Silicon carbide tray defect detection method and system based on image recognition

Abstract

The invention relates to the technical field of image processing, in particular to a silicon carbide tray defect detection method and system based on image recognition, and the method comprises the steps: obtaining a gray image of a silicon carbide tray to construct an observation matrix, and initializing a low-rank matrix, a sparse matrix and a Lagrange multiplier matrix for representing error accumulation; iterative optimization is executed; and repeating the iteration steps until a termination condition is met, and taking the final sparse matrix as a defect detection result. Through the technical scheme of the invention, the problems of false detection and missing detection caused by misjudging the background texture as the defect can be reduced, and the accuracy and robustness of the defect detection result of the silicon carbide tray are improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for detecting defects in silicon carbide trays based on image recognition. Background Technology

[0002] With the rapid development of third-generation semiconductor technology, silicon carbide (SiC) epitaxial wafers have been widely used in core fields such as new energy vehicles, photovoltaic inverters, and high-voltage power transmission. In the silicon carbide epitaxial growth process, the silicon carbide tray, as a key consumable material supporting the wafer, directly affects the growth quality of the wafer and the cleanliness of the reaction chamber. Because the tray is repeatedly used in high-temperature and highly corrosive environments, it is prone to cracking, leading to wafer scrap. Therefore, defect detection on the surface of the silicon carbide tray has become a necessary step in the semiconductor manufacturing equipment industry.

[0003] Currently, surface defect detection of silicon carbide trays typically employs traditional image processing methods based on machine vision. These methods utilize industrial cameras to acquire grayscale images of the tray surface, extract features by calculating pixel grayscale gradients or using edge detection operators, or perform binarization segmentation of the image based on a global threshold. They mainly rely on the significant differences in grayscale values ​​or gradient intensity between the defect area and the background area to locate and identify surface anomalies.

[0004] However, even on a silicon carbide tray surface without defects, there are texture features such as annular grooves left by processing. At the same time, silicon carbide material itself has high light absorption properties, which causes grayscale images to contain a lot of noise, making it difficult to effectively identify real crack defects from complex, highly textured backgrounds, resulting in inaccurate defect detection results. Summary of the Invention

[0005] To address the technical problem of inaccurate defect detection results for silicon carbide trays, this application provides a silicon carbide tray defect detection method and system based on image recognition, which can reduce the problem of false detection and missed detection caused by misjudging background texture as defects, and improve the accuracy and robustness of silicon carbide tray defect detection results.

[0006] In a first aspect, this application provides a method for detecting defects in silicon carbide trays based on image recognition. The method includes: acquiring a grayscale image of the silicon carbide tray to construct an observation matrix; initializing a low-rank matrix, a sparse matrix, and a Lagrange multiplier matrix for characterizing error accumulation; performing iterative optimization, including: in each iteration, updating the low-rank matrix using a singular value thresholding operator, and using the difference between the observation matrix and the low-rank matrix as a background matrix; acquiring the eigenvalues ​​of the structure tensor matrix of each pixel in the background matrix, calculating the saliency of each pixel based on the eigenvalues, and calculating an adaptive threshold for each pixel based on the saliency, wherein the adaptive threshold is negatively correlated with the saliency coefficient; performing a thresholding operation on the background matrix using the adaptive threshold to update the sparse matrix and the Lagrange multiplier matrix; repeating the above iterative steps until a termination condition is met, and using the final sparse matrix as the defect detection result.

[0007] During the iterative optimization process, an adaptive threshold is calculated based on saliency to update the sparse matrix, the background texture is determined based on the low-rank matrix, and local features are differentially filtered based on the adaptive threshold. While stripping away highly repetitive background, defect signals with specific structural features are selectively retained, thereby achieving adaptive separation of defects on the surface of silicon carbide trays and obtaining accurate defect detection results.

[0008] Preferably, the step of acquiring a grayscale image of the silicon carbide tray to construct an observation matrix includes: normalizing the grayscale image and using the normalized grayscale image as the observation matrix.

[0009] Preferably, obtaining the eigenvalues ​​of the structure tensor matrix of each pixel in the background matrix includes: calculating the horizontal gradient in the horizontal direction and the vertical gradient in the vertical direction of each pixel in the background matrix; constructing the structure tensor matrix of each pixel based on the horizontal gradient and the vertical gradient; and performing eigenvalue decomposition on the structure tensor matrix to obtain a first eigenvalue and a second eigenvalue, wherein the first eigenvalue is greater than or equal to the second eigenvalue.

[0010] The horizontal and vertical gradients of pixels in the background matrix are calculated and a structural tensor matrix is ​​constructed for eigenvalue decomposition. Since eigenvalues ​​can quantify the intensity of grayscale changes in local image texture in the main direction and its orthogonal directions, this provides a data basis for distinguishing between cracks with linear structural features and noise with isotropic characteristics.

[0011] Preferably, the saliency is the product of the linearity factor and the structural energy factor; the linearity factor characterizes the directional consistency of the local texture and is positively correlated with the difference between the first eigenvalue and the second eigenvalue; the structural energy factor characterizes the texture intensity of the local texture and is positively correlated with the sum of the first eigenvalue and the second eigenvalue.

[0012] The linearity factor characterizes the directional consistency of the texture, and the structure energy factor characterizes the texture intensity. The product of the two is used as the significance, so that only real defects that have both linear morphology and texture intensity can obtain a high significance value, thus achieving effective suppression of artifact noise that resembles cracks in shape but lacks intensity.

[0013] Preferably, the method for calculating the linearity factor includes: calculating the difference between the first eigenvalue and the second eigenvalue, and the sum of the first eigenvalue and the second eigenvalue, and using the ratio of the difference to the sum as the linearity factor; the method for calculating the structural energy factor includes: calculating the product of the sum and a preset energy sensitivity coefficient, and normalizing the product to obtain a significance coefficient.

[0014] Preferably, the adaptive threshold is calculated as follows: a basic denoising threshold is set; a threshold adjustment coefficient is calculated, wherein the threshold adjustment coefficient is negatively correlated with significance; and the product of the threshold adjustment coefficient and the basic denoising threshold is used as the adaptive threshold.

[0015] Preferably, the step of using an adaptive threshold to perform thresholding on the background matrix to update the sparse matrix includes: calculating the absolute value of each pixel in the background matrix; calculating the difference between the absolute value and the adaptive threshold of the corresponding pixel, and taking the maximum value between the difference and 0; and multiplying the maximum value by the sign of the corresponding pixel in the background matrix as the value of the corresponding pixel in the updated sparse matrix.

[0016] An adaptive threshold is used to perform thresholding on the background matrix to update the sparse matrix, so that the updated sparse matrix only retains abnormal grayscale changes that exceed the adaptive threshold, thus transforming the continuous background residual into a sparse data distribution that contains only defect information.

[0017] Preferably, the method for calculating the background matrix includes: calculating the difference between the observation matrix and the updated low-rank matrix, and the sum of the ratios of the Lagrange multiplier matrix and the penalty parameter in the previous iteration, to obtain the background matrix.

[0018] Preferably, the initialization of the low-rank matrix, the sparse matrix, and the Lagrange multiplier matrix used to characterize error accumulation includes: initializing the low-rank matrix as the observation matrix; and initializing the sparse matrix and the Lagrange multiplier matrix as all-zero matrices.

[0019] In a second aspect, this application also provides a silicon carbide tray defect detection system based on image recognition, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the silicon carbide tray defect detection method based on image recognition according to the first aspect of this application is implemented.

[0020] The technical solution of this application has the following beneficial technical effects: A decomposition model containing low-rank and sparse matrices is constructed. In the iterative optimization, the singular value threshold operator is used to remove highly repetitive background textures. At the same time, the linearity factor and structure energy factor are used to construct saliency. Then, the adaptive threshold of each pixel is determined based on the saliency, realizing the dynamic adjustment of the threshold when updating the sparse matrix. This accurately separates the complex texture background and crack defects on the surface of silicon carbide trays, effectively reducing the false detection and false negative problem of misjudging background textures as defects in strong noise and low contrast environments. This achieves highly accurate and robust surface defect detection. Attached Figure Description

[0021] Figure 1 This is a flowchart of a silicon carbide tray defect detection method based on image recognition according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of a grayscale image of a silicon carbide tray according to an embodiment of this application.

[0023] Figure 3 This is a rendering of the defect detection results of a silicon carbide tray according to an embodiment of this application.

[0024] Figure 4 This is a structural block diagram of a silicon carbide tray defect detection system based on image recognition according to an embodiment of this application. Detailed Implementation

[0025] According to a first aspect of this application, this application provides a silicon carbide tray defect detection method based on image recognition, which is used to monitor surface defects of silicon carbide trays to ensure the surface quality of silicon carbide trays. Figure 1 This is a flowchart of a silicon carbide tray defect detection method based on image recognition according to an embodiment of this application. Figure 1 As shown, the image recognition-based silicon carbide tray defect detection method includes steps S101 to S104, which are described in detail below.

[0026] S101, acquire a grayscale image of the silicon carbide tray to construct an observation matrix.

[0027] In one embodiment, a grayscale image of a silicon carbide tray is acquired using a pre-deployed industrial camera, covering the entire surface area of ​​the tray. To facilitate subsequent numerical calculations, the grayscale image needs to be standardized, i.e., normalized, and used as the observation matrix.

[0028] Understandably, grayscale values ​​in a grayscale image typically range from 0 to 255. Normalization distributes grayscale values ​​within the range of 0 to 1. Subsequently, the two-dimensional image data is reconstructed into an observation matrix. The observation matrix is ​​the same size as the grayscale image, and each value in the observation matrix corresponds to the normalized grayscale value at the corresponding location in the grayscale image, providing standardized data input for subsequent algorithms.

[0029] S102 initializes the low-rank matrix, the sparse matrix, and the Lagrange multiplier matrix used to characterize error accumulation.

[0030] In one embodiment, after establishing the observation matrix, based on robust principal component analysis theory, the observation matrix is ​​regarded as a low-rank matrix representing the background texture. and the sparse matrix representing defect noise Composition, that is ,in, For the observation matrix, It is a low-rank matrix. This is a sparse matrix. Initial values ​​need to be assigned to each matrix to initiate the iterative optimization process.

[0031] Specifically, the low-rank matrix is ​​initialized as the observation matrix; the sparse matrix and the Lagrange multiplier matrix are initialized as all-zero matrices.

[0032] Understandingly, initializing the low-rank matrix to the observation matrix itself implies that, before starting iterative optimization, it is assumed that the grayscale image consists entirely of background and contains no defect noise; initializing the sparse matrix to an all-zero matrix implies that the initial assumption is that there are no defects. The Lagrange multiplier matrix is ​​used to calculate the constraint error during iteration and is used to characterize the cumulative amount of constraint error during iteration. Its zero initialization ensures that iterative optimization can start from a feasible solution.

[0033] S103, perform iterative optimization, including: in each iteration, update the low-rank matrix using the singular value threshold operator, and use the difference between the observation matrix and the low-rank matrix as the background matrix; obtain the eigenvalues ​​of the structure tensor matrix of each pixel in the background matrix, calculate the saliency of each pixel based on the eigenvalues, and calculate the adaptive threshold of each pixel based on the saliency, wherein the adaptive threshold is negatively correlated with the saliency coefficient; use the adaptive threshold to perform thresholding operations on the background matrix to update the sparse matrix and the Lagrange multiplier matrix.

[0034] In one embodiment, by performing iterative optimization, the low-rank matrix and the sparse matrix are gradually corrected so that the sparse matrix can accurately reflect the defect noise.

[0035] First, the low-rank matrix is ​​updated using the singular value thresholding operator. See [link to documentation]. Figure 2This is a schematic diagram of a silicon carbide tray grayscale image according to an embodiment of this application. The grayscale image contains a circular or elongated groove-shaped background texture, and the background texture has high repetition. The highly repetitive background texture corresponds to the large singular values ​​of the observation matrix, and a pure low-rank matrix is ​​reconstructed by preserving the large singular values.

[0036] The specific operation process is as follows: Construct a temporary matrix by combining the observation matrix, sparse matrix, and Lagrange multiplier matrix of the current iteration. According to the update rule of the Alternating Direction Multiplier Method (ADMM), this temporary matrix... The following relationship must be satisfied: ; in, It is the sparse matrix from the previous iteration. For the observation matrix, It is the multiplier matrix from the previous round. It is a penalty parameter. For temporary matrices... Perform singular value decomposition on the temporary matrix. Decomposed into left singular matrices Singular value matrix And right singular matrix In the singular value matrix, large singular values ​​correspond to high-energy, highly repetitive components (i.e., circular or elongated groove-shaped backgrounds) in the grayscale image; small singular values ​​correspond to cluttered noise or sparse outliers in the grayscale image. Therefore, the singular value matrix... In this process, for any singular value, if the singular value is greater than a singular value threshold, the difference between the singular value and the singular value threshold is used as the updated singular value; if the singular value is not greater than the singular value threshold, the singular value is set to 0 as the updated singular value. The updated low-rank matrix is ​​then reconstructed using the updated singular value matrix. : ;in, , For left singular matrices and right singular matrices, This is the updated singular value matrix.

[0037] Next, in order to obtain the data to be analyzed containing defect information, the background matrix needs to be calculated based on the updated low-rank matrix. The background matrix is ​​obtained by calculating the difference between the observation matrix and the updated low-rank matrix, and the sum of the ratios of the Lagrange multiplier matrix and the penalty parameter in the previous iteration.

[0038] No. Background matrix of the next iteration Satisfying the relation: ; In the formula, For the observation matrix, For the first The low-rank matrix after the next iteration. This is the Lagrange multiplier matrix from the previous round. The penalty parameter is exemplified by a value of [value to be filled in]. Background matrix It contains all the components that the background model cannot explain, including information on defects such as cracks that need to be found, as well as a large amount of noise information.

[0039] To further distinguish between cracks and noise in the background matrix, structural tensor analysis is introduced. The horizontal gradient in the horizontal direction and the vertical gradient in the vertical direction of each pixel in the background matrix are calculated; a structural tensor matrix for each pixel is constructed based on the horizontal and vertical gradients; eigenvalue decomposition is performed on the structural tensor matrix to obtain a first eigenvalue and a second eigenvalue, wherein the first eigenvalue is greater than or equal to the second eigenvalue.

[0040] It should be noted that, for any pixel in the background matrix, the structure tensor matrix of that pixel can be constructed. : ; In the formula, and These are the horizontal and vertical gradients of the pixel, respectively. The first eigenvalue is obtained by performing eigenvalue decomposition on the structure tensor matrix. Second eigenvalue ,and , The intensity of change represents the direction of the most drastic local gray-level change, and is located perpendicular to this direction. It represents the intensity of change in the vertical direction.

[0041] After obtaining the first and second feature values, saliency needs to be calculated to quantify the probability that a pixel belongs to a crack. The saliency is the product of the linearity factor and the structure energy factor; the linearity factor represents the directional consistency of the local texture and is positively correlated with the difference between the first and second feature values; the structure energy factor represents the texture intensity of the local texture and is positively correlated with the sum of the first and second feature values.

[0042] Specifically, the method for calculating the linearity factor includes: calculating the difference between the first eigenvalue and the second eigenvalue, and the sum of the first eigenvalue and the second eigenvalue, and using the ratio of the difference to the sum as the linearity factor; the method for calculating the structural energy factor includes: calculating the product of the sum and a preset energy sensitivity coefficient, and normalizing the product to obtain a significance coefficient.

[0043] It should be noted that significance The following relationship must be satisfied: ; In the formula, For linearity factor, To prevent the default positive number with a denominator of zero, the value is set to 1 for example; For structural energy factor, The preset energy sensitivity coefficient is 5, for example.

[0044] It should be noted that the linearity factor is used to quantify the directional consistency of local texture in terms of geometric shape. When a local region exhibits obvious linear characteristics, the first eigenvalue is much larger than the second eigenvalue, causing the linearity factor to approach 1; conversely, if the local area presents a chaotic scattered pattern, the linearity factor approaches 0. The structure energy factor is used to quantify the intensity of local texture changes. When the grayscale changes in a local region are drastic, i.e., the crack has a certain depth, the structure energy factor approaches 1; when the grayscale changes are weak, the structure energy factor approaches 0. That is, only when the image structure has both directionality and sufficient intensity of change is the significance... Only when it approaches 1 can it effectively filter out noise that resembles cracks but lacks strength.

[0045] Furthermore, an adaptive threshold is dynamically set for each pixel based on the calculated saliency. Specifically, a base denoising threshold is set; a threshold adjustment coefficient is calculated, which is negatively correlated with saliency; and the product of the threshold adjustment coefficient and the base denoising threshold is used as the adaptive threshold.

[0046] For any pixel, its adaptive threshold The following relationship must be satisfied: ; In the formula, The basic denoising threshold is, for example, set to a value of [value]. , These are the width and height of the grayscale image, respectively. The image size is approximately 0.031; The structure sensitivity coefficient is exemplarily set to 3. In regions of high saliency, the adaptive threshold decreases rapidly, allowing subtle grayscale changes to be preserved; in noisy regions of low saliency, the adaptive threshold remains high, thereby eliminating noise.

[0047] Subsequently, an adaptive threshold is used to perform a threshold operation on the background matrix to update the sparse matrix. Specifically, this includes: calculating the absolute value of each pixel in the background matrix; calculating the difference between the absolute value and the adaptive threshold of the corresponding pixel, and taking the maximum value between the difference and 0; and multiplying the maximum value by the sign of the corresponding pixel in the background matrix as the value of the corresponding pixel in the updated sparse matrix.

[0048] No. The sparse matrix after the second iteration update The following relationship must be satisfied: ; In the formula, This is a sign function used to extract the sign direction. For the first Background matrix of the next iteration The function is a maximum value function. If the absolute value of any pixel in the background matrix exceeds the adaptive threshold of that pixel, the excess portion is retained as a defect signal and stored in the sparse matrix; otherwise, the value of that pixel in the sparse matrix is ​​set to 0.

[0049] Finally, to ensure that the sum of the decomposed low-rank matrix and sparse matrix can accurately reconstruct the observation matrix, the reconstruction error generated in each iteration must be accumulated and penalized. The update of the Lagrange multiplier matrix in the next iteration satisfies the following relationship: ; In the formula, For the first The Lagrange multiplier matrix after the next iteration update The Lagrange multiplier matrix from the previous iteration. The penalty parameter is exemplified by a value of [value to be filled in]. . This represents the reconstruction residual of the current iteration, which is the deviation between the original image data and the sum of the current decomposition results.

[0050] Understandably, if the current decomposition result With observation matrix There is a significant deviation, and this deviation will be amplified. The increased value is accumulated in the Lagrange multiplier matrix. In the next iteration calculating the background matrix, the enlarged Lagrange multiplier matrix exerts a stronger corrective force, forcing the algorithm to adjust in the direction of reducing error. Thus, through this error backpropagation mechanism, it continuously self-corrects during iteration, ultimately ensuring that the decomposition result satisfies [the required parameters]. The constraints ensured the accuracy of the defect detection results.

[0051] S104, repeat the above iterative steps until the termination condition is met, and use the final sparse matrix as the defect detection result.

[0052] In one embodiment, iteration stops when the error is less than a preset convergence tolerance or the number of iterations reaches the upper limit. The error characterizes the deviation between the observed matrix and the sum of the decomposed low-rank matrix and sparse matrix. For example, the convergence tolerance is set to a value... The iteration limit is set to 100 times; the final output sparse matrix is ​​the defect detection result. Further, binarization and connected component analysis can be performed on the sparse matrix to output the final defect detection result.

[0053] Specifically, the final defect detection result output includes: Since the grayscale image has already been normalized in S101, the values ​​in the sparse matrix are distributed between 0 and 1. The closer the value is to 1, the higher the probability of a defect existing at that location. The segmentation threshold is determined using the maximum inter-class variance method, and the sparse matrix is ​​binarized. Pixels in the sparse matrix greater than the segmentation threshold are marked as 1, and the rest are marked as 0, generating a binarized mask image. Please refer to [link to relevant documentation]. Figure 3 This is an effect diagram of the defect detection results of a silicon carbide tray according to an embodiment of this application; the white area in the binarized mask image is the defect area.

[0054] According to a second aspect of this application, this application also provides a silicon carbide tray defect detection system based on image recognition. Figure 4 This is a structural block diagram of a silicon carbide tray defect detection system based on image recognition, according to an embodiment of this application. Figure 4 As shown, the system 50 includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the image recognition-based silicon carbide tray defect detection method according to the first aspect of this application. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface. Their configuration and functions are known in the art and will not be described further here.

[0055] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application.

Claims

1. A method for detecting defects in silicon carbide trays based on image recognition, characterized in that, The detection method includes: acquiring a grayscale image of a silicon carbide tray to construct an observation matrix, and initializing a low-rank matrix, a sparse matrix, and a Lagrange multiplier matrix for characterizing error accumulation; Performing iterative optimization includes: in each iteration, updating the low-rank matrix using the singular value thresholding operator, and using the difference between the observation matrix and the low-rank matrix as the background matrix; obtaining the eigenvalues ​​of the structure tensor matrix of each pixel in the background matrix, calculating the saliency of each pixel based on the eigenvalues, and calculating the adaptive threshold of each pixel based on the saliency, wherein the adaptive threshold is negatively correlated with the saliency coefficient; and performing thresholding operations on the background matrix using the adaptive threshold to update the sparse matrix and the Lagrange multiplier matrix. Repeat the above iterative steps until the termination condition is met, and use the final sparse matrix as the defect detection result.

2. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The process of acquiring a grayscale image of the silicon carbide tray to construct an observation matrix includes: The grayscale image is normalized, and the normalized grayscale image is used as the observation matrix.

3. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The process of obtaining the eigenvalues ​​of the structure tensor matrix of each pixel in the background matrix includes: Calculate the horizontal gradient in the horizontal direction and the vertical gradient in the vertical direction for each pixel in the background matrix; construct the structure tensor matrix for each pixel based on the horizontal gradient and the vertical gradient; The eigenvalue decomposition of the structure tensor matrix yields the first eigenvalue and the second eigenvalue, where the first eigenvalue is greater than or equal to the second eigenvalue.

4. The image recognition-based silicon carbide tray defect detection method according to claim 3, characterized in that, The significance is the product of the linearity factor and the structural energy factor; The linearity factor characterizes the directional consistency of local texture and is positively correlated with the difference between the first eigenvalue and the second eigenvalue. The structural energy factor characterizes the texture intensity of the local texture and is positively correlated with the sum of the first and second eigenvalues.

5. The image recognition-based silicon carbide tray defect detection method according to claim 4, characterized in that, The method for calculating the linearity factor includes: calculating the difference between the first eigenvalue and the second eigenvalue, and the sum of the first eigenvalue and the second eigenvalue, and using the ratio of the difference to the sum as the linearity factor; The method for calculating the structural energy factor includes: calculating the product of the sum and a preset energy sensitivity coefficient, and then normalizing the product to obtain a significance coefficient.

6. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The adaptive threshold is calculated as follows: Set the basic noise reduction threshold; Calculate the threshold adjustment coefficient, which is negatively correlated with significance; The product of the threshold adjustment coefficient and the basic denoising threshold is used as the adaptive threshold.

7. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The step of using an adaptive threshold to perform thresholding on the background matrix to update the sparse matrix includes: calculating the absolute value of each pixel in the background matrix; Calculate the difference between the absolute value and the adaptive threshold of the corresponding pixel, and take the maximum value between the difference and 0; multiply the maximum value by the sign of the corresponding pixel in the background matrix, and use the result as the value of the corresponding pixel in the updated sparse matrix.

8. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The method for calculating the background matrix includes: The background matrix is ​​obtained by calculating the difference between the observation matrix and the updated low-rank matrix, and the sum of the ratios of the Lagrange multiplier matrix and the penalty parameter in the previous iteration.

9. The image recognition-based silicon carbide tray defect detection method according to claim 1, characterized in that, The initialization of the low-rank matrix, sparse matrix, and Lagrange multiplier matrix used to characterize error accumulation includes: initializing the low-rank matrix as the observation matrix; and initializing the sparse matrix and Lagrange multiplier matrix as all-zero matrices.

10. A silicon carbide tray defect detection system based on image recognition, characterized in that, It includes a processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the image recognition-based silicon carbide tray defect detection method according to any one of claims 1 to 9.

Images