A wafer image recognition method, device, equipment and application

Abstract

The present invention discloses a wafer image recognition method, device, equipment and application, which relate to the fields of semiconductor technology and image recognition technology. It includes traversing the wafer image to be denoised to obtain a set of defect points, and judging the defect type of the wafer image to be denoised. If it is of the Near-full or Random type, it is added to the image training set; if it is of the None type, DBSCAN clustering structure with a neighborhood radius of 2 is used for denoising; for other types, first, the OPTICS clustering structure with a neighborhood radius of 2 is used to obtain the reachable distance of each defect point in the defect point set of the wafer image to be denoised, calculate the average reachable distance, select a suitable clustering structure for denoising, and add the denoised wafer image to the image training set. The image training set of this application maximally removes the noise points on the surface of the wafer image, retains the complete defect features, greatly improves the accuracy of defect pattern classification of the wafer image, and promotes the further development of the field of wafer image defect recognition.

Description

Technical Field

[0001] The present invention relates to the fields of semiconductor technology and image recognition technology, and particularly to a method, device, equipment and application for wafer image recognition. Background Art

[0002] A wafer map is the result presented by electrical testing of wafers during the semiconductor production process. The defective wafers with failed tests constitute the defect features of the wafer map. Accurately identifying the defect patterns of the wafer map helps to locate the root causes of problems in the production process, thereby improving production efficiency and product yield. Traditional recognition of wafer map defect patterns is completed by professional engineers, which is a process consuming a large amount of time and huge manpower, and human subjective judgment often leads to errors. With the increase in the number of wafers produced, manual recognition can no longer meet the current needs. The development of machine learning has realized the automation of wafer map defect pattern recognition. Scholars use mathematical methods (such as Radon transform, Hough transform, etc.) to extract the features of the wafer map, and input the obtained features into machine learning models such as support vector machines and decision trees for training and prediction, which speeds up the recognition speed and improves the classification accuracy. With the development of hardware technology and the maturity of deep learning frameworks, deep learning has achieved outstanding results in the field of computer vision. Therefore, more and more scholars have begun to apply deep learning in the field of wafer maps, and deep learning has reached the highest level in the field of wafer map defect pattern recognition and classification.

[0003] The existing noise reduction methods for wafer maps can be divided into two categories. One is noise reduction based on spatial filtering, and the other is noise reduction based on density clustering. However, the parameter settings of the above methods are fixed and not suitable for all types of data. For a wafer map to be denoised, improper parameter settings will lead to poor denoising effects and even serious loss of defect features.

[0004] Such as Figure 1 shown, in the noise reduction based on spatial filtering, the most commonly used method is median filtering. Its specific implementation is to select a filtering window with an odd value size to coincide with the image, read the gray values of the overlapping part and sort them, and use the median after sorting as the gray value of the center point of the overlapping part. It can be obtained from Figure 1 that median filtering with this parameter setting will cause the loss of defect features of the Scratch type.

[0005] Such as Figure 2As shown, in noise reduction based on density clustering, the most commonly used is DBSCAN (Density-Based Spatial Clustering of Applications with Noise), which is a representative classical algorithm based on density clustering. It defines a cluster as the largest set of density-connected points and can divide regions with high density into clusters in a random noise space. It can be seen from the figure that although the wafer map after clustering and noise reduction does not damage its features, there are still a large number of noise points on the surface of the wafer map.

[0006] As can be seen from the above, how to find a noise reduction method that can not only remove a large number of noise points but also retain complete defect features is a problem to be solved at present. Summary of the Invention

[0007] The purpose of the present invention is to provide a wafer image recognition method to solve the problem that the noise reduction method in the existing wafer map recognition process cannot both remove a large number of noise points and retain complete defect features.

[0008] To solve the above technical problems, the present invention provides a wafer image recognition method, including:

[0009] S1. Traverse the wafer image to be noise-reduced to obtain a set of defect points;

[0010] S2. Determine the type of the wafer image to be noise-reduced;

[0011] S3. If it is of the Near-full or Random defect type, add the wafer image to be noise-reduced to the image training set;

[0012] S4. If it is of the None defect type, based on the set of defect points, use the DBSCAN clustering structure with a neighborhood radius of 2 for noise reduction, and add the noise-reduced wafer image to the image training set;

[0013] S5. Otherwise, use the OPTICS clustering structure with a neighborhood radius of 2 to obtain the reachability distance of each defect point in the set of defect points of the wafer image to be noise-reduced;

[0014] S6. Replace the values of the reachability distances that do not satisfy that there are no 3 sample points found within a neighborhood radius of 2 for the defect points with 0;

[0015] S7. Calculate the average reachability distance based on the reachability distances, compare the average reachability distance with 1. If the average reachability distance is greater than or equal to 1, set the neighborhood radius to 1, otherwise set the neighborhood radius to ;

[0016] S8. Determine the type of the wafer image to be denoised. If it is of the Scratch or Edge-ring defect type, use the DBSCAN clustering structure corresponding to the neighborhood radius in S7 for denoising, and add the denoised wafer image to the image training set;

[0017] S9. Otherwise, use the OPTICS clustering structure corresponding to the neighborhood radius in S7 for denoising, and add the denoised wafer image to the image training set;

[0018] S10. Use the image training set to train the deep model to obtain a trained wafer image training model.

[0019] Preferably, the traversing of the wafer image to be denoised to obtain the defect point set includes:

[0020] Locate the wafer image to be denoised in the Cartesian coordinate system to form a two-dimensional matrix;

[0021] Traverse the two-dimensional matrix, traversing from top to bottom and then from left to right in sequence, record the coordinates of the defect points to obtain the defect point set.

[0022] Preferably, the denoising using the DBSCAN clustering structure with a neighborhood radius of 2 includes:

[0023] Input the wafer image to be denoised into the DBSCAN clustering structure to obtain multiple clusters;

[0024] Obtain the defect points that are not clustered into any cluster and mark them as noise points;

[0025] Based on the noise points, change the corresponding defect values to 1 to obtain the denoised wafer image.

[0026] Preferably, the denoising using the OPTICS clustering structure corresponding to the neighborhood radius in S7 includes:

[0027] Input the wafer image to be denoised into the OPTICS clustering structure to obtain multiple clusters;

[0028] Obtain the defect points that are not clustered into any cluster and mark them as noise points;

[0029] Based on the noise points, change the corresponding defect values to 1 to obtain the denoised wafer image.

[0030] Preferably, the formula for calculating the average reachability distance based on the reachability distance is:

[0031]

[0032] where, is the average reachability distance, is the number of defect points, , , is the reachable distance of the defect point.

[0033] Preferably, training the deep model using the image training set to obtain a wafer image training model that has completed training includes:

[0034] Forward training the model parameters using a convolutional neural network. During the training process, the BatchNormalization strategy is used to prevent the model from overfitting, the Relu activation function is used to prevent gradient disappearance, the fully connected layer is used to classify the defect types of the wafer map, the cross-entropy loss function is used to calculate the error, and the Adam optimizer is used to perform backpropagation on the model parameters until the loss converges.

[0035] Preferably, the calculation formula of the cross-entropy loss function is:

[0036]

[0037] where, is the number of categories, is the number of wafer images in a batch size, is the true value of the wafer image, is the predicted probability of the deep model, is the wafer image category, is the th wafer image.

[0038] The present invention also provides a wafer image recognition device, including:

[0039] A traversal module for traversing the wafer image to be denoised to obtain a defect point set;

[0040] An image judgment module for judging the type of the wafer image to be denoised;

[0041] A first defect module, if it is of the Near-full or Random defect type, adding the wafer image to be denoised to the image training set;

[0042] A second defect module, if it is of the None defect type, based on the defect point set, using a DBSCAN clustering structure with a neighborhood radius of 2 for denoising, and adding the denoised wafer image to the image training set;

[0043] A first denoising module for using an OPTICS clustering structure with a neighborhood radius of 2 to obtain the reachable distance of each defect point in the defect point set of the wafer image to be denoised;

[0044] An accessible distance replacement module, configured to replace the value of the accessible distance that does not satisfy that 3 sample points cannot be found within the neighborhood with a radius of 2 with 0;

[0045] A neighborhood radius reset module, configured to calculate the average accessible distance based on the accessible distance, compare the size of the average accessible distance with 1, if the average accessible distance is greater than or equal to 1, set the neighborhood radius to 1, otherwise set the neighborhood radius to ;

[0046] A second noise reduction module, configured to determine the type of the wafer image to be denoised, if it is a Scratch or Edge-ring defect type, use the DBSCAN clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module to perform noise reduction, and add the denoised wafer image to the image training set;

[0047] A third noise reduction module, which uses the OPTICS clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module to perform noise reduction, and adds the denoised wafer image to the image training set;

[0048] A training module, which uses the image training set to train a deep model to obtain a trained wafer image training model.

[0049] The present invention also provides a wafer image recognition device, including:

[0050] A memory, configured to store a computer program;

[0051] A processor, configured to implement the steps of a wafer image recognition method as described above when executing the computer program.

[0052] The present invention also provides an application of the wafer image recognition method as described above in the field of semiconductor technology.

[0053] A wafer image recognition method provided by the present invention performs noise reduction by introducing a DBSCAN clustering structure and an OPTICS clustering structure with a neighborhood radius of 2, calculates the neighborhood radius, maximally removes the noise points on the surface of the wafer image, and at the same time retains the complete defect features. By calculating the average accessible distance, the distribution characteristics of the defect density of the wafer image are considered, greatly improving the accuracy of the defect mode classification of the wafer image, and realizing noise reduction that can both remove a large number of noise points and retain the complete defect features. Description of the Drawings

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings;

[0055] Figure 1 It is the noise reduction effect diagram of the Scratch type of the wafer map using median filtering (3*3 filtering window);

[0056] Figure 2 It is the noise reduction effect diagram of the Center defect type using the DBSCAN (neighborhood radius = 2, minimum number of sample points = 3) clustering algorithm;

[0057] Figure 3 It is the flowchart of the first specific embodiment of the wafer image recognition method provided by the present invention;

[0058] Figure 4 It is the schematic diagram of the Center defect type;

[0059] Figure 5 It is the reachable distance distribution diagram of the Center defect point set;

[0060] Figure 6 It is the Center defect type diagram after noise reduction;

[0061] Figure 7 It is the structural block diagram of a wafer image recognition device provided by an embodiment of the present invention. Detailed implementation manners

[0062] The core of the present invention is to provide a wafer image recognition method, device, equipment and application, realizing a noise reduction method that can both remove a large number of noise points and retain complete defect features in the process of wafer image recognition, and improving the recognition accuracy of the wafer map.

[0063] In order to enable those skilled in the art to better understand the solution of the present invention, the following will further elaborate on the present invention in conjunction with the drawings and specific implementation manners. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0064] Please refer to Figure 3 , Figure 3 It is the flowchart of the specific embodiment of the wafer image recognition method provided by the present invention; the specific operation steps are as follows:

[0065] Step S1: Traverse the wafer image to be denoised to obtain a set of defect points;

[0066] Locate the wafer image to be denoised in the Cartesian coordinate system to form a two-dimensional matrix;

[0067] Traverse the two-dimensional matrix, traversing from top to bottom and from left to right in sequence, record the coordinates of the defect points to obtain a set of defect points.

[0068] Step S2: Determine the type of the wafer image to be denoised;

[0069] Step S3: If it is of the Near-full or Random defect type, add the wafer image to be denoised to the image training set;

[0070] Step S4: If it is of the None defect type, based on the set of defect points, use the DBSCAN clustering structure with a neighborhood radius of 2 to denoise, and add the denoised wafer image to the image training set;

[0071] Step S5: Otherwise, use the OPTICS clustering structure with a neighborhood radius of 2 to obtain the reachability distance of each defect point in the set of defect points of the wafer image to be denoised;

[0072] Among them, the denoising using the DBSCAN clustering structure with a neighborhood radius of 2 includes:

[0073] Input the wafer image to be denoised into the DBSCAN clustering structure to obtain multiple clusters;

[0074] Obtain the defect points that are not clustered into any cluster and mark them as noise points;

[0075] Based on the noise points, change the corresponding defect value to 1 to obtain the denoised wafer image.

[0076] Step 6: Replace the values of the reachability distances that do not satisfy that there are no 3 sample points found within a neighborhood radius of 2 for the defect points with 0;

[0077] Step S7. Calculate the average reachability distance based on the reachability distances, compare the average reachability distance with 1. If the average reachability distance is greater than or equal to 1, set the neighborhood radius to 1, otherwise set the neighborhood radius to ;

[0078] The formula for calculating the average reachability distance based on the reachability distances is:

[0079]

[0080] Among them, is the average reachability distance, is the number of defect points, , , is the reachable distance of the defect point.

[0081] Step S8: Determine the type of the wafer image to be denoised. If it is of the Scratch or Edge-ring defect type, use the DBSCAN clustering structure corresponding to the neighborhood radius in S7 for denoising, and add the denoised wafer image to the image training set;

[0082] Step S9: Otherwise, use the OPTICS clustering structure corresponding to the neighborhood radius in S7 for denoising, and add the denoised wafer image to the image training set;

[0083] The denoising using the OPTICS clustering structure corresponding to the neighborhood radius in S7 includes:

[0084] Input the wafer image to be denoised into the OPTICS clustering structure to obtain multiple clustering clusters;

[0085] Obtain the defect points that are not clustered into any cluster and mark them as noise points;

[0086] Based on the noise points, change the corresponding defect values to 1 to obtain the denoised wafer image.

[0087] Step S10: Use the image training set to train the deep model to obtain the trained wafer image training model;

[0088] Use the convolutional neural network to train the model parameters forward. During the training process, use the BatchNormalization strategy to prevent the model from overfitting, use the Relu activation function to prevent gradient disappearance, use the fully connected layer to classify the wafer map defect types, use the cross-entropy loss function to calculate the error, and use the Adam optimizer to perform backpropagation on the model parameters until the loss converges;

[0089] The calculation formula of the cross-entropy loss function is:

[0090]

[0091] where, is the number of classes, is the number of wafer images in a batch size, is the true value of the wafer image, is the prediction probability of the deep model, is the wafer image class, is the th wafer image.

[0092] This embodiment provides a method for wafer image recognition, which maximally removes the noise points on the surface of the wafer image while preserving the complete defect features. The denoised dataset promotes the learning of the deep model. Compared with using the traditional fixed-parameter denoising method, the accuracy of the defect pattern classification of the wafer image is greatly improved, which helps engineers quickly locate the root cause of the problem, thereby improving the production efficiency and product yield of wafer manufacturing, and achieving denoising that can remove a large number of noise points and retain the complete defect features.

[0093] Based on the above embodiment, as Figure 4 shown, this embodiment selects the Center defect type as an example for expansion and description, as follows:

[0094] Step 201: Input the defect type (Center) of the wafer image to be denoised. If it does not belong to the Near-full or Random type, continue to execute Step 202;

[0095] Step 202: Locate the wafer image to be denoised in the Cartesian coordinate system and regard it as a two-dimensional matrix;

[0096] Step 203: Traverse the pixel values of the wafer image, traverse the rows from top to bottom in sequence, and traverse the columns from left to right, record the coordinates of the defect points, and form a defect point set;

[0097] Step 204: Convert the defect point set into a two-dimensional numpy.array format;

[0098] Step 205: If the type does not belong to the None defect type, execute Step 206;

[0099] Step 206: Use the OPTICS clustering structure (neighborhood radius = 2, minimum number of sample points = 3) to obtain the reachability distance of each defect point in the defect point set;

[0100] Step 207: Replace the value with UNDEFINED (the defect point cannot find 3 defect points in the neighborhood with a radius of 2) with 0;

[0101] Step 208: Accumulate the reachability distances of all defect points, and then divide by the number of defect points to calculate the average reachability distance (MeanRD);

[0102] Step 209: As Figure 5 shown, using the reachability distance distribution diagram of the Center defect point set, compare the size of MeanRD with 1. It can be obtained from the figure that the MeanRD of this wafer image defect point set is 1.191, and its value is greater than 1. Therefore, the neighborhood radius parameter for the following steps is set to 1;

[0103] Step 210: Determine that the Center defect type does not belong to the Scratch or Edge-ring type, and use the OPTICS clustering structure (neighborhood radius = 1, minimum number of sample points = 3) for noise reduction;

[0104] Traverse all the wafer images in the wafer image dataset in sequence, perform noise reduction according to steps 201 to 210, and finally form a noise-reduced image training set;

[0105] As Figure 6 shown, Figure 6 it is a Center defect type map after noise reduction.

[0106] Step 211: Use the noise-reduced image training set to train the deep model to obtain a trained wafer image training model.

[0107] This embodiment provides a wafer recognition method. The dataset after noise reduction promotes the learning of the deep model. Through multiple experiments, it is proved that the accuracy of using this method on the validation set and the test set is as high as 98.50%. Compared with using the traditional fixed-parameter noise reduction method, the classification accuracy is improved by nearly 2%, greatly improving the accuracy of wafer image defect pattern classification, maximizing the removal of noise points on the surface of the wafer image, while retaining the complete defect features, and realizing noise reduction that can both remove a large number of noise points and retain the complete defect features.

[0108] Please refer to Figure 7 , Figure 7 which is a structural block diagram of a wafer image recognition device provided by an embodiment of the present invention; the specific device may include:

[0109] A traversal module for traversing the wafer image to be noise-reduced to obtain a set of defect points;

[0110] An image judgment module for judging the type of the wafer image to be noise-reduced;

[0111] A first defect module, if it is a Near-full or Random defect type, add the wafer image to be noise-reduced to the image training set;

[0112] A second defect module, if it is a None defect type, based on the set of defect points, use the DBSCAN clustering structure with a neighborhood radius of 2 for noise reduction, and add the noise-reduced wafer image to the image training set;

[0113] A first noise reduction module for obtaining the reachable distance of each defect point in the defect point set of the wafer image to be noise-reduced by using the OPTICS clustering structure with a neighborhood radius of 2;

[0114] A reachable distance replacement module, configured to replace the value of the reachable distance that does not satisfy the condition that 3 sample points cannot be found within the neighborhood with a radius of 2 with 0;

[0115] A neighborhood radius reset module, configured to calculate the average reachable distance based on the reachable distance, compare the average reachable distance with 1, if the average reachable distance is greater than or equal to 1, set the neighborhood radius to 1, otherwise set the neighborhood radius to ;

[0116] A second noise reduction module, configured to determine the type of the wafer image to be denoised, if it is of the Scratch or Edge-ring defect type, use the DBSCAN clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module to perform noise reduction, and add the denoised wafer image to the image training set;

[0117] A third noise reduction module, which uses the OPTICS clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module to perform noise reduction, and adds the denoised wafer image to the image training set;

[0118] A training module, which uses the image training set to train a deep model to obtain a trained wafer image training model.

[0119] A wafer image recognition device in this embodiment is used to implement the foregoing wafer image recognition method. Therefore, the specific implementation manners in a wafer image recognition device can be seen in the embodiment part of the foregoing wafer image recognition method. For example, the traversal module 100, the image model judgment module 200, the reachable distance replacement module 300, the neighborhood radius reset module 400, the noise reduction module 500, and the training module 600 are respectively used to implement steps S1, S2, S3, S4, S5, S6, S7, S8, S9, and S10 in the foregoing wafer image recognition method. Therefore, its specific implementation manners can refer to the descriptions of the corresponding parts of each embodiment and will not be elaborated here.

[0120] A specific embodiment of the present invention further provides a wafer image recognition device, including: a memory, configured to store a computer program; a processor, configured to implement the steps of the foregoing wafer image recognition method when executing the computer program.

[0121] A specific embodiment of the present invention further provides an application of the foregoing wafer image recognition method in the field of semiconductor technology.

[0122] The various embodiments in this specification are described in a progressive manner. The key points of each embodiment are the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the description of the method part.

[0123] Those skilled in the art may further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0124] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of the two. The software modules can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0125] The above has introduced in detail a wafer image recognition method, device, equipment, and application provided by the present invention. Specific examples are used herein to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A wafer image recognition method, characterized in that, Including: S1. Traverse the wafer image to be denoised to obtain a set of defect points; S2. Determine the type of the wafer image to be denoised; S3. If it is of the Near-full or Random defect type, add the wafer image to be denoised to the image training set; S4. If it is of the None defect type, based on the set of defect points, use a DBSCAN clustering structure with a neighborhood radius of 2 to perform denoising, and add the denoised wafer image to the image training set; S5. Otherwise, use an OPTICS clustering structure with a neighborhood radius of 2 to obtain the reachability distance of each defect point in the set of defect points of the wafer image to be denoised; S6. Replace the values of the reachability distances that do not satisfy that there are no 3 sample points in the neighborhood with a radius of 2 with 0; S7. Calculate the average reachable distance based on the reachable distances, compare the average reachable distance with 1. If the average reachable distance is greater than or equal to 1, set the neighborhood radius to 1; otherwise, set the neighborhood radius to ; S8. Determine the type of the wafer image to be denoised. If it is of the Scratch or Edge-ring defect type, use a DBSCAN clustering structure corresponding to S7 in terms of neighborhood radius to perform denoising, and add the denoised wafer image to the image training set; S9. Otherwise, use an OPTICS clustering structure corresponding to S7 in terms of neighborhood radius to perform denoising, and add the denoised wafer image to the image training set; S10. Use the image training set to train a deep model to obtain a trained wafer image training model.

2. The wafer map recognition method according to claim 1, characterized in that, The traversing the wafer image to be denoised to obtain a set of defect points includes: Locate the wafer image to be denoised in the Cartesian coordinate system to form a two-dimensional matrix; Traverse the two-dimensional matrix, traversing from top to bottom and then from left to right in sequence, and record the coordinates of the defect points to obtain a set of defect points.

3. The wafer map recognition method according to claim 1, characterized in that, The using a DBSCAN clustering structure with a neighborhood radius of 2 to perform denoising includes: Input the wafer image to be denoised into the DBSCAN clustering structure to obtain multiple clusters; Obtain the defect points that are not clustered into any cluster and mark them as noise points; Based on the noise points, change the corresponding defect values to 1 to obtain a denoised wafer image.

4. The wafer map recognition method according to claim 1, characterized in that, The using an OPTICS clustering structure corresponding to S7 in terms of neighborhood radius to perform denoising includes: Input the wafer image to be denoised into the OPTICS clustering structure to obtain multiple clusters; Obtain the defect points that are not clustered into any cluster and mark them as noise points; Based on the noise points, change the corresponding defect values to 1 to obtain a denoised wafer image.

5. The wafer map recognition method according to claim 1, characterized in that, The formula for calculating the average reachability distance based on the reachability distance is: Among them, is the average accessible distance, is the number of defect points, , , is the accessible distance of the defect point.

6. The wafer image recognition method according to claim 1, characterized in that, The using the image training set to train a deep model to obtain a trained wafer image training model includes: Use a convolutional neural network to train the model parameters forward. During the training process, use the BatchNormalization strategy to prevent the model from overfitting, use the Relu activation function to prevent gradient disappearance, use a fully connected layer to classify the defect types of the wafer images, use the cross-entropy loss function to calculate the error, and use the Adam optimizer to perform backpropagation on the model parameters until the loss converges.

7. The wafer image recognition method according to claim 6, characterized in that, The formula for the cross-entropy loss function is: Among them, is the number of categories, is the number of wafer images in a batch size, is the ground truth of the wafer image, is the predicted probability of the deep model, is the wafer image category, is the th wafer image.

8. A wafer image recognition device, characterized in that, Including: A traversing module for traversing the wafer image to be denoised to obtain a set of defect points; An image judgment module for judging the type of the wafer image to be denoised; The first defect module, if it is of the Near-full or Random defect type, adds the wafer image to be denoised to the image training set; The second defect module, if it is of the None defect type, denoises based on the defect point set using a DBSCAN clustering structure with a neighborhood radius of 2, and adds the denoised wafer image to the image training set; The first denoising module is used to obtain the reachable distance of each defect point in the defect point set of the wafer image to be denoised using an OPTICS clustering structure with a neighborhood radius of 2; The reachable distance replacement module is used to replace the value of the reachable distance that does not satisfy that there are no 3 sample points found within a neighborhood with a radius of 2 with 0; A neighborhood radius reset module is used to calculate an average reachable distance based on the reachable distances, compare the average reachable distance with 1. If the average reachable distance is greater than or equal to 1, the neighborhood radius is set to 1; otherwise, the neighborhood radius is set to ; The second denoising module is used to determine the type of the wafer image to be denoised. If it is of the Scratch or Edge-ring defect type, it denoises using a DBSCAN clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module, and adds the denoised wafer image to the image training set; The third denoising module denoises using an OPTICS clustering structure corresponding to the neighborhood radius and the neighborhood radius reset module, and adds the denoised wafer image to the image training set; The training module uses the image training set to train a deep model to obtain a wafer image training model after the training is completed.

9. A wafer image recognition device, characterized in that, Including: A memory for storing a computer program; A processor, when executing the computer program, implements the steps of the method for identifying a wafer image according to any one of claims 1 to 7.

10. An application of the wafer image recognition method according to any one of claims 1-7 in the field of semiconductor technology.

Images