A method, system, device and medium for identifying and classifying silicon wafer surface defects

Abstract

The present invention provides a method, system, device and medium for identifying and classifying silicon wafer surface defects, which belongs to the field of semiconductors and includes a method for identifying and classifying silicon wafer surface defects. The method is based on machine vision and the assistance of traditional image processing algorithms. By grayscale processing and defect feature image identification of the silicon wafer surface image, silicon wafer defects are classified into five types according to the degree of harm: edge collapse, cracks, scratches, dirt, and shallow line marks. At the same time, scratches are divided into shallow scratches and deep scratches. Through the improvement of the deep learning classification model, the types of defects on the silicon wafer surface can be effectively identified and classified, and shallow scratches, dirt and shallow line mark silicon wafers in defective silicon wafers can be identified. The surface defects of the silicon wafer can be eliminated through a subsequent polishing process, thereby reducing the scrap rate of silicon wafers and improving the production capacity of the enterprise.

Description

Technical Field

[0001] The present invention belongs to the field of semiconductor technology, and in particular relates to a method, system, equipment and medium for identifying and classifying surface defects of silicon wafers. Background Art

[0002] Silicon wafer surface defect detection is a very important link in the semiconductor industry's industrial line, especially the defect detection of semiconductor-grade polished wafers, which has a great impact on entering the subsequent process stages.

[0003] At present, most semiconductor manufacturers use machine vision to detect semiconductor-level silicon wafer defects, find silicon wafers with abnormal surface grayscale values, and rely on manual re-inspection of defect types to diagnose whether the silicon wafers are good products. Some also use traditional image algorithms to perform grayscale analysis on silicon wafers, perform morphological analysis of defect features, determine defect types, obtain defect sample images, and then use CNN classification networks to classify defects into edge collapse, dirt, and scratches.

[0004] Although the method of defect classification through CNN classification network uses deep learning to classify samples, after a demand survey of a large number of silicon wafer manufacturers, it was found that the CNN network's classification of silicon wafer defects was not accurate enough and there were misclassifications. The defects were only divided into three types: edge chipping, dirt and scratches, which greatly increased the scrap rate of silicon wafers, reduced silicon wafer production capacity, and affected economic benefits. Summary of the invention

[0005] In order to solve the problem that traditional image processing is not accurate enough in classifying silicon wafer defects, the present invention provides a method, system, device and medium for identifying and classifying silicon wafer surface defects.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for identifying and classifying silicon wafer surface defects comprises the following steps:

[0008] Acquire the front and back images of the silicon wafer and the edge image of the silicon wafer;

[0009] Inputting the front and back images of the silicon wafer and the edge images of the silicon wafer into a deep learning classification model, and outputting the surface defect types of the silicon wafer; the surface defect types of the silicon wafer are divided into edge collapse, cracks, scratches, dirt and shallow line marks according to the degree of damage;

[0010] The deep learning classification model includes an improved ResNet34 network and an improved Unet semantic segmentation model; the improved ResNet34 network is used to process the front and back images of silicon wafers, the improved ResNet34 network includes a ResNet34 benchmark network, the traditional convolution part in the ResNet34 benchmark network is replaced by deep DW convolution and GS convolution, and the EMA attention mechanism is introduced into the ResNet34 benchmark network; the improved Unet semantic segmentation model includes a Unet semantic segmentation benchmark model, and residual connection and CBAM attention mechanism are added to the Unet semantic segmentation benchmark model.

[0011] Preferably, the scratches are divided into shallow scratches and deep scratches according to the scratch depth. Scratches with straight dotted lines or intermittent light lines are shallow scratches, and scratches with straight black lines are deep scratches.

[0012] Preferably, before inputting the front and back images of the silicon wafer and the edge images of the silicon wafer into the deep learning classification model, it also includes constructing a training set, and training the deep learning classification model through the training set, wherein obtaining the training set includes:

[0013] Scan the front and back sides of the silicon wafer to obtain an image of the silicon wafer with surface defects and an image of a normal silicon wafer, divide the image of the silicon wafer with surface defects into multiple area blocks, and obtain defect area images of set sizes in the multiple area blocks as image one.

[0014] A side image of a silicon wafer with edge defects is obtained, and the side image of the silicon wafer is cut according to a set size as image 2.

[0015] Grayscale processing is performed on the normal silicon wafer image, image one and image two to obtain the grayscale interval of the normal silicon wafer image and set the grayscale threshold range; multiple defect samples whose grayscale values ​​exceed the grayscale threshold range are obtained from image one, and the multiple defect samples are used as training sets.

[0016] Preferably, the grayscale processing is performed by a binarization method, specifically comprising the following steps:

[0017] Obtain the histograms of the normal silicon wafer image, image one, and image two, divide the pixels into different bins according to the grayscale range, and count the number of pixels in each bin.

[0018] Normalize the histogram so that the number of pixels in each bin is divided by the total number of pixels.

[0019] Iteratively calculate the proportion of pixels of different gray levels in the image and the average gray level.

[0020] Calculate the variance of foreground pixels and background pixels, and select the threshold that maximizes the variance as the global threshold.

[0021] Preferably, the step of acquiring a plurality of defect samples whose grayscale values ​​exceed a grayscale threshold range from the image 1 specifically includes:

[0022] Perform mean filtering on the image, perform edge detection on the filtered image, and divide the image into multiple regions by using a regiongrowing function, screen the multiple regions according to the grayscale mean, and extract the image region that meets the conditions;

[0023] For the obtained image area, the image features are extracted through morphological analysis methods, and the erosion operation is used to further smooth the area edges. The unconnected areas are merged through expansion to restore the parts lost due to erosion. The features of the defects are extracted through opening and closing operations to obtain a connected domain. The features of the connected domain are calculated and screened to remove non-defective areas. In the remaining connected domain, a fixed-size image is intercepted with the center of gravity as the center as the defect sample.

[0024] Preferably, the ResNet34 benchmark network consists of 34 convolutional layers, including 16 basic blocks, and the residual block consists of two 3×3 convolutional layers and a skip connection.

[0025] Preferably, the improved Unet semantic segmentation model sets the loss function to a DICE+BCE composite loss function.

[0026] The present invention also provides a system for identifying and classifying silicon wafer surface defects, comprising:

[0027] The image acquisition module is used to acquire the front and back images of the silicon wafer and the edge image of the silicon wafer.

[0028] A deep learning module is used to establish a deep learning classification model including an improved ResNet34 network and an improved Unet semantic segmentation model; the improved ResNet34 network is used to process the front and back images of silicon wafers, the improved ResNet34 network includes a ResNet34 benchmark network, the traditional convolution part in the ResNet34 benchmark network is replaced by deep DW convolution and GS convolution, and the EMA attention mechanism is introduced into the ResNet34 benchmark network; the improved Unet semantic segmentation model includes a Unet semantic segmentation benchmark model, and residual connection and CBAM attention mechanism are added to the Unet semantic segmentation benchmark model.

[0029] The classification and recognition module is used to input the front and back images of the silicon wafer and the edge images of the silicon wafer into the deep learning classification model and output the surface defect type of the silicon wafer.

[0030] The present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the method for identifying and classifying surface defects of silicon wafers.

[0031] The present invention also provides a computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is loaded by a processor, it can execute the steps described in the method for identifying and classifying silicon wafer surface defects.

[0032] The present invention provides a method, system, device and medium for identifying and classifying silicon wafer surface defects, which have the following beneficial effects:

[0033] The present invention proposes a method for identifying and classifying surface defects of silicon wafers. The front and back images of silicon wafers and the edge images of silicon wafers are respectively input into an improved ResNet34 network and an improved Unet semantic segmentation model. The improved ResNet34 network is replaced with deep DW convolution and GS convolution, which reduces the number of model parameters and the amount of calculation, and the EMA attention mechanism is added to improve the model performance; residual connection and CBAM attention mechanism are added to the Unet semantic segmentation benchmark model to obtain an improved Unet semantic segmentation model. After the input image is processed by the model, the surface defect type of the silicon wafer is output. Silicon wafer defects are divided into five types according to the degree of damage: edge collapse, cracks, scratches, dirt, and shallow line marks. For small defects, small defect silicon wafers that are restored to good products after subsequent process treatment can reduce the scrap rate of silicon wafers and improve the production capacity and benefits of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 The present invention is a flow chart of a method for identifying and classifying silicon wafer surface defects. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.

[0037] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0038] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.

[0039] Example

[0040] The present invention provides a method for identifying and classifying silicon wafer surface defects. Figure 1 As shown, specifically including:

[0041] S1. Scan the front and back sides of the silicon wafer by using a fixed light source and a line scan camera to obtain images of the silicon wafer with surface defects and normal silicon wafer images. Divide the image of the silicon wafer with surface defects into six area blocks, and obtain a 155*155 defect area image in each area block as image one.

[0042] S2. For tiny defects occurring at the edge, a side image of the silicon wafer is obtained by using a ring light source and a line scan camera, and the side image of the silicon wafer is cropped to 768x10000 as image 2.

[0043] S3, grayscale processing is performed on the normal silicon wafer image, image 1 and image 2. Grayscale processing is the process of grayscale image binarization, which converts the grayscale in the image into binary (black and white). It usually involves selecting a threshold value, setting all pixel values ​​in the image greater than or equal to the threshold value to white (or 1), and setting all pixel values ​​less than the threshold value to black (or 0). Specifically, the following steps are included:

[0044] Calculate the histogram of the image, divide the pixels into different bins according to the grayscale range, and count the number of pixels in each bin.

[0045] Normalize the histogram so that the number of pixels in each bin is divided by the total number of pixels.

[0046] Iteratively calculate the proportion of pixels of different gray levels in the image and the average gray level.

[0047] Calculate the variance of foreground pixels and background pixels, and select the threshold that maximizes the variance as the global threshold.

[0048] Define the grayscale interval of a normal silicon wafer image as the grayscale threshold.

[0049] S4. Processing image 1 through an algorithm, identifying and locating defective areas, and obtaining defective samples on the surface of the problematic silicon wafer that exceed the grayscale threshold range.

[0050] Image 1 is processed by mean filtering, and the edge detection and screening are performed on the filtered image. The image is divided into multiple regions by the regiongrowing function, and multiple regions are screened according to the grayscale mean to extract the regions that meet the conditions.

[0051] Morphological analysis method processing is a mathematical method for image analysis and processing, mainly used for morphological feature extraction and image morphological operation. For example, in Matlab, some morphological processing functions can be used for image morphological operation. For the obtained image area, morphological analysis methods can be used to extract features from the image. The erosion_circle function can be used to erode the image to further smooth the edge of the area. The dilation_circle function is used for dilation operation to merge unconnected small areas to restore the parts lost due to erosion. The features of the defects can be extracted through opening and closing operations to obtain a connected domain. The features of the connected domain are calculated and screened to remove non-defective areas. In the remaining connected domain, a fixed-size image is intercepted with the center of gravity as the center as the defect sample. Obtain defect samples on the surface of the problem silicon wafer that exceed the grayscale threshold range.

[0052] The defect types of defective samples are divided into chipping, cracks, scratches, dirt and shallow line marks according to the degree of damage. Scratch defects are divided into shallow scratches and deep scratches according to the depth of the scratches. Lines with straight dotted lines or intermittent light lines are shallow scratches, and lines with black straight lines are deep scratches.

[0053] S5. Input the front and back images of the silicon wafer and the edge image of the silicon wafer into the deep learning classification model to identify and output the surface defect type of the silicon wafer.

[0054] The adopted ResNet34 model consists of 34 convolutional layers, including 16 basic blocks, and the residual block consists of two 3×3 convolutional layers and a skip connection. The specific network structure is as follows: The input layer of ResNet34 consists of a standard convolutional layer and 64 convolutional filters, with a total of 16 basic blocks, each of which consists of two convolutional layers and a cross-layer connection. The traditional convolution part in the baseline network model is replaced by deep convolution and GS convolution, and deep convolution performs convolution operations on each input channel separately. GS convolution can retain as many hidden connections between each channel as possible with lower time complexity. The two reduce the computational complexity and achieve more efficient results compared with standard convolution. The attention mechanism is introduced into the model. The attention mechanism still introduces two tensors, one is the output of the 1×1 branch, and the other is the output of the 3×3 branch. The global spatial information is encoded in the output of the 1×1 branch using two-dimensional global average pooling. Before the channel feature joint activation mechanism, directly converting the output of the smallest branch to the shape of the corresponding dimension can help the model better focus on the key areas in the image and improve the classification accuracy.

[0055] The specific processing steps of the improved ResNet34 model include:

[0056] Data augmentation,selected 155*155 grayscale images, processed the grayscale images by horizontal flipping, vertical flipping, and 90°, 180°, and 270° rotation, and expanded the defect images to 12150 images of each defect,component image dataset.

[0057] Divide the data set into the ratio of 10000:1000:1150 = training set: validation set: test set.

[0058] Load the dataset, use datasets.ImageFolder to load the training and validation datasets, and create a data loader.

[0059] Load the pre-trained ResNet34 model and modify the last layer to adapt to the new classification task. Define the loss function and optimizer using the cross entropy loss function and the Adam optimizer. In each epoch, perform the training and validation steps, record the loss and accuracy, and save the model when the best accuracy is achieved on the validation set. Save the model each time, and finally, plot the loss and accuracy curves during the training and validation process, and load the test set for prediction.

[0060] The grayscale processed image 2 is used as a sample, residual connections are added to the Unet semantic segmentation, the CNN network in the feature extraction part is replaced with the Resnet network, the CBAM attention mechanism is added to the network, and the loss function is set to the DICE+BCE composite loss function to obtain an improved Unet semantic segmentation model. The improved Unet semantic segmentation model is used to distinguish defect areas from background areas, the model marks defect contours and defect labels to make defect samples, and then the segmentation network model is trained to identify and classify silicon wafer edge defects.

[0061] S6. The deep learning classification model analyzes the silicon wafer images to be processed and classifies and identifies the surface defects of the silicon wafers.

[0062] The present invention also provides a system for identifying and classifying silicon wafer surface defects, comprising:

[0063] The image acquisition module is used to acquire the front and back images of the silicon wafer and the edge image of the silicon wafer.

[0064] A deep learning module is used to establish a deep learning classification model including an improved ResNet34 network and an improved Unet semantic segmentation model; the improved ResNet34 network is used to process the front and back images of silicon wafers, the improved ResNet34 network includes a ResNet34 benchmark network, the traditional convolution part in the ResNet34 network is replaced by deep DW convolution and GS convolution, and the EMA attention mechanism is introduced into the ResNet34 network; the improved Unet semantic segmentation model includes a Unet semantic segmentation benchmark model, and residual connections and CBAM attention mechanisms are added to the Unet semantic segmentation benchmark model.

[0065] The classification and recognition module is used to input the front and back images of the silicon wafer and the edge images of the silicon wafer into the deep learning classification model and output the surface defect type of the silicon wafer.

[0066] Each module in the above-mentioned system for identifying and classifying silicon wafer surface defects can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above modules.

[0067] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps in an embodiment of a method for identifying and classifying surface defects of a silicon wafer. The specific implementation method can be found in the method embodiment, which will not be described in detail here.

[0068] Furthermore, the present invention also provides a non-temporary computer-readable storage medium containing instructions, and a computer program is stored on the storage medium. For example, a memory containing instructions, the above instructions can be executed by a processor of a computer device to complete the above method. For example, a non-temporary computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device. When the computer program is executed by the processor, the steps in an embodiment of a method for identifying and classifying surface defects of a silicon wafer can be implemented. The specific implementation method can be found in the method embodiment, which will not be repeated here.

[0069] It will be appreciated by those skilled in the art that embodiments of the present invention may provide methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0070] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as a combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0071] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0072] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps for the functions specified in one or more boxes.

[0073] It should be pointed out that the specific implementation methods described above can enable those skilled in the art to understand the invention more comprehensively, but do not limit the invention in any way. Therefore, although the invention has been described in detail in this specification and embodiments, those skilled in the art should understand that the invention can still be modified or replaced by equivalents; and all technical solutions and improvements that do not deviate from the spirit and scope of the invention are included in the protection scope of the patent for the invention. Any figure mark in the claims should not be regarded as limiting the claims involved. Any simple change or equivalent replacement of the technical solution that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belongs to the protection scope of the present invention.

Claims

1. A method for identifying and classifying silicon wafer surface defects, characterized in that: The following steps are involved: Acquire the front and back images of the silicon wafer and the edge image of the silicon wafer; Inputting the front and back images of the silicon wafer and the edge images of the silicon wafer into a deep learning classification model, and outputting the surface defect types of the silicon wafer; the surface defect types of the silicon wafer are divided into edge collapse, cracks, scratches, dirt and shallow line marks according to the degree of damage; The deep learning classification model includes an improved ResNet34 network and an improved Unet semantic segmentation model; the improved ResNet34 network is used to process the front and back images of the silicon wafer, the improved ResNet34 network includes the ResNet34 benchmark network, the traditional convolution part in the ResNet34 benchmark network is replaced by deep DW convolution and GS convolution, and the EMA attention mechanism is introduced into the ResNet34 benchmark network; the improved Unet semantic segmentation model is used to process the edge image of the silicon wafer, the improved Unet semantic segmentation model includes the Unet semantic segmentation benchmark model, the residual is added to the Unet semantic segmentation benchmark model, the CNN network of the feature extraction part in the Unet semantic segmentation benchmark model is replaced by the Resnet network, the CBAM attention mechanism is added to the network, and the loss function is set to the DICE+BCE composite loss function.

2. A method for identifying and classifying silicon wafer surface defects according to claim 1, characterized in that: The scratches are divided into shallow scratches and deep scratches according to the depth of the scratches. The scratches with straight dotted lines or intermittent light lines are shallow scratches, and the scratches with straight black lines are deep scratches.

3. The method for identifying and classifying silicon wafer surface defects according to claim 1, characterized in that: Before inputting the front and back images of the silicon wafer and the edge images of the silicon wafer into the deep learning classification model, a training set is constructed to train the deep learning classification model through the training set, wherein obtaining the training set includes: Scan the front and back sides of the silicon wafer to obtain an image of the silicon wafer with surface defects and an image of a normal silicon wafer, divide the image of the silicon wafer with surface defects into a plurality of area blocks, and obtain defect area images of a set size in each of the plurality of area blocks as image 1; Acquire an edge image of a silicon wafer having edge defects, and cut the edge image of the silicon wafer according to a set size as image 2; Grayscale processing is performed on the normal silicon wafer image, image one and image two to obtain the grayscale interval of the normal silicon wafer image and set the grayscale threshold range; multiple defect samples whose grayscale values ​​exceed the grayscale threshold range are obtained from image one, and the multiple defect samples are used as training sets.

4. A method for identifying and classifying silicon wafer surface defects according to claim 3, characterized in that: The grayscale processing is performed by a binarization method, specifically comprising the following steps: Obtain the histograms of the normal silicon wafer image, image 1, and image 2, divide the pixels into different bins according to the grayscale range, and count the number of pixels in each bin; Normalize the histogram so that the number of pixels in each bin is divided by the total number of pixels; Iteratively calculate the proportion of pixels of different gray levels in the image and the average gray level; Calculate the variance of foreground pixels and background pixels, and select the threshold that maximizes the variance as the global threshold.

5. The method for identifying and classifying silicon wafer surface defects according to claim 3, characterized in that: The step of obtaining a plurality of defect samples whose grayscale values ​​exceed a grayscale threshold range from the image 1 specifically includes: Perform mean filtering on the image, perform edge detection on the filtered image, and divide the image into multiple regions by using a regiongrowing function, screen the multiple regions according to the grayscale mean, and extract the image region that meets the conditions; For the obtained image area, the image features are extracted through morphological analysis methods, and the erosion operation is used to further smooth the area edges. The unconnected areas are merged through expansion to restore the parts lost due to erosion. The features of the defects are extracted through opening and closing operations to obtain a connected domain. The features of the connected domain are calculated and screened to remove non-defective areas. In the remaining connected domain, a fixed-size image is intercepted with the center of gravity as the center as the defect sample.

6. A method for identifying and classifying silicon wafer surface defects according to claim 1, characterized in that: The ResNet34 baseline network consists of 34 convolutional layers, including 16 basic blocks, and the residual block consists of two 3×3 convolutional layers and a skip connection.

7. The method for identifying and classifying silicon wafer surface defects according to claim 1, characterized in that: The improved Unet semantic segmentation model sets the loss function as the DICE+BCE composite loss function.

8. A system for identifying and classifying silicon wafer surface defects, characterized in that: include: An image acquisition module, used to acquire front and back images of the silicon wafer and edge images of the silicon wafer; A deep learning module, for establishing a deep learning classification model including an improved ResNet34 network and an improved Unet semantic segmentation model; the improved ResNet34 network is used to process the front and back images of silicon wafers, the improved ResNet34 network includes a ResNet34 benchmark network, the traditional convolution part in the ResNet34 benchmark network is replaced by deep DW convolution and GS convolution, and an EMA attention mechanism is introduced into the ResNet34 benchmark network; the improved Unet semantic segmentation model is used to process the edge image of the silicon wafer, the improved Unet semantic segmentation model includes a Unet semantic segmentation benchmark model, a residual is added to the Unet semantic segmentation benchmark model, the CNN network of the feature extraction part in the Unet semantic segmentation benchmark model is replaced by a Resnet network, a CBAM attention mechanism is added to the network, and the loss function is set to a DICE+BCE composite loss function; The classification and recognition module is used to input the front and back images of the silicon wafer and the edge images of the silicon wafer into the deep learning classification model and output the surface defect type of the silicon wafer.

9. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is loaded into a processor, the computer program can execute the steps of the method according to any one of claims 1 to 7.

Images