Wafer defect detection background suppression and defect feature enhancement method
By employing frequency domain background reconstruction and removal techniques and an autoencoder network with fused attention, the problem of background interference in wafer defect detection was solved, achieving efficient background suppression and defect feature enhancement, thereby improving detection accuracy and speed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-14
- Publication Date
- 2026-06-23
AI Technical Summary
In existing technologies for wafer defect detection, severe background interference leads to increased model computation and reduced inference speed, making it unsuitable for the real-time detection needs of industrial production lines. Furthermore, research on background suppression and defect enhancement is insufficient.
By employing frequency domain background reconstruction and removal techniques, combined with an autoencoder network that integrates attention, we can achieve accurate localization and feature weighting enhancement of defect regions. Furthermore, through an adaptive balance adjustment strategy, we can synergistically optimize background suppression and defect enhancement.
It improves the accuracy and efficiency of wafer defect detection, adapts to the real-time detection needs of industrial production lines, and achieves high-precision background suppression and defect feature enhancement.
Smart Images
Figure CN122265334A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wafer defect detection technology, and in particular to a method for background suppression and defect feature enhancement in wafer defect detection. By frequency domain background reconstruction and removal, autoencoded defect enhancement with fused attention, adaptive balance adjustment and quantitative evaluation, the method achieves accurate suppression of backgrounds such as periodic textures and uniform noise in wafer images and significant enhancement of defect features such as scratches and dirt, thereby improving the accuracy and efficiency of wafer defect detection. Background Technology
[0002] Wafer defect detection is a core process for detecting micro- or nano-defects on or inside wafers and ensuring chip yield.
[0003] Wafer defect detection is a core quality inspection process in semiconductor manufacturing. It refers to the technical process of using optical, image processing, machine vision, artificial intelligence and other technologies to scan and analyze silicon wafers after each process of wafer production, such as photolithography, etching, deposition and polishing, to identify various physical or process defects on or inside the wafer surface, and to accurately detect and determine the location, size, type and quantity of defects.
[0004] In actual wafer manufacturing, the background of wafer images is highly interfering and overlaps significantly with defect features.
[0005] Wafer manufacturing operates on a mass production model, with a single inspection device processing thousands of wafer images with a resolution of 1024×1024 or higher per day. This requires the inspection algorithm to possess both sub-micron level inspection accuracy and millisecond-level single-image processing speed. Directly inputting the raw images into the inspection model without prior background suppression and defect enhancement significantly increases the model's computational load and reduces inference speed. Furthermore, background interference increases the difficulty of model training and decreases accuracy, making it unsuitable for the real-time inspection needs of industrial production lines.
[0006] Background separation, background suppression, and background removal are classic research directions in the fields of machine vision and image analysis. After decades of development, a complete technical system has been formed. However, research on background separation specifically for wafer defect detection images is seriously lacking. Summary of the Invention
[0007] To address the shortcomings of existing technologies, this invention provides a method for background suppression and defect feature enhancement in wafer defect detection. Facing severe background interference such as periodic textures and uniform noise in wafer defect detection images, this method first employs frequency domain background reconstruction and removal techniques to obtain a preliminary background-removed image. Then, based on a pixel-level annotated defect dataset, an autoencoder network architecture incorporating a spatial attention mechanism is designed to achieve precise localization and weighted feature enhancement of defect regions, while simultaneously suppressing residual background noise. Building upon this, a dynamic adjustment strategy for both background and defect regions is constructed to adaptively balance and fine-tune the enhancement parameters, thereby achieving synergistic optimization of background suppression and defect enhancement in wafer defect detection images. This approach provides high-quality image preprocessing support for high-precision wafer defect detection tasks, effectively improving the accuracy and efficiency of wafer defect detection.
[0008] The technical solution of this invention is: a method for background suppression and defect feature enhancement in wafer defect detection, comprising the following steps:
[0009] S1) Frequency Domain Background Reconstruction and Removal
[0010] In wafer defect detection scenarios, the image background is mostly periodic texture or uniform noise, which manifests as low-frequency components in the frequency domain, while defects are mostly edges or abrupt texture changes, manifesting as high-frequency components. Frequency domain filtering is used to separate the background from the defects, achieving accurate background removal.
[0011] S101) Data Acquisition
[0012] For the five types of defect images in wafer inspection scenarios, all wafer images are standardized to PNG format and archived in folders according to defect category.
[0013] S102) Data Preprocessing
[0014] To ensure a fixed resolution of 1024×1024 for each image, the impact of size differences on annotation accuracy is eliminated. A unified image input format is implemented, mapping three-dimensional color channels to one-dimensional grayscale channels. This eliminates interference from color and pixel value scale differences on subsequent frequency domain transformation and feature extraction, adapting to the industrial scene characteristics of wafer grayscale inspection images.
[0015] S103) Frequency Domain Transformation and Background Estimation
[0016] The spatial domain image is transformed to the frequency domain, and low-frequency components are extracted to reconstruct the background image, achieving accurate background estimation and providing a foundation for subsequent background removal. This step is the core of frequency domain background suppression, adapting to the low-frequency distribution characteristics of periodic textures or uniform noise in the wafer background.
[0017] S104) Image difference yields a preliminary background-removed image.
[0018] By performing pixel-by-pixel differencing on the image, the background component in the original normalized image is removed, while the high-frequency feature components of the defect are preserved, resulting in a preliminary de-background image and achieving initial background suppression. The differencing operation directly eliminates the pixel value contribution of the background, highlighting the grayscale difference between the defect and the background.
[0019] S2) Enhancement of the deficiency in fusion attention
[0020] To address the issues of residual background noise and weak defect features in the initial background removal image, a training dataset is constructed through pixel-level annotation. By combining spatial attention mechanism and autoencoder, the defect region is accurately located and its features are weighted and enhanced. At the same time, residual background noise is suppressed, thus completing the secondary enhancement of defect features.
[0021] S201) Sample labeling
[0022] We construct a pixel-level defect region annotation dataset to provide supervision information for attention-integrated autoencoders, enabling accurate segmentation of defect regions and background regions, and adapting to the characteristics of small size and irregular shape of wafer defects.
[0023] S202), Enhancement of autoencoding defects in fusion attention mechanisms
[0024] Based on an autoencoder framework, a spatial attention mechanism is embedded to achieve feature extraction, localization, and weighted enhancement of defect regions through supervised learning, while suppressing residual background noise and outputting images with enhanced defect features.
[0025] S3), Adaptive Balance Adjustment
[0026] A dynamic adjustment strategy for the background and defect dual regions is constructed. By adaptively adjusting the brightness and contrast, the "degree of background purification" and "degree of defect feature prominence" are balanced to achieve a visual effect of "bright and blurred background and dark and clear defect", while ensuring that the brightness is controllable under the premise of enhancing the contrast of defects.
[0027] S301), Background-Defect Dual-Region Dynamic Parameter Adjustment
[0028] Based on the defect region localization results obtained in the fusion attention stage, differentiated brightness and blur adjustments are performed on the background region and the defect region to achieve visual weakening of the background and visual highlighting of the defect.
[0029] S302), Fine-tuning of weights and enhancement parameters
[0030] The parameters such as brightness coefficient, contrast coefficient, and blur kernel size in S301 are adaptively fine-tuned to ensure that while the contrast of defects is enhanced, the brightness of the defect area remains within a controllable range.
[0031] S4) Quantitative evaluation indicators
[0032] The results of the original image after processing through steps S1 to S3 are comprehensively evaluated from both quantitative and qualitative perspectives to verify the effectiveness of the proposed method in background suppression and defect feature enhancement.
[0033] S401) Quantitative Indicators
[0034] The effect of background suppression and defect feature enhancement is evaluated using defect contrast (CR). Based on the defect region and background region divided by pixel-level mask, the gray-scale mean of the two regions is calculated, and the contrast is quantified by the ratio of the mean values.
[0035] S402), Qualitative effect
[0036] By visually comparing the original image with the image results after processing through steps S1, S2, and S3, observe whether the background is clean and whether the defect edges are clear. Attached Figure Description
[0037] Figure 1 This is a flowchart of a method for background suppression and defect feature enhancement in wafer defect detection. Detailed Implementation
[0038] The following is in conjunction with the appendix Figure 1 The present invention provides a method for background suppression and defect feature enhancement in wafer defect detection, comprising the following steps:
[0039] S1) Frequency Domain Background Reconstruction and Removal
[0040] In wafer defect detection scenarios, the image background is mostly periodic texture or uniform noise, which manifests as low-frequency components in the frequency domain, while defects are mostly edge or texture abrupt changes, which manifest as high-frequency components.
[0041] By utilizing the frequency domain distribution difference between the background and defects in the wafer inspection image, the spatial domain image is mapped to the frequency domain through Fourier transform, achieving accurate separation of the background and defects. Finally, the background is removed through differential operation to obtain a preliminary background-removed image.
[0042] S101) Data Acquisition
[0043] For five typical defects in wafer inspection scenarios, namely dirt, poor etching, cracks, poor substrate, and peeling residue, a total of 365 defect images were collected from real industrial sites.
[0044] All wafer images are standardized to PNG format and archived in folders according to defect category.
[0045] S102) Data Preprocessing
[0046] Ensure that the resolution of each image is fixed at 1024×1024 to eliminate the impact of size differences on annotation accuracy.
[0047] By unifying the image input format, mapping the three-dimensional color channels to one-dimensional grayscale channels, traversing all acquired images, performing weighted calculations on the color images to obtain grayscale images, and finally obtaining a set of grayscale images with uniform size and channels.
[0048] Pixel value normalization maps pixel values to the [0,1] interval through a linear transformation.
[0049] S103) Frequency Domain Transformation and Background Estimation
[0050] A Fast Fourier Transform (FFT) is performed on the normalized image to obtain the frequency domain matrix. Spectrum centering is then applied to shift the low-frequency components to the center of the image for easier filtering. A Gaussian low-pass filter is then used to extract the low-frequency components in the frequency domain. An inverse Fourier transform is then performed on the obtained low-frequency components to obtain the spatial background image.
[0051] The discrete spatial image is converted into a frequency domain complex matrix using the Fast Fourier Transform (FFT), decomposing the image's brightness and texture information into components of different frequencies. This yields the frequency domain complex matrix F(u,v) and the corresponding spectrum |F(u,v)|.
[0052] Performing the above cyclic shift operation on the frequency domain complex matrix F(u,v) yields the centered frequency domain matrix F. shift (u,v), where the central region of the spectrum is the low-frequency component and the edge region is the high-frequency component.
[0053] A Gaussian low-pass filter (GLPF) extracts low-frequency components, calculates the center coordinates of the centered frequency domain matrix, sets the optimal σ based on the background texture characteristics of the wafer image, and generates a frequency domain matrix with F. shift An element-wise multiplication operation is performed on Gaussian low-pass filters H(u,v) of the same size as (u,v) to obtain a frequency domain background matrix F containing only low-frequency components. bg (u,v).
[0054] S104) Image difference yields a preliminary background-removed image.
[0055] The normalized image and the reconstructed background image are subjected to pixel-by-pixel absolute value difference to output a preliminary background-removed image. At this point, background noise in the image is significantly suppressed, and defect features are initially highlighted, but residual background noise and weak defect features may still exist.
[0056] S2) Enhancement of the deficiency in fusion attention
[0057] To address the issues of residual background noise and weak defect features in the initial background removal image, a training dataset is constructed through pixel-level annotation. By combining spatial attention mechanism and autoencoder, the defect region is accurately located and its features are weighted and enhanced. At the same time, residual background noise is suppressed, thus completing the secondary enhancement of defect features.
[0058] S201) Sample labeling
[0059] Pixel-level category division is achieved through binary masks, dividing image pixels into defect regions and background regions of no interest, providing labels for region weight allocation in the attention mechanism and feature learning of the autoencoder.
[0060] For five typical defects in wafer inspection scenarios—dirt, poor etching, cracks, substrate defects, and residual peeling—LabelMe polygon annotation tool was used to precisely annotate the defect areas of each image pixel by pixel. During annotation, closed contours were drawn point by point along the defect edges. When an image contained multiple defects, each defect was drawn and matched with its corresponding label. Background and areas of no interest did not require separate annotation and were classified as non-defect areas by default. After annotation, a JSON format annotation file with the same name as the original image was generated, recording information such as defect contour coordinates and label categories. This file was stored in the same path as the original image to ensure a one-to-one correspondence between files.
[0061] S202), Enhancement of autoencoding defects in fusion attention mechanisms
[0062] Based on an autoencoder framework and embedding a spatial attention mechanism, this method achieves feature extraction, localization, and weighted enhancement of defect regions through supervised learning, while suppressing residual background noise and outputting an image with enhanced defect features. The autoencoder is responsible for the nonlinear mapping and reconstruction of features, while the spatial attention mechanism is responsible for assigning high weights to defect regions.
[0063] An autoencoder consists of an encoder and a decoder. The encoder maps the input image to a low-dimensional feature space through convolutional layers, extracting the core features of defects. The decoder maps the low-dimensional features back to the original-size image through deconvolutional layers, reconstructing the defect features. By minimizing the loss between the input image and the reconstructed image, unsupervised learning of features is achieved, adapting to the complex feature extraction of wafer defects.
[0064] The encoder performs multiple rounds of convolution, activation, and pooling operations on the initial background-removed input image to gradually reduce the feature map size, increase the number of channels, and extract low-dimensional defect features.
[0065] To meet the feature extraction requirements of wafer defects, the encoder undergoes five rounds of downsampling to achieve a size reduction from 1024 to 512 to 256 to 128 to 64 to 32.
[0066] The spatial attention module is the core innovation of this network. It is embedded in the middle layer of the decoder and is used to generate a 1024×1024×1 spatial weight map. It assigns high weights to defect regions and low weights to background regions, thereby enhancing defect features and suppressing background noise.
[0067] The decoder performs multiple rounds of deconvolution and activation operations on the low-dimensional features to gradually restore the feature map size, reduce the number of channels, and reconstruct the image features.
[0068] The decoder and encoder have a symmetrical structure. The core of the decoder is upsampling, ReLU activation, skip connections and spatial attention module embedding. By gradually increasing the spatial size of the feature map and reducing the number of channels, the shallow detail features of the encoder are fused by skip connections, and the features of the defect area are enhanced by the spatial attention module. Finally, a 1024×1024×1 defect enhancement image is reconstructed.
[0069] The decoder includes 5 upsampling operations, achieving a size increment of 32→64→128→256→512→1024, which corresponds one-to-one with the 5 coding blocks of the encoder.
[0070] S3), Adaptive Balance Adjustment
[0071] A dynamic adjustment strategy for the background and defect dual regions is constructed. By adaptively adjusting the brightness and contrast, the "degree of background purification" and "degree of defect feature prominence" are balanced to achieve a visual effect of "bright and blurred background and dark and clear defect", while ensuring that the brightness is controllable under the premise of enhancing the contrast of defects.
[0072] S301), Background-Defect Dual-Region Dynamic Parameter Adjustment
[0073] Based on the defect region localization results obtained in the fusion attention stage, differentiated brightness and blur adjustments are performed on the background region and the defect region to achieve visual weakening of the background and visual highlighting of the defect.
[0074] The image is divided into a background region and a defect region by pixel-level region segmentation. Brightness adjustment and blurring operations are performed on the two regions respectively. The blurring operation adopts Gaussian blur, while the background region is smoothed at the same time.
[0075] Linear brightness enhancement is applied to the pixel values of the background area, followed by Gaussian blur to suppress texture noise in the background area.
[0076] Linear brightness reduction is applied to the pixel values of the defective area, followed by contrast enhancement, to further improve the grayscale difference between the defective edge and the interior.
[0077] The adjusted background and defect areas are then fused pixel-by-pixel using a mask to obtain a preliminary adaptively adjusted image.
[0078] S302), Fine-tuning of weights and enhancement parameters
[0079] The parameters such as brightness coefficient, contrast coefficient, and blur kernel size in S301 are adaptively fine-tuned to ensure that while the contrast of defects is enhanced, the brightness of the defect area remains within a controllable range.
[0080] S4) Quantitative evaluation indicators
[0081] The results of the original image after processing through steps S1 to S3 are comprehensively evaluated from both quantitative and qualitative perspectives to verify the effectiveness of the proposed method in background suppression and defect feature enhancement.
[0082] S401) Quantitative Indicators
[0083] The effect of background suppression and defect feature enhancement is evaluated using defect contrast (CR). Based on the defect region and background region divided by pixel-level mask, the gray-scale mean of the two regions is calculated, and the contrast is quantified by the ratio of the mean values.
[0084]
[0085] Where, μ def μ is the mean gray value of the defect region in the optimal image. bg This represents the average grayscale value of the background area.
[0086] S402), Qualitative effect
[0087] By visually comparing the original image with the image results after processing through steps S1, S2, and S3, observe whether the background is clean and whether the defect edges are clear.
[0088] The above embodiments and description are intended only to illustrate the core concept and preferred implementation of the present invention. It is clearly stated that, without departing from the basic concept and scope of protection of the present invention, there are many possible variations and optimizations. All such variations and improvements are considered to be included within the scope of protection defined by the claims of the present invention.
Claims
1. A method for background suppression and defect feature enhancement in wafer defect detection, characterized in that, Includes the following steps: S1) Frequency Domain Background Reconstruction and Removal In wafer defect detection scenarios, the image background is mostly periodic texture or uniform noise, which manifests as low-frequency components in the frequency domain, while defects are mostly edges or abrupt texture changes, manifesting as high-frequency components. Frequency domain filtering is used to separate the background from the defects, achieving accurate background removal. S101) Data Acquisition To address five typical defects in wafer inspection scenarios, a total of 365 defect images were collected from real industrial sites. S102) Data Preprocessing A unified image input format is used to eliminate the interference of color and pixel value scale differences on subsequent frequency domain transformation and feature extraction, thus adapting to the industrial scene characteristics of wafer grayscale inspection images. S103) Frequency Domain Transformation and Background Estimation A Fast Fourier Transform (FFT) is performed on the normalized image to obtain the frequency domain matrix. Spectrum centering is then applied to shift the low-frequency components to the center of the image for easier filtering. A Gaussian low-pass filter is then used to extract the low-frequency components in the frequency domain. An inverse Fourier transform is then performed on the obtained low-frequency components to obtain the spatial background image. S104) Image difference yields a preliminary background-removed image. By performing pixel-by-pixel differencing on the image, the background component in the original normalized image is removed while the high-frequency feature components of the defects are retained, resulting in a preliminary background-removed image and achieving initial background suppression. S2) Enhancement of the deficiency in fusion attention Based on the initial denoised image, the spatial attention mechanism and autoencoder are combined to achieve accurate localization and feature weighting enhancement of the defect region, while suppressing residual background noise and completing the secondary enhancement of defect features. S201) Sample labeling LabelMe software was used to complete pixel-level binary mask annotation, achieving accurate segmentation of defective regions (labeled with a value of 1) and background regions of no interest (labeled with a value of 0) in the training dataset. S202), Enhancement of autoencoding defects in fusion attention mechanisms Based on an autoencoder framework, a spatial attention mechanism is embedded to achieve feature extraction, localization, and weighted enhancement of defective regions through supervised learning. S3), Adaptive Balance Adjustment A comprehensive evaluation index is constructed to balance the "degree of background purification" and the "degree of prominence of defect features," and the parameters at each stage are adaptively adjusted to output the optimal image. S301), Background-Defect Dual-Region Dynamic Parameter Adjustment Based on the defect region localization results obtained in the fusion attention stage, differentiated brightness and blur adjustments are performed on the background region and the defect region. S302), Fine-tuning of weights and enhancement parameters The parameters such as brightness coefficient, contrast coefficient, and blur kernel size in S301 are adaptively fine-tuned to ensure that while the contrast of defects is enhanced, the brightness of the defect area remains within a controllable range. S4) Quantitative evaluation indicators To verify the effectiveness of the plan, it is necessary to evaluate it from both quantitative indicators and qualitative effects. S401) Quantitative Indicators The difference in grayscale mean between the defect region and the background region is calculated using the defect contrast ratio (CR) to evaluate the effectiveness of background suppression and defect feature enhancement. S402), Qualitative effect By visually comparing the original image with the image results after processing through steps S1, S2, and S3, observe whether the background is clean and whether the defect edges are clear.