Improved wafer graph identification method and system based on generative adversarial network
By optimizing the lightweight classification model and building a deep convolutional generation adversarial network model, the problems of poor recognition effect and data imbalance in wafer diagram fault pattern recognition are solved, and higher recognition accuracy and stronger model generalization capabilities are achieved.
Patent Information
- Application Number
- CN202510068125.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-16
- Publication Date
- 2025-06-06
AI Technical Summary
The prior art has the problem of poor recognition effect in wafer pattern recognition, especially in high feature dimensions and multi-noise environments, and the data imbalance problem is difficult to solve.
By optimizing the lightweight classification model, increasing the depth of the convolution layer and combining the global average pooling layer, a deep convolution generation adversarial network model is constructed, and the wafer graph recognition method is improved to solve the problem of data imbalance.
It improves the accuracy of wafer diagram recognition, reduces the computational complexity, enhances the generalization ability of the model, and improves the yield and reliability of wafer manufacturing.
Smart Images
Figure CN120107745A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of semiconductor technology, and in particular to an improved wafer image recognition method and system based on a generative adversarial network. Background Art
[0002] Since 1996, there have been many studies on wafer image fault pattern recognition. With the development of machine learning (ML), various machine learning-based recognizers have been used to identify random and systematic defects. Unlabeled unsupervised machine learning, such as adaptive resonance theory 1 (ART1) and K-means clustering, extracts features through similarity and clustering analysis; labeled supervised machine learning, such as support vector machine (SVM) and decision tree, has also been applied to wafer image fault pattern recognition. However, wafer images have high feature dimensions and much noise. The recognition effect of labeled supervised machine learning depends on the accuracy of manual labeling. Therefore, this type of method has great limitations and may even seriously affect the performance of the recognizer.
[0003] In recent years, deep learning research has achieved unprecedented development, and scholars have begun to try to solve the problem of wafer image fault pattern recognition based on the deep learning framework. Research on wafer image fault pattern recognition based on deep learning mainly involves the fields of wafer image feature clustering, feature extraction, wafer image data augmentation, and wafer image classification. The main applied networks include convolutional neural networks, generative adversarial networks, and autoencoders.
[0004] However, the recognition effect of convolutional neural networks is highly dependent on the quality of the dataset, and the multi-classification problem under unbalanced datasets remains challenging because the model tends to classify samples with minority classes as majority classes to minimize the loss function.
[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention
[0006] In response to the problems in the related art, the present invention proposes an improved wafer image recognition method and system based on a generative adversarial network to overcome the above-mentioned technical problems existing in the existing related art.
[0007] To this end, the specific technical solution adopted by the present invention is as follows:
[0008] According to one aspect of the present invention, a wafer image recognition method based on a generative adversarial network is provided, and the wafer image recognition method comprises the following steps:
[0009] S1. Use the pre-configured stacked convolutional layers to increase the depth of the convolutional layer of the lightweight classification model, and combine the global average pooling layer to optimize the lightweight classification model;
[0010] S2. Based on the optimized lightweight classification model, a convolutional neural network model is constructed, the wafer image is used as input, and the convolutional neural network model is trained;
[0011] S3. Based on the trained convolutional neural network model, the pre-configured generative adversarial network model is improved to construct a deep convolutional generative adversarial network model, and the wafer image is recognized through the deep convolutional generative adversarial network model.
[0012] Optionally, the method of optimizing the lightweight classification model based on the lightweight classification model by using stacked convolutional layers and shrinking the lightweight classification model in combination with a global average pooling layer includes the following steps:
[0013] S11. Use multiple convolutional layers to directly stack and increase the depth of the convolutional layer in the lightweight classification model.
[0014] S12. The feature vector is obtained by replacing the fully connected layer with the global average pooling layer to reduce the parameters in the light quantization classification model.
[0015] Optionally, the convolutional neural network model includes a convolution layer, a normalization layer, an activation layer, a maximum pooling layer, a global average pooling layer, a Softmax layer and a classification layer.
[0016] Optionally, constructing a convolutional neural network model based on the reduced lightweight classification model, and inputting the wafer image into the convolutional neural network model for training comprises the following steps:
[0017] S21, extracting feature maps based on the convolution layer and processing the feature maps;
[0018] S22, based on the processed feature map, optimizing the features in the feature map through the activation layer, and reducing the dimension of the feature map by using the maximum pooling layer;
[0019] S23, the reduced feature map re-enters other convolutional layers, and then enters the normalization layer, activation layer and maximum pooling layer in sequence;
[0020] S24. After repeating step S23 twice, the feature map enters the last convolution layer, and then enters the normalization layer, activation layer, and global average pooling layer in sequence, and is classified using the Softmax layer to obtain the output result.
[0021] Optionally, extracting the feature map based on the convolution layer and preprocessing the feature map includes the following steps:
[0022] S211, based on the convolutional neural network model, entering the convolution layer and using random calculation to extract features to obtain a feature map;
[0023] S212: Normalize based on the extracted feature map and standardize the statistical distribution of the feature map.
[0024] Optionally, the step of optimizing features in the feature map by an activation layer based on the processed feature map and reducing the dimension of the feature map by a maximum pooling layer comprises the following steps:
[0025] S221, the activation layer optimizes the features in the feature map using an activation function;
[0026] S222. Use a maximum pooling layer to reduce the dimension of the feature map, and the pooling window size in the maximum pooling layer is 2×2, and the step size is 2.
[0027] Optionally, the expression of the activation function is:
[0028]
[0029] In the formula, x represents the input parameter, α represents the coefficient, and LeakyReLU represents the activation function.
[0030] Optionally, the method of improving the generative adversarial network model based on the trained convolutional neural network model, constructing a deep convolutional generative adversarial network model, and recognizing the wafer image through the deep convolutional generative adversarial network model includes the following steps:
[0031] S31, a generator based on a deep convolutional generative adversarial network model, receiving a one-dimensional random Gaussian vector, passing through multiple deconvolution layers and performing normalization processing, to generate a forged image of a wafer image;
[0032] S32, the discriminator based on the deep convolutional generative adversarial network model, connects the activation function after each convolution layer, and uses the Sigmoid function to perform the classification task;
[0033] S33. Use the discriminator of the deep convolutional generative adversarial network model and optimize the generator by combining the forged image of the wafer image.
[0034] Optionally, the method of optimizing the generator by using the discriminator of the deep convolutional generative adversarial network model in combination with the forged image of the wafer image comprises the following steps:
[0035] S331, training a discriminator based on the wafer image and the forged image of the wafer image;
[0036] S332: If the discriminator identifies the forged image of the wafer image as a false sample, the weight of the generator is updated.
[0037] According to another aspect of the present invention, there is also provided an improved wafer image recognition system based on a generative adversarial network, the wafer image recognition system comprising a lightweight classification model optimization module, a convolutional neural network model construction module and a deep convolutional generative adversarial network model construction module;
[0038] The lightweight classification model optimization module is used to increase the depth of the convolution layer of the lightweight classification model by using the pre-configured stacked convolution layer, and optimize the lightweight classification model in combination with the global average pooling layer;
[0039] The convolutional neural network model building module is used to build a convolutional neural network model based on the optimized lightweight classification model, take the wafer image as input, and train the convolutional neural network model;
[0040] The deep convolutional generative adversarial network model construction module is used to improve the pre-configured generative adversarial network model based on the trained convolutional neural network model, construct a deep convolutional generative adversarial network model, and recognize wafer image through the deep convolutional generative adversarial network model.
[0041] The beneficial effects of the present invention are:
[0042] The present invention achieves the improvement of wafer image recognition accuracy, the reduction of computational complexity, and the enhancement of model generalization ability by optimizing the lightweight classification model and constructing a deep convolutional generative adversarial network, which is helpful to improve the wafer manufacturing yield and reliability and promote the continuous advancement of semiconductor manufacturing technology. The present invention adopts model modification and data sampling to solve the data imbalance problem in the prior art. The former improves the classification accuracy by adjusting the weights of misclassified data samples; the latter is used to adjust data samples according to the distribution of samples among different categories by using undersampling, oversampling or data enhancement. Among them, data enhancement accounts for a relatively high proportion in the current research. The data enhancement method can not only solve the data imbalance problem, but also prevent overfitting. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, 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 paying creative work.
[0044] Figure 1 is a flow chart of a wafer image recognition method improved based on a generative adversarial network according to an embodiment of the present invention;
[0045] Figure 2is a principle block diagram of a wafer image recognition system improved based on a generative adversarial network according to an embodiment of the present invention;
[0046] Figure 3 is a structural schematic diagram of a deep convolutional neural network model in an improved wafer image recognition method based on a generative adversarial network according to an embodiment of the present invention;
[0047] Figure 4 is a schematic diagram of the structure of a deep convolutional generative adversarial network in an improved wafer image recognition method based on a generative adversarial network according to an embodiment of the present invention;
[0048] Figure 5 is an iterative process diagram of the accuracy of the DCNN model in the improved wafer image recognition method based on the generative adversarial network according to an embodiment of the present invention;
[0049] Figure 6 is an iterative process diagram of the loss function of the DCNN model in the improved wafer image recognition method based on the generative adversarial network according to an embodiment of the present invention;
[0050] Figure 7 1 is a schematic diagram of a confusion matrix structure of a DCNN model training set classification in a wafer image recognition method improved based on a generative adversarial network according to an embodiment of the present invention;
[0051] Figure 8 It is a schematic diagram of the confusion matrix structure of the DCNN model verification set classification in the wafer image recognition method improved based on the generative adversarial network according to an embodiment of the present invention.
[0052] In the figure:
[0053] 1. Lightweight classification model optimization module; 2. Convolutional neural network model construction module; 3. Deep convolutional generative adversarial network model construction module. DETAILED DESCRIPTION
[0054] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0055] According to an embodiment of the present invention, an improved wafer image recognition method and system based on a generative adversarial network are provided.
[0056] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 , Figure 3-Figure 8As shown, according to one embodiment of the present invention, a wafer image recognition method based on a generative adversarial network improvement is provided, and the wafer image recognition method includes the following steps:
[0057] S1. Use the pre-configured stacked convolutional layers to increase the depth of the convolutional layer of the lightweight classification model, and combine the global average pooling layer to optimize the lightweight classification model.
[0058] In one embodiment, the lightweight classification model is optimized by using stacked convolutional layers, and the lightweight classification model is reduced by combining a global average pooling layer, including the following steps:
[0059] S11. Use multiple convolutional layers to directly stack and increase the depth of the convolutional layer in the lightweight classification model.
[0060] S12. The feature vector is obtained by replacing the fully connected layer with the global average pooling layer to reduce the parameters in the light quantization classification model.
[0061] It should be explained that the present invention introduces a lightweight classification model that has low hardware requirements and takes less time to train, while achieving good classification performance. In terms of improving the expressiveness of the model, the depth of the convolutional layer is mainly increased by stacking convolutional layers, and the depth of the convolutional layer is increased by directly superimposing multiple convolutional layers. However, the model parameters increase exponentially with the depth of the model, but when the model reaches a certain depth, the reduction of its generalization error is not obvious. This article sets the convolutional layer to have 64 convolution kernels to obtain 64 feature maps to extract different features of the input image; in terms of reducing model parameters, the global average pooling is used instead of the fully connected layer to obtain the feature vector, which can greatly reduce the dimension of the output feature vector and avoid the huge amount of parameters introduced by the fully connected layer.
[0062] S2. Based on the optimized lightweight classification model, a convolutional neural network model is constructed, the wafer image is used as input, and the convolutional neural network model is trained.
[0063] In one embodiment, the convolutional neural network model consists of four convolutional layers, four normalization layers, four activation layers, three maximum pooling layers, one global average pooling layer, one Softmax layer and one classification layer.
[0064] In one embodiment, the steps of constructing a convolutional neural network model based on the reduced lightweight classification model and inputting a wafer image into the convolutional neural network model for training include the following steps:
[0065] S21. Extract feature maps based on the convolutional layer and process the feature maps.
[0066] In one embodiment, the extracting feature map based on the convolutional layer and preprocessing the feature map comprises the following steps:
[0067] S211, based on the convolutional neural network model, entering the convolution layer and using random calculation to extract features to obtain a feature map;
[0068] S212: Normalize based on the extracted feature map and standardize the statistical distribution of the feature map.
[0069] S22. Based on the processed feature map, the features in the feature map are optimized through the activation layer, and the dimension of the feature map is reduced using the maximum pooling layer.
[0070] In one embodiment, the step of optimizing the features in the feature map by an activation layer based on the processed feature map and reducing the dimension of the feature map by a maximum pooling layer comprises the following steps:
[0071] S221, the activation layer optimizes the features in the feature map using an activation function;
[0072] S222. Use a maximum pooling layer to reduce the dimension of the feature map, and the pooling window size in the maximum pooling layer is 2×2, and the step size is 2.
[0073] In one embodiment, the expression of the activation function is:
[0074]
[0075] In the formula, x represents the input parameter, α represents the coefficient, and LeakyReLU represents the activation function.
[0076] S23, the reduced feature map re-enters other convolutional layers, and then enters the normalization layer, activation layer and maximum pooling layer in turn.
[0077] S24. After repeating step S23 twice, the feature map enters the last convolution layer, and then enters the normalization layer, activation layer, and global average pooling layer in sequence, and is classified using the Softmax layer to obtain the output result.
[0078] It should be explained that based on the principle of reducing model parameters and improving model expression ability, the convolutional neural network model structure designed for wafer image fault pattern recognition is composed of four convolutional layers, four normalization layers, four ReLU layers, three maximum pooling layers, one global average pooling layer, one Softmax layer and one classification layer, as shown in Figure 3As shown in the figure, Input represents input, Conv represents convolution layer, Max pool represents maximum pooling layer, Global avg pool represents global average pooling layer, Softmax represents activation function, and Output represents output.
[0079] After the image is input into the network, the training process is as follows:
[0080] 1) First, enter the convolution layer to extract features. Multiple layers of convolution obtain multiple feature maps. The convolution layer is a calculation method for deep learning. The convolution layer uses random calculation to extract features by default.
[0081] 2) Then, the feature maps are batch normalized and the statistical distribution of the samples is standardized to reduce the differences between samples and speed up the training and convergence of the model.
[0082] 3) The activation layer applies the ReLU nonlinear activation function to the feature map to achieve complex feature expression.
[0083] 4) The maximum pooling layer is used to reduce the dimension of the feature map, thereby reducing the computational complexity. The pooling window size is 2×2, the step size is 2, and the size of the feature map obtained after the maximum pooling layer is only 1 / 4 of the original size.
[0084] 5) After the above steps, the reduced feature map will enter the new convolution layer, and then enter the batch normalization layer, activation layer (ReLU layer) and maximum pooling layer in sequence.
[0085] 6) After repeating the above operation twice, it enters the final convolution layer. After normalization and activation function, the feature map enters the global average pooling layer, and finally enters the Softmax layer for classification. Softmax is the activation function, and the formula for the output result is obtained.
[0086] S3. Based on the trained convolutional neural network model, the pre-configured generative adversarial network model is improved to construct a deep convolutional generative adversarial network model, and the wafer image is recognized through the deep convolutional generative adversarial network model.
[0087] In one embodiment, the method of improving the generative adversarial network model based on the trained convolutional neural network model, constructing a deep convolutional generative adversarial network model, and recognizing the wafer image through the deep convolutional generative adversarial network model includes the following steps:
[0088] S31. A generator based on a deep convolutional generative adversarial network model receives a one-dimensional random Gaussian vector, passes it through multiple deconvolution layers and performs normalization processing to generate a forged image of a wafer image.
[0089] S32. The discriminator based on the deep convolutional generative adversarial network model connects the activation function after each convolution layer and uses the Sigmoid function to perform the classification task.
[0090] S33. Use the discriminator of the deep convolutional generative adversarial network model and optimize the generator by combining the forged image of the wafer image.
[0091] In one embodiment, the method of optimizing the generator by using the discriminator of the deep convolutional generative adversarial network model in combination with the forged image of the wafer image comprises the following steps:
[0092] S331, training a discriminator based on the wafer image and the forged image of the wafer image;
[0093] S332: If the discriminator identifies the forged image of the wafer image as a false sample, the weight of the generator is updated.
[0094] It should be explained that, starting from the goal of wafer image data augmentation, the present invention proposes a new network structure, namely the Deep Convolutional Generative Adversarial Network (DCGAN), by improving the structure and training algorithm of the generative adversarial network. Compared with the basic GAN structure, the discriminator D and the generator G in the DCGAN use convolutional layers and deconvolution layers instead of the MLP structure. Such a network is more suitable for image data augmentation tasks. While satisfying the difficulty of distinguishing generated samples from real samples, it is also beneficial for the classifier to learn the classification boundaries on the synthesized image data. The specific network structure is as follows: Figure 4 As shown in the figure, Dconv represents the deconvolution layer and Conv represents the convolution layer.
[0095] The generator G in DCGAN receives a one-dimensional random Gaussian vector of size 100 as input, and then applies multiple deconvolution layers to upconvert the vector into a two-dimensional random noise image. All deconvolution layers are followed by ReLU activation functions, and batch normalization is used in these layers to stabilize the learning process. The output image of the last layer is the same size as the original real sample after scale normalization, both of which are 128×128. This is the noise image of the first period of generator training. This image will simulate the image features of the original sample during the training process, and finally evolve into a fake image that is difficult to distinguish between true and false. The layers of the generator network are shown in Table 1.
[0096] Table 1 Generator G network structure table
[0097] Layer number type Convolution kernel size Step Length Activation Function Input Structure Output structure 1 Fully connected layer 3 1 \ [100,] [32,32,128] 2 Deconvolution layer 3 2 ReLU [32,32,128] [64,64,64] 3 Deconvolution layer 3 1 ReLU [64,64,64] [64,64,32] 4 Deconvolution layer 3 2 ReLU [64,64,32] [128,128,1]
[0098] The discriminator D in DCGAN is a convolutional neural network that performs classification tasks. A LeakyReLU activation function is connected to each convolutional layer. Dropout operations and batch normalization are used in the first three convolutional layers to stabilize the learning process. The output of the discriminator D uses a Sigmoid function for classification. The discriminator network structure is shown in Table 2.
[0099] Table 2 Discriminator D network structure table
[0100]
[0101] Among them, the training of DCGAN is divided into two steps:
[0102] 1) Train the discriminator based on fake data and real data separately, with the training goal that the discriminator can effectively distinguish true from false. In this step, the generator has not been trained yet. The discriminator updates the weights through back propagation when it misclassifies (misclassifying real samples as fake samples or misclassifying fake samples as real samples).
[0103] 2) Train the generator. When the discriminator identifies a forged image as a fake sample, the weights of the generator are updated.
[0104] In addition, the DCNN model is first trained according to the above parameters, where the training iteration process is as follows: Figure 5-Figure 6 As shown, Figure 5 represents the iterative process of accuracy, Figure 6 Represents the iterative process of the loss function; the classification performance of the recognition model in the present invention is reflected in Figure 7-Figure 8 In the confusion matrix of Figure 7 represents the confusion matrix of the training set classification, Figure 8 Represents the confusion matrix of the validation set classification, where the horizontal axis labels represents the label and the vertical axis pridictions represents the predicted value.
[0105] According to another aspect of the present invention, a wafer image recognition system improved based on a generative adversarial network is also provided. Figure 2 As shown, the wafer image recognition system includes a lightweight classification model optimization module 1, a convolutional neural network model construction module 2 and a deep convolutional generative adversarial network model construction module 3;
[0106] The lightweight classification model optimization module 1 is used to increase the depth of the convolution layer of the lightweight classification model by using the pre-configured stacked convolution layer, and optimize the lightweight classification model in combination with the global average pooling layer;
[0107] The convolutional neural network model construction module 2 is used to construct a convolutional neural network model based on the optimized lightweight classification model, take the wafer image as input, and train the convolutional neural network model;
[0108] The deep convolutional generative adversarial network model construction module 3 is used to improve the pre-configured generative adversarial network model based on the trained convolutional neural network model, construct a deep convolutional generative adversarial network model, and recognize the wafer image through the deep convolutional generative adversarial network model.
[0109] To summarize, with the aid of the above-mentioned technical scheme of the present invention, the present invention optimizes the lightweight classification model and constructs a deep convolutional generative adversarial network to achieve the improvement of wafer image recognition accuracy, the reduction of computational complexity, and the enhancement of model generalization ability, which in turn helps to improve wafer manufacturing yield and reliability, and promote the continuous progress of semiconductor manufacturing technology. Model modification and data sampling are used to solve the data imbalance problem in the prior art. The former improves the classification accuracy by adjusting the weights of misclassified data samples; the latter is used to adjust data samples according to the distribution of samples between different categories by using undersampling, oversampling or data enhancement. Among them, data enhancement accounts for a high proportion in the current research. The data enhancement method can not only solve the data imbalance problem, but also prevent overfitting.
[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. An improved wafer image recognition method based on generative adversarial network, characterized in that: The wafer image recognition method comprises the following steps: S1. Use the pre-configured stacked convolutional layers to increase the depth of the convolutional layer of the lightweight classification model, and combine the global average pooling layer to optimize the lightweight classification model; S2. Based on the optimized lightweight classification model, a convolutional neural network model is constructed, the wafer image is used as input, and the convolutional neural network model is trained; S3. Based on the trained convolutional neural network model, the pre-configured generative adversarial network model is improved to construct a deep convolutional generative adversarial network model, and the wafer image is recognized through the deep convolutional generative adversarial network model.
2. According to claim 1, the improved wafer image recognition method based on generative adversarial network is characterized in that: The method of optimizing the lightweight classification model by using stacked convolutional layers based on the lightweight classification model and shrinking the lightweight classification model by combining the global average pooling layer includes the following steps: S11. Use multiple convolutional layers to directly stack and increase the depth of the convolutional layer in the lightweight classification model. S12. The feature vector is obtained by replacing the fully connected layer with the global average pooling layer to reduce the parameters in the light quantization classification model.
3. The improved wafer image recognition method based on generative adversarial network according to claim 1 is characterized in that: The convolutional neural network model includes a convolution layer, a normalization layer, an activation layer, a maximum pooling layer, a global average pooling layer, a Softmax layer and a classification layer.
4. The improved wafer image recognition method based on generative adversarial network according to claim 3 is characterized in that: The method of constructing a convolutional neural network model based on the reduced lightweight classification model and inputting a wafer image into the convolutional neural network model for training includes the following steps: S21, extracting feature maps based on the convolution layer and processing the feature maps; S22, based on the processed feature map, optimizing the features in the feature map through the activation layer, and reducing the dimension of the feature map by using the maximum pooling layer; S23, the reduced feature map re-enters other convolutional layers, and then enters the normalization layer, activation layer and maximum pooling layer in sequence; S24. After repeating step S23 twice, the feature map enters the last convolution layer, and then enters the normalization layer, activation layer, and global average pooling layer in sequence, and is classified using the Softmax layer to obtain the output result.
5. The improved wafer image recognition method based on generative adversarial network according to claim 4 is characterized in that: The extracting feature map based on the convolution layer and preprocessing the feature map include the following steps: S211, based on the convolutional neural network model, entering the convolution layer and using random calculation to extract features to obtain a feature map; S212: Normalize based on the extracted feature map and standardize the statistical distribution of the feature map.
6. The improved wafer image recognition method based on generative adversarial network according to claim 4 is characterized in that: The process of optimizing the features in the feature map by using an activation layer based on the processed feature map and reducing the dimension of the feature map by using a maximum pooling layer includes the following steps: S221, the activation layer optimizes the features in the feature map using an activation function; S222. Use a maximum pooling layer to reduce the dimension of the feature map, and the pooling window size in the maximum pooling layer is 2×2, and the step size is 2.
7. The improved wafer image recognition method based on generative adversarial network according to claim 6 is characterized in that: The expression of the activation function is: In the formula, x represents the input parameter, α represents the coefficient, and LeakyReLU represents the activation function.
8. The improved wafer image recognition method based on generative adversarial network according to claim 1 is characterized in that: The method of improving the generative adversarial network model based on the trained convolutional neural network model, constructing a deep convolutional generative adversarial network model, and recognizing the wafer image through the deep convolutional generative adversarial network model includes the following steps: S31, a generator based on a deep convolutional generative adversarial network model, receiving a one-dimensional random Gaussian vector, passing through multiple deconvolution layers and performing normalization processing, to generate a forged image of a wafer image; S32, the discriminator based on the deep convolutional generative adversarial network model, connects the activation function after each convolution layer, and uses the Sigmoid function to perform the classification task; S33. Use the discriminator of the deep convolutional generative adversarial network model and optimize the generator by combining the forged image of the wafer image.
9. The improved wafer image recognition method based on generative adversarial network according to claim 8, characterized in that: The method of optimizing the generator by using the discriminator of the deep convolutional generative adversarial network model and combining the forged image of the wafer image comprises the following steps: S331, training a discriminator based on the wafer image and the forged image of the wafer image; S332: If the discriminator identifies the forged image of the wafer image as a false sample, the weight of the generator is updated.
10. An improved wafer image recognition system based on a generative adversarial network, used to implement an improved wafer image recognition method based on a generative adversarial network as claimed in any one of claims 1 to 9, characterized in that: The wafer image recognition system includes a lightweight classification model optimization module, a convolutional neural network model building module, and a deep convolutional generative adversarial network model building module; The lightweight classification model optimization module is used to increase the depth of the convolution layer of the lightweight classification model by using the pre-configured stacked convolution layer, and optimize the lightweight classification model in combination with the global average pooling layer; The convolutional neural network model building module is used to build a convolutional neural network model based on the optimized lightweight classification model, take the wafer image as input, and train the convolutional neural network model; The deep convolutional generative adversarial network model construction module is used to improve the pre-configured generative adversarial network model based on the trained convolutional neural network model, construct a deep convolutional generative adversarial network model, and recognize wafer image through the deep convolutional generative adversarial network model.