Defect detection method and system based on large-size single crystal silicon rod
By using the adaptive reconstruction elliptic Gaussian kernel technique, the problem of stripe blurring in the detection of large-size single-crystal silicon rods by the traditional Gaussian difference algorithm is solved, achieving accurate preservation and efficient detection of defect features, and improving detection accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAANXI XINGYAO JUYUAN TECHNOLOGY CO LTD
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-09
AI Technical Summary
When processing large-size single-crystal silicon rods, the traditional Gaussian difference algorithm causes the stripes to become blurred due to the circular Gaussian kernel, resulting in artifacts and making it difficult to accurately preserve defect features. The shortcomings of the infrared light source and mechanical motion system in the existing technology also lead to insufficient detection accuracy.
An adaptive reconstruction elliptical Gaussian kernel is adopted. By analyzing the local extension of pixels, local stripe intensity and anisotropic scale ratio, the shape of the filter kernel is dynamically adjusted to adapt to the stripe direction for smoothing, reduce the blur in the vertical direction, and improve the sensitivity of defect edge detection.
It effectively suppresses stripe interference, improves the high-frequency signal-to-noise ratio and anti-interference robustness of defect detection, and ensures quality assessment and process traceability in the manufacturing process of single crystal silicon rods.
Smart Images
Figure CN122175967A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor device technology. More specifically, this invention relates to a defect detection method and system based on large-size single-crystal silicon rods. Background Technology
[0002] Single-crystal silicon rods are core components of semiconductor etching materials and are important materials in chip manufacturing. These single-crystal silicon rods are grown using the Czochralski method and are silicon substrates used in etching processes. In the production process of Czochralski single-crystal silicon rods, high-precision defect detection of the surface and shallow surface of the single-crystal silicon rod is a key step to ensure product yield. However, under the Czochralski process, periodic growth stripes inevitably form inside the silicon rod. At the same time, due to thermal field fluctuations and other reasons, micro-defects such as initial dislocation clusters are easily generated inside the crystal.
[0003] In order to accurately identify the initial dislocation clusters on the surface of single-crystal silicon rods, the existing technology usually uses the difference of Gaussian algorithm for background suppression and defect extraction. This algorithm filters and subtracts the image by using Gaussian smoothing kernels of different scales, which can smooth out the gradual background and highlight the abrupt features to a certain extent. It has high computational efficiency and good industrial applicability.
[0004] Traditional Gaussian difference algorithms use isotropic circular Gaussian kernels globally to smooth images of single-crystal silicon rod surfaces. However, the growth stripes on the surface of large-sized single-crystal silicon rods pulled by the Czochralski method have strong directionality, typically exhibiting alternating light and dark gradient textures along the pulling direction. If an isotropic circular Gaussian kernel is used to extract the background containing strong directional textures, the kernel will indiscriminately average across the horizontal boundaries of the stripes during filtering, causing excessive blurring in the vertical direction of the growth stripes. This results in a large number of stripe contour artifacts remaining in the residual image after subtracting two Gaussian smoothed images of different scales. These high-contrast artifacts can interfere with or even mask the true, minute gray-level abrupt changes of early dislocation clusters, ultimately leading to a large number of false positives and false negatives when thresholding based on the Gaussian difference residual image.
[0005] In related technologies, for example, Chinese patent document CN109406451B discloses a silicon rod surface defect detection device and method. In the defect detection stage, it mainly relies on the coordinated operation of a silicon rod rotation module and a triaxial detection mechanism motion module. It uses an infrared light source to transmit light through the silicon rod and a linear array camera to acquire infrared image signals of the internal structure of the silicon rod in real time. These signals are then transmitted to a control module for image analysis and processing. By identifying abnormal textures and contours in the image, it determines whether there are structural defects such as microcracks, dislocations, or twins inside the silicon rod, thereby achieving panoramic automated scanning and detection of the entire silicon rod. Its drawback is that this patent document relies on the uniform penetration of the infrared light source inside the silicon rod and the absolute stability of the multi-axis mechanical motion system, and does not consider the characteristics of single-crystal silicon rods. Due to its large thickness and uneven distribution of residual stress in the internal lattice, the infrared beam scattering is non-uniform. Furthermore, its strategy of using a linear array camera to scan images line by line and extract defect features lacks dynamic compensation and adaptive noise reduction mechanisms to address the unavoidable differences in surface roughness, slight fluctuations in transmittance, and image stitching misalignment caused by motor vibration during complex mechanical scanning. This static image recognition logic can easily lead the system to misjudge normal transmittance fluctuations and dark spots as internal dislocations or twin defects, or cause real, low-contrast microcracks to be masked and missed in motion stitching artifacts. It is difficult to achieve high-precision and robust detection of defects in large-size single-crystal silicon rods in complex industrial scanning environments. Summary of the Invention
[0006] To address the technical problem that the circular Gaussian kernel in the traditional Gaussian difference algorithm easily causes stripe blurring and artifacts, making it difficult to accurately preserve the defect features of single-crystal silicon rods while suppressing directional stripe background, this invention provides solutions in the following aspects.
[0007] In a first aspect, the present invention provides a defect detection method based on a large-size single-crystal silicon rod, comprising: acquiring an image of the surface of the single-crystal silicon rod, performing preprocessing to obtain a grayscale image of the surface of the single-crystal silicon rod; determining the local region of each pixel, constructing a structure tensor matrix of the local region of each pixel in the image, determining the local extension degree of the pixel based on the maximum and minimum values of the structure tensor matrix of the local region of the pixel, and determining the local stripe intensity of the pixel based on the ratio between the sum of the absolute values of the gradient magnitudes of the local region of the pixel and the average value of the sum of the absolute values of the gradient magnitudes of the local regions of all pixels, combined with the local extension degree of the pixel; determining the local stripe intensity of each pixel. The local pointiness of a pixel is determined by considering the variance of gray values of all pixels within its micro-neighborhood, combined with the variance of gray values of all pixels in the outermost ring of the micro-neighborhood. Based on the local stripe intensity and the local pointiness, the anisotropic scale ratio of the pixel is determined. Then, based on the anisotropic scale ratio and a preset Gaussian kernel standard deviation, the major and minor axis standard deviations of the adaptively reconstructed elliptical Gaussian kernel are determined. Finally, based on these standard deviations, the grayscale image of the single-crystal silicon rod surface is differentially processed to obtain the detection result of the single-crystal silicon rod.
[0008] This invention establishes the inherent logical relationship between pixel-level gradient spatial distribution and macroscopic growth stripe trends by analyzing the local extension degree and local stripe intensity of pixels; it distinguishes between regular stripe backgrounds and isolated defect targets by calculating the local point-like degree and anisotropic scale ratio; and it constructs an adaptive reconstruction elliptical Gaussian kernel, enabling the filtering process to perform anisotropic smoothing in accordance with the stripe direction. While effectively eliminating directional stripe interference, it significantly weakens the cross-boundary blurring effect perpendicular to the stripe direction, thereby improving the sensitivity and accuracy of capturing subtle defect edges.
[0009] Preferably, the method for obtaining the degree of local extension is as follows: the difference between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the first value, the sum between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the second value, the ratio between the first value and the second value is calculated, and the result is taken as the degree of local extension.
[0010] This invention evaluates the local extension of pixels, so that pixels along the direction of clear stripes are calculated to have a larger local extension, while isolated defect points in non-directional areas are calculated to have a smaller local extension, thus providing a basis for accurately distinguishing highly anisotropic growth stripes and abnormal micro defects.
[0011] Preferably, the method for obtaining the local stripe intensity is as follows: calculate the ratio between the sum of the absolute values of the gradient magnitudes of the local regions of a pixel and the average value of the sum of the absolute values of the gradient magnitudes of the local regions of all pixels, perform maximum and minimum value normalization on the ratio to obtain a third value, calculate the product of the third value and the degree of local extension, and use the result as the local stripe intensity.
[0012] This invention evaluates the local stripe intensity of pixels, enabling the calculation of a large local stripe intensity for both high-contrast significant defects and obvious stripe structures, effectively filtering out interference from the smooth reflections and low-frequency noise on the surface of single-crystal silicon rods.
[0013] Preferably, the method for obtaining the local point-like degree is as follows: the difference between the variance of gray values of all pixels in the micro-neighborhood of a pixel and the variance of gray values of all pixels in the outermost ring of the micro-neighborhood of a pixel is counted as the fourth value, the fourth value is normalized by the maximum and minimum values, and the result is used as the local point-like degree.
[0014] This invention evaluates the local dot degree of pixels, enabling the calculation of a large local dot degree only for isolated and high-contrast real dot defects, thus distinguishing defect features from complex striped backgrounds.
[0015] Preferably, the method for obtaining the anisotropic scale ratio is as follows: the inverse of the local point-like degree is normalized to the maximum and minimum values to obtain a fifth value, the product of the fifth value and the local fringe intensity is calculated, and the result is used as the anisotropic scale ratio.
[0016] This invention evaluates the anisotropic scale ratio of pixels, so that pixels in dot-shaped regions with higher defect probabilities have an anisotropic scale ratio closer to 0, thereby guiding the filter kernel to tend towards an isotropic circles to preserve omnidirectional edges, while pixels in pure stripe regions have an anisotropic scale ratio close to 1 to cause the filter kernel to elongate, providing a deformation direction for the reconstruction of the adaptive elliptical Gaussian kernel.
[0017] Preferably, the method for obtaining the standard deviation of the major axis and the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel is as follows: calculate the sum of the anisotropic scale ratio and 1, calculate the product of the sum and the preset standard deviation of the Gaussian kernel, use the result as the standard deviation of the major axis of the adaptively reconstructed elliptical Gaussian kernel, calculate the ratio between the preset standard deviation of the Gaussian kernel and the sum, and use the result as the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel.
[0018] Preferably, the differential processing of the grayscale image of the single-crystal silicon rod surface to obtain the detection result of the single-crystal silicon rod includes: determining the major axis standard deviation and minor axis standard deviation of two different elliptical Gaussian kernels for adaptive reconstruction of pixels based on two different preset Gaussian kernel standard deviations; performing convolution integral operation on the grayscale image of the single-crystal silicon rod surface based on the major axis standard deviation and minor axis standard deviation of the two different elliptical Gaussian kernels to obtain two sets of Gaussian smoothed images; subtracting the two sets of Gaussian smoothed images to obtain a residual image of the single-crystal silicon rod surface; and responding to a pixel in the residual image having a residual absolute value greater than an anomaly detection threshold, then the pixel belongs to a dislocation cluster defect on the surface of the single-crystal silicon rod.
[0019] Preferably, the acquisition of the surface image of the monocrystalline silicon rod includes: using an industrial line scan camera in conjunction with a ring diffuse reflection light source to scan and capture the surface of the monocrystalline silicon rod to obtain the surface image of the monocrystalline silicon rod.
[0020] Preferably, obtaining the grayscale image of the surface of the monocrystalline silicon rod includes: performing grayscale processing on the surface image of the monocrystalline silicon rod to obtain the grayscale image of the surface of the monocrystalline silicon rod.
[0021] Secondly, the present invention provides a defect detection system based on large-size single-crystal silicon rods, 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 above-mentioned defect detection method based on large-size single-crystal silicon rods is implemented.
[0022] By adopting the above technical solution, the defect detection method based on large-size single-crystal silicon rods is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of terminal devices based on the memory and processor, facilitating their use.
[0023] The beneficial effects of this invention are as follows: This invention solves the technical problem that traditional Gaussian difference algorithms are prone to aliasing and distortion of stripe features and weak defect features when processing images of large single-crystal silicon rods with directional stripe features due to the isotropic smoothing of circular Gaussian kernels by introducing an adaptive reconstruction mechanism based on anisotropic scale ratio.
[0024] This invention constructs a discrimination method that can distinguish between regular growth textures and random defect morphologies by mining the local extension degree and local stripe intensity features of pixels. On this basis, it further evolves the local point-like degree that reflects the significance of defects and maps it to anisotropic scale ratio. The index describing the micro-spatial gradient distribution is coupled with the index of macro-structural trend, realizing the identification and stripping of striped background.
[0025] This invention can calculate the major axis standard deviation and minor axis standard deviation for each pixel in the grayscale image of a single-crystal silicon rod surface, which are consistent with the local texture directionality. By dynamically adjusting the flattened shape of the elliptical Gaussian kernel, the smoothing force is distributed in different geometric dimensions. The smoothing span is maximized in the direction parallel to the stripes to suppress noise, while the blur radius is converged in the direction perpendicular to the stripes to preserve the edge sharpness of defects. Ultimately, this invention improves the high-frequency signal-to-noise ratio and anti-interference robustness of defect detection, providing a solid visual feedback guarantee for quality assessment and process traceability in the manufacturing process of single-crystal silicon rods. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the defect detection method based on large-size single-crystal silicon rods in this invention; Figure 2 This is a comparison chart of the difference residuals corresponding to the circular Gaussian kernel of the traditional algorithm and the difference residuals corresponding to the adaptive elliptical Gaussian kernel of this invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0029] This invention discloses a defect detection method based on large-size single-crystal silicon rods, referring to... Figure 1 This includes steps S001-S005: S001: Obtain the grayscale image of the surface of the single-crystal silicon rod.
[0030] Specifically, an industrial line scan camera is used in conjunction with a ring diffuse reflection light source to scan and photograph the surface of a monocrystalline silicon rod to obtain an image of the monocrystalline silicon rod surface. The image of the monocrystalline silicon rod surface is then processed into grayscale to obtain a grayscale image of the monocrystalline silicon rod surface.
[0031] S002: Determine the local extension degree of the pixel and the local stripe intensity.
[0032] It should be noted that traditional Gaussian difference algorithms use isotropic circular Gaussian kernels globally to extract backgrounds containing directional textures. This causes the circular Gaussian kernels to indiscriminately average across the horizontal boundaries of the stripes during filtering, resulting in excessive blurring. Consequently, a large number of stripe contour artifacts remain in the residual image. According to the theory of local gradient distribution in images, the gradient accumulation changes of growth stripes on the surface of a single-crystal silicon rod exhibit significant physical differences in the direction perpendicular to and parallel to the stripe direction. Therefore, this invention determines the local extension degree and local stripe intensity of pixels to characterize the gray-level distribution in the local neighborhood of a pixel, which exhibits a single-dimensional extension attribute and the possibility that it belongs to a region of pure growth stripes with directionality.
[0033] Specifically, for a pixel, a neighborhood of 11 pixels with a side length centered on the pixel is selected as the local region of the pixel. The partial derivatives of the image in the horizontal and vertical directions of each pixel in the local neighborhood are calculated respectively, and a second-order structure tensor matrix of the local neighborhood of the pixel is constructed. The maximum and minimum values of the structure tensor matrix of the local region of the pixel are obtained by solving the structure tensor matrix. In other embodiments, the implementer can set the construction method of the local region of the pixel according to the actual situation.
[0034] Specifically, the method for obtaining the degree of local extension is as follows: the difference between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the first value, the sum between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the second value, the ratio between the first value and the second value is calculated, and the result is taken as the degree of local extension.
[0035] It should be noted that the larger the ratio of the first value to the second value, the greater the difference between the physical feature values of the direction of maximum gradient change and the direction of minimum gradient change in the local area of the pixel. This indicates that the gray-scale distribution in the local area of the pixel exhibits a stronger single-dimensional extension attribute, and therefore the greater the degree of local extension of the pixel.
[0036] Specifically, the method for obtaining the local fringe intensity is as follows: calculate the ratio between the sum of the absolute values of the gradient magnitudes of the local regions of a pixel and the average of the sum of the absolute values of the gradient magnitudes of the local regions of all pixels, perform maximum and minimum value normalization on this ratio to obtain a third value, calculate the product of the third value and the degree of local extension, and use the result as the local fringe intensity.
[0037] It should be noted that the greater the local extension, the greater the likelihood that the pixel belongs to a region of growth stripes with strong directionality. Therefore, the greater the local stripe intensity of the pixel, the greater the third value. This indicates that the absolute value of the gradient amplitude and the relative value of the local region of the pixel are greater in the grayscale image of the entire single-crystal silicon rod surface. This means that the directional texture features in the local region of the pixel are more obvious and reliable visually, thus making the greater the local extension more credible. Therefore, the greater the local stripe intensity of the pixel.
[0038] S003: Determine the local point-likeness of the pixel and the anisotropic scale ratio.
[0039] It should be noted that in the single-crystal silicon rod detection scenario, tiny initial dislocation clusters may randomly overlap on macroscopic growth stripes with strong contrast, causing the macroscopic stripe gradient in this specific region to dominate in the local feature space. This leads to the Gaussian smoothing kernel being easily overextended in the overlapping region, causing the real dislocation cluster defect signal to be smoothed and masked by the filtering kernel, resulting in a decrease in the recognition accuracy of initial dislocation clusters. According to the local statistical theory of images, the variance of a pure texture has spatial consistency at the local scale, while point-like abrupt changes will disrupt the variance balance between the central region and the edge region. Therefore, this invention determines the local point-like degree and anisotropic scale ratio of the pixel to characterize the significance of the abrupt change of the pixel relative to the local stripe background.
[0040] Specifically, for a pixel, a neighborhood with a side length of 9 pixels centered on the pixel is selected as the micro-neighborhood of the pixel. In other embodiments, implementers can set the construction method of the micro-neighborhood of the pixel according to the actual situation.
[0041] Specifically, the method for obtaining the local point-like degree is as follows: the difference between the variance of gray values of all pixels in the micro-neighborhood of a pixel and the variance of gray values of all pixels in the outermost ring of the micro-neighborhood of a pixel is counted as the fourth value. The fourth value is then normalized to its maximum and minimum values, and the result is used as the local point-like degree.
[0042] It should be noted that the larger the fourth value, the greater the difference between the variance of the gray value set of all pixels in the micro-neighborhood and the variance of the gray value set of all pixels in the ring-shaped region of the micro-neighborhood boundary. This indicates that the pixel contributes more to the variance of the gray value of its micro-neighborhood, and that the pixel is more likely to be a dislocation cluster patch overlapping on a high-contrast growth stripe. Therefore, the pixel has a greater degree of local point-likeness.
[0043] Specifically, the method for obtaining the anisotropic scale ratio is as follows: the inverse of the local point-like degree is normalized by the maximum and minimum values to obtain the fifth value, the product of the fifth value and the local fringe intensity is calculated, and the result is used as the anisotropic scale ratio.
[0044] It should be noted that the greater the intensity of the local stripes, the closer the pixel is to the center of the smoothly growing stripes. The higher the physical requirement for smoothing in accordance with the stripe extension, the greater the anisotropic scale ratio of the pixel and the greater the degree of local point-likeness. The smaller the fifth value, the greater the credibility of the distortion caused by the underlying coarse texture of the underlying local stripe directional intensity of the pixel. Therefore, the smaller the anisotropic scale ratio of the pixel, the better the high-frequency mutation characteristics of dislocation groups are protected at the algorithm parameter level.
[0045] S004: Determine the standard deviation of the major axis and the standard deviation of the minor axis of the elliptical Gaussian kernel for adaptive reconstruction of pixels.
[0046] It should be noted that after obtaining the anisotropic scale ratio of the pixel, this invention will adaptively reconstruct and optimize the standard deviations of two Gaussian kernels of different sizes preset in the Gaussian difference algorithm according to the final anisotropic scale ratio of the pixel, to obtain the major axis standard deviation and minor axis standard deviation of the two sets of adaptively reconstructed elliptical Gaussian kernels of the pixel, which are used for the subsequent acquisition of the residual image of the single crystal silicon rod surface, so that the two-dimensional smoothing filter kernel can perform anisotropic blurring along the physical direction of the stripes, thereby eliminating the contour while limiting the blurring in the vertical direction.
[0047] Specifically, the method for obtaining the standard deviation of the major axis and the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel is as follows: calculate the sum of the anisotropic scale ratio and 1, calculate the product of this sum and the preset standard deviation of the Gaussian kernel, and use the result as the standard deviation of the major axis of the adaptively reconstructed elliptical Gaussian kernel; calculate the ratio between the preset standard deviation of the Gaussian kernel and this sum, and use the result as the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel.
[0048] It should be noted that the present invention sets the preset Gaussian kernel standard deviation to 1.5 and 12, which are used to initially filter out high-frequency micro-machining noise and to extract the background outline of macroscopic coarse growth stripes, respectively. In other embodiments, implementers can set the Gaussian kernel standard deviation according to the specific implementation situation.
[0049] It should be noted that the larger the anisotropic scale ratio, the purer and less defect-free the directional features of the stripes corresponding to the pixel's position in the image. This results in a higher filtering requirement that conforms to the stripe extension of the pixel. Consequently, the larger the standard deviation of the major axis of the elliptical Gaussian kernel adaptively reconstructed by the pixel, the stronger the smoothing effect of the elliptical Gaussian kernel in the direction parallel to the stripes. Conversely, the smaller the standard deviation of the minor axis of the elliptical Gaussian kernel adaptively reconstructed by the pixel, the weaker the cross-boundary blurring effect of the elliptical Gaussian kernel in the direction perpendicular to the stripes. This ensures that the algorithm can improve the sensitivity of fine edge extraction while suppressing interference.
[0050] S005: Obtain the test results for the single-crystal silicon rod.
[0051] Specifically, the test results for obtaining the single-crystal silicon rod include: The standard deviation of the Gaussian kernel used for initial filtering of high-frequency micro-machining noise is designated as the first Gaussian kernel standard deviation and set to 1.5. The standard deviation of the Gaussian kernel used for extracting the background contour of macroscopic coarse growth stripes is designated as the second Gaussian kernel standard deviation and set to 12. In other embodiments, implementers can set the first and second Gaussian kernel standard deviations according to specific implementation conditions. For example, when the accuracy requirement for capturing micro-early dislocation cluster defects is relatively strict, the first Gaussian kernel standard deviation can be appropriately reduced to retain more weak defect mutation features. When the high-frequency noise left by machining on the surface of the single crystal silicon rod is relatively serious, the first Gaussian kernel standard deviation can be appropriately increased to improve the filtering effect of artifact noise. When the physical spacing of the growth stripes on the surface of the single crystal silicon rod is relatively wide, the second Gaussian kernel standard deviation can be appropriately increased to extract the macroscopic background contour more completely. When the growth stripes on the surface of the single crystal silicon rod are relatively fine, the second Gaussian kernel standard deviation can be appropriately reduced to make the extracted background contour more accurately fit the local stripe undulation changes.
[0052] Based on the standard deviations of the first and second Gaussian kernels, the major and minor axis standard deviations of two different elliptical Gaussian kernels are determined for the adaptive reconstruction of pixels. Convolution integration is performed on the grayscale image of the single-crystal silicon rod surface based on these two different elliptical Gaussian kernels to obtain a first and second Gaussian smoothed image. The difference between the first and second Gaussian smoothed images is calculated to obtain a residual image of the single-crystal silicon rod surface. If the absolute value of the residual of a pixel in the residual image is greater than the anomaly detection threshold, then the pixel belongs to a dislocation cluster defect on the surface of the single-crystal silicon rod. In this embodiment, the anomaly detection threshold is set to 35. In other embodiments, the implementer can set the anomaly detection threshold according to the specific implementation situation. For example, when the detection sensitivity requirement for weak initial dislocation cluster defects is high, the anomaly detection threshold can be appropriately reduced to reduce the false alarm rate. When the background artifact noise interference on the single-crystal silicon rod surface is strong or the accuracy requirement for the defect detection result is strict, the anomaly detection threshold can be appropriately increased to effectively reduce the false alarm rate.
[0053] like Figure 2 As shown in the figure, this diagram compares the difference residuals corresponding to the circular Gaussian kernel of the traditional algorithm and the adaptive elliptical Gaussian kernel of this invention. The horizontal axis represents the pixel position, and the vertical axis represents the absolute value of the residual. Near pixel position 65, the curve corresponding to this invention forms a sharp and obvious single peak, clearly exceeding the anomaly detection threshold and accurately and intuitively indicating the presence of dislocation cluster defects. Outside this range, in the background area, the curve corresponding to this invention is relatively flat, effectively removing strong directional growth stripe interference. In contrast, the curve corresponding to the traditional algorithm, unable to adapt to complex anisotropic textures, produces continuous and highly disruptive low-frequency fluctuation artifacts across the entire pixel range, and its response at defects fails to effectively exceed the threshold. This demonstrates that the adaptive elliptical Gaussian kernel of this invention overcomes the stripe interference that traditional Gaussian difference algorithms struggle to handle, achieving accurate removal and detection of weak defects in complex backgrounds.
[0054] The present invention also discloses a defect detection system based on large-size single-crystal silicon rods, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the defect detection method based on large-size single-crystal silicon rods according to the present invention is implemented.
[0055] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
Claims
1. A defect detection method based on large-size single-crystal silicon rods, characterized in that, include: Images of the surface of a single-crystal silicon rod are acquired and preprocessed to obtain a grayscale image of the single-crystal silicon rod surface. Determine the local region of each pixel, construct the structure tensor matrix of the local region of each pixel in the image, determine the local extension degree of the pixel based on the maximum and minimum values of the structure tensor matrix of the local region of the pixel, and determine the local stripe intensity of the pixel based on the ratio between the sum of the absolute values of the gradient magnitudes of the local region of the pixel and the average value of the sum of the absolute values of the gradient magnitudes of the local regions of all pixels, combined with the local extension degree of the pixel. The micro-neighborhood of each pixel is determined. Based on the variance of gray values of all pixels within the micro-neighborhood, and combined with the variance of gray values of all pixels in the outermost ring of the micro-neighborhood, the local pointiness of the pixel is determined. Based on the local stripe intensity and the local pointiness of the pixel, the anisotropic scale ratio of the pixel is determined. Based on the anisotropic scale ratio of the pixels and the preset standard deviation of the Gaussian kernel, the standard deviation of the major axis and the standard deviation of the minor axis of the elliptical Gaussian kernel for adaptive reconstruction of the pixels are determined. Based on the standard deviation of the major axis and the standard deviation of the minor axis of the elliptical Gaussian kernel adaptively reconstructed by pixels, the grayscale image of the single-crystal silicon rod surface is differentially processed to obtain the detection result of the single-crystal silicon rod.
2. The defect detection method based on large-size single-crystal silicon rods according to claim 1, characterized in that, The method for obtaining the degree of local extension is as follows: the difference between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the first value, the sum between the maximum and minimum values of the structure tensor matrix of the local region of the pixel is taken as the second value, the ratio between the first value and the second value is calculated, and the result is taken as the degree of local extension.
3. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The method for obtaining the local fringe intensity is as follows: calculate the ratio between the sum of the absolute values of the gradient magnitudes of the local regions of a pixel and the average value of the sum of the absolute values of the gradient magnitudes of the local regions of all pixels, perform maximum and minimum value normalization on the ratio to obtain a third value, calculate the product of the third value and the degree of local extension, and use the result as the local fringe intensity.
4. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The method for obtaining the local point-like degree is as follows: the difference between the variance of gray values of all pixels in the micro-neighborhood of a pixel and the variance of gray values of all pixels in the outermost ring of the micro-neighborhood of a pixel is counted as the fourth value. The fourth value is normalized by the maximum and minimum values, and the result is used as the local point-like degree.
5. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The method for obtaining the anisotropic scale ratio is as follows: the inverse of the local point degree is normalized by the maximum and minimum values to obtain the fifth value, the product of the fifth value and the local fringe intensity is calculated, and the result is used as the anisotropic scale ratio.
6. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The method for obtaining the standard deviation of the major axis and the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel is as follows: calculate the sum of the anisotropic scale ratio and 1, calculate the product of the sum and the preset standard deviation of the Gaussian kernel, use the result as the standard deviation of the major axis of the adaptively reconstructed elliptical Gaussian kernel, calculate the ratio between the preset standard deviation of the Gaussian kernel and the sum, and use the result as the standard deviation of the minor axis of the adaptively reconstructed elliptical Gaussian kernel.
7. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The differential processing of the grayscale image of the single-crystal silicon rod surface to obtain the detection result of the single-crystal silicon rod includes: determining the major axis standard deviation and minor axis standard deviation of two different elliptical Gaussian kernels for adaptive reconstruction of pixels based on two different preset Gaussian kernel standard deviations; performing convolution integral operation on the grayscale image of the single-crystal silicon rod surface based on the major axis standard deviation and minor axis standard deviation of the two different elliptical Gaussian kernels to obtain two sets of Gaussian smoothed images; subtracting the two sets of Gaussian smoothed images to obtain the residual image of the single-crystal silicon rod surface; and determining that if the absolute value of the residual of a pixel in the residual image is greater than the anomaly detection threshold, then the pixel belongs to the dislocation cluster defect on the surface of the single-crystal silicon rod.
8. The defect detection method based on a large-size single-crystal silicon rod according to claim 1, characterized in that, The acquisition of images of the surface of the monocrystalline silicon rod includes: using an industrial line scan camera in conjunction with a ring diffuse reflection light source to scan and capture images of the surface of the monocrystalline silicon rod to obtain images of the surface of the monocrystalline silicon rod.
9. The defect detection method based on large-size single-crystal silicon rods according to claim 1, characterized in that, The process of obtaining a grayscale image of the surface of a single-crystal silicon rod includes: performing grayscale processing on the surface image of the single-crystal silicon rod to obtain a grayscale image of the surface of the single-crystal silicon rod.
10. A defect detection system based on large-size single-crystal silicon rods, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the defect detection method based on a large-size single-crystal silicon rod according to any one of claims 1-9.