Method and apparatus for training a multi-defect segmentation model of wafer images

Abstract

This specification provides a method and apparatus for training a multi-defect segmentation model for wafer images. One specific implementation of the method includes: acquiring a sample wafer image containing m defects, and m defect regions labeled for each of the m defects; determining a label image containing multiple channels based on the m defect regions, wherein a label image for a single channel represents a single defect region; and training the multi-defect segmentation model by using the sample wafer image as input and the label image as the desired output.

Description

Technical Field

[0001] This specification relates to the field of semiconductor integrated circuit manufacturing, and more particularly to a method and apparatus for training a multi-defect segmentation model of a wafer image. Background Technology

[0002] A wafer is a chip used to manufacture semiconductor circuits, and its raw material is silicon. High-purity polycrystalline silicon is dissolved, doped with silicon seed crystals, and then slowly pulled out to form cylindrical single-crystal silicon. After grinding, polishing, and slicing, the silicon ingot forms a silicon wafer. During wafer fabrication, chemical vapor deposition, optical development, chemical mechanical polishing, single-crystal pulling, slicing, grinding, polishing, layer addition, photolithography, doping, heat treatment, and dicing can all introduce defects to the wafer surface. To prevent defective wafers from entering the packaging process, it is necessary to detect surface defects, assist in wafer sorting, and analyze the causes of defects to improve the manufacturing process.

[0003] In semiconductor manufacturing, a single wafer may sometimes contain several defects simultaneously, with the defect patterns partially overlapping or not. Defect image segmentation, as an intermediate processing step in defect detection systems, plays a crucial bridging role between defect image preprocessing and advanced machine vision analysis, and is a fundamental and critical technology. Its purpose is to divide an image into regions with distinct characteristics and extract the regions of interest, transforming the input image into a segmented image. This is essential for feature extraction, target recognition and classification, and subsequent high-level processing. Therefore, how to segment defect regions from wafer images is of significant importance and value for defect detection and recognition. Summary of the Invention

[0004] The embodiments of this specification describe a method and apparatus for training a multi-defect segmentation model for wafer images. By introducing multi-channel labeled images as the desired output, a multi-defect segmentation model capable of outputting multiple wafer defect images can be trained, with each wafer defect image including a single defect region, thereby achieving the segmentation of multiple defects in the wafer image.

[0005] According to a first aspect, a method for training a multi-defect segmentation model for a wafer image is provided, comprising: acquiring a sample wafer image containing m defects, and m defect regions labeled for the m defects respectively; determining a label image containing multiple channels based on the m defect regions, wherein the label image of a single channel shows a single defect region; and training the multi-defect segmentation model by taking the sample wafer image as input and the label image as the desired output.

[0006] In some implementations, determining a label image containing multiple channels based on the m defect regions includes: for a first defect region among the m defect regions, determining a first channel label image in the multiple channels of the label image that corresponds to the first defect region; setting the pixel value of the region in the first channel label image that corresponds to the first defect region as a first pixel value; setting the pixel value of the region in the first channel label image that corresponds to the non-defect region corresponding to the first defect region as a second pixel value, wherein the first pixel value and the second pixel value are different.

[0007] In some implementations, the first pixel value is 1 and the second pixel value is 0.

[0008] In some implementations, obtaining the sample wafer image containing m defects includes: obtaining m sample defect images, wherein each sample defect image corresponds to one defect; and superimposing the m sample defect images to obtain the sample wafer image.

[0009] In some implementations, the number of channels in the aforementioned label image is determined by statistical analysis of the number of defects contained in multiple historical multi-defect wafer images.

[0010] In some implementations, the m defects contained in the above sample wafer image are of the same type or multiple types of defects; and the m defects contained in the above sample wafer image include overlapping defects and / or non-overlapping defects.

[0011] In some implementations, the aforementioned multi-defect segmentation model is a U-Net neural network model.

[0012] In some implementations, training the multi-defect segmentation model includes: inputting the sample wafer image into the multi-defect segmentation model to obtain multiple feature maps; determining the difference loss between the multiple feature maps and the label images of the multiple channels; and adjusting the parameters of the multi-defect segmentation model with the goal of minimizing the difference loss.

[0013] According to a second aspect, a method for segmenting a wafer defect image from a wafer image is provided, comprising: acquiring a target wafer image; inputting the target wafer image into a multi-defect segmentation model to obtain multiple feature maps output by the multi-defect segmentation model, wherein each feature map corresponds to a defect region; and determining a wafer defect image of the target wafer image based on the pixel values ​​of the multiple feature maps, wherein a single wafer defect image includes a single defect region.

[0014] According to a third aspect, an apparatus for training a multi-defect segmentation model for a wafer image is provided, comprising: an acquisition unit configured to acquire a sample wafer image containing m defects, and m defect regions labeled for the m defects respectively; a determination unit configured to determine a label image containing multiple channels based on the m defect regions, wherein the label image of a single channel indicates a single defect region; and a training unit configured to train the multi-defect segmentation model by taking the sample wafer image as input and the label image as the desired output.

[0015] According to a fourth aspect, an apparatus for segmenting a wafer defect image from a wafer image is provided, comprising: an image acquisition unit configured to acquire a target wafer image; an input unit configured to input the target wafer image into a multi-defect segmentation model obtained based on the method described in the first aspect, to obtain multiple feature maps output by the multi-defect segmentation model, wherein each feature map corresponds to a defect region; and a defect image determination unit configured to determine a wafer defect image of the target wafer image based on the pixel values ​​of the multiple feature maps, wherein a single wafer defect image includes a single defect region.

[0016] According to a fifth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in any implementation of the first or second aspect.

[0017] According to a sixth aspect, a computing device is provided, including a memory and a processor, characterized in that the memory stores executable code, and when the processor executes the executable code, it implements the method described in any implementation of the first or second aspect.

[0018] The method for training a multi-defect segmentation model for wafer images according to the embodiments of this specification first acquires a sample wafer image containing m defects, and m defect regions labeled for each of the m defects. Then, based on the m defect regions, a label image containing multiple channels is determined, wherein a label image for a single channel represents a single defect region. Finally, the sample wafer image is used as input, and the label image containing multiple channels is used as the desired output to train the multi-defect segmentation model for the wafer image. Because a multi-channel label image is introduced during training, a multi-defect segmentation model capable of outputting multiple wafer defect images can be trained, with each wafer defect image including a single defect region, thereby achieving the segmentation of multiple defects in the wafer image. Attached Figure Description

[0019] Figure 1 A schematic diagram is shown illustrating one application scenario in which the embodiments of this specification can be applied;

[0020] Figure 2 A flowchart illustrating a method for training a multi-defect segmentation model of wafer images according to one embodiment is shown.

[0021] Figure 3 Schematic diagrams of three wafer images (a), (b), and (c) are shown;

[0022] Figure 4 This diagram illustrates how a first defect region in a sample wafer image is mapped to a first channel label image.

[0023] Figure 5 A schematic diagram of the network structure of a U-Net neural network model is shown;

[0024] Figure 6 A flowchart illustrating a method for segmenting a wafer defect image from a wafer image according to one embodiment is shown;

[0025] Figure 7 A schematic block diagram of an apparatus for training a multi-defect segmentation model of wafer images according to one embodiment is shown;

[0026] Figure 8 A schematic block diagram of an apparatus for segmenting a wafer defect image from a wafer image according to one embodiment is shown. Detailed Implementation

[0027] The technical solutions provided in this specification will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this specification can be combined with each other.

[0028] As mentioned earlier, segmenting defect regions from wafer images is of great significance and value for defect detection and identification. Therefore, embodiments of this specification provide a method for training a multi-defect segmentation model for wafer images. Figure 1 This diagram illustrates an application scenario in which embodiments of this specification can be applied. For example... Figure 1 As shown, in Figure 1In the application scenario shown, a sample wafer image containing m defects can be acquired, along with m defect regions labeled for each defect, where m ≥ 0 and m is an integer. Depending on the specific needs, the sample wafer image can be preprocessed, such as image enhancement or scaling. Then, based on the m defect regions, a label image containing n channels can be determined, where n ≥ 0 and n ≥ m, and n is an integer. A single-channel label image represents a single defect region. The sample wafer image is then used as input to a multi-defect segmentation model, which outputs n feature maps. The label image containing n channels is taken as the desired output, and the difference between the n feature maps output by the multi-defect segmentation model and the label image containing n channels is calculated. The model parameters of the multi-defect segmentation model are then adjusted based on the calculated difference.

[0029] See also Figure 2 , Figure 2 A flowchart illustrating a method for training a multi-defect segmentation model of wafer images according to one embodiment is shown. It will be understood that this method can be executed by any device, apparatus, platform, or cluster of devices with computing and processing capabilities. Figure 2 As shown, the method for training a multi-defect segmentation model for wafer images may include the following steps:

[0030] Step 201: Obtain a sample wafer image containing m defects, and m defect regions labeled for each of the m defects.

[0031] In this embodiment, a sample wafer image containing m defects can be acquired, along with m defect regions labeled for each of the m defects, where m is a positive integer greater than or equal to 0. For example, a defect region can be manually labeled in the sample wafer image for each defect, marking its location. Here, defects can refer to various types of wafer surface defects that may occur, such as wafer surface redundancy, crystal defects, mechanical damage (scratch patterns), etc. Wafer surface redundancy can include nanoscale particles, micron-scale dust, residues from related processes, etc. Crystal defects are often caused by uneven heating during crystal growth; compared to other wafer surface defects, crystal defects on the wafer surface are more readily apparent. Mechanical damage generally occurs during wafer manufacturing processes such as polishing and slicing, caused by chemical mechanical abrasion.

[0032] In some implementations, the m defects contained in the sample wafer image can be of the same type or multiple types. Furthermore, the m defects in the sample wafer image can include overlapping defects, non-overlapping defects, or both. Here, an overlapping defect can refer to a defect whose portion overlaps with other defects. For example, such as... Figure 3 As shown, Figure 3 Schematic diagrams of three wafer images (a), (b), and (c) are shown, wherein wafer image (a) includes two overlapping defects, wafer image (b) includes one non-overlapping defect, and wafer image (c) includes two overlapping defects and one non-overlapping defect. It can be understood that... Figure 3 The boundary region in the image does not contain information about the wafer image; therefore, in this embodiment, the boundary region does not need to be processed. Figure 3 The examples shown are only used to explain the forms of overlapping and non-overlapping defects in wafer images, and are not intended to limit the number of defects, the size of defect regions, the number of overlapping defects, etc. in wafer images.

[0033] In some implementations, a sample wafer image containing m defects can be obtained as follows: First, obtain m sample defect images, where each sample defect image corresponds to one defect, and the m sample defect images are of the same size. Then, superimpose the m sample defect images to obtain the sample wafer image. As an example, the pixel values ​​of the defective regions in the sample defect images can be set to 1, and the pixel values ​​of the non-defective regions can be set to 0. Then, the pixel values ​​of pixels at the same position in the m sample defect images can be superimposed to obtain the sample wafer image containing m defects.

[0034] Step 202: Based on m defect regions, determine a label image containing multiple channels.

[0035] In this embodiment, a label image containing multiple channels can be determined based on m defect regions corresponding to m defects in a sample wafer image. Here, the number of channels in the label image is greater than or equal to m. As an example, the number of channels in the label image can be set manually based on experience. As another example, the number of channels in the label image can also be determined by statistical analysis of the number of defects contained in multiple historical multi-defect wafer images. For example, the maximum number of defects contained in multiple historical multi-defect wafer images can be counted, and this maximum value can be used as the number of channels in the label image. For yet another example, if it is statistically determined that X% of the wafer images in multiple historical multi-defect wafer images contain no more than Y defects, then Y can be used as the number of channels in the label image. Here, the specific values ​​of X and Y can be set according to actual needs.

[0036] Here, the sample wafer image and the label image are of equal size. A single channel of the label image represents a single defect region. For example, taking a label image with n channels (n≥m) as an example, firstly, based on the m defect regions marked in the sample wafer image, m defect images are generated, each containing one defect region. In this example, the defect images and label images are of equal size, N×M. As an example, when generating the defect images, the pixel value of the defect region in each defect image is 1, and the pixel value of the non-defect region is 0. Then, the m defect images are used as label images with m channels, and (nm) images with a pixel value of 0 and a size of N×M are used as label images with (nm) channels, thus obtaining a label image with n channels.

[0037] In some implementations, an initial label image containing multiple channels can be pre-set, and the label image and the sample wafer image can be the same size. Furthermore, step 202 above can also be performed as follows:

[0038] 1) For any first defect region among m defect regions, determine the first channel label image in multiple channels of the label image to correspond to the first defect region.

[0039] 2) Set the pixel value of the region corresponding to the first defect region in the first channel label image to the first pixel value.

[0040] 3) Set the pixel value of the region in the first channel label image that corresponds to the non-defect region corresponding to the first defect region as the second pixel value. The first pixel value and the second pixel value can be any different value. For example, the first pixel value can be 1, and the second pixel value can be 0.

[0041] For example, such as Figure 4 As shown, Figure 4 This diagram illustrates the mapping of a first defect region 4011 in sample wafer image 401 to a first channel label image 402. Figure 4 In the example shown, the sample wafer image 401 includes a first defect region 4011 and a non-defect region 4012 corresponding to the first defect region 4011. The pixel value of region 4021 in the first channel label image 402, corresponding to the first defect region 4011, is set to 1. The pixel value of region 4022 in the first channel label image 402, corresponding to the non-defect region 4012, is set to 0. Thus, information about the first defect region 4011 can be displayed in the first channel label image 402.

[0042] Similarly, a label image with m channels can represent m defect regions. Since the number of channels in a label image is greater than m, for a label image in a channel that does not correspond to a defect region, the pixel value of that channel can be set to the second pixel value. For example, suppose m is 3, and the m defect regions are region 1, region 2, and region 3. Suppose the label image has 4 channels, namely channel A, channel B, channel C, and channel D. And suppose channels A, B, and C represent region 1, region 2, and region 3 respectively. Then, since channel D does not correspond to a defect region, the pixel value of channel D is set to the second pixel value.

[0043] Step 203: Use the sample wafer image as input and the label image as the desired output to train the multi-defect segmentation model.

[0044] In this embodiment, a sample wafer image can be input into a multi-defect segmentation model to obtain a prediction result containing multiple channels output by the multi-defect segmentation model for the sample wafer image. Then, a preset loss function can be used to calculate the difference between this prediction result and a label image containing multiple channels. Based on the calculated difference, the model parameters of the multi-defect segmentation model are adjusted, and training ends when a preset training termination condition is met. For example, the preset training termination condition may include, but is not limited to, at least one of the following: training time exceeds a preset duration; the number of training iterations exceeds a preset number; the calculated difference is less than a preset difference threshold. Various implementation methods can be used to adjust the model parameters of the multi-defect segmentation model based on the calculated difference. For example, the BP (Back Propagation) algorithm or the SGD (Stochastic Gradient Descent) algorithm can be used to adjust the model parameters of the multi-defect segmentation model.

[0045] In some implementations, the method for training the multi-defect segmentation model described above can specifically include the following: First, the sample wafer image is input into the multi-defect segmentation model, which outputs multiple feature maps. Here, the number of feature maps output by the model is the same as the number of channels in the label image. Then, the difference loss between the multiple feature maps and the label image with multiple channels is determined; for example, the difference loss can be cross-entropy loss, mean squared error loss, mean absolute error loss, etc. Then, the model parameters of the multi-defect segmentation model are adjusted with the goal of minimizing the difference loss.

[0046] In some implementations, the multi-defect segmentation model can be various neural network models, such as the U-Net neural network model. U-Net is a convolutional neural network method. The name U-Net comes from its network structure, which, when visualized, resembles the letter U. Figure 5 As shown, Figure 5 This diagram illustrates the network structure of a U-Net neural network model. This structure, also known as an encoder-decoder structure, involves the encoder gradually compressing information into a low-dimensional representation, which the decoder then decodes back to the original image size. One of the key features of the U-Net architecture is skip connections. The network first performs convolutions and pooling on the image, resulting in a series of features of different sizes. Then, the smallest feature map is sequentially upsampled or deconvolved to obtain the next-level feature map, which is then concatenated with previous feature maps of the same level (this skip connection preserves more dimensional information, allowing subsequent layers to have both shallow and deep features). The concatenated feature map is then convolved and upsampled again. After multiple concatenations and upsampling, a prediction result with the same size as the input image is finally obtained. This can be understood as… Figure 5 The network structure of the U-Net architecture shown is merely illustrative and not a limitation on the network structure of the U-Net architecture. In practice, network structures of different depths can be set according to actual needs.

[0047] The above describes the training process of the multi-defect segmentation model. Because multi-channel labeled images are introduced during training, a multi-defect segmentation model capable of outputting multiple wafer defect images can be trained, with each wafer defect image containing a single defect region. The resulting multi-defect segmentation model can segment multiple defects in an input wafer image and output corresponding wafer defect images, each containing a single defect region.

[0048] based on Figure 2 The multi-defect segmentation model trained using the method shown can be used to segment multiple defects in a wafer image. See also... Figure 6 , Figure 6 A flowchart illustrating a method for segmenting a wafer defect image from a wafer image according to one embodiment is shown. It will be understood that this method can be executed by any device, apparatus, platform, or cluster of devices with computing and processing capabilities. Figure 6 The main body and execution of the method shown Figure 2 The main components of the methods shown can be the same or different. For example... Figure 6 As shown, the method for segmenting wafer defect images from wafer images may include the following steps:

[0049] Step 601: Obtain the target wafer image. Here, the target wafer image can refer to the image that needs to be segmented for defects. The target wafer image may include one or more defects, or it may not contain any defects.

[0050] Step 602, input the target wafer image based on Figure 2 The method shown yields a multi-defect segmentation model, which outputs multiple feature maps. Each feature map corresponds to a defect region.

[0051] Step 603: Determine the wafer defect image of the target wafer image based on the pixel values ​​of multiple feature images, wherein a single wafer defect image includes a single defect region.

[0052] For example, the multiple feature maps output by the multi-defect segmentation model can be feature maps processed by the sigmoid activation function. Therefore, the pixel values ​​in the feature maps range from (0,1). In this case, a threshold (e.g., 0.5) can be set to binarize each feature map, and feature maps with non-zero pixel values ​​can be used as wafer defect images. If, after binarization, all or almost all pixel values ​​of a feature map are 0, then that feature map cannot be used as a wafer defect image. Thus, if there is only one defect in the target wafer image, one wafer defect image is obtained. If there are m defects in the target wafer image, m wafer defect images are obtained, where m is a positive integer greater than 0, and m is less than or equal to the number of feature maps output by the multi-defect segmentation model.

[0053] Reviewing the above process, in the embodiments of this specification, since multi-channel labeled images are introduced during the training process, a multi-defect segmentation model that can output multiple wafer defect images can be trained, and each wafer defect image includes a single defect region, thereby achieving the segmentation of multiple defects in the wafer image through the multi-defect segmentation model.

[0054] According to another embodiment, an apparatus for training a multi-defect segmentation model of a wafer image is provided. The apparatus for training a multi-defect segmentation model of a wafer image can be deployed in any device, platform, or cluster of devices with computing and processing capabilities.

[0055] Figure 7 A schematic block diagram of an apparatus for training a multi-defect segmentation model of wafer images according to one embodiment is shown. Figure 7 As shown, the apparatus 700 for training a multi-defect segmentation model for wafer images includes: an acquisition unit 701 configured to acquire a sample wafer image containing m defects, and m defect regions labeled for the m defects respectively; a determination unit 702 configured to determine a label image containing multiple channels based on the m defect regions, wherein the label image of a single channel shows a single defect region; and a training unit 703 configured to train a multi-defect segmentation model by taking the sample wafer image as input and the label image as the desired output.

[0056] In some optional implementations of this embodiment, the determining unit 702 is further configured to: for the first defect region among the m defect regions, determine that the first channel label image among the multiple channels of the label image corresponds to the first defect region; set the pixel value of the region in the first channel label image corresponding to the first defect region as the first pixel value; set the pixel value of the region in the first channel label image corresponding to the non-defect region corresponding to the first defect region as the second pixel value, wherein the first pixel value and the second pixel value are different.

[0057] In some optional implementations of this embodiment, the first pixel value is 1 and the second pixel value is 0.

[0058] In some optional implementations of this embodiment, the acquisition unit 701 is further configured to: acquire m sample defect images, wherein each sample defect image corresponds to a defect; and superimpose the above m sample defect images to obtain the above sample wafer image.

[0059] In some optional implementations of this embodiment, the number of channels in the aforementioned label image is determined by statistical analysis of the number of defects contained in multiple historical multi-defect wafer images.

[0060] In some optional implementations of this embodiment, the m defects contained in the above sample wafer image are of the same type or multiple types of defects; and the m defects contained in the above sample wafer image include overlapping defects and / or non-overlapping defects.

[0061] In some optional implementations of this embodiment, the above-mentioned multi-defect segmentation model is a U-Net neural network model.

[0062] In some optional implementations of this embodiment, the training unit 703 is further configured to: input the above sample wafer image into the above multi-defect segmentation model to obtain multiple feature maps; determine the difference loss between the above multiple feature maps and the label images of the above multiple channels; and adjust the parameters of the above multi-defect segmentation model with the goal of minimizing the difference loss.

[0063] According to another embodiment, an apparatus for segmenting a wafer defect image from a wafer image is provided. The apparatus for segmenting a wafer defect image from a wafer image can be deployed in any device, platform, or cluster of devices with computing and processing capabilities.

[0064] Figure 8 A schematic block diagram of an apparatus for segmenting a wafer defect image from a wafer image according to one embodiment is shown. Figure 8As shown, the apparatus 800 for segmenting wafer defect images from wafer images includes: an image acquisition unit 801 configured to acquire a target wafer image; and an input unit 802 configured to input the target wafer image into a wafer image based on... Figure 2 The multi-defect segmentation model obtained by the method shown yields multiple feature maps output by the multi-defect segmentation model, wherein each feature map corresponds to a defect region; the defect image determination unit 803 is configured to determine the wafer defect image of the target wafer image based on the pixel values ​​of the multiple feature maps, wherein a single wafer defect image includes a single defect region.

[0065] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform... Figure 2 or Figure 6 The method described.

[0066] According to another embodiment, a computing device is also provided, including a memory and a processor, characterized in that the memory stores executable code, and when the processor executes the executable code, it implements... Figure 2 or Figure 6 The method described.

[0067] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0068] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, processor-executed software modules, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a multi-defect segmentation model for wafer images, comprising: Obtain a sample wafer image containing m defects, and m defect regions labeled for each of the m defects, wherein the m defects are of the same type or multiple types of defects; Based on the m defect regions, a label image containing multiple channels is determined, wherein a label image with a single channel represents a single defect region, the number of channels n of the label image is greater than or equal to m, and the sample wafer image and the label image are of equal size; and, the determination of the label image containing multiple channels based on the m defect regions includes: Based on the m defect regions marked in the sample wafer image, generate m defect images, wherein the pixel value of the defect region in each defect image is the first pixel value, and the pixel value of the non-defect region is the second pixel value. The m defect images are used as label images for m channels, and (nm) images with pixel values ​​of the second pixel value are used as label images for (nm) channels, resulting in n-channel label images; The sample wafer image is used as input, and the label image is used as the desired output to train a multi-defect segmentation model.

2. The method according to claim 1, wherein, The process of determining a label image containing multiple channels based on the m defect regions includes: For the first defect region among the m defect regions, determine the first channel label image among the multiple channels of the label image that corresponds to the first defect region; Set the pixel value of the region in the first channel label image that corresponds to the first defect region to the first pixel value; The pixel value of the region in the first channel label image that corresponds to the non-defect region corresponding to the first defect region is set as the second pixel value, wherein the first pixel value is different from the second pixel value.

3. The method according to claim 2, wherein, The first pixel value is 1, and the second pixel value is 0.

4. The method according to claim 1, wherein, The process of obtaining a sample wafer image containing m defects includes: Obtain m sample defect images, where each sample defect image corresponds to one defect; The m sample defect images are superimposed to obtain the sample wafer image.

5. The method according to claim 1, wherein, The number of channels in the label image is determined by statistical analysis of the number of defects contained in multiple historical multi-defect wafer images.

6. The method according to claim 1, wherein, The m defects contained in the sample wafer image are of the same type or multiple types of defects; and the m defects contained in the sample wafer image include overlapping defects and / or non-overlapping defects.

7. The method according to claim 1, wherein, The multi-defect segmentation model is a U-Net neural network model.

8. The method according to claim 1, wherein, The trained multi-defect segmentation model includes: The sample wafer image is input into the multi-defect segmentation model to obtain multiple feature maps; Determine the difference loss between the multiple feature maps and the label images of the multiple channels; The parameters of the multi-defect segmentation model are adjusted with the goal of minimizing the difference loss.

9. A method for segmenting a wafer defect image from a wafer image, comprising: Acquire the target wafer image; The target wafer image is input into a multi-defect segmentation model obtained based on any one of the methods in claims 1-8 to obtain multiple feature maps output by the multi-defect segmentation model, wherein each feature map corresponds to a defect region; Based on the pixel values ​​of the multiple feature maps, a wafer defect image of the target wafer image is determined, wherein a single wafer defect image includes a single defect region.

10. An apparatus for training a multi-defect segmentation model of a wafer image, comprising: The acquisition unit is configured to acquire a sample wafer image containing m defects, and m defect regions marked for each of the m defects, wherein the m defects are of the same type or multiple types of defects; The determining unit is configured to determine a label image containing multiple channels based on the m defect regions, wherein a label image with a single channel represents a single defect region, the number of channels n of the label image is greater than or equal to m, and the sample wafer image and the label image are of equal size; and the determination of the label image containing multiple channels based on the m defect regions includes: Based on the m defect regions marked in the sample wafer image, generate m defect images, wherein the pixel value of the defect region in each defect image is the first pixel value, and the pixel value of the non-defect region is the second pixel value. The m defect images are used as label images for m channels, and (nm) images with pixel values ​​of the second pixel value are used as label images for (nm) channels, resulting in n-channel label images; The training unit is configured to take the sample wafer image as input and the label image as the desired output to train a multi-defect segmentation model.

Images