embodiments of the present application relate to a wafer defect detection method, and in particular to a wafer defect detection method based on the RefineDet model, comprising the following steps: first, collecting a wafer image; then, filtering the wafer image; next, de-averaging and normalizing the filtered wafer image in turn to obtain a processed image; then, performing data augmentation on the processed image to obtain an image data set; next, dividing the image data set into a training set, a validation set, and a test set, annotating the defects in the training set and the validation set to obtain an annotated data set, and using the annotated data set to train the RefineDet network model; finally, inputting the test set into the trained RefineDet network model to detect the type and location of the defect in the wafer image. The wafer defect detection method provided in the present application, based on the RefineDet network model, can effectively identify wafer defects and is suitable for the detection of small targets.

Description

Technical Field

[0001] The embodiments of the present application relate to a wafer defect detection method, and specifically to a wafer defect detection method based on the RefineDet model. Background Art

[0002] Semiconductor wafer manufacturing processes are crucial to the development of various industries, including communications, industrial manufacturing, and aerospace. Defects can occur on the wafer surface during processes such as crystal growth and chemical vapor deposition. Crystal defects can significantly impact device performance, making defect suppression a crucial task during crystal growth. Rapidly identifying and distinguishing different types of silicon carbide wafer defects is crucial. Furthermore, due to the relatively complex optical properties of wafers, it is difficult to directly analyze the causes of wafer defect formation through modeling. The primary approach involves classifying wafer defects and then modeling them using physical and mathematical models.

[0003] Currently, wafer defect detection technology is primarily performed through manual inspection or machine vision-based inspection. Manual inspection suffers from low efficiency and uncertain detection accuracy. Machine vision-based inspection utilizes contour extraction, cropping, and morphological transformation methods, followed by template matching, significantly improving efficiency and accuracy.

[0004] Common deep learning networks for object detection include R-CNN, Fast R-CNN, Faster R-CNN, SSD, and R-FCN. Convolutional neural networks, which are hierarchical networks composed of neurons with learnable weights and bias constants, have achieved promising results in defect detection. However, these models are more suitable for large defects. In wafer defect detection, some defects are very small and exhibit significant spatial variation. Current mainstream object detection networks are not very effective for detecting small objects. Summary of the Invention

[0005] The embodiment of the present application provides a wafer defect detection method based on the RefineDet model, which can detect defects with large spatial position changes and tiny defects, thereby improving the accuracy of wafer defect detection.

[0006] To solve the above technical problems, an embodiment of the present application provides a wafer defect detection method based on the RefineDet model, comprising the following steps: first, collecting wafer images; then, filtering the wafer images; next, de-averaging and normalizing the filtered wafer images in turn to obtain processed images; then, performing data augmentation on the processed images to obtain an image dataset; dividing the image dataset into a training set, a validation set, and a test set, annotating the defects in the training set and the validation set to obtain an annotated dataset, and using the annotated dataset to train the RefineDet network model; finally, inputting the test set into the trained RefineDet network model to detect the types and locations of defects in the wafer images.

[0007] In some exemplary embodiments, the image data in the image data set is input into the trained RefineDet network model, and VGG-16 is used as the backbone network of RefinedDet to detect the types and locations of defects in wafer images; the RefineDet network model includes a refinement module, a transmission connection module, and a target detection module connected in sequence; the refinement module is used to filter the positive samples with negative confidence scores greater than a threshold in the image data in the image data set, and extract the positions and sizes of the remaining positive samples to obtain adjusted positive samples; the target detection module is used to take the adjusted positive samples as input and predict the position, size, and category label of the defects through regression processing; the transmission connection module is used to connect the refinement module and the target detection module.

[0008] In some exemplary embodiments, an industrial camera is used to capture wafer images, including: using an area array CCD industrial camera with a resolution of 3384×2710 to scan the wafer surface to obtain the wafer image.

[0009] In some exemplary embodiments, a wafer image is filtered, including: classifying a neighborhood to be filtered based on pixel values; performing weighted summation on the neighborhood to be filtered according to the category to which the neighborhood to be filtered belongs to obtain a denoised image; and performing image enhancement on the denoised image using a histogram.

[0010] In some exemplary embodiments, r is used to represent the grayscale value of the denoised image. In L digital images with different grayscale levels, the value interval of r is set to [0, L-1], and h(r k )=n k Represents the pixel grayscale r of level k k The total number of pixels is n k The total number of pixels in an image is represented as Nall, and the normalized gray level probability value is represented by p(r k), then the probability of gray level rk appearing in an image is expressed as:

[0011]

[0012] The discrete form of the histogram of the equalized image is expressed as:

[0013]

[0014] Among them, H(r) is the corresponding transformation function, P r (r) is the probability function of the image gray level.

[0015] In some exemplary embodiments, the wafer image after filtering is subjected to de-averaging processing, including: subtracting the mean of the corresponding dimension from each dimension, so that each dimension of the input data is centered around 0.

[0016] In some exemplary embodiments, normalization processing is performed on the filtered wafer image, including: inputting a set of batch data, a scaling variable, and an offset variable; calculating a mean and a variance respectively; and normalizing each feature in the batch data to a mean of 0 and a variance of 1, wherein a conversion function for the normalization processing is:

[0017]

[0018] The standardized features are post-processed using scaling and offset variables.

[0019] In some exemplary embodiments, data augmentation is performed on the processed image, including: defining the rotation angle as 90°, 180°, and 270°, taking the processed image as the original image, and rotating the original image three times respectively; flipping the original image vertically or horizontally to form mirror image data of the original image about the horizontal or vertical centerline.

[0020] In some exemplary embodiments, defects in an image dataset are labeled to obtain a labeled dataset, and the labeled dataset is used to train a RefineDet network model, including: dividing the image dataset into a training set, a validation set, and a test set, wherein the training set and the validation set are used to train the RefineDet network model, and the test set is used to test the trained RefineDet network model; labeling the defect locations and types in the training set and the validation set to obtain a labeled dataset, and using the labeled dataset to train the RefineDet network model.

[0021] In some exemplary embodiments, the image dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0022] The technical solution provided by the embodiments of the present application has at least the following advantages:

[0023] An embodiment of the present application provides a wafer defect detection method based on the RefineDet model, which includes the following steps: first, collecting wafer images; then, filtering the wafer images; next, de-averaging and normalizing the filtered wafer images in turn to obtain processed images; then, performing data augmentation on the processed images to obtain an image dataset; dividing the image dataset into a training set, a validation set, and a test set, annotating the defects in the training set and the validation set to obtain an annotated dataset, and using the annotated dataset to train the RefineDet network model; finally, inputting the test set into the trained RefineDet network model to detect the types and locations of defects in the wafer images.

[0024] The present application discloses a wafer defect detection method, which uses a wafer detection model to perform defect detection on a target wafer image to obtain a wafer defect detection result. The wafer defect detection model is established based on a convolutional neural network, and the wafer defect detection result includes whether there are defects in the wafer image, as well as the defect category and location information. The present application adopts a coarse-to-fine positioning method, and at the same time introduces feature fusion into the detection network, which maintains the advantages of a single frame in speed and has significantly improved the detection accuracy, which can effectively improve the detection effect of small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] One or more embodiments are exemplarily described by the pictures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0026] Figure 1 A schematic diagram of a process flow of a wafer defect detection method based on the RefineDet model provided in an embodiment of the present application;

[0027] Figure 2 This is a diagram of the RefineDet network model structure provided in the embodiment of the present application;

[0028] Figure 3 Schematic diagram of the TCB transformation process of the connection module in the RefineDet network model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] As can be seen from the background technology, the existing wafer defect model is more suitable for large-sized defects, and the detection effect for small targets is not very ideal.

[0030] In order to solve the above technical problems, the embodiment of the present application provides a wafer defect detection method based on the RefineDet model, comprising the following steps: first, collecting wafer images; then, filtering the wafer images; next, de-averaging and normalizing the filtered wafer images in turn to obtain processed images; then, performing data augmentation on the processed images to obtain an image data set; dividing the image data set into a training set, a validation set, and a test set, annotating the defects in the training set and the validation set to obtain an annotated data set, and using the annotated data set to train the RefineDet network model; finally, inputting the test set into the trained RefineDet network model to detect the types and positions of defects in the wafer images. By providing a wafer defect detection method based on the RefineDet model, the present application can detect defects with large spatial position changes and tiny defects, thereby improving the accuracy of wafer defect detection.

[0031] The following detailed description of the various embodiments of the present application is provided in conjunction with the accompanying drawings. However, those skilled in the art will appreciate that many technical details are provided in the various embodiments of the present application to facilitate a better understanding of the present application. However, even without these technical details and the various variations and modifications based on the following embodiments, the technical solutions claimed in the present application can still be implemented.

[0032] See Figure 1 , an embodiment of the present application provides a wafer defect detection method based on the RefineDet model, comprising the following steps:

[0033] Step S1: Acquire wafer images.

[0034] Step S2: filter the wafer image.

[0035] Step S3: De-averaging and normalizing the filtered wafer image in sequence to obtain a processed image.

[0036] Step S4: performing data augmentation on the processed image to obtain an image dataset;

[0037] Step S5: Divide the image dataset into a training set, a validation set, and a test set, annotate the defects in the training set and the validation set to obtain an annotated dataset, and use the annotated dataset to train the RefineDet network model.

[0038] Step S6: Input the test set into the trained RefineDet network model to detect the defect types and locations of the wafer images.

[0039] In some embodiments, the image data in the image data set is input into the trained RefineDet network model, and VGG-16 is used as the backbone network of RefinedDet to detect the type and location of defects in the wafer image; the RefineDet network model includes a refinement module (ARM), a transmission connection module (TCB) and an object detection module (ODM) connected in sequence; the refinement module is used to filter the positive samples with negative confidence scores greater than a threshold in the image data in the image data set, and extract the position and size of the remaining positive samples to obtain adjusted positive samples; the object detection module is used to take the adjusted positive samples as input and predict the position, size and category label of the defect through regression processing; the transmission connection module is used to connect the refinement module and the object detection module.

[0040] Specifically, when building the RefineDet neural network model, VGG16 is used as the backbone network for feature extraction. This includes conv_fc6 and conv_fc7, which are converted from the two fully connected layers fc6 and fc7, as well as the expanded conv6_1 and conv6_2. conv4_3, conv5_3, conv_fc7, and conv6_2 are used as the detection layers. For the input image to be detected, the Anchor Refine Module (ARM) extracts features from the convolutional layers to generate feature maps of different sizes. This allows for a rough estimation of the defect location and type, and filters out some invalid candidate regions to reduce the classifier's search space and roughly adjust the location and size of the retained regions.

[0041] Figure 2 Figure 2 shows the RefineDet network model structure diagram. Figure 2 As shown in the figure, the input object is extracted through the convolutional layer in the Anchor Refinement Module (ARM) to obtain feature maps of different sizes.

[0042] The specific methods for designing and matching the positive samples (anchors) are as follows: Feature layers with different pixel stride sizes are selected on VGG-16 and associated with anchors of different scales for prediction. The matching strategy is established using an intersection-over-union (IOU) threshold. The Anchor Refinement Module (ARM) specifically filters out anchors with negative confidence scores greater than a threshold and extracts the positions and sizes of the remaining anchors. The Object Detection Module (ODM) specifically takes the adjusted anchors as input and further uses regression processing to obtain a more accurate target location and predict multi-level labels. The Transfer Connection Module (TCB) specifically establishes a connection between the ARM and ODM, converting the different layers of the ARM features into the form required by the ODM, allowing the ODM to share the ARM features, improving detection accuracy, and further implementing the prediction of the defect location, size, and category labels in the ODM. Finally, non-maximum suppression is applied, the Jaccard similarity coefficient overlap rate is limited to 0.45, and the top 200 high-confidence anchors are retained to produce the final detection results.

[0043] In some embodiments, an industrial camera is used to capture a wafer image in step S1, including: using an area array charge coupled device (CCD) industrial camera with a resolution of 3384×2710 to scan the wafer surface to obtain a wafer image.

[0044] Specifically, the area array CCD industrial camera is combined with an independently built programmable automatic loading slide to convert the image (optical image) of the wafer surface scanned by the camera into a digital signal and send it to the computer.

[0045] In some embodiments, filtering the wafer image in step S2 includes the following steps:

[0046] Step S201: Classify the neighborhood to be filtered based on pixel values.

[0047] Step S202: performing weighted summation on the neighborhood to be filtered according to the category to which the neighborhood to be filtered belongs, and obtaining a denoised image.

[0048] Step S203: perform image enhancement on the denoised image using the histogram.

[0049] Step S2 is mainly the process of filtering the acquired wafer image using Pytorch as a deep learning framework. Specifically, the acquired wafer image is subjected to bilateral filtering. Generally, the difference between adjacent pixels in the image is not large, but this inference does not hold true at the edge. Therefore, it is necessary to first segment or classify the neighborhood to be filtered based on the pixel value, and then give the category to which the point belongs a relatively high weight, and then perform neighborhood weighted summation to obtain the result. Its mathematical expression formula is as follows:

[0050]

[0051] In the above formula, the integral sign precedes the normalization factor, which weights all pixels. c and s are the closeness and similarity functions, respectively. x represents the desired point, and f(x) represents the pixel value at that point. f(x) and h(x) are the images before and after filtering, respectively. After image denoising, histogram enhancement is performed to improve the contrast of the original image.

[0052] In some embodiments, r is used to represent the grayscale value of the denoised image. In L digital images with different grayscale levels, the value interval of r is set to [0, L-1], and h(r k )=n k Represents the pixel grayscale r of level k k The total number of pixels is n k The total number of pixels in an image is represented as Nall, and the normalized gray level probability value is represented by p(r k ), then the probability of gray level rk appearing in an image is expressed as:

[0053]

[0054] The discrete form of the histogram of the equalized image is expressed as:

[0055]

[0056] Among them, H(r) is the corresponding transformation function, and Pr(r) is the probability function of the image gray level.

[0057] In some embodiments, in step S3, the wafer image after filtering is de-averaged, including: subtracting the mean of the corresponding dimension from each dimension, so that each dimension of the input data is centered on 0.

[0058] Specifically, in order to accelerate the convergence of parameters during the back propagation process in the training of the neural network and prevent the problem of easy fitting, the present application performs de-averaging on the wafer image data after filtering.

[0059] In order to solve the problem of slow convergence or gradient explosion of the neural network, which makes the network unable to continue training, the data in all dimensions should be normalized so that the variation range is within a certain range.

[0060] In some embodiments, step S3 normalizes the filtered wafer image, including: inputting a set of batch data, a scaling variable, and an offset variable; calculating the mean and variance respectively; and normalizing each feature in the batch data to a mean of 0 and a variance of 1, wherein the conversion function of the normalization is:

[0061]

[0062] The standardized features are post-processed using scaling and offset variables.

[0063] It should be noted that during the training process, the mean and variance of each small batch need to be retained, and unbiased estimates of the mean and variance are used for standardization during the testing phase.

[0064] In some embodiments, performing data augmentation on the processed image in step S4 includes the following steps:

[0065] Step S401: define the rotation angles as 90°, 180°, and 270°, take the processed image as the original image, and rotate the original image three times respectively.

[0066] Step S402: flip the original image vertically or horizontally to form mirror image data of the original image about the horizontal or vertical center line.

[0067] Specifically, the process of step S4 is mainly to perform data augmentation on the preprocessed image to expand the data volume, thereby improving the performance of the algorithm, so that the wafer defect detection algorithm provided by this application has better robustness to input targets of different sizes and shapes.

[0068] In some embodiments, step S5 labels defects in the image dataset to obtain a labeled dataset, and uses the labeled dataset to train the RefineDet network model, including the following steps:

[0069] Step S501: Divide the image dataset into a training set, a validation set, and a test set. The training set and the validation set are used to train the RefineDet network model, and the test set is used to test the trained RefineDet network model.

[0070] Step S502: Label the defect locations and types of the training set and the validation set to obtain a labeled data set, and use the labeled data set to train the RefineDet network model.

[0071] In some embodiments, the image dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0072] Specifically, in step S501, the augmented images are divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The open-source annotation software LabelImg is used to annotate the defect locations and types. The dataset is then formatted according to the standard VOC data format, and the corresponding environment is configured for training. Caffe-ssd and LabelImg are used, and the generated bounding boxes are formatted as XML. After labeling the data using LabelImg, the dataset is converted to VOC format using Caffe-ssd, which serves as the input for the training model.

[0073] Therefore, the wafer defect detection method based on the RefineDet model proposed in this application first collects wafer images; then, the obtained wafer images are filtered to improve the contrast of the images; next, the images are de-meaned and normalized to speed up the convergence speed in model training; then, the processed images are augmented to expand the data volume; next, LabelImg is used to label the defects in the wafer images, and the obtained data set is used to train the RefineDet network model to obtain the corresponding weight model; finally, the trained RefineDet neural network model is used to detect the types and locations of defects in the wafer images.

[0074] It should be noted that when constructing the RefineDet neural network model, this application uses VGG16 as the backbone network for feature extraction, including conv_fc6 and conv_fc7 converted from the two fully connected layers fc6 and fc7, as well as the expanded conv6_1 and conv6_2. Conv4_3, conv5_3, conv_fc7, and conv6_2 are used as detection layers. For the input image to be detected, feature maps of different sizes are obtained after feature extraction at the convolutional layer in ARM. The location and type of defects can be roughly estimated, and some invalid candidate areas can be filtered out to reduce the search space of the classifier and roughly adjust the location and size of the retained area.

[0075] After establishing the backbone network, feature fusion is performed. First, a 5×5 feature map passes through a connection block (TCB) to obtain a rectangular block of the corresponding size. Next, the 10×10 gray feature map passes through the TCB module to obtain another feature map. The TCB module now performs a deconvolution upsampling operation, and the same process continues. Feature map fusion is used to strengthen the connection between different layers. Through this connection, the feature maps in the detection network all incorporate features of different scales and semantic strengths, ensuring that the feature maps of the detection layer can detect defects of different scales.

[0076] Figure 3 The following is a schematic diagram of the TCB transformation process of the connection module in the RefineDet network model. Taking the fusion of the high-level feature map CONV4_3 of the ODM module and the low-level feature map conv3_3 of the ARM module as an example, the process is as follows: Figure 3 As shown in the figure, the convolution kernel size is 3×3, the number of channels is 256, and the deconvolution kernel size is 4×4, the stride is 2, and the number of channels is 256. Feature maps of different layers vary in size, so CONV4_3 is first scaled to the same size as the conv3_3 feature map through deconvolution. These maps are then fused through element-wise addition to produce CONV3_3. Finally, ODM performs more accurate regression on the fused wafer defect candidate regions and uses the non-maximum suppression algorithm to obtain detection results for wafer defects of different scales.

[0077] Specifically, first, the RefineDet model of this application uses conv4_3 as the initial detection layer. In order to make the model more suitable for detecting small targets, the step size is set to 2. Moving one point on the feature map is equivalent to moving 2 pixels on the original image, and the aspect ratio is 1:1.

[0078] Then, the RefineDet network adopts an end-to-end training method. The loss function of the entire network consists of two parts: ARM and ODM. The ARM part includes the binary classification loss Lbcls and the regression loss Lreg, and the ODM part includes the multi-classification loss Lmcls and the regression loss Lreg. The loss function of the entire network is:

[0079]

[0080] N arm and N odm Refers to the number of positive sample anchors in ARM and ODM respectively, p i Refers to the predicted i-th anchor, which is the confidence of a target, x i Refers to the coordinates of the i-th anchor predicted after ARM refinement, c i is the predicted b in ODM boxThe object category, t i is the position coordinate of the candidate region box predicted by ODM, l i * is the true category label of the i-th anchor, g i * is the actual position and size of the i-th anchor; L b represents the binary classification loss, L m represents the multi-classification loss, L r Represents the regression loss. [l i * >=1] means that if the negative confidence is greater than a threshold θ, then the value 1 is returned, otherwise 0 is returned.

[0081] Finally, the parameters of the two convolutional layers (conv6_1 and conv6_2) added to RefineDet based on VGG-16 are randomly initialized using the Xavier method. ARM filters out anchors with negative confidence scores greater than a threshold and extracts the positions and sizes of the remaining anchors. ODM then outputs the top 400 high-confidence anchors for each detection image. Finally, non-maximum suppression is applied with a Jaccard overlap limit of 0.45, and the top 200 high-confidence anchors are retained to produce the final detection results.

[0082] Based on the above technical solution, an embodiment of the present application provides a wafer defect detection method based on the RefineDet model, which includes the following steps: first, collecting wafer images; then, filtering the wafer images; next, de-averaging and normalizing the filtered wafer images in turn to obtain processed images; then, performing data augmentation on the processed images to obtain an image data set; dividing the image data set into a training set, a validation set, and a test set, annotating the defects in the training set and the validation set to obtain an annotated data set, and using the annotated data set to train the RefineDet network model; finally, inputting the test set into the trained RefineDet network model to detect the types and locations of defects in the wafer images.

[0083] The present application discloses a wafer defect detection method, which uses a wafer detection model to perform defect detection on a target wafer image to obtain a wafer defect detection result. The wafer defect detection model of the present application is established based on a convolutional neural network, and the wafer defect detection result includes whether there are defects in the wafer image, as well as the defect category and location information. The present application adopts a coarse-to-fine positioning method, and at the same time introduces feature fusion into the detection network, which maintains the advantages of a single frame in speed and has significantly improved the detection accuracy, which can effectively improve the detection effect of small targets.

[0084] Those skilled in the art will appreciate that the above-described embodiments are specific examples for implementing the present application, and that in actual applications, various changes in form and detail may be made thereto without departing from the spirit and scope of the present application. Any person skilled in the art may make changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be subject to the scope defined in the claims.

Claims

1. A wafer defect detection method based on the RefineDet model, characterized in that: The following steps are involved: Acquire wafer images; performing filtering processing on the wafer image; De-averaging and normalizing the filtered wafer image in sequence to obtain a processed image; performing data augmentation on the processed image to obtain an image dataset; Dividing the image dataset into a training set, a validation set, and a test set, annotating defects in the training set and the validation set to obtain an annotated dataset, and using the annotated dataset to train the RefineDet network model; Inputting the test set into the trained RefineDet network model to detect the types and locations of defects in the wafer image; Input the image data in the image dataset into the trained RefineDet network model, use VGG-16 as the backbone network of RefineDet, and detect the defect types and locations of wafer images; The RefineDet network model includes a refinement module, a transmission connection module, and a target detection module connected in sequence; The refinement module is used to filter out positive samples whose negative confidence scores are greater than a threshold in the image data in the image dataset, and extract the positions and sizes of the remaining positive samples to obtain adjusted positive samples; The target detection module is used to take the adjusted positive samples as input and predict the position, size, and category label of the defect through regression processing; The transmission connection module is used to connect the refinement module and the target detection module.

2. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: Use industrial cameras to capture wafer images, including: The wafer surface is scanned using an area array CCD industrial camera with a resolution of 3384×2710 to obtain a wafer image.

3. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: Performing filtering processing on the wafer image, comprising: Based on the pixel value, the neighborhood to be filtered is,classified; According to the category of the neighborhood to be filtered, the neighborhood to be filtered is weighted summed to obtain the denoised image; The denoised image is enhanced using a histogram.

4. The wafer defect detection method based on the RefineDet model according to claim 3, characterized in that: Let r represent the gray value of the denoised image. In the digital images with L different gray levels, let the value range of r be [0, L-1]. Indicates the pixel grayscale of level k The total number of pixels is The total number of pixels in an image is represented as Nall, and the normalized grayscale probability value is expressed as Indicates that the gray level in an image The probability of occurrence is expressed as: The discrete form of the histogram of the equalized image is expressed as: Among them, H(r) is the corresponding transformation function, P r (r) is the probability function of the image gray level.

5. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: The filtered wafer image is de-averaged, including: Subtract the mean of the corresponding dimension from each dimension so that each dimension of the input data is centered around 0.

6. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: Normalize the filtered wafer image, including: Input a set of batch data, scaling variables and offset variables; calculate the mean and variance respectively; Each feature in the batch data is standardized with a mean of 0 and a variance of 1. The transformation function of the standardized processing is: The standardized features are post-processed using scaling and offset variables.

7. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: Performing data augmentation on the processed image, including: The rotation angles are defined as 90°, 180°, and 270°, and the processed image is used as the original image, and the original image is rotated three times respectively; The original image is flipped vertically or horizontally to form mirror image data of the original image about a horizontal or vertical center line.

8. The wafer defect detection method based on the RefineDet model according to claim 1, characterized in that: Annotating defects in the image dataset to obtain an annotated dataset, and using the annotated dataset to train a RefineDet network model, including: Dividing the image dataset into a training set, a validation set, and a test set, wherein the training set and the validation set are used to train the RefineDet network model, and the test set is used to test the trained RefineDet network model; The defect locations and types of the training set and the validation set are labeled to obtain a labeled data set, and the labeled data set is used to train the RefineDet network model.

9. The wafer defect detection method based on the RefineDet model according to claim 8, characterized in that: The image dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

Images