Wafer defect intelligent detection and classification method based on deep learning

Abstract

This invention discloses a deep learning-based intelligent detection and classification method for wafer defects, belonging to the field of computer vision technology. First, a multi-scale pyramid decomposition is performed on the original wafer grayscale image. Then, a Fast Fourier Transform is performed on each layer of the image in the frequency domain. An adaptive sliding window detector and Hough transform are used to accurately identify and locate frequency peaks representing periodic background. Next, an adaptive notch filter is constructed and applied to suppress these periodic background frequencies, and non-periodic defect signals are explicitly enhanced by calculating frequency domain residuals. Finally, the multi-scale processed defect enhancement residual maps are fused to generate the final defect enhancement map, which serves as the input to the deep learning detection network. This invention solves the problem in existing technologies where low-contrast micro-defects are easily obscured by the strong periodic background texture of the wafer, leading to low detection accuracy and poor model generalization ability.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to a method for intelligent detection and classification of wafer defects based on deep learning. Background Technology

[0002] In the semiconductor chip manufacturing process, the surface quality of the wafer is a key factor determining the performance and yield of the final product. With the continuous miniaturization of integrated circuit manufacturing processes, the circuit patterns on wafers are becoming increasingly dense and complex, and defect sizes are shrinking to submicron or even nanometer levels. These tiny defects, such as micro-interruptions, bridging, particles, and scratches, are often submerged within the highly periodic circuit patterns (e.g., memory cell arrays, logic gate arrays) on the wafer surface. This makes traditional image processing-based detection methods, such as threshold segmentation, edge detection, or template matching, prone to high false alarm or false negative rates. These methods often struggle to effectively distinguish between "normal periodic background textures" and "abnormal non-periodic defect signals."

[0003] In recent years, deep learning techniques, especially convolutional neural networks (CNNs) and the Transformer architecture, have made significant progress in image recognition and object detection, and are gradually being introduced into wafer defect detection. These methods automatically extract image features in the spatial domain by learning from large amounts of labeled data, thereby identifying defects.

[0004] However, existing deep learning-based wafer defect detection methods still face some inherent technical challenges in practical applications:

[0005] 1. In wafer images, the energy of periodic circuit patterns is often much higher than that of minute defects. When deep learning models (such as CNNs) extract features through multiple layers of convolution and pooling operations, features of strong background textures gradually dominate during downsampling, causing low-contrast micro-defect features to be obscured by background information, making them difficult to detect and classify effectively. This obfuscation makes it difficult for the model to accurately locate and distinguish defects in complex backgrounds.

[0006] 2. Current deep learning models utilize attention mechanisms or image enhancement techniques, such as channel attention, spatial attention, or histogram equalization, primarily operating in the spatial domain. While these methods can improve feature representation to some extent, they lack the ability to perceive the fundamental difference between "periodic" and "non-periodic" elements in an image. This means the model still requires significant learning resources and data to indirectly "learn" to ignore periodic backgrounds, rather than fundamentally decoupling them from defects. This indirect learning approach leads to insufficient generalization ability and performance degradation when faced with new wafer types or process variations.

[0007] 3. To detect defects at different scales, existing methods typically employ structures such as Feature Pyramid Networks (FPNs) to fuse multi-scale features. However, these pyramid structures still simultaneously present periodic background and defects at different scales. They lack a dedicated mechanism to "remove" background information; they merely propagate and fuse background and defect information across different scales. This not only increases the learning burden on the network but also fails to fundamentally solve the problem of background-defect confusion.

[0008] In summary, existing technologies fail to effectively decouple the background from micro-defects when processing strongly periodic background textures in wafer images, thus limiting the detection accuracy and generalization ability of deep learning models for low-contrast micro-defects.

[0009] Therefore, it is necessary to propose a deep learning-based intelligent detection and classification method for wafer defects to solve the above problems. Summary of the Invention

[0010] The purpose of this invention is to solve the problem that existing technologies fail to effectively decouple the background from micro-defects when processing strongly periodic background textures in wafer images, thus limiting the detection accuracy and generalization ability of deep learning models for low-contrast micro-defects.

[0011] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0012] A deep learning-based intelligent detection and classification method for wafer defects includes the following steps: a. acquiring the original wafer grayscale image;

[0013] b. Input the original wafer grayscale image into the frequency domain decoupled multi-period texture contrast enhancement module for preprocessing to obtain a defect enhancement map;

[0014] c. Input the enhanced defect map into a pre-trained deep learning detection network to classify and locate defects;

[0015] d. Output the detection results.

[0016] Furthermore, the internal processing flow of the frequency domain decoupled multi-period texture contrast enhancement module in step b includes:

[0017] b1. Image Multi-Scale Pyramid Decomposition: Construct a Gaussian pyramid on the original wafer grayscale image to obtain image layers at different scales. ,in Indicates the number of levels in the pyramid;

[0018] b2. Frequency Domain Transformation and Background Peak Detection: For each image layer... Perform a Fast Fourier Transform to obtain its spectrum. And calculate the amplitude spectrum. Then, an adaptive sliding window detector and Hough transform are used to identify discrete peak points representing the periodic background in the amplitude spectrum, forming a peak point set. ;

[0019] b3. Adaptive Notch Filter Construction and Background Suppression: For the aforementioned set of peak points Construct a notch filter at each peak point. And combine all the individual notch filters into a total filter. Then the total filter Applied to the spectrum Obtain the filtered spectrum ;

[0020] b4. Frequency Domain Residual Learning and Defect Augmentation: Calculating the Original Spectrum With the filtered spectrum The difference value is used to obtain the residual spectrum. Then, the residual spectrum Performing an inverse Fourier transform yields the defect-enhanced residual map in the spatial domain. ;

[0021] b5. Multi-scale feature fusion: Enhance the residual maps of defects generated from each layer of the pyramid. Upsample to the original input image size, and then... Convolution is used to fuse channels and generate the final defect enhancement map. .

[0022] Furthermore, in step b1, the Gaussian pyramid is generated by performing continuous Gaussian blurring and smoothing downsampling operations on the original image.

[0023] Further, in step b2, the adaptive sliding window detector, within a ring-shaped region centered on the origin of the spectrum, for a point... If its amplitude Larger than its neighborhood The magnitude of all other points within If the value is multiple times higher, it is determined to be a potential peak point; the Hough transform is used to verify whether the potential peak points form a straight line or a concentric circle equally spaced distribution pattern in the frequency domain.

[0024] Furthermore, the neighborhood The size is Pixel; the The coefficient of multiple The value is 5.

[0025] Furthermore, in step b3, the notch filter A Gaussian notch filter is used, and its expression is: ,in It is a point to peak point Euclidean distance, It is the notch radius.

[0026] Furthermore, the notch radius The value is 2 pixels.

[0027] Further, in step b4, the residual spectrum... After performing the inverse Fourier transform, the real part of the result is taken as the defect enhancement residual map. .

[0028] Furthermore, in step b5, the upsampling operation employs bilinear interpolation or bicubic interpolation, and the channel fusion is performed by stitching together the data along the channel dimension, followed by a... Convolutional layer implementation.

[0029] Furthermore, the deep learning detection network is a detector based on a convolutional neural network or a Transformer architecture.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. This invention explicitly and adaptively removes strongly periodic background textures in the frequency domain, making low-contrast micro-defects that were originally submerged in the background clearly visible in the generated defect enhancement image. This reduces the difficulty for subsequent deep learning detection networks to identify defects and significantly improves the detection recall and average precision of micro-defects.

[0032] 2. This invention decouples background and defects based on the inherent physical characteristics of wafer images, rather than relying solely on data-driven implicit learning. Therefore, even when faced with wafer images taken from different batches, under different process conditions, or under different lighting conditions, as long as the frequency characteristics of their periodic background texture remain consistent or detectable, the background can be robustly suppressed, effectively preventing the model from overfitting to a specific background pattern and significantly improving the model's generalization ability.

[0033] 3. The frequency domain decoupled multi-period texture contrast enhancement module proposed in this invention, as an independent front-end preprocessing unit, can be seamlessly integrated into any existing deep learning-based detection framework, whether it is a detector based on a convolutional neural network (CNN) or a Transformer architecture. It does not require modification of the original detection network structure or training method, reducing the cost and complexity of technology deployment and application. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0035] A deep learning-based intelligent detection and classification method for wafer defects, the overall process of which includes the following steps:

[0036] 1. Obtain the original wafer grayscale image;

[0037] 2. The original wafer grayscale image is input into the frequency domain decoupled multi-period texture contrast enhancement module for preprocessing to obtain a defect enhancement map;

[0038] 3. Input the enhanced defect map into a pre-trained deep learning detection network to classify and locate defects;

[0039] 4. Output the detection results.

[0040] The frequency domain decoupling multi-period texture contrast enhancement module described in step 2 has the following internal processing flow:

[0041] Module structure and processing flow

[0042] Step 2.1: Multi-scale Pyramid Decomposition of the Image To effectively capture the periodic textures of different sizes and frequencies that may exist in the wafer image, this step first decomposes the input raw wafer image... A Gaussian pyramid is constructed. By performing successive Gaussian blurring and smoothing downsampling operations on the original image, a series of image layers with different resolutions are generated. .in, Indicates the number of layers in the pyramid, for example, when When the image is at its original size, This refers to an image that has undergone one downsampling, such as reducing its size by half, and so on. The goal is to ensure that subsequent frequency domain processing can cover all levels of periodic texture, from fine high-resolution layers to coarse low-resolution layers, thereby achieving more comprehensive background suppression.

[0043] Step 2.2: Frequency Domain Transformation and Background Peak Detection for each layer of image obtained in Step 2.1 Perform the following sub-steps:

[0044] Sub-step 2.2.1: Apply Fast Fourier Transform to the current layer image By applying the two-dimensional fast Fourier transform (2D-FFT), its frequency spectrum is obtained by transforming it from the spatial domain to the frequency domain. In the frequency domain, periodic background textures in an image appear as discrete, highly concentrated peaks on the spectrum, while aperiodic defects appear as more diffuse, low-energy broadband components.

[0045] Sub-step 2.2.2: First, calculate the spectrum. amplitude spectrum To facilitate analysis, the amplitude spectrum is centered so that the low-frequency components are located at the center of the spectrum. Then, an adaptive sliding window detector is used to identify the frequency peaks representing the periodic background. This detector eliminates the central DC component and excessively high-frequency noise, sliding within a ring-shaped region centered on the origin of the spectrum. The detection window. For each point within the window. If its amplitude Greater than the magnitude of all other points within its current window times (e.g.) If the value is 5, then this point is preliminarily identified as a potential background peak point.

[0046] To further and more accurately distinguish between true periodic background peaks and occasional high-frequency noise, this invention introduces a verification mechanism based on the Hough transform. For the initially identified set of potential peak points, these points are projected into Hough space. If these peak points form distinct straight lines or concentric circles in Hough space, and exhibit an evenly spaced distribution pattern in the frequency domain, they are ultimately confirmed as periodic background peak points. In this way, the final set of background peak points is constructed. This adaptive recognition mechanism does not require manual setting of specific frequency values ​​and can dynamically adjust according to the image content.

[0047] Step 2.3: Construction of Adaptive Notch Filter and Background Suppression

[0048] Sub-step 2.3.1: Notch filter design for each periodic background peak detected in step 2.2.2. This invention constructs a corresponding notch filter. A notch filter aims to suppress specific frequency components within a small neighborhood centered on a peak point, while preserving other frequencies. This invention preferably employs a Gaussian notch filter, the mathematical expression of which is: ,in Representing frequency domain points to peak point Euclidean distance, It is the cutoff frequency or notch radius of the notch filter, used to control the width of the suppression region. The selection needs to be adaptively set according to the periodic fluctuation range of the wafer process, for example... It can be set to 2 pixels to cover the peak point and the small amount of frequency diffusion around it. Compared with the ideal notch filter, the Gaussian notch filter has a smoother transition band and can effectively avoid the ringing effect introduced during the inverse transform.

[0049] Sub-step 2.3.2: Combined Filtering. All notch filters designed for individual background peaks are combined into a single overall notch filter through a product operation. This combined filter exhibits low-passband suppression at all identified periodic background frequencies, while providing high-passband transmission in other frequency regions. The combined filter is then applied to the original spectrum. Perform filtering: The filtered spectrum The frequency components of the periodic background were significantly suppressed, while the frequency components corresponding to the defect signal were preserved to the maximum extent.

[0050] Step 2.4: Frequency Domain Residual Learning and Defect Augmentation

[0051] Sub-step 2.4.1: Calculating the residual spectrum. To maximize the highlighting of the defect signal, this invention calculates the difference between the original spectrum and the filtered spectrum to generate the residual spectrum: Its physical meaning lies in: because Since most of the periodic background energy has been suppressed by the notch filter, the residual spectrum is obtained by subtracting the spectrum containing the suppressed background from the original spectrum. This primarily includes the background energy "removed" by the notch filter from the original spectrum, as well as unsuppressed or minimally suppressed aperiodic defect signals. More importantly, this subtraction operation maximizes the frequency difference between the background and the defect, making the defect signal more prominent in energy relative to the background.

[0052] Sub-step 2.4.2: Inverse transform and spatial domain reconstruction of the residual spectrum Perform an inverse Fourier transform (IFFT) to convert it back to the spatial domain, and obtain the defect enhancement residual map. To obtain a realistic image, the real part of the inverse transform result is usually taken. In the original image, the periodic background texture has been significantly reduced, while the relative intensity and contrast of the micro-defect areas have been significantly enhanced, making the defects stand out clearly from the background.

[0053] Step 2.5: Multi-scale feature fusion

[0054] The defect-enhanced residual maps generated for each layer of the pyramid in step 2.4.2 are used. An upsampling operation is performed to restore the resolution to the original input image size. Then, these upsampled residual images are concatenated along the channel dimension to form a multi-channel feature map. Finally, through a... The convolutional layer performs channel fusion and dimensionality reduction on the concatenated feature maps to generate the final single-channel enhanced feature map. .Should The image contains rich multi-scale defect information, and the periodic background has been effectively suppressed. It will be used as input to the subsequent deep learning detection network, replacing the original image.

[0055] The frequency-domain decoupled multi-period texture contrast enhancement module proposed in this invention is based on the coordinated operation of the following five key components:

[0056] 1. Multi-scale frequency domain decomposition unit: By using Gaussian pyramids to separate periodic backgrounds and potential defects at different resolution levels, it ensures that periodic textures of various scales can be effectively processed.

[0057] 2. Adaptive background peak detection unit: Combining sliding window detection and Hough transform, it intelligently and without human intervention identifies discrete high-energy peaks in the spectrum that represent periodic background.

[0058] 3. Notch filter suppression unit: For the identified background frequencies, a precise notch filter is constructed and applied to suppress them, so as to preserve non-periodic defect frequencies to the maximum extent.

[0059] 4. Frequency Domain Residual Learning Unit: By calculating the difference between the original spectrum and the filtered spectrum, the contrast between the background and the defect is explicitly amplified in the form of residuals, highlighting the defect signal.

[0060] 5. Multi-scale fusion unit: Integrates defect enhancement information obtained from processing at different scales to generate a more comprehensive and more salient enhancement feature map.

[0061] Example 1

[0062] This embodiment uses the detection of micro-interruption defects in a DRAM memory array as an example to illustrate the implementation process of the present invention in detail. DRAM wafers typically have highly repeating and periodic grid-like circuit patterns.

[0063] Step 1: Data Preparation

[0064] First, a dataset of grayscale images of DRAM wafers containing micro-interruption defects was collected. The images were uniformly sized to 256×256 pixels and represented using 8-bit grayscale. Additionally, if required by the detection network, a batch of defect-free wafer images was prepared as normal samples for background learning during training. All images underwent brightness normalization.

[0065] Step 2: Construct a frequency domain decoupled multi-period texture contrast enhancement module. In a deep learning framework, such as PyTorch, the frequency domain decoupled multi-period texture contrast enhancement module is implemented as an independent component of the data preprocessing pipeline.

[0066] Pyramid Construction: Set the number of layers in the Gaussian pyramid to L=3. For the input 256x256 pixel original image... Perform Gaussian blur and 2x downsampling operations, for example, using... The kernel, with a standard deviation of 1, yields a 128x128 pixel image. .right Repeat the same operation to obtain a 64x64 pixel image. .

[0067] FFT and Adaptive Peak Detection: For each image layer The complex spectrum is obtained by applying 2D Fast Fourier Transform. Subsequently, the amplitude spectrum was calculated. And shift the zero-frequency component to the center of the spectrum. For adaptive background peak identification, set the sliding window size. Pixel, Energy Multiplier Threshold =5. Excluding the central region with a radius less than 5 pixels and the high-frequency noise region at the edges with a radius greater than half the image size, iterate through each point. If a certain point... amplitude Greater than its Points whose amplitude is five times that of other points in the neighborhood are marked as candidate peak points. Next, the coordinates of these candidate peak points are input into the Hough transform algorithm. For the grid pattern of a DRAM wafer, periodic background peaks typically appear in the frequency domain as symmetrical points about the origin, located on a straight line passing through the origin and spaced equally. The Hough transform detects whether these points are collinear and verifies their equidistant nature by calculating the distance between them. If these conditions are met, they are confirmed as periodic background peak points and added to the set. For example, an accumulator threshold can be set in Hough space to identify straight lines with a sufficient number of collinear peak points.

[0068] Notch filtering: for a set For each background peak point in the image, construct a Gaussian notch filter. The notch radius... Set to 2 pixels. Combine all single-notch filters into a single filter through a product operation. Then, Acting on the original spectrum Obtain the filtered spectrum .

[0069] Frequency Domain Residuals and Inverse Transform: Calculating the Residual Spectrum .right Perform an inverse Fourier transform and take the real part of the result to obtain the defect enhancement residual map. .

[0070] Multi-scale fusion: and Upsampled to 256x256 pixels using bilinear interpolation. The features are concatenated along the channel dimension to form a 3-channel feature map. Finally, through a... Convolutional layers reduce the dimensionality of the 3-channel feature map to a single-channel final enhanced feature map. .

[0071] Step 3: Training and Testing of Deep Learning Detection Network

[0072] Network selection: A pre-trained ResNet-50 was selected as the backbone network, and a standard Faster R-CNN detector was constructed by combining the Feature Pyramid Network (FPN) and the Region Proposal Network (RPN).

[0073] Training: The augmented map generated in step 2 The input to Faster R-CNN is used for training. During training, the parameters of the frequency domain decoupled multi-period texture contrast enhancement module are fixed, used only in the preprocessing stage, and do not participate in backpropagation; they only update the parameters of the Faster R-CNN detection network. A stochastic gradient descent (SGD) optimizer is used with an initial learning rate of 0.005, a batch size of 8, and training for 20 epochs.

[0074] Comparative experiment: A control group was set up, in which the original wafer image was directly input into the Faster R-CNN detector with the same configuration and training parameters for training.

[0075] Results Analysis: The performance of the two models was evaluated on independent test sets. The mean accuracy (mAP) of the method of this invention is expected to be 5%-10% higher than that of the control group, and the recall rate for small-sized, low-contrast defects such as micro-broken lines will be significantly improved, for example, by 10%-15%.

[0076] Example 2

[0077] This embodiment aims to demonstrate the effectiveness of the present invention in detecting scratches and particle defects on the surface of logic chip wafers, and employs different deep learning detection networks. The circuit patterns of logic chips are typically more complex than those of DRAM, and their periodicity may be less regular, but strong local periodic structures still exist.

[0078] Step 1: Data Preparation. Collect a batch of logic chip wafer images containing scratches and particle defects. The image size is uniformly 512×512 pixels, 8-bit grayscale. Normalize the images.

[0079] Step 2: Constructing a frequency domain decoupled multi-period texture contrast enhancement module. Similar to Example 1, this module is constructed, but the parameters are fine-tuned according to the characteristics of the logic chip.

[0080] Pyramid construction: The number of pyramid layers L=4 is set to accommodate potentially more complex periodic structures and a wider range of defect scales, resulting in four layers of images: 512x512, 256x256, 128x128, and 64x64.

[0081] FFT and Adaptive Peak Detection: FFT is performed on each image layer. In adaptive background peak recognition, considering the potential complexity due to the periodicity of the logic chip, the sliding window size is adjusted. Adjusted to Pixel, Energy Multiplier Threshold The value was adjusted to 4 to slightly relax the peak identification criteria and capture more possible periodic frequencies. The verification mechanism of the Hough transform is still used to ensure that the identified peaks correspond to the true periodic structure.

[0082] Notch filtering: notch radius The resolution was adjusted to 3 pixels. This is because variations in the logic chip's manufacturing process can cause slight fluctuations in the periodic frequency, and a slightly larger notch radius can more robustly cover these fluctuations.

[0083] Frequency domain residuals and inverse transform: The process is the same as in Example 1.

[0084] Multi-scale fusion: Upsampled to 512x512 pixels, and Perform channel splicing. Finally, through a... Convolutional layers are fused into a single-channel enhanced feature map. .

[0085] Step 3: Training and Testing of Deep Learning Detection Network

[0086] Network selection: YOLOv5 large model was chosen as the detection network because it balances speed and accuracy and is suitable for real-time detection scenarios.

[0087] Training: The augmented map generated in step 2 The training was performed using the YOLOv5 as input. The optimizer was Adam, with an initial learning rate of 0.01, a batch size of 16, and a training duration of 100 epochs.

[0088] Comparative experiment: A control group was set up, and the original images were directly input into the YOLOv5 detector with the same configuration and training parameters for training.

[0089] Results Analysis: The detection performance of the two models for scratches and particles was compared on independent test sets. The method of this invention is expected to outperform the control group in both mean precision and recall for both scratch and particle defects, especially in low-contrast scratch detection, where the recall rate is expected to improve by more than 10%.

[0090] Example 3

[0091] This embodiment explores the application of the present invention in the detection of defects in extremely low contrast, and combines it with a lighter-weight deep learning network to verify its effectiveness in resource-constrained scenarios.

[0092] Step 1: Data Preparation. Acquire a batch of wafer images containing minute bridging defects. The image size is uniformly 1024×1024 pixels, 16-bit grayscale, to capture finer grayscale information. Normalize the images and convert them to 8 bits.

[0093] Step 2: Construct a frequency domain decoupled multi-period texture contrast enhancement module. Optimize the module parameters to address extreme low contrast defects and higher image resolution.

[0094] Pyramid construction: The number of pyramid layers L=5 is set to process richer frequency information in high-resolution images, resulting in five layers of images: 1024x1024, 512x512, 256x256, 128x128, and 64x64.

[0095] FFT and Adaptive Peak Detection: Perform FFT on each image layer. In adaptive background peak detection, maintain the sliding window size. However, the energy multiplier threshold K was adjusted to 6. Increasing the K value was to more rigorously screen out the highest-energy, most typical periodic background peaks, avoiding the false suppression of weak frequency components that may be related to defects. The Hough transform still plays a crucial role here, ensuring that the identified peaks do indeed correspond to periodic structures on the wafer, for example, by detecting at least six collinear and equally spaced candidate peaks to confirm a periodic line.

[0096] Notch filtering: Because minute bridging defects may be close to some high-frequency background noise frequencies, the notch radius... Set to 1.5 pixels to achieve finer and narrower frequency suppression, maximizing the protection of frequency information related to defects.

[0097] Frequency Domain Residuals and Inverse Transform: Calculating the Residual Spectrum .right Perform an inverse Fourier transform and take the real part of the result to obtain the defect enhancement residual map. .

[0098] Multi-scale fusion: Upsampled to 1024x1024 pixels using bicubic interpolation. These upsampled residual images are then concatenated along the channel dimension to form a 5-channel feature map. Finally, through a... The convolutional layer fuses the 5-channel feature maps into a single-channel final enhanced feature map. .

[0099] Step 3: Training and Testing of Deep Learning Detection Network

[0100] Network Selection: Considering the high sensitivity required for detecting extremely low-contrast defects and the potential deployment on edge devices, this embodiment selects the RetinaNet detector based on a lightweight backbone network (such as MobileNetV3-Large). RetinaNet can effectively handle small targets and has relatively high computational efficiency.

[0101] Training: The augmented map generated in step 2 The input to RetinaNet is used for training. During training, the parameters of the frequency domain decoupled multi-period texture contrast enhancement module remain fixed. The optimizer uses SGD with an initial learning rate of 0.0025, a batch size of 4, and a training duration of 50 epochs. To address the sparsity of subtle bridging defects, focal loss is used during training to balance positive and negative samples.

[0102] Comparative experiment: A control group was set up, in which the original images were directly input into a RetinaNet detector with the same configuration and training parameters for training.

[0103] Results Analysis: The detection performance of the two models for subtle bridging defects was compared on independent test sets. The proposed method is expected to show a significant improvement of over 15% in recall and F1 score for subtle bridging defects compared to the control group, particularly in improving the ability to identify defects with ambiguous boundaries. Parameter Tuning Suggestions:

[0104] The number of pyramid layers, L, should be set based on the minimum and maximum size range of the periodic pattern in the wafer image. If the image resolution is high and the periodic structure is complex, the number of layers can be increased to capture frequency information at more scales.

[0105] Energy multiple threshold for adaptive peak detection : A higher value indicates a more stringent identification of background peaks, meaning only the most energetic typical periodic background is suppressed. This is suitable for cases where the periodicity is very regular and the defects do not interfere with the main periodic frequency. Conversely, a lower value indicates a more stringent identification of background peaks. The value will identify more frequency components as background, potentially suppressing some weak periodic or high-frequency noise associated with defects, but it may also inadvertently impair defect signals. It is recommended to perform a grid search or Bayesian optimization on the validation set to find the optimal balance.

[0106] notch radius Notch radius This determines the width of the suppression frequency for each notch filter. For wafer patterns with very regular periods and high frequency concentration, The size should be small, such as 1 to 2 pixels, to avoid over-suppression; for wafers with slight periodic fluctuations or frequency diffusion, The size can be increased appropriately, for example, by 3 to 5 pixels, to ensure complete coverage of background frequencies. Similarly, this requires careful adjustment on the validation set through experimentation.

[0107] Hough Transform Parameters: Parameters such as the accumulator threshold, angle, and distance resolution of the Hough transform should be adjusted according to the geometric characteristics of the wafer pattern to ensure accurate identification of straight lines or circular patterns representing periodic structures. For example, for the grid pattern of a DRAM wafer, the angular resolution can be set to a small value, such as 0.5 degrees, the distance resolution can be set to 1 pixel, and the accumulator threshold is set according to the number and intensity of periodic peaks in the image to filter out inconspicuous straight lines.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the content of the present invention's specification shall also be included within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent detection and classification method for wafer defects, characterized in that, Includes the following steps: a. Obtain the original wafer grayscale image; b. Input the original wafer grayscale image into the frequency domain decoupled multi-period texture contrast enhancement module for preprocessing to obtain a defect enhancement map; c. Input the enhanced defect map into a pre-trained deep learning detection network to classify and locate defects; d. Output the detection results.

2. The intelligent detection and classification method for wafer defects based on deep learning according to claim 1, characterized in that, The internal processing flow of the frequency domain decoupling multi-period texture contrast enhancement module in step b includes: b1. Image Multi-Scale Pyramid Decomposition: Construct a Gaussian pyramid on the original wafer grayscale image to obtain image layers of different scales. ,in Indicates the number of levels in the pyramid; b2. Frequency Domain Transformation and Background Peak Detection: For each image layer... Perform a Fast Fourier Transform to obtain its spectrum. And calculate the amplitude spectrum. Then, an adaptive sliding window detector and Hough transform are used to identify discrete peak points representing the periodic background in the amplitude spectrum, forming a peak point set. ; b3. Adaptive Notch Filter Construction and Background Suppression: For the aforementioned set of peak points Construct a notch filter at each peak point. And combine all the individual notch filters into a total filter. Then the total filter Applied to the spectrum Obtain the filtered spectrum ; b4. Frequency Domain Residual Learning and Defect Augmentation: Calculating the Original Spectrum With the filtered spectrum The difference value is used to obtain the residual spectrum. Then, the residual spectrum Performing an inverse Fourier transform yields the defect-enhanced residual map in the spatial domain. ; b5. Multi-scale feature fusion: Enhance the residual maps of defects generated from each layer of the pyramid. Upsample to the original input image size, and then... Convolution is used to fuse channels and generate the final defect enhancement map. .

3. The intelligent detection and classification method for wafer defects based on deep learning according to claim 2, characterized in that, In step b1, the Gaussian pyramid is generated by performing continuous Gaussian blurring and smoothing downsampling operations on the original image.

4. The intelligent detection and classification method for wafer defects based on deep learning according to claim 2, characterized in that, In step b2, the adaptive sliding window detector, within a ring-shaped region centered on the origin of the spectrum, for a point... If its amplitude Larger than its neighborhood The magnitude of all other points within If the value is multiples of a certain value, it is determined to be a potential peak point; the Hough transform is used to verify whether the potential peak points form a straight line or a concentric circle equally spaced distribution pattern in the frequency domain.

5. The intelligent detection and classification method for wafer defects based on deep learning according to claim 4, characterized in that, The neighborhood The size is Pixel; the The coefficient of multiple The value is 5.

6. The intelligent detection and classification method for wafer defects based on deep learning according to claim 2, characterized in that, In step b3, the notch filter A Gaussian notch filter is used, and its expression is: ,in It is a point to peak point Euclidean distance, It is the notch radius.

7. The intelligent detection and classification method for wafer defects based on deep learning according to claim 6, characterized in that, The notch radius The value is 2 pixels.

8. The intelligent detection and classification method for wafer defects based on deep learning according to claim 2, characterized in that, In step b4, the residual spectrum After performing the inverse Fourier transform, the real part of the result is taken as the defect enhancement residual map. .

9. The intelligent detection and classification method for wafer defects based on deep learning according to claim 2, characterized in that, In step b5, the upsampling operation uses bilinear interpolation or bicubic interpolation, and the channel fusion is performed by stitching together the data along the channel dimension, followed by a... Convolutional layer implementation.

10. The intelligent detection and classification method for wafer defects based on deep learning according to claim 1, characterized in that, The deep learning detection network is a detector based on a convolutional neural network or a Transformer architecture.

Images