Defect detection model establishment method and device, equipment and defect detection method

Abstract

The application relates to a defect detection model establishing method and device, and a defect detection method. The method comprises the following steps: acquiring a wafer image; performing data preprocessing on the wafer image to obtain a model training data set; processing the model training data set based on a micro-defect enhancement and a multi-scale field of view network to obtain defect-specific features; performing quantum state coding processing and feature reconstruction on the defect-specific features by using a quantum-inspired sparse feature reconstruction network to obtain an initial model and generate result features of the initial model; and optimizing the initial model based on the result features by using a dynamic focusing fine-tuning loss function to obtain a detection model. The defect detection model establishing method can greatly reduce the manpower and material resources cost of defect detection, and improve the detection efficiency and recognition accuracy.

Description

Technical Field

[0001] This application relates to the field of semiconductor testing technology, and in particular to a method, apparatus, equipment and defect detection method for establishing a defect detection model. Background Technology

[0002] During the wafer manufacturing process, various defects may occur due to a variety of factors (such as improper process control, material defects, etc.). Figure 1 and Figure 2 As shown, bubbles are a common type of defect. The presence of defects can severely affect the electrical performance and mechanical strength of chips, and even lead to chip failure. Therefore, accurate defect detection is crucial for ensuring product quality.

[0003] Traditional defect detection relies primarily on manual visual inspection. Since most defects are extremely small and difficult for the human eye to detect, misjudgments are common. Defects often require magnification of the image many times before they can be clearly seen. Therefore, this method is inefficient, costly, and susceptible to human error. Furthermore, manual inspection requires constantly zooming in on the image, resulting in high time and labor costs.

[0004] In addition, although existing technologies employ detection methods such as bright field, dark field, and electron beam imaging that can cover most defect detection scenarios, they cannot balance detection accuracy, sensitivity, and speed during the detection process. Summary of the Invention

[0005] Therefore, it is necessary to provide a defect detection model establishment method, apparatus, equipment, and defect detection method that can improve detection accuracy, sensitivity, and detection efficiency in response to the above-mentioned technical problems.

[0006] To achieve the above objectives, this application provides a method for establishing a defect detection model, comprising the following steps:

[0007] Acquire wafer images;

[0008] The wafer image is preprocessed to obtain the model training dataset;

[0009] Based on micro-defect enhancement and multi-scale vision network, the training dataset of the model is processed to obtain defect-specific features;

[0010] A quantum-inspired sparse feature reconstruction network is used to perform quantum state encoding and feature reconstruction on the defect-specific features to obtain an initial model and generate the resulting features of the initial model.

[0011] A dynamic focusing fine-tuning loss function is used to optimize the initial model based on the resulting features to obtain a detection model.

[0012] In one embodiment, the data preprocessing of the image data of the wafer to obtain the model training dataset includes:

[0013] The wafer image is divided into multiple unit images using multidimensional data segmentation.

[0014] Defect samples are introduced into the unit image to perform anomaly fusion;

[0015] Construct a coordinate mapping relationship between the unit image and the wafer image.

[0016] In one embodiment, the processing of the model training dataset based on micro-defect enhancement and multi-scale vision networks includes:

[0017] Morphological filtering is performed on the model training dataset to augment the data in the model training dataset;

[0018] Multi-scale feature extraction is performed on the training dataset of the model after data augmentation using convolutional kernels of different sizes;

[0019] The low-level details and high-level semantic information of features extracted by convolutional kernels of different sizes are fused to obtain intermediate features;

[0020] A functional adapter is used to adapt the intermediate features to transform them into the defect-specific features.

[0021] In one embodiment, performing morphological filtering on the model training dataset to augment the model training dataset includes:

[0022] Morphological filters are used to perform dilation, erosion, and opening / closing operations on the training dataset of the model.

[0023] In one embodiment, the quantum-inspired sparse feature reconstruction network performs quantum state encoding and feature reconstruction on the defect-specific features to obtain an initial model and generate result features, including:

[0024] The defect-specific features are quantum-inspired feature encodings.

[0025] The defect-specific features, after being quantum-inspired feature encoding, are subjected to quantum gate transformation;

[0026] A sparse coding and feature reconstruction network is used to sparsely encode and reconstruct the defect-specific features after quantum gate transformation to obtain the initial model and generate the resulting features.

[0027] In one embodiment, acquiring the wafer image further includes:

[0028] Obtain defect information from the wafer image;

[0029] The method employs a dynamically focused, fine-tuned loss function to optimize the initial model based on the resulting features, thereby obtaining a detection model, including:

[0030] The generated result features are filtered using a discriminator;

[0031] The difference between the selected result features and the defect information is calculated using the dynamic focusing fine-tuning loss function, and then the adjustment parameters are obtained.

[0032] The initial model is optimized based on the adjustment parameters.

[0033] In one embodiment, before employing a dynamically focused fine-tuned loss function to optimize the initial model based on the resulting features to obtain the detection model, the method further includes:

[0034] Defect and invalidity filtering is performed on the resulting features.

[0035] On the other hand, this application also provides a defect model building apparatus, including:

[0036] The acquisition module is used to acquire wafer images;

[0037] The preprocessing module is used to preprocess the wafer image to obtain the model training dataset;

[0038] The feature enhancement and extraction module is used to process the model training dataset based on micro-defect enhancement and multi-scale vision network to obtain defect-specific features;

[0039] The feature reconstruction module is used to reconstruct the network using quantum-inspired sparse features, perform quantum state encoding processing and feature reconstruction on the defect-specific features, so as to obtain an initial model and generate the result features of the initial model.

[0040] The model optimization module is used to optimize the initial model based on the result features by employing dynamic focusing and fine-tuning of the loss function to obtain the detection model.

[0041] On the other hand, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the defect detection model establishment method as described in any of the preceding claims.

[0042] On the other hand, this application also provides a defect detection method, including the following steps:

[0043] Provide a defect detection model established by the defect detection model establishment method described in any of the above;

[0044] Load the image to be tested and input the image to be tested into the defect detection model to output the defect detection result.

[0045] The aforementioned defect detection model establishment method, apparatus, equipment, and defect detection method preprocess wafer images to obtain a model training dataset. The training dataset is then processed using a synergistic approach of micro-defect enhancement and a multi-scale vision network to extract defect-specific features. This effectively captures and enhances the features of minute defects, improving detection sensitivity and accuracy. A quantum-inspired sparse feature reconstruction network encodes the extracted defect-specific features using quantum states and reconstructs the features to obtain an initial model and generate its resulting features. This effectively compresses the feature space and improves training efficiency. Dynamic focusing and fine-tuning of the loss function optimizes the initial model based on the resulting features, resulting in a detection model. This enhances the robustness and generalization ability of the detection model, significantly improving its reliability. In short, this defect detection model optimizes wafer surface defect identification, greatly reducing manpower and material costs while improving detection efficiency and accuracy. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is an optical image of a wafer in the prior art;

[0048] Figure 2 For containing Figure 1 An enlarged optical image of the area selected by the dashed line in the middle;

[0049] Figure 3 This is a flowchart illustrating a defect detection model establishment method in one embodiment;

[0050] Figure 4 This is a schematic optical image of a wafer image obtained in another embodiment;

[0051] Figure 5 for Figure 4Optical image of a wafer image obtained from a multidimensional data segmentation process;

[0052] Figure 6 for Figure 5 Optical image of the unit image obtained in the process;

[0053] Figure 7 This is a coordinate diagram illustrating the coordinate mapping relationship between a cell image and a wafer image in one embodiment.

[0054] Figure 8 This is a schematic diagram of the process of fusing low-level details and high-level semantic information of features extracted by convolutional kernels of different sizes to obtain intermediate features in one embodiment.

[0055] Figure 9 This is a schematic diagram illustrating quantum-inspired feature encoding of defect-specific features in one embodiment;

[0056] Figure 10 This is a schematic diagram illustrating the process of using a sparse coding and feature reconstruction network in one embodiment to perform sparse coding and feature reconstruction on the defect-specific features after quantum gate transformation.

[0057] Figure 11 This is a schematic diagram of the curve changes after introducing a balance factor and an adjustment factor into the dynamic focus fine-tuning loss function in one embodiment.

[0058] Figure 12 This is a schematic diagram of an optical image used for defect-invalid filtering of the resulting features in one embodiment.

[0059] Figure 13 This is a schematic diagram of the coordinates of the fitted elliptic curve after performing defect invalid filtering on the resulting features in one embodiment. Detailed Implementation

[0060] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0062] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish one element from another.

[0063] Spatial relation terms such as “below,” “under,” “below,” “under,” “above,” “above,” etc., are used herein to describe the relationship between one element or feature shown in the figure and other elements or features. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, the element or feature described as “below,” “under,” or “below” will be oriented “above” the other element or feature. Therefore, the exemplary terms “below” and “under” can include both above and below orientations. Furthermore, the device may also include other orientations (e.g., rotated 90 degrees or other orientations), and the spatial descriptive terms used herein will be interpreted accordingly.

[0064] It should be noted that when one element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through an intermediary element. Furthermore, in the following embodiments, "connection" should be understood as "electrical connection," "communication connection," etc., if there is transmission of electrical signals or data between the connected objects.

[0065] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0066] In one embodiment, such as Figure 3 As shown, a method for establishing a defect detection model is provided, including the following steps:

[0067] Step S1: Obtain wafer image;

[0068] Step S2: Perform data preprocessing on the wafer image to obtain the model training dataset;

[0069] Step S3: Based on micro-defect enhancement and multi-scale vision network, process the model training dataset to obtain defect-specific features;

[0070] Step S4: Reconstruct the network using quantum-inspired sparse features, perform quantum state encoding on the defect-specific features and feature reconstruction to obtain the initial model and generate the resulting features of the initial model;

[0071] Step S5: Employ dynamic focusing to fine-tune the loss function and optimize the initial model based on the result features to obtain the detection model.

[0072] In the aforementioned defect detection model establishment method, wafer images are preprocessed to obtain a model training dataset. The training dataset is then processed using a combination of micro-defect enhancement and a multi-scale vision network to obtain defect-specific features. This effectively captures and enhances the features of minute defects, improving defect detection sensitivity and accuracy. A quantum-inspired sparse feature reconstruction network is used to encode the extracted defect-specific features using quantum states and reconstruct the features to obtain an initial model and generate its resulting features. This effectively compresses the feature space and improves model training efficiency. Finally, the initial model is optimized based on the resulting features by dynamically focusing and fine-tuning the loss function to obtain the defect detection model. This enhances the robustness and generalization ability of the defect detection model, effectively improving its reliability.

[0073] Specifically, please refer to Figures 4 to 7 Perform steps S1 to S2 to obtain a wafer image; perform data preprocessing on the wafer image to obtain a model training dataset.

[0074] For example, the resolution of the acquired wafer image is 4000×4000.

[0075] In one embodiment, obtaining a wafer image further includes:

[0076] Obtain defect information from wafer images, including the location coordinates and types of defects on the wafer image.

[0077] In one embodiment, the wafer image is preprocessed to obtain a model training dataset, including:

[0078] Multidimensional data segmentation is used to divide the wafer image into multiple unit images; for example, such as Figures 4 to 6 As shown, the original 4000×4000 resolution wafer image is divided into 4×4 unit images with a resolution of 1000×1000. This magnifies defects in the wafer image and expands the data volume. This segmentation method can divide the wafer image from multiple dimensions, ensuring that each unit image contains relatively complete local features, providing a foundation for subsequent defect detection.

[0079] Defect samples are introduced into the unit images to perform anomaly fusion; that is, when using the acquired wafer images for sample training, normal samples (i.e., unit images) are the main focus, and a small number of defect samples (including sample images with defects) are dynamically injected for fusion training. By introducing defect samples, the training model's ability to identify different types of defects is enhanced, which can better adapt to various defect situations that may occur in actual production. This helps the model to provide a reference in the feature generation stage, guides the detection model to generate more accurate and effective features, and enhances the robustness and generalization ability of the obtained detection model.

[0080] Construct the coordinate mapping relationship between the cell image and the wafer image. For example, Figure 7 As shown, constructing the coordinate mapping relationship between the cell image and the wafer image can transform the coordinates of defect information in the cell image back to the coordinates in the original wafer image, including:

[0081] Assign two-dimensional coordinates to the unit image ,in and These represent the row and column indices of the cell image within the wafer image, respectively.

[0082] Original defect coordinates Multiply by the reduction factor In addition, the offset of the cell image in the wafer image The restored defect coordinates are obtained. .in, , , , By restoring the coordinates of defects in the cell image, it is ensured that after a defect is detected in the cell image, it can be accurately located to the corresponding position in the original wafer image, facilitating subsequent analysis and processing.

[0083] Specifically, please refer to Figure 8 Step S3 is executed, which involves processing the model training dataset based on micro-defect enhancement and multi-scale vision network to obtain defect-specific features.

[0084] In one embodiment, such as Figure 8 As shown, based on micro-defect enhancement and multi-scale vision networks, the model training dataset is processed, including:

[0085] Morphological filtering is applied to the model training dataset to augment the data. Specifically, morphological filtering is used to process the model training dataset with structuring elements of different sizes, processing unit images at multiple scales. This can highlight defect features of different sizes and effectively enhance the boundary and shape structure of tiny defects.

[0086] The process involves performing morphological filtering on the model training dataset to augment the data. This includes applying dilation, erosion, and opening / closing operations to the training dataset using morphological filters. Dilation increases the brightness of features to make them more prominent, erosion reduces the brightness of the background to make features stand out more, opening removes small noise points, and closing fills small holes.

[0087] Multi-scale feature extraction is performed on the data-augmented model training dataset using convolutional kernels (Conv) of different sizes. For example, the data-augmented model training dataset is used with convolutional kernels of different sizes (1×1, 3×3, 5×5), and multi-scale defect feature maps are extracted according to different number of point intervals, which contain semantic information from low level (contour edge) to high level (size, shape, distribution pattern).

[0088] The low-level details and high-level semantic information of features extracted by convolutional kernels of different sizes are fused to obtain intermediate features; feature maps of different scales are fused by upsampling, downsampling and lateral connection to retain low-level details and high-level semantic information, thereby improving the accuracy of detection.

[0089] Functional adapters are used to adapt intermediate features, transforming them into defect-specific features. These adapters can adapt and transform intermediate features from different models or algorithms, making them more suitable for defect detection tasks. This helps the detection model better adapt to various complex scenarios and data distributions, improving its generalization ability and enabling it to achieve better performance in different defect detection tasks.

[0090] Specifically, please refer to Figures 9 to 10 Step S4 is executed, where the network is reconstructed using quantum-inspired sparse features, and the defect-specific features are used for quantum state encoding and feature reconstruction to obtain the initial model and generate the resulting features of the initial model.

[0091] In one embodiment, a quantum-inspired sparse feature reconstruction network is used to perform quantum state encoding and feature reconstruction on defect-specific features to obtain an initial model and generate resulting features, including:

[0092] Quantum-inspired feature encoding of defect-specific features; for example, such as Figure 9As shown, the extracted defect-specific features are encoded as the amplitude and phase of qubits. An amplitude encoding method is used to encode the feature vector into the amplitude of the qubits. The binary pixel data of the defect-specific features are encoded as multidimensional quantum states. For example, color, texture, and shape correspond to different quantum superposition states, thus constructing a high-dimensional feature space. The quantum state encoding can be represented as... , , , , For amplitude, , For phase.

[0093] Wherein, it is assumed that the original defect-specific feature matrix is After normalization, we get The normalization formula is: ,in, yes Modulus length: Therefore, after normalization The normalized feature vector Encoded as a quantum state: ,because , ,Right now , ,but radian, (Assuming the phase is 0), therefore, the quantum state can be represented as: , .

[0094] The defect-specific features after quantum-inspired feature encoding are subjected to quantum gate transformation; quantum gates such as rotation gates and CNOT gates are applied to perform nonlinear transformations on qubits to enrich the feature space;

[0095] The quantum gate transformation of the aforementioned qubits includes at least two levels of transformation. The first level of transformation involves using a rotation gate and a CNOT gate. Assume the rotation angle is... Then the quantum state after the first layer of transformation is obtained as Performing a second-level rotation gate transformation on the matrix after the first-level transformation yields the following quantum state: The initial matrix is ​​obtained by quantum state encoding and two-level quantum gate transformation. The quantum transformation (QE) is simulated to capture the potential relationship between different defect features, which effectively enriches the feature space.

[0096] A sparse coding and feature reconstruction network is employed to sparsely encode and reconstruct the defect-specific features after quantum gate transformation, thereby obtaining an initial model and generating the resulting features. For example... Figure 10As shown, A(z) represents the sparse coding inspired by quantum heuristics, and ξ i Represented as reconstructed features, a sparse coding algorithm is used to map high-dimensional features to a low-dimensional sparse feature space, removing redundant information. At the same time, normal wafer samples are used as a reference to retain the most useful information for defect identification, avoiding the problem of increasing the computational cost of model training and inference by directly processing high-dimensional features.

[0097] Furthermore, a sparse coding and feature reconstruction network is employed to sparsely encode and reconstruct the defect-specific features after quantum gate transformation to obtain the initial model and generate the resulting features. This also includes adaptive noise enhancement and suppression of the defect-specific features. First, the noise distribution characteristics of the unit image are analyzed, and the intensity and type of noise are automatically adjusted according to the image content. Second, the noise level is appropriately enhanced in suspected defect areas to simulate defect variations under different lighting conditions or imaging devices, thereby enhancing the robustness of the model. Finally, noise is suppressed in the background area to maintain the clarity and recognizability of the unit image.

[0098] Specifically, please refer to Figures 11 to 13 Then, in step S5, the loss function is dynamically focused and fine-tuned to optimize the initial model based on the result features to obtain the detection model.

[0099] In one embodiment, a dynamic focusing fine-tuning loss function is employed to optimize the initial model based on the resulting features, thereby obtaining a detection model, including:

[0100] A discriminator is used to filter the generated features; the discriminator is used to evaluate the features to select the features that are closest to the defect information on the wafer image.

[0101] The difference between the selected features and the defect information is calculated using a dynamic focusing fine-tuning loss function, and then the adjustment parameters are obtained.

[0102] The initial model is optimized based on the adjusted parameters.

[0103] The optimization of the initial model based on the adjustment parameters includes:

[0104] Based on the Dynamic Focusing Fine-Tuning Loss (DFR-Loss), the positive and negative samples, as well as the easy and difficult samples, are balanced and adjusted. In wafer image defect detection, the DFR-Loss loss function is used to solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult samples in classification. The DFR-Loss loss function can effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult samples, thereby improving the robustness and detection accuracy of the model. By dynamically adjusting the weights of the loss function during training, the model can better focus on the difficult-to-classify defect samples, thereby improving the ability to identify small defects.

[0105] For example, such as Figure 11 As shown, based on the dynamic focusing fine-tuning loss function, a balance adjustment is made for positive and negative samples and easy and difficult samples, including:

[0106] A balancing factor is introduced into the dynamic focusing fine-tuning loss function to balance the weights of positive and negative samples;

[0107] An adjustment factor is introduced into the dynamic focusing fine-tuning loss function to balance the weights of easy and difficult samples. By introducing the adjustment factor, the loss contribution of easily classified defect samples can be reduced, allowing the model to focus more on difficult-to-classify samples, which is particularly useful for cases with subtle defects and small datasets.

[0108] The formula for the dynamic focus fine-tuning loss function is expressed as follows: ,in: For the model, wafer samples The predicted probability, As a balancing factor, it is used to balance the weights of defective samples and normal samples. When the value is large, the model will pay more attention to the loss of positive samples (defective samples). This is a regulating factor used to adjust the weights of easy and difficult samples. When the value is large, the loss contribution of easily classified samples decreases, while the loss contribution of difficult-to-classify samples increases. Therefore, the loss function can be dynamically adjusted to differentiate between easily classified and difficult-to-classify defective samples. If the value increases (i.e., the probability of the sample being correctly classified is higher), then... As the value decreases, the sample loss is reduced. A decrease in the value (meaning a higher probability that the sample is misclassified) indicates that... As the value increases, the sample loss is amplified. This can be addressed by adjusting... This can balance the overall contribution of defective samples and normal samples. By adjusting... This allows the model to better focus on defect samples that are difficult to classify, thereby improving the model's robustness and detection accuracy, and solving the class imbalance problem.

[0109] In one embodiment, before employing a dynamically focused fine-tuned loss function and optimizing the initial model based on the resulting features to obtain the detection model, the following steps are also included:

[0110] Defect invalidity filtering is performed on the resulting features to ensure that the detection points are located within the wafer. Specific steps include:

[0111] Use the Open Source Computer Vision Library (OpenCV) to merge masks and fit elliptical curves;

[0112] The center coordinates and axis lengths are obtained; where the center coordinates are... The minor and major axes are represented as The minor semi-axis is represented as ;

[0113] Perform perspective transformation, let ;

[0114] Filtering invalid data using geometric judgment, if If the detection point is within the ellipse, it is considered valid defect data. Then it is considered invalid data and will be filtered out.

[0115] It should be understood that, although Figure 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 3 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0116] In one embodiment, this application also provides a defect detection model building apparatus, including: an acquisition module, a preprocessing module, a feature enhancement and extraction module, a feature reconstruction module, and a model optimization module. The acquisition module acquires a wafer image; the preprocessing module preprocesses the wafer image to obtain a model training dataset; the feature enhancement and extraction module processes the model training dataset based on micro-defect enhancement and a multi-scale vision network to obtain defect-specific features; the feature reconstruction module utilizes a quantum-inspired sparse feature reconstruction network to perform quantum state encoding and feature reconstruction on the defect-specific features to obtain an initial model and generate the resulting features of the initial model; and the model optimization module employs a dynamically focused fine-tuned loss function to optimize the initial model based on the resulting features to obtain a detection model.

[0117] For example, the acquisition module is also used to acquire defect information of the wafer image.

[0118] Specific limitations regarding the defect detection model establishment device can be found in the limitations of the wafer defect detection model establishment method described above, and will not be repeated here. Each module in the aforementioned defect detection model establishment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0119] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described defect model establishment method embodiments.

[0120] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0121] In one embodiment, this application also provides a defect detection method, comprising the following steps:

[0122] Provides a defect detection model established using the defect detection model establishment method described above;

[0123] Load the image to be tested and input it into the defect detection model to output the defect detection results.

[0124] The process involves loading new images for testing, preprocessing these images to extract their features, and then adapting these new features to the defect detection model to output detection results. This significantly reduces manpower and material costs while improving detection efficiency and recognition accuracy.

[0125] In the description of this specification, references to terms such as "some embodiments," "other embodiments," and "ideal embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative descriptions of the above terms do not necessarily refer to the same embodiments or examples.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A defect detection model establishment method characterized by comprising: The method comprises the following steps: obtaining a wafer image; performing data preprocessing on the wafer image to obtain a model training data set; processing the model training data set based on a micro-defect enhancement and multi-scale field of view network to obtain defect-specific features; performing quantum state encoding processing and feature reconstruction on the defect-specific features by using a quantum-inspired sparse feature reconstruction network to obtain an initial model and generate result features of the initial model, wherein the quantum-inspired sparse feature reconstruction network performs quantum state encoding processing and feature reconstruction on the defect-specific features to obtain an initial model and generate result features, and the method comprises the following steps: quantum-inspired feature encoding of the defect-specific features, wherein the quantum-inspired feature encoding of the defect-specific features comprises encoding the extracted defect-specific features into amplitudes and phases of quantum bits, encoding feature vectors into amplitudes of quantum bits by using an amplitude encoding method, and encoding binary pixel data of the defect-specific features into multi-dimensional quantum states; performing quantum gate transformation on the defect-specific features after quantum-inspired feature encoding; performing sparse coding and feature reconstruction on the defect-specific features after quantum gate transformation by using a sparse coding and feature reconstruction network to obtain the initial model and generate the result features; optimizing the initial model based on the result features by using a dynamic focus fine-tuning loss function to obtain a detection model, the method of obtaining a wafer image further comprises: obtaining defect information of the wafer image; the method of optimizing the initial model based on the result features by using a dynamic focus fine-tuning loss function to obtain a detection model comprises: screening the generated result features by using a discriminator; calculating differences between the screened result features and the defect information by using the dynamic focus fine-tuning loss function to obtain adjustment parameters; The initial model is optimized based on the aforementioned adjustment parameters, wherein the dynamic focus fine-tuning loss function is expressed as follows: ,in: For the model, wafer samples The predicted probability, This is a balancing factor used to balance the weights of defective and normal samples. This is an adjustment factor used to adjust the weights of easy and difficult samples. 2.The defect detection model establishing method of claim 1, wherein, the method of performing data preprocessing on image data of the wafer to obtain a model training data set comprises: segmenting the wafer image into a plurality of unit images by using data multi-dimensional segmentation; introducing defect samples in the unit images to perform abnormal fusion; constructing a coordinate mapping relationship between the unit images and the wafer image. 3.The method of claim 1, wherein, the method of processing the model training data set based on a micro-defect enhancement and multi-scale field of view network comprises: performing a morphological filtering operation on the model training data set to perform data enhancement on the model training data set; performing multi-scale feature extraction on the model training data set after data enhancement by using convolution kernels of different sizes; fusing low-level details and high-level semantic information of features extracted by convolution kernels of different sizes to obtain intermediate features; performing feature adaptation on the intermediate features by using a function adapter to convert the intermediate features into the defect-specific features.

4. The defect detection model establishing method according to claim 3, wherein the method of performing a morphological filtering operation on the model training data set to perform data enhancement on the model training data set comprises: performing inflation, corrosion and open-close operation processing on the model training data set by using a morphological filter. 5.The defect detection model establishing method of claim 1, wherein, Before the adopting a dynamic focus fine-tuning loss function, based on the result feature, optimizing the initial model to obtain a detection model, the method further comprises: Performing defect invalid filtering on the result feature.

6. A defect model building apparatus characterized by comprising: Comprise: An acquisition module, configured to acquire a wafer image; A preprocessing module, configured to perform data preprocessing on the wafer image to obtain a model training data set; A feature enhancement and extraction module, configured to process the model training data set based on a micro-defect enhancement and multi-scale field of view network to obtain defect-specific features; A feature reconstruction module, configured to perform quantum state coding processing and feature reconstruction on the defect-specific features by using a quantum-inspired sparse feature reconstruction network to obtain an initial model and generate result features of the initial model; A model optimization module, configured to adopt a dynamic focus fine-tuning loss function, based on the result features, to optimize the initial model to obtain a detection model. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the defect detection model establishment method according to any one of claims 1-5.

8. A defect detection method characterized by, Comprise the following steps: Providing a defect detection model established by the defect detection model establishment method according to any one of claims 1-5; Loading a to-be-tested image and inputting the to-be-tested image into the defect detection model to output a defect detection result.

Images