A method for real-time defect detection of silicon carbide epitaxial growth

Abstract

This invention relates to silicon carbide defect detection and discloses a real-time defect detection method for silicon carbide epitaxial growth. It improves detection efficiency based on a layered imaging strategy and a three-dimensional attention mechanism. The layered imaging strategy performs a global initial inspection to quickly screen defect-free samples and avoid ineffective fine-tuning. Switching to high-resolution, small-field-of-view imaging acquires local images, providing targeted data for subsequent detection and improving efficiency. The three-dimensional attention mechanism strengthens key dimensions through three correlation angles: global statistical pooling, local peak pooling, and temporal trend pooling. It combines dimensional mutual information to mine correlations and generates dimensional attention weights through dual-branch weight fusion, suppressing redundancy and strengthening key features. Image preprocessing optimizes local image quality, and temporal attention is introduced into a CNN-LSTM hybrid model to obtain spatiotemporal features. A lightweight classifier outputs defect categories and confidence levels to improve real-time defect detection efficiency.

Description

Technical Field

[0001] This invention relates to silicon carbide defect detection, and more specifically to a real-time defect detection method for silicon carbide epitaxial growth. Background Technology

[0002] Silicon carbide is widely used in high-temperature, high-pressure, and high-frequency applications. Defects in its epitaxial growth process directly affect device performance and reliability. Therefore, real-time and accurate defect detection is a key step in silicon carbide production.

[0003] Traditional silicon carbide epitaxial defect detection methods are mostly offline detection methods, which have problems such as detection lag, inability to provide real-time feedback of process parameters, and difficulty in achieving full-process quality control. At the same time, traditional defect detection methods often rely solely on spatial or temporal features, without fully integrating the correlation between the two to improve the ability of feature extraction and screening, and lack efficient attention mechanisms, making it easy for key defect features to be masked by redundant information.

[0004] Therefore, designing a real-time defect detection method for silicon carbide epitaxial growth to improve the efficiency and accuracy of real-time defect detection of silicon carbide epitaxial wafers is of great practical significance. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a real-time defect detection method for silicon carbide epitaxial growth. Through a layered imaging detection strategy of global imaging for initial inspection and local imaging for fine inspection, real-time defect detection is achieved during the silicon carbide epitaxial growth process. At the same time, an innovative three-dimensional attention mechanism is introduced into the defect detection model to improve the accuracy of defect detection.

[0006] The technical solution to achieve the objective of this invention is as follows:

[0007] A method for real-time defect detection in silicon carbide epitaxial growth includes the following steps:

[0008] The X-ray imaging unit acquires global images of the silicon carbide epitaxial growth process using a low-resolution, full-field-of-view imaging method during the initial stage of silicon carbide epitaxial growth. And a global image of silicon carbide epitaxial growth. Defect detection was performed to obtain preliminary global inspection results; subsequently, based on the preliminary global inspection results, the epitaxial growth region was determined, and local images of silicon carbide epitaxial growth were acquired using high-resolution, small-field-of-view imaging. ;

[0009] An adaptive median filtering algorithm is used to analyze local images of silicon carbide epitaxial growth. Noise suppression is performed to obtain a denoised local image. Denoising local images based on adaptive histogram equalization Adjusting the grayscale distribution eliminates local grayscale unevenness, resulting in a local grayscale image. The Laplacian operator is used to process local grayscale images. Edge enhancement is performed to obtain an enhanced local image. To highlight the outline of the defect;

[0010] Collecting continuous Frame of local images of silicon carbide epitaxial growth And obtained through image preprocessing Frame Enhancement Local Image Let it be the input sequence A temporal attention mechanism is introduced to calculate the attention weights of each frame and then weighted to obtain a weighted input sequence. The weighted input sequence is fed into a CNN-LSTM hybrid model, and spatial features are extracted from the weighted input sequence in the CNN sub-network. Temporal features are obtained by extracting features from the weighted input sequence in the LSTM subnetwork. Spatial features With time series characteristics Fusion yields fusion characteristics Based on fusion features Global statistical pooling, local peak pooling, and temporal trend pooling are used to obtain multi-dimensional vectors. Inter-dimensional mutual information and a dual-branch attention network are introduced to obtain optimized features. ; optimize features The input is fed into a lightweight classifier, which outputs the final detection result.

[0011] Furthermore, the X-ray imaging unit employs a layered imaging strategy, acquiring global images of silicon carbide epitaxial growth through low-resolution, full-field-of-view imaging. A pre-trained YOLOv8 model was used to analyze the global image of silicon carbide epitaxial growth. Defect detection is performed to obtain a global preliminary inspection result; this result is then evaluated to avoid subsequent invalid inspections; if no defects are detected in the global preliminary inspection, the X-ray imaging unit switches to a high-resolution, small-field-of-view imaging mode to acquire local images of the silicon carbide epitaxial growth. ;

[0012] Furthermore, local images of silicon carbide epitaxial growth. Image preprocessing is performed to obtain enhanced local images. This includes the following steps:

[0013] An adaptive median filtering algorithm is used to dynamically adjust the window size based on the image noise density, while an edge protection mechanism is introduced to protect the local image of silicon carbide epitaxial growth. Noise suppression is performed to obtain a denoised local image. ;

[0014] Denoising Local Images Based on Adaptive Histogram Equalization The image is divided into blocks and grayscale correction is performed. Then, the boundary bilinear interpolation method is used to eliminate the boundaries of the sub-blocks in the equalized image to obtain the grayscale local image. ;

[0015] Using the Laplacian operator on grayscale local images While performing edge enhancement, an edge threshold is set to prevent over-enhancement of other areas, resulting in an enhanced local image. To highlight the outline of the defect;

[0016] Furthermore, local images of silicon carbide epitaxial growth are obtained using an X-ray imaging unit. At the same time, considering that the silicon carbide epitaxial growth process is affected by the environment, resulting in local images of the silicon carbide epitaxial growth process. The system contains noise, including airflow noise, vibration noise, and electronic noise; based on this, an adaptive window selection and edge protection mechanism are used to effectively suppress noise and preserve details.

[0017] First, a local image of silicon carbide epitaxial growth. Divided into A uniform non-overlapping sub-block The ratio between the gray-level variance and the gray-level mean within each sub-block is calculated as the noise density of the sub-block; for example, for the Sub-block The average gray value is calculated based on the gray values ​​of all pixels within the sub-block and denoted as . The gray-scale variance corresponding to this sub-block is then calculated and denoted as . The noise density is obtained by calculating the ratio between the gray-level variance and the gray-level mean, denoted as . Noise density The larger the size, the more likely it is to be a sub-block. The more severe the internal noise; then according to The size of the filter window used by different sub-blocks is dynamically adjusted to account for noise density. Low-noise regions with noise density less than or equal to the noise density threshold are used The size of the filtering window is used to process the sub-blocks; for noise density In high-noise regions where the noise density exceeds the threshold, the window size varies with the noise density. Linear increase;

[0018] An edge protection mechanism is introduced based on the adaptive median filtering algorithm. Before processing pixels through the filtering window, the average gray-level difference between the center pixel and its neighboring pixels is first calculated. When the average gray-level difference is greater than or equal to the gray-level threshold, the original gray-level value of the pixel is directly retained without median filtering; when the average gray-level difference is less than the gray-level threshold, median filtering is performed. The final denoised local image is denoted as... ;

[0019] Furthermore, based on adaptive histogram equalization, denoising local images are... The image is divided into blocks and grayscale correction is performed. Then, the boundary bilinear interpolation method is used to eliminate the sub-block boundaries in the equalized image.

[0020] First, denoise the local image. Convert to grayscale and follow the steps in the previous step. A uniform non-overlapping sub-block The denoised local image is obtained by dividing it in a certain way. On A uniform non-overlapping sub-block ; then regarding the first Sub-block The original grayscale histogram is obtained by counting the number of 256 gray levels within the sub-block. The grayscale cropping threshold for a sub-block is calculated based on the number of pixels within that sub-block; subsequently, for each grayscale level within that sub-block, when... Greater than When the grayscale level exceeds the limit, all the excess portion is cropped out to obtain the total grayscale level excess; the total grayscale level excess is then calculated. The grayscale histogram is obtained by uniformly distributing the grayscale values ​​across 256 grayscale levels. The cropped grayscale histogram is standardized, and the cumulative probability of each grayscale level is calculated. Then, based on the grayscale levels... The cumulative probability of the grayscale value within the sub-block Correction is performed to obtain the corrected grayscale value. ;

[0021] Perform the above operations on a sub-block basis to eliminate noise in local images. The local gray-level unevenness phenomenon was initially corrected in the image. Because the grayscale adjustment range varies among different sub-blocks, abrupt grayscale changes occur at the boundaries of each sub-block, leading to issues with the initial corrected image. Sub-block boundaries exist; a bilinear interpolation method is used to calculate the interpolated grayscale value of the boundary pixels of adjacent sub-blocks according to distance weights. Interpolation is performed on the boundaries of all sub-blocks to obtain the final grayscale local image. ;

[0022] Furthermore, the Laplacian operator is used to process local grayscale images. Edge enhancement is performed, while edge threshold constraints are applied to avoid amplifying noise.

[0023] First, the Laplacian operator is used to process the local grayscale image. The edge image is obtained by convolutional calculation of each pixel in the image, using a Laplacian convolution kernel of type 1. The Laplacian convolution kernel calculates the difference between the center pixel and its neighboring pixels in the top, bottom, left, right, and four diagonal directions. It is more sensitive to edges, but also more sensitive to background noise.

[0024] To avoid the impact of background noise on edge enhancement, the Laplacian operator is used to process local grayscale images. Edge threshold constraints are introduced at the time; based on grayscale local images gray mean Calculate the corresponding edge threshold, which is set to the grayscale mean. Multiply by 0.1; during edge image calculation, only pixels whose convolution result is greater than the edge threshold are retained as valid edges, and the grayscale value of the remaining pixels is set to 0; the edge image With grayscale local image Superimposed and fused to obtain enhanced local images In preserving local grayscale images Enhance the defect edges based on the background and internal grayscale information of the defect;

[0025] Furthermore, outputting the defect type and corresponding confidence level using the CNN-LSTM hybrid model includes the following steps:

[0026] Collecting continuous A frame of local images of silicon carbide epitaxial growth is obtained, and noise reduction, grayscale correction, and edge enhancement are performed through the aforementioned three image preprocessing steps. The frame-enhanced local image is denoted as the input sequence. A temporal attention mechanism is introduced, which calculates and normalizes the Pearson correlation coefficient between frames to obtain the attention weights. The weighted input sequence is obtained by multiplying the enhanced local image of each frame in the input sequence with the corresponding attention weight. ;

[0027] Weighted input sequence The input is fed into a CNN-LSTM hybrid model, where the CNN and LSTM work in parallel. Spatial features are obtained by extracting features from the weighted input sequence within the CNN sub-network. Temporal features are obtained by extracting features from the weighted input sequence in the LSTM subnetwork. Spatial features With time series characteristics Fusion yields fusion characteristics ;

[0028] Based on fusion features Global statistical pooling, local peak pooling, and temporal trend pooling are used to obtain multi-dimensional vectors. Inter-dimensional mutual information and a dual-branch attention network are introduced to obtain optimized features. ;

[0029] Optimize features The input is fed into a lightweight classifier, which outputs the final detection result.

[0030] Furthermore, a temporal attention mechanism is introduced, which assigns attention weights by calculating the feature correlation between different frames; for the frame The previous frame As the reference frame, and Flattening them respectively yields the eigenvectors, denoted as... and Calculate the Pearson correlation coefficient between the two. The degree of silicon carbide epitaxial growth defect evolution between adjacent frames is represented by the Pearson correlation coefficient of each frame in the input sequence calculated according to the above method. Then, the attention weight corresponding to each frame is generated by linear normalization. Then the current frame Compared with the previous reference frame The weighted input sequence is obtained by weighting. ;

[0031] Furthermore, the weighted input sequence The input is fed into a CNN-LSTM hybrid model, where the CNN and LSTM work in parallel. Spatial features are obtained by extracting features from the weighted input sequence within the CNN sub-network. ; Weighted input sequence Each frame is flattened to obtain the corresponding feature vector, let the first frame be denoted as the first frame. frame The corresponding feature vector is The corresponding dimension is Then, all frames are averaged and fused to obtain the single-frame fused spatial feature input. ;

[0032] Input single-frame fused spatial features Spatial features are extracted from the input into the CNN subnetwork. The CNN subnetwork consists of two convolutional layers and one fully connected layer, where the kernel size of the first convolutional layer is [size missing]. It is used to capture fine-grained features such as defect edges and textures, and the feature signals are enhanced by the ReLU activation function; through a The max pooling layer reduces the dimensionality of the features before feeding them into the second convolutional layer, whose kernel size is [missing value]. This covers a wider pixel range to aggregate mesoscale morphological information; it is then input again into a... The dimensionality is further reduced in the max pooling layer, and finally input into a fully connected layer to filter out the semantic features most critical for defect classification and output them as spatial features. , dimension Each dimension of a spatial feature corresponds to a spatial attribute;

[0033] Weighted input sequence in LSTM subnetwork Temporal features are obtained by feature extraction. ,pass The convolutional kernel reduces the dimensionality of the weighted input sequence, and then an average pooling operation is used to compress it to obtain the temporal feature sequence. , dimension The LSTM subnetwork consists of two LSTM hidden layers and one fully connected layer. The two LSTM hidden layers capture the long-term and short-term evolution trends of defects during silicon carbide epitaxial growth. Each of the two LSTM hidden layers has 128 hidden units. Subsequently, the output of the LSTM layers is compressed through a fully connected layer to obtain the temporal features. , dimension Each dimension of the temporal feature corresponds to a temporal attribute;

[0034] Spatial features output by the CNN subnetwork Temporal characteristics of LSTM subnetwork output By splicing and blending, the fusion features are obtained. The dimension of the fused features is ;

[0035] Furthermore, considering that the fused features obtained by direct splicing and fusion have insufficient feature discrimination, causing key defect features to be masked by secondary features and redundancy between dimensions, a three-level dimensional attention mechanism is designed to filter key features in the fused features and suppress the interference of redundant dimensions.

[0036] First, based on fusion features Global statistical pooling is used to calculate the average value of all features within a dimension as the global response intensity for each dimension. This is because the features are fused. The dimension is Therefore, global statistical pooling is actually a single-element summation, the first... Global response strength of dimension Values ​​and fusion features No. The eigenvalues ​​of each dimension have a one-to-one correspondence;

[0037] To compensate for the neglect of local details by global statistical pooling, a 3-element sliding window method is used, taking the maximum value within the window to generate the first element. Local peak components of dimension To enhance the local features of the defect;

[0038] For the temporal feature component of the fused features, temporal trend pooling is used to capture the dynamic changes in dimensions during defect evolution. The mean of the inter-frame feature change rate is used to calculate the dimensionality of the first feature. Time-series trend components of dimension ;

[0039] After obtaining the first through the above three-dimensional attention mechanism Global response strength of dimension Local peak components and time series trend components After defining three components, a min-max normalization method is used to map the three components to the [0,1] interval to eliminate dimensional differences. Then, the three components are concatenated to obtain a first component. Multidimensional components of a dimension After obtaining 512 multi-dimensional components using the above method, the fused features are obtained by splicing them together. Corresponding multi-dimensional vector Its dimensions are ;

[0040] Subsequently, the correlation between the feature distributions of the two dimensions is further measured by calculating the mutual information between the two dimensions, based on the dimension. Corresponding multi-dimensional components The probability is obtained by discretizing the data into 8 intervals using histogram estimation and then calculating the frequency. Calculate the corresponding single-dimensional information entropy Subsequently, based on multi-dimensional components The component in falls in the first intervals and The same dimension component falls in the first Dimensions of joint frequency calculation for each interval and dimensions Two-dimensional joint information entropy Then calculate the dimensions. and dimensions mutual information After calculating the mutual information between dimensions, the average mutual information for each dimension is calculated. ;

[0041] Then, dimensional attention weights are obtained through a dual-branch attention network. The convolution kernel will convert multi-dimensional vectors Dimensional compression to Then, the vector values ​​are mapped to the [0,1] interval using the Sigmoid activation function to obtain the dimension-adaptive weights, denoted as... The mutual information weights are obtained by mapping the average mutual information to the [0,1] interval using the Sigmoid function. Adaptive weights for dimensions Mutual information weights Fusion generation of dimensional attention weights The fused features are then weighted dimension by dimension to enhance key features and obtain optimized features. Optimize features The dimension is ;

[0042] Furthermore, we will optimize the features. The input is fed into a lightweight classifier, which outputs the final detection result; the optimized features are then processed through a fully connected layer. The mapping is performed as low-dimensional features, and the confidence scores of various defects in the local image of silicon carbide epitaxial growth are obtained through a softmax function; based on the confidence vector... The maximum confidence level is selected as the corresponding defect category and corresponding confidence level as the final detection result;

[0043] Compared with existing technologies, this invention improves real-time detection efficiency based on a layered imaging strategy and a three-dimensional attention mechanism. The layered imaging strategy uses low-resolution full-field imaging for global initial detection, quickly screening defect-free samples and avoiding ineffective fine detection; then it switches to high-resolution small-field imaging to acquire local images, providing targeted data for subsequent detection and improving detection efficiency. The three-dimensional attention mechanism strengthens key dimensions through three correlation angles: global statistical pooling, local peak pooling, and temporal trend pooling. It combines dimensional mutual information to mine correlation relationships and generates dimensional attention weights through dual-branch weight fusion, suppressing redundancy and strengthening key features. Combined with image preprocessing to optimize local image quality, temporal attention is introduced into the CNN-LSTM hybrid model to obtain spatiotemporal features, and a lightweight classifier outputs defect categories and confidence scores to further improve real-time defect detection efficiency. Attached Figure Description

[0044] Figure 1 Flowchart of a real-time defect detection method for silicon carbide epitaxial growth;

[0045] Figure 2 Flowchart of a local image preprocessing method for silicon carbide epitaxial growth;

[0046] Figure 3This is a flowchart of the defect detection process using a CNN-LSTM hybrid model.

[0047] Figure 4 A flowchart for obtaining optimized features based on a three-dimensional attention mechanism. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] like Figure 1 As shown in the figure, a specific embodiment of the present invention discloses a method for real-time defect detection in silicon carbide epitaxial growth, comprising the following steps:

[0050] The X-ray imaging unit acquires global images of the silicon carbide epitaxial growth process using a low-resolution, full-field-of-view imaging method during the initial stage of silicon carbide epitaxial growth. And a global image of silicon carbide epitaxial growth. Defect detection was performed to obtain preliminary global inspection results; subsequently, based on the preliminary global inspection results, the epitaxial growth region was determined, and local images of silicon carbide epitaxial growth were acquired using high-resolution, small-field-of-view imaging. ;

[0051] An adaptive median filtering algorithm is used to analyze local images of silicon carbide epitaxial growth. Noise suppression is performed to obtain a denoised local image. Denoising local images based on adaptive histogram equalization Adjusting the grayscale distribution eliminates local grayscale unevenness, resulting in a local grayscale image. The Laplacian operator is used to process local grayscale images. Edge enhancement is performed to obtain an enhanced local image. To highlight the outline of the defect;

[0052] Collecting continuous Frame of local images of silicon carbide epitaxial growth And obtained through image preprocessing Frame Enhancement Local Image Let it be the input sequence A temporal attention mechanism is introduced to calculate the attention weights of each frame and then weighted to obtain a weighted input sequence. The weighted input sequence is fed into a CNN-LSTM hybrid model, and spatial features are extracted from the weighted input sequence in the CNN sub-network. Temporal features are obtained by extracting features from the weighted input sequence in the LSTM subnetwork. Spatial features With time series characteristics Fusion yields fusion characteristics Based on fusion features Global statistical pooling, local peak pooling, and temporal trend pooling are used to obtain multi-dimensional vectors. Inter-dimensional mutual information and a dual-branch attention network are introduced to obtain optimized features. ; optimize features The input is fed into a lightweight classifier, which outputs the final detection result.

[0053] Furthermore, the X-ray imaging unit employs a layered imaging strategy, acquiring global images of silicon carbide epitaxial growth through low-resolution, full-field-of-view imaging. A pre-trained YOLOv8 model was used to analyze the global image of silicon carbide epitaxial growth. Defect detection is performed to obtain a preliminary global inspection result. This result is then evaluated. If a defect is detected, it indicates the silicon carbide has a defect, and no further operations are performed. If no defect is detected, the X-ray imaging unit switches to high-resolution, small-field-of-view imaging to acquire local images of the silicon carbide epitaxial growth. By using low-resolution global initial screening, samples with known defects can be quickly removed, avoiding the waste of time on invalid samples in subsequent high-resolution imaging and feature extraction. At the same time, the global initial screening determines the epitaxial growth region, making the subsequent continuous frames more targeted and improving the efficiency of defect detection.

[0054] Furthermore, such as Figure 2 As shown, considering the local images of silicon carbide epitaxial growth obtained by the X-ray imaging unit during the silicon carbide epitaxial growth process. Quality issues arise due to the silicon carbide growth environment, thus requiring image preprocessing; local images of silicon carbide epitaxial growth are analyzed. Image preprocessing is performed to obtain enhanced local images. This includes the following steps:

[0055] An adaptive median filtering algorithm is used to dynamically adjust the window size based on the image noise density, while an edge protection mechanism is introduced to protect the local image of silicon carbide epitaxial growth. Noise suppression is performed to obtain a denoised local image. ;

[0056] Denoising Local Images Based on Adaptive Histogram Equalization The image is divided into blocks and grayscale correction is performed. Then, the boundary bilinear interpolation method is used to eliminate the boundaries of the sub-blocks in the equalized image to obtain the grayscale local image. ;

[0057] Using the Laplacian operator on grayscale local images While performing edge enhancement, an edge threshold is set to prevent over-enhancement of other areas, resulting in an enhanced local image. To highlight the outline of the defect;

[0058] Furthermore, local images of silicon carbide epitaxial growth are obtained using an X-ray imaging unit. At the same time, considering that the silicon carbide epitaxial growth process is affected by the environment, resulting in local images of the silicon carbide epitaxial growth process. The image contains noise, including airflow noise, vibration noise, and electronic noise. In traditional median filtering algorithms, a fixed window is used to process the image. When a small window is used, it can only suppress electronic noise and cannot eliminate large-sized noise generated by airflow noise and vibration noise. When a large window is used, it will blur the details of small-sized noise. Based on this, an adaptive window selection and edge protection mechanism are used to effectively suppress noise and preserve details.

[0059] First, a local image of silicon carbide epitaxial growth. Divided into A uniform non-overlapping sub-block The ratio between the gray-level variance and the gray-level mean within each sub-block is calculated as the noise density of the sub-block; for example, for the Sub-block The average gray value is calculated based on the gray values ​​of all pixels within the sub-block and denoted as . The gray-scale variance corresponding to this sub-block is then calculated and denoted as . The noise density is obtained by calculating the ratio between the gray-level variance and the gray-level mean, denoted as . The calculation formula is as follows:

[0060] ;

[0061] Noise density The larger the size, the more likely it is to be a sub-block. The more severe the internal noise; then according to The size of the filter window used by different sub-blocks is dynamically adjusted; for example, if the noise density threshold is set to 0.25, the noise density... Low-noise regions with noise density less than or equal to the noise density threshold are used The size of the filtering window is used to process the sub-blocks; for noise density In high-noise regions where the noise density exceeds the threshold, the window size varies with the noise density. It increases linearly, as shown in the following formula:

[0062] ,

[0063] in, Indicates window size. Indicates rounding down;

[0064] Considering the significant variation in pixel grayscale values ​​at defect edges, directly applying median filtering would blur the defect edges, affecting subsequent defect detection. Therefore, an edge protection mechanism is introduced based on the adaptive median filtering algorithm. Before processing pixels through the filtering window, the average grayscale difference between the center pixel and its neighboring pixels is calculated. When the average grayscale difference is greater than or equal to the grayscale threshold, the original grayscale value of the pixel is retained without median filtering; when the average grayscale difference is less than the grayscale threshold, median filtering is performed.

[0065] Local images of silicon carbide epitaxial growth obtained through the above steps. After adaptive median filtering, image noise was suppressed, and defect edge details were preserved through an edge protection mechanism, resulting in a denoised local image, denoted as . ;

[0066] Furthermore, during the silicon carbide epitaxial growth process, the X-ray imaging unit exhibits systematic errors such as uneven intensity distribution, leading to localized errors in the acquired images of the silicon carbide epitaxial growth process. There is a problem with grayscale differences; based on this, adaptive histogram equalization is used to denoise local images. The image is divided into blocks and grayscale correction is performed. Then, the boundary bilinear interpolation method is used to eliminate the sub-block boundaries in the equalized image.

[0067] First, denoise the local image. Convert to grayscale and follow the steps in the previous step. A uniform non-overlapping sub-block The denoised local image is obtained by dividing it in a certain way. On A uniform non-overlapping sub-block ; then regarding the first Sub-block The original grayscale histogram is obtained by counting the number of 256 gray levels within the sub-block. The grayscale cropping threshold for a sub-block is calculated based on the number of pixels within that sub-block. The calculation formula is as follows:

[0068] ,

[0069] in, Indicates the first Sub-block The clipping threshold, This is for cropping and limiting the histogram box height, which is the ratio of the histogram box height to the ideal uniform distribution histogram. Indicates a sub-block The number of pixels within the sub-block; subsequently, for each gray level within that sub-block, when Greater than When the grayscale level exceeds the limit, all excess parts are cropped out to obtain the total grayscale excess. The calculation formula is as follows:

[0070] ;

[0071] Total grayscale level excess The grayscale histogram is obtained by uniformly distributing the grayscale values ​​across 256 grayscale levels. The cropped grayscale histogram is standardized, and the cumulative probability of each grayscale level is calculated using the following formula:

[0072] ,

[0073] in, Represents grayscale level The cumulative probability; then based on the gray level The cumulative probability of the grayscale value within the sub-block Correction is performed to obtain the corrected grayscale value. The calculation formula is as follows:

[0074] ;

[0075] Perform the above operations on a sub-block basis to eliminate noise in local images. The local gray-level unevenness phenomenon was initially corrected in the image. Because the grayscale adjustment range varies among different sub-blocks, abrupt grayscale changes occur at the boundaries of each sub-block, leading to issues with the initial corrected image. Sub-block boundaries exist; using bilinear interpolation, the interpolated grayscale value is calculated for the boundary pixels of adjacent sub-blocks based on distance weights. The calculation formula is as follows:

[0076] ,

[0077] in, Indicates adjacent sub-blocks and sub-blocks The interpolated boundary pixel values, Indicates a sub-block The grayscale of the boundary pixels, Indicates a sub-block The grayscale of the boundary pixels, Indicates the current pixel to the sub-block Distance from the center Indicates the current pixel to the sub-block The distance from the center; the boundaries of all sub-blocks are interpolated using the bilinear interpolation method described above to obtain the final grayscale local image. ;

[0078] Furthermore, the grayscale local image obtained after denoising and grayscale correction The edges of defects in images can become blurred, making it difficult for defect detection models to accurately extract defect edge features. Furthermore, traditional edge enhancement algorithms not only enhance defect edges but also amplify residual noise in the background. Therefore, this paper employs the Laplacian operator to refine the grayscale local image. Edge enhancement is performed, while edge threshold constraints are applied to avoid amplifying noise.

[0079] First, the Laplacian operator is used to process the local grayscale image. The edge image is obtained by convolving each pixel in the image, and the calculation formula is as follows:

[0080] ,

[0081] in, This represents the edge image after Laplacian convolution. This indicates the Laplacian convolution kernel used, specifically:

[0082] ,

[0083] This Laplacian convolution kernel calculates the difference between the center pixel and its neighboring pixels in the top, bottom, left, right, and four diagonal directions. It is more sensitive to edges, but also more sensitive to background noise.

[0084] To avoid the impact of background noise on edge enhancement, the Laplacian operator is used to process local grayscale images. Edge threshold constraints are introduced at the time; based on grayscale local images gray mean Calculate the corresponding edge threshold, which is set to the grayscale mean. Multiply by 0.1; during edge image calculation, only pixels whose convolution result is greater than the edge threshold are retained as valid edges, and the grayscale value of the remaining pixels is set to 0; the edge image With grayscale local image Superimposed and fused to obtain enhanced local images In preserving local grayscale images Enhance the defect edges based on the background and internal grayscale information of the defect;

[0085] By preprocessing local images of silicon carbide epitaxial growth to optimize image quality, the original images are transformed into enhanced local images with high recognition, providing high-quality input data for the CNN-LSTM hybrid model, effectively extracting defect features, and improving the defect detection accuracy.

[0086] Furthermore, such as Figure 3 As shown, the steps for outputting the defect type and corresponding confidence level using a CNN-LSTM hybrid model include:

[0087] Collecting continuous A frame of local images of silicon carbide epitaxial growth is obtained, and noise reduction, grayscale correction, and edge enhancement are performed through the aforementioned three image preprocessing steps. The frame-enhanced local image is denoted as the input sequence. A temporal attention mechanism is introduced, which calculates and normalizes the Pearson correlation coefficient between frames to obtain the attention weights. The weighted input sequence is obtained by multiplying the enhanced local image of each frame in the input sequence with the corresponding attention weight. ;

[0088] Weighted input sequence The input is fed into a CNN-LSTM hybrid model, where the CNN and LSTM work in parallel. Spatial features are obtained by extracting features from the weighted input sequence within the CNN sub-network. Temporal features are obtained by extracting features from the weighted input sequence in the LSTM subnetwork. Spatial features With time series characteristics Fusion yields fusion characteristics ;

[0089] Based on fusion features Global statistical pooling, local peak pooling, and temporal trend pooling are used to obtain multi-dimensional vectors. Inter-dimensional mutual information and a dual-branch attention network are introduced to obtain optimized features. ;

[0090] Optimize features The input is fed into a lightweight classifier, which outputs the final detection result.

[0091] Furthermore, continuous data collection A frame of local images of silicon carbide epitaxial growth is obtained, and noise reduction, grayscale correction, and edge enhancement are performed through the aforementioned three image preprocessing steps. The frame-enhanced local image is denoted as the input sequence. Because the evolution of defects in silicon carbide epitaxial growth has a strong temporal correlation, directly inputting the input sequence into a CNN-LSTM hybrid model for feature extraction would ignore this temporal correlation. Therefore, a temporal attention mechanism is introduced, which allocates attention weights by calculating the feature correlation between different frames. Specifically, for the first frame... frame The previous frame As the reference frame, and Flattening them respectively yields the eigenvectors, denoted as... and Calculate the Pearson correlation coefficient between the two. The degree of silicon carbide epitaxial growth defect evolution between adjacent frames is represented by the following formula:

[0092] ,

[0093] in, This indicates the size of the enhanced local image, i.e., the dimension of the feature vector. Representing the eigenvector The Dimension value, Representing the eigenvector The Dimension value, and respectively, feature vectors and The mean of the values ​​is calculated; after obtaining the Pearson correlation coefficient for each frame in the input sequence using the method described above, the attention weights for each frame are generated through linear normalization. The calculation formula is as follows:

[0094] ,

[0095] in, This represents the maximum value in the Pearson correlation coefficient. This represents the minimum value in the Pearson correlation coefficient; subsequently, the current frame... Compared with the previous reference frame The weighted input sequence is obtained by weighting. The calculation formula is as follows:

[0096] ;

[0097] Furthermore, the weighted input sequence The input is fed into a CNN-LSTM hybrid model, where the CNN and LSTM work in parallel. Spatial features are obtained by extracting features from the weighted input sequence within the CNN sub-network. For example, a weighted input sequence Each frame is flattened to obtain the corresponding feature vector, let the first frame be denoted as the first frame. frame The corresponding feature vector is The corresponding dimension is Then, all frames are averaged and fused to obtain the single-frame fused spatial feature input. The calculation formula is as follows:

[0098] ;

[0099] Input single-frame fused spatial features Spatial features are extracted from the input into the CNN subnetwork. The CNN subnetwork consists of two convolutional layers and one fully connected layer, where the kernel size of the first convolutional layer is [size missing]. It is used to capture fine-grained features such as defect edges and textures, and the feature signals are enhanced by the ReLU activation function; through a The max pooling layer reduces the dimensionality of the features before feeding them into the second convolutional layer, whose kernel size is [missing value]. This covers a wider pixel range to aggregate mesoscale morphological information; it is then input again into a... The dimensionality is further reduced in the max pooling layer, and finally input into a fully connected layer to filter out the semantic features most critical for defect classification and output them as spatial features. , dimension Each dimension of a spatial feature corresponds to a spatial attribute;

[0100] Weighted input sequence in LSTM subnetwork Temporal features are obtained by feature extraction. ,pass The convolutional kernel reduces the dimensionality of the weighted input sequence, and then an average pooling operation is used to compress it to obtain the temporal feature sequence. , dimension The LSTM subnetwork consists of two LSTM hidden layers and one fully connected layer. The two LSTM hidden layers capture the long-term and short-term evolution trends of defects during silicon carbide epitaxial growth, and the output of the LSTM layers is then compressed by the fully connected layer to obtain the temporal features. , dimension Each dimension of the temporal feature corresponds to a temporal attribute;

[0101] Spatial features output by the CNN subnetwork Temporal characteristics of LSTM subnetwork output By splicing and blending, the fusion features are obtained. The calculation formula is as follows:

[0102] ,

[0103] The dimension of the fusion feature is ;

[0104] Furthermore, such as Figure 4As shown, considering the insufficient feature discrimination of fused features obtained by direct splicing and fusion, which leads to key defect features being masked by secondary features and redundancy between dimensions, a three-level dimensional attention mechanism is designed from the perspectives of overall, local, and dynamic correlation to filter key features in the fused features and suppress the interference of redundant dimensions. Optimized features are obtained by utilizing inter-dimensional mutual information and a dual-branch attention network. ;

[0105] First, based on fusion features Global statistical pooling is used to calculate the average value of all features within a dimension as the global response intensity for each dimension. This is because the features are fused. The dimension is Therefore, global statistical pooling is actually a single-element summation, the first... Global response strength of dimension Values ​​and fusion features No. The eigenvalues ​​of each dimension have a one-to-one correspondence, and the calculation method is as follows:

[0106] ,

[0107] in, Indicates a dimension index;

[0108] To compensate for the neglect of local details by global statistical pooling, enhance the feature signals of local defect regions, and further narrow down the range of key dimensions, a 3-element sliding window method is adopted, taking the maximum value within the window to generate the first element. The local peak component of the dimension is calculated as follows:

[0109] ,

[0110] In this case, zeros are added at the boundary positions, i.e. ;

[0111] To address the dimensionality of the temporal feature component in the fused features, temporal trend pooling is employed to capture the dynamic changes in dimensions during defect evolution. The temporal trend component is calculated using the mean of the inter-frame feature change rate, thus distinguishing between the static background dimension and the dynamic defect dimension and identifying the temporal evolution characteristics of secondary defects. The calculation method for the time series trend component of the dimension is as follows:

[0112] ,

[0113] in, Indicates the number of frames. Representing time-series feature sequences Upper Temporal characteristics of frames; furthermore, when hour, ;

[0114] After obtaining the first through the above three-dimensional attention mechanism Global response strength of dimension Local peak components and time series trend components After defining three components, a min-max normalization method is used to map the three components to the [0,1] interval to eliminate dimensional differences. Then, the three components are concatenated to obtain a first component. Multidimensional components of a dimension The calculation formula is as follows:

[0115] ,

[0116] in, , and All are normalized. The components of the dimension; after obtaining a total of 512 multi-dimensional components through the above method, the fused features are obtained by splicing them together. Corresponding multi-dimensional vector Its dimensions are ;

[0117] After eliminating the dimensional differences of the three components through min-max normalization and concatenating them to obtain a multi-dimensional vector, this multi-dimensional vector contains the global, local, and dynamic features of the defect. Subsequently, the synergistic and redundant relationships between dimensions are mined by calculating the mutual information between dimensions. Combined with a dual-branch attention network, dimension weights are generated, and the fused features are weighted dimension by dimension to obtain optimized features that enhance the key dimensions, providing high-discrimination feature inputs for the classifier.

[0118] Subsequently, the correlation between the feature distributions of the two dimensions is further measured by calculating the mutual information between the two dimensions. The formula for calculating its corresponding single-dimensional information entropy is as follows:

[0119] ,

[0120] in, For dimension Corresponding multi-dimensional components The probability is obtained by discretizing the data into 8 intervals using the histogram estimation method and then calculating the frequency. The calculation formula is as follows:

[0121] ,

[0122] in, Representing multi-dimensional components The component in falls in the first The frequency of each interval, A smoothing factor that is a small positive number not equal to zero; then the dimension is calculated. and dimensions The two-dimensional joint information entropy is calculated using the following formula:

[0123] ,

[0124] in, Let be the joint probability of the multi-dimensional components corresponding to the two dimensions. The calculation formula is as follows:

[0125] ,

[0126] in, Representing multi-dimensional components The component in falls in the first intervals and The same dimension component falls in the first The joint frequency of each interval; then the dimension is calculated. and dimensions The mutual information is calculated using the following formula:

[0127] ;

[0128] After calculating the mutual information between dimensions, calculate the average mutual information for each dimension. To assess the synergy of each dimension, the calculation formula is as follows:

[0129] ;

[0130] Then, dimensional attention weights are obtained through a dual-branch attention network. The convolution kernel will convert multi-dimensional vectors Dimensional compression to Then, the vector values ​​are mapped to the [0,1] interval using the Sigmoid activation function to obtain the dimension-adaptive weights, denoted as... The mutual information weights are obtained by mapping the average mutual information to the [0,1] interval using the Sigmoid function. Adaptive weights for dimensions Mutual information weights Fusion generation of dimensional attention weights The fused features are then weighted dimension by dimension to enhance key features and obtain optimized features. The calculation formula is as follows:

[0131] ,

[0132] in, The branch balance coefficients of a two-branch attention network are... The temperature coefficient of the softmax function is used to optimize the features. The dimension is For example, The value is 0.5. The value is 0.8;

[0133] Furthermore, the optimized features, after being filtered through a three-dimensional attention mechanism, have had redundant information removed and key features strengthened, thus optimizing the features. The input is fed into a lightweight classifier, which outputs the final detection result; the optimized features are then processed through a fully connected layer. The confidence scores of various defects in the local image of silicon carbide epitaxial growth are obtained by mapping to low-dimensional features and using a softmax function. The calculation formula is as follows:

[0134] ,

[0135] in, A confidence vector representing various types of defects. To output the weight matrix with dimension 1 , The weight matrix is ​​trainable and has a dimension of . , The deviation term has a dimension of , Indicates the number of defect categories; based on the confidence vector The maximum confidence level is selected as the corresponding defect category and corresponding confidence level as the final detection result;

[0136] This invention discloses a real-time defect detection method for silicon carbide epitaxial growth. It improves real-time detection efficiency based on a layered imaging strategy and a three-level dimensional attention mechanism. The layered imaging strategy uses low-resolution full-field imaging for initial global detection, quickly screening defect-free samples and avoiding ineffective fine-tuning. Then, it switches to high-resolution small-field imaging to acquire local images, providing targeted data for subsequent detection and improving efficiency. The three-level dimensional attention mechanism strengthens key dimensions through three correlation angles: global statistical pooling, local peak pooling, and temporal trend pooling. It combines dimensional mutual information mining to uncover correlations and generates dimensional attention weights through dual-branch weight fusion, suppressing redundancy and strengthening key features. Image preprocessing optimizes local image quality, and temporal attention is introduced into a CNN-LSTM hybrid model to obtain spatiotemporal features. A lightweight classifier outputs defect categories and confidence scores to further improve real-time defect detection efficiency.

[0137] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time defect detection in silicon carbide epitaxial growth, characterized in that, Includes the following steps: A layered imaging strategy is used to acquire global and local images, and defect-free samples are quickly screened based on the global image. An adaptive median filtering algorithm is used to suppress noise in the local image to obtain a denoised local image; adaptive histogram equalization is used to adjust the gray-level distribution of the denoised local image to obtain a gray-level local image; and the Laplacian operator is used to enhance the edges of the gray-level local image to obtain an enhanced local image. Continuously enhanced local images are acquired as input sequences, and a temporal attention mechanism is introduced to obtain a weighted input sequence. Spatial and temporal features are extracted and fused using a CNN-LSTM hybrid model to obtain fused features. Based on the fusion features, global statistical pooling, local peak pooling, and temporal trend pooling are used to obtain multi-dimensional vectors. Inter-dimensional mutual information and a dual-branch attention network are introduced to obtain optimized features. The optimized features are input into a lightweight classifier, which outputs the final detection result. The method for calculating the weighted input sequence includes: The feature correlation is calculated using the frame preceding the current frame in the input sequence as the reference frame. The feature correlation is normalized to obtain the attention weight of each frame. The current frame and the reference frame are weighted and summed to update the input sequence to obtain the weighted input sequence. The method for obtaining a multi-dimensional vector based on fusion features using global statistical pooling, local peak pooling, and temporal trend pooling includes: calculating the average value of all features using global statistical pooling to obtain the global response intensity; using a 3-element sliding window method to take the maximum value within the window to generate the local peak component of each dimension; using temporal trend pooling to calculate the temporal trend component of each dimension based on the mean of the inter-frame feature change rate; after obtaining the three components—global response intensity, local peak component, and temporal trend component—the method uses min-max normalization to map the three components to the [0,1] interval and concatenates them to obtain the multi-dimensional vector. The process of introducing inter-dimensional mutual information and a dual-branch attention network to obtain optimized features includes: calculating the inter-dimensional mutual information and the average mutual information of each dimension of a multi-dimensional vector; generating dimensional attention weights through a dual-branch attention network; and using the dimensional attention weights to weight the fused features dimension by dimension to obtain optimized features. The calculation of inter-dimensional mutual information of a multi-dimensional vector includes: discretizing the multi-dimensional vector into 8 intervals using histogram estimation, obtaining the probability distribution based on the frequency, and calculating the single-dimensional information entropy; calculating the joint information entropy of the two dimensions based on the joint frequency of the two dimensions; and adding the two single-dimensional information entropies and subtracting the joint information entropy of the two dimensions to obtain the inter-dimensional mutual information. The generation of dimensional attention weights includes: compressing a multi-dimensional vector to one dimension and normalizing it to obtain dimensional adaptive weights; normalizing the average mutual information to obtain mutual information weights; and fusing the dimensional adaptive weights and mutual information weights through a softmax function to obtain dimensional attention weights.

2. The real-time defect detection method for silicon carbide epitaxial growth as described in claim 1, wherein the layered imaging strategy is characterized in that, Includes the following steps: The system acquires global images using low-resolution full-field imaging, uses a pre-trained model to perform defect detection on the global images to obtain preliminary inspection results, and then switches to high-resolution small-field imaging to acquire local images based on the preliminary inspection results.

3. The method for real-time defect detection in silicon carbide epitaxial growth as described in claim 1, wherein the method for acquiring the grayscale local image is characterized in that, Includes the following steps: The denoised local image is divided into blocks and grayscale correction is performed based on adaptive histogram equalization. The boundary bilinear interpolation method is used to eliminate the boundaries of the sub-blocks to obtain the grayscale local image.

4. The method for real-time defect detection in silicon carbide epitaxial growth as described in claim 1, wherein the method for acquiring enhanced local images is characterized in that, Includes the following steps: The edge image is obtained by convolving the grayscale local image with the Laplacian operator. An edge threshold constraint is introduced to filter the effective edges. The edge image is then superimposed and fused with the grayscale local image to obtain the enhanced local image.

5. The method for real-time defect detection in silicon carbide epitaxial growth as described in claim 4, wherein the method for obtaining the fusion feature is characterized in that, Includes the following steps: Spatial features are obtained by extracting features from the weighted input sequence through a CNN sub-network, and temporal features are obtained by extracting features from the weighted input sequence through an LSTM sub-network. The spatial features and temporal features are then concatenated and fused to obtain the fused features.

Images