Sub-pixel level sensing fiber path gaussian extraction method and system

Abstract

This specification provides a subpixel-level Gaussian extraction method and system for sensing fiber paths. The method is used for path detection of sensing fibers laid on a silicon wafer surface, including: determining a visual illumination scheme based on material properties; acquiring an image based on the visual illumination scheme; denoising the image using bilateral filtering; extracting edge information of the distributed sensing fiber using subpixel edge detection technology based on the Canny algorithm; closing edge pairs based on the edge information; and extracting skeleton information using Gaussian line detection to obtain the path of the distributed sensing fiber.

Description

Technical Field

[0001] This document relates to the field of computer technology, and in particular to a sub-pixel-level sensing fiber optic path Gaussian extraction method and system. Background Technology

[0002] Distributed sensing optical fibers (FSIFs) possess controllable ultra-high spatial resolution, small mass and diameter, corrosion resistance, electrical insulation, and high sensitivity. Furthermore, due to their relatively flexible and tough properties, FIFs exhibit good adaptability to the shape of structural surfaces, making them commonly used for monitoring strain, stress deformation, and temperature changes. In wafer inspection systems based on FIFs, to ensure that the measurement point data corresponds precisely to the wafer surface location, providing a data foundation for investigating the causes of defects such as warpage, accurate data reconstruction of the sensing fiber path on the wafer surface is necessary.

[0003] Distributed sensing fiber optics are typically used in large-scale, long-distance applications such as tunnels, slopes, and mines. Their spatial resolution is usually 0.5m or 1m, and precise positioning of each measuring point is not required. However, this is clearly unsuitable for precision measurement applications in modern semiconductor technology. To apply distributed sensing fiber optics to precision measurement, a method is needed to accurately obtain the spatial location of the fiber optic measuring points. In recent years, machine vision technology has matured significantly. As a typical non-contact inspection technology, it is widely used in various applications requiring precise monitoring due to its high precision and intelligence, such as defect detection, image restoration, and medical imaging, achieving excellent results. In existing technologies, one approach uses the Forstner feature extraction operator to extract fiber optic feature points from portions of an image, stitches adjacent images together to obtain a panoramic image of the component, and uses the Otsu's method to extract the desired regions. Finally, a Fast Fourier Transform is used to remove texture noise, and a line extraction operator is used to obtain the complete fiber optic path. Existing technologies also provide an improved SIFT algorithm, along with methods and empirical formulas for setting relevant parameters. Through actual surface defect detection, the robustness and anti-interference capabilities of the SIFT algorithm, as well as the correctness and feasibility of the parameter setting methods, are verified. This demonstrates that the SIFT algorithm has significant advantages in detecting dents and spots, especially in noisy image interference, exhibiting a large advantage in reducing false detection rates for cracks. Furthermore, existing image processing-based packaging surface defect detection methods address the issue of low accuracy. This paper describes a method for detecting surface defects on packaging boxes. CCD cameras are used to capture images of the packaging box surface. Binarization is employed to distinguish between the background and target images. Differential processing with density correction is used to separate the background and foreground. Image redrawing algorithms are used to detect image edges, eliminate noise, and obtain clear image edge features. Thresholding is used to define various defects in the image to extract defect features. Based on the extracted defect features, a defect classifier is designed to classify and identify defect content, thus completing the detection of surface defects on the packaging box. Existing technologies can use seed growth methods to extract edges from solar cell electrodes, or the OTSU method can be used to detect surface defects. However, this method is mostly used when the global grayscale is uniform and there is relatively little interference information.

[0004] In summary, existing technologies use line extraction operators from machine vision to extract fiber optic paths, but their detection targets are fiber optic paths within composite material interlayers. They primarily address the problem of multi-image path recognition and stitching, without addressing fiber optic path detection on highly reflective surfaces, and they do not provide information on the algorithm's recognition accuracy. Existing technologies use the SIFT algorithm to detect surface defects in products. While this algorithm can achieve high-precision category recognition in noisy and complex images, it is not focused on high-precision measurement and therefore cannot handle fiber optic path data reconstruction. Existing technologies utilize defect classifiers for product packaging defect detection, but their requirements for image edge detection accuracy are not high, failing to meet the requirements for fiber optic position data reconstruction. Existing technologies employ the seed growth method for edge extraction of solar cell electrodes. All three methods are for detecting specific types of defects and are not suitable for continuous fiber optic path extraction. Existing technologies use the OTSU method to detect image surface defects, but this method is unsuitable for situations with uneven global grayscale and a large amount of interference. Summary of the Invention

[0005] The purpose of this invention is to provide a sub-pixel-level Gaussian extraction method and system for sensing fiber optic paths, aiming to solve the above-mentioned problems in the prior art.

[0006] This invention provides a sub-pixel-level Gaussian extraction method for sensing fiber paths, used for path detection of small-diameter sensing fibers laid on the surface of a silicon wafer. The method includes:

[0007] A visual lighting scheme is determined based on the material properties, and an image is acquired based on the visual lighting scheme.

[0008] The image is denoised by bilateral filtering, and the edge information of the distributed sensing fiber is extracted by subpixel edge detection technology based on the Canny algorithm.

[0009] The edge pairs are closed based on the edge information, and the skeleton information is extracted using the Gaussian line detection method to obtain the path of the distributed sensing fiber.

[0010] This invention provides a sub-pixel-level sensing fiber optic path Gaussian extraction system for the above-mentioned method, the system comprising:

[0011] An industrial camera, connected to a computer, is used to acquire images of inspected silicon wafers and transmit them to the computer.

[0012] A computer is used to determine a visual illumination scheme based on material properties, acquire the image, reduce noise in the image through bilateral filtering, extract edge information of the distributed sensing fiber using sub-pixel edge detection technology based on the Canny algorithm, close edge pairs based on the edge information, and extract skeleton information using Gaussian line detection method to obtain the path of the distributed sensing fiber.

[0013] A worktable is used to fix the industrial camera above the silicon wafer to be inspected;

[0014] A square shadowless light source is used to provide visual illumination for a silicon wafer to be inspected, positioned at its center, based on a defined visual illumination scheme.

[0015] The embodiments of the present invention can accurately detect the distributed sensing fiber optic path and extract the fiber optic measurement points, and have high robustness and high precision, which can meet the requirements of accurate extraction of fiber optic measurement point coordinates and high anti-interference. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the sub-pixel-level sensing fiber path Gaussian extraction method according to an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the system hardware side view according to an embodiment of the present invention;

[0019] Figure 3 This is a top view schematic diagram of the system hardware according to an embodiment of the present invention;

[0020] Figure 4 This is a preferred processing flowchart of the sub-pixel-level sensing fiber path Gaussian extraction method according to an embodiment of the present invention. Detailed Implementation

[0021] To address the aforementioned problems in the prior art, this invention provides a method and apparatus for extracting the coordinate sequence of measurement points along the length of a sensing fiber based on Gaussian line detection, thereby reconstructing strain detection data from a distributed sensing fiber on a wafer surface. Based on the optical characteristics and surface height difference between the silicon wafer and the fiber, a square shadowless light source is used for illumination in a dark environment to improve the distinguishability of the sensing fiber on the wafer surface. After reducing image noise through bilateral filtering, a preprocessed image is obtained. Then, sub-pixel edge detection technology based on the Canny algorithm is used, and feature filtering is applied to remove interfering edges, obtaining the target area of ​​the distributed sensing fiber. Finally, the fiber path is extracted using the Gaussian line detection method. Based on the actual number of sensing fiber measurement points, the path can be segmented, and the coordinate values ​​of the fiber measurement points can be extracted.

[0022] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0023] Method Implementation Examples

[0024] According to an embodiment of the present invention, a sub-pixel-level Gaussian extraction method for sensing fiber paths is provided, characterized in that it is used for path detection of small-diameter sensing fibers laid on the surface of a silicon wafer. Figure 1 This is a flowchart of the sub-pixel-level sensing fiber path Gaussian extraction method according to an embodiment of the present invention, as follows: Figure 1 As shown, the sub-pixel-level sensing fiber path Gaussian extraction method according to an embodiment of the present invention specifically includes:

[0025] Step 101: Determine a visual lighting scheme based on material properties, and acquire an image based on the visual lighting scheme;

[0026] Step 102: Denoise the image by bilateral filtering and extract the edge information of the distributed sensing fiber using sub-pixel edge detection technology based on the Canny algorithm.

[0027] Step 103: Close the edge pairs based on the edge information, and use the Gaussian line detection method to extract the skeleton information to obtain the path of the distributed sensing fiber.

[0028] Step 101 specifically includes: using a square shadowless light source to illuminate the wafer from all sides in a dark environment, and using a camera with sufficient resolution to acquire an image that is non-reflective and has high distinction between the optical fiber and the silicon wafer surface.

[0029] Step 102 specifically includes:

[0030] First, open the acquired image at its original size without stretching. Convert the color image to a grayscale image. Set the region to be processed within the grayscale image, and draw a rectangular Region of Interest (ROI) to be processed, denoted as region1. Draw multiple regions for other unused fiber optic segments, and then merge the drawn regions, denoted as region2. Obtain the image processing area. On the grayscale image, [the image processing area is then defined]. The region requiring fiber optic path identification is obtained, and the preprocessed image is obtained.

[0031] According to Formulas 1 and 2, noise is removed from the image using bilateral filtering:

[0032] Formula 1;

[0033] Formula 2;

[0034] Among them, parameters and To smooth the parameters, and These are pixels ( )and( The grayscale values ​​of () are normalized after the weights are calculated. For pixels ( The grayscale value after noise reduction;

[0035] The image is smoothed by convolving it with a Gaussian filter, with a size of [value missing]. The generation equation for the Gaussian filter kernel is shown in Equation 3:

[0036] Formula 3;

[0037] According to Equation 4, the gradient images in the x and y directions are obtained using a Sobel filter, and then the gradient intensity is calculated. and gradient direction :

[0038] Formula 4;

[0039] in, and Sobel operators and In the image Convolution of window A;

[0040] Based on Equation 5, an algorithm that calculates non-maximum suppression and hysteresis thresholding is used to link edge points into edges, thereby detecting points in the previous image. amplitude If greater than If it is an edge point, it will be immediately accepted as an edge point, and the output image point will be... The grayscale value is set to 255, while the amplitude is less than... Points that are not accepted are rejected, while other points that are connected to accepted edge points are also accepted as edges.

[0041] Formula 5;

[0042] The subpixel edge coordinates are finally obtained by fitting the quadratic polynomial shown in Formula 6. ;

[0043] Formula 6;

[0044] in, and The x and y coordinates of the current integer edge points. and These are the gradient values ​​to the left and right of the edge point. The gradient value at the edge point. The distance from adjacent pixels to the edge point;

[0045] Regions are selected based on area characteristics. For each input region, the area characteristics are calculated. If the calculated characteristics of each region are within the limits (6000, 2e+006), the region will be output.

[0046] Close the XLD outline and then fill it into a region, denoted as region3.

[0047] An opening operation is performed on region3, and the fiber edge is extracted by subtracting the eroded region3 from the preprocessed image.

[0048] Step 103 specifically includes:

[0049] The region is formed by merging sub-pixel level edges into edge pairs and then closing them. Image interference is reduced by threshold selection and image opening operation (erosion processing). Finally, the skeleton path is extracted using a Gaussian line detection method. Specifically, the Gaussian line detection method includes:

[0050] The partial derivative of the convolution of the image with a Gaussian mask is used to determine the position of each pixel in the image. and The Taylor quadratic polynomial parameters in the direction are used to calculate the line direction of each pixel.

[0051] The second partial derivative perpendicular to the direction of the line In this process, pixels exhibiting local maxima are marked as skeleton points, and a hysteresis thresholding operation is performed, accepting values ​​where the second derivative is greater than 1. The line point, rejecting the second derivative being less than All other line points that are adjacent to an accepted point are marked as skeleton points, and the discovered line points are eventually connected to form skeleton lines.

[0052] Based on Formula 7, the contrast of grayscale values ​​of the lines to be extracted is calculated. and With the selected Value calculation parameters and :

[0053] Formula 7;

[0054] Among them, parameters The amount of smoothing to be performed by the Gaussian mask is determined, which is directly proportional to the smoothness of the image but inversely proportional to the accuracy of line positioning. The extraction result of the skeleton line breaks down its path into a set of points, providing coordinate points for data reconstruction.

[0055] While extracting the skeleton using Gaussian detection, the line width of the XLD contour line is also extracted. The Parabolic mode is selected using LineMode to merge the subpixel-level skeleton lines into a continuous skeleton line, and finally, smoothing is performed.

[0056] After executing step 103, the following processing is also performed:

[0057] The sub-pixel-level skeleton coordinates extracted by Gaussian detection are obtained by the get_contour_xld operator. Then, the coordinates of the skeleton at interval d are filtered out according to the spatial resolution of the actual system and the interval d of the fiber optic data acquisition, so as to achieve a one-to-one correspondence between the fiber optic detection data and the actual position of the silicon wafer.

[0058] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] The detection object in this embodiment of the invention is a sensing optical fiber whose surface is laid on the surface of a silicon wafer.

[0060] like Figure 2-3As shown, the system architecture of the method in this embodiment of the invention consists of a computer (not shown), an industrial camera 1, a 12mm focal length lens 2, a worktable 3, a shadowless light source 4, a sensing fiber optic cable 5 for the object being detected, and a silicon wafer 6. The industrial camera 1 is mounted on the worktable and can move in the x, y, and z directions to adjust the imaging area and distance. The object being detected is the sensing fiber optic cable 5 laid on the silicon wafer 6, located below the lens. Multiple LEDs are distributed around the shadowless light source 4 to illuminate the silicon wafer 6 from all sides. Simultaneously, the entire system operates in a dark environment to avoid issues such as wafer reflection caused by surrounding light sources. This is because the sensing fiber optic cable 5 is made of highly transparent silicon dioxide material with an extremely small diameter, and the silicon wafer surface is smooth with significant specular reflection. Therefore, side lighting can fully utilize the height difference between the sensing fiber optic cable 5 and the silicon wafer 6, ensuring that the side lighting only reflects light through the sensing fiber optic cable 5 into the lens. Due to the smooth surface of the silicon wafer 6, almost no light is reflected into the lens from areas where the sensing fiber optic cable 5 is not laid.

[0061] like Figure 4 As shown, the method of this embodiment of the invention specifically includes the following processes:

[0062] Image acquisition: The computer acquires high-resolution original images of the sensing fiber optic cable 5 and the silicon wafer 6 from the industrial lens via a gigabit network cable.

[0063] Image preprocessing:

[0064] 1. First, open image 1 at its original size without stretching for subsequent image processing.

[0065] 2. Convert color image 1 to grayscale image to obtain image 2.

[0066] 3. Set the region to be processed in image 2. First, draw the rectangular ROI region to be processed, denoted as region1. Additionally, since only some segments of the optical fibers in region1 require path identification for data reconstruction, multiple regions need to be drawn for the other unused optical fiber segments. Then, merge the drawn regions; the merged region is denoted as region2.

[0067] 4. Obtain the image processing area. In image 2, from: The region requiring fiber optic path identification can be obtained, resulting in image 3.

[0068] Filtering and noise reduction:

[0069] In visual processing systems, dust and other particles are easily present on the surface of real objects, and noise may be generated during image acquisition and transmission, affecting image quality and interfering with the extraction of target information. This is especially true in edge detection, where the gray-scale distribution characteristics of random noise often lead to its being identified as edges by detection algorithms. Therefore, preprocessing the acquired raw image is essential. Image display quality can be enhanced through noise reduction and contrast adjustment to facilitate subsequent feature extraction and other processing steps. Commonly used traditional noise reduction methods include mean filtering, median filtering, and Gaussian filtering. This method uses bilateral filtering, a typical nonlinear filter that not only eliminates random noise but also minimizes edge blurring. It achieves its effect by smoothing pixels in homogeneous regions while retaining edge pixels with higher contrast, as defined in Equations 1 and 2.

[0070] Formula 1;

[0071] Formula 2;

[0072] Parameters in the formula and To smooth the parameters, and These are pixels ( )and( The grayscale values ​​of () are normalized after the weights are calculated. For pixels ( The grayscale value after noise reduction.

[0073] Edge extraction:

[0074] Since the data reading of the sensing fiber corresponds to the fiber length information, in order to ensure the consistency of data reconstruction, the path of the distributed sensing fiber on the silicon wafer surface is extracted along the length direction. Due to the extremely high accuracy requirements of the detection results, the sub-pixel edge detection technology based on the Canny algorithm is used to detect the edge of the fiber. After forming the region by closing the edge pairs, the skeleton information is extracted by the Gaussian line detection method to obtain the sensing fiber path.

[0075] Sub-pixel edge detection based on the Canny algorithm: Edge detection, as one of the technologies with high practical application value in the field of machine vision, has significant research importance. Traditional edge detection algorithms achieve pixel-level accuracy. However, with the advancement of technology and the development of the semiconductor industry, the accuracy required for industrial inspection has rapidly increased. Traditional pixel-level edge detection cannot meet the detection accuracy requirements of the sensing fiber optic cable. Therefore, this method chooses a sub-pixel-level detection technique based on the Canny algorithm to further divide pixels and thus improve image resolution.

[0076] 1. Use the edges_sub_pix operator to extract subpixel contours from image 3. The filter is Canny.

[0077] The specific principle is as follows: First, a Gaussian filter is convolved with the image to smooth it, with a size of... The generation equation for the Gaussian filter kernel is shown in Equation 3.

[0078] Formula 3;

[0079] Then, the gradient images in the x and y directions are obtained using a Sobel filter, and the gradient strength is then calculated. and gradient direction As shown in Formula 4, where and Sobel operators and In the image Convolution of window A.

[0080] Formula 4;

[0081] Edge points are linked into edges using algorithms that compute non-maximum suppression and hysteresis-like thresholding. Points in the previous image are then detected. amplitude If greater than If it is an edge point, it will be immediately accepted as an edge point, and the output image point will be... The grayscale value is set to 255, while the amplitude is less than... Points that do not meet the criteria are rejected, while other points that are connected to accepted edge points are also accepted as edges. This is expressed in Formula 5.

[0082] Formula 5;

[0083] Finally, the sub-pixel edge coordinates were obtained by fitting the quadratic polynomial formula 6. and The x and y coordinates of the current integer edge points. and These are the gradient values ​​to the left and right of the edge point. The gradient value at the edge point. The distance from adjacent pixels to edge points is the final result. .

[0084] Formula 6

[0085] 2. Select regions based on area features. For each input region, calculate its area feature. If the calculated feature of each region is within the limit (6000, 2e+006), that region will be output.

[0086] 3. First, close the XLD outline, then fill it into a region, denoted as region3.

[0087] 4. Perform an opening operation on region3. By subtracting the eroded region3 from image 3, the fiber edge can be extracted.

[0088] Gaussian detection for skeleton extraction:

[0089] The region is formed by merging the sub-pixel level edges into edge pairs and then closing them. The image interference is further reduced by threshold selection and image opening operation (i.e., erosion processing). Finally, the skeleton path is extracted using the Gaussian line detection method.

[0090] This detection method determines the position of each pixel in an image by taking the partial derivative of the convolution of the image with a Gaussian mask. and The Taylor quadratic polynomial parameters in the direction are used to calculate the line direction of each pixel. The second-order partial derivative perpendicular to the line direction... In this process, pixels exhibiting local maxima are marked as skeleton points. Similar to the hysteresis thresholding operation in Equation 5, it accepts second derivatives greater than... The line point, rejecting the second derivative being less than All other line points adjacent to an accepted point are also marked as skeleton points. Finally, the discovered line points are connected to form a skeleton line. Parameters and The contrast of the grayscale values ​​of the lines to be extracted can be determined. and With the selected The value is calculated as shown in Formula 7, where the parameter The amount of smoothing to be performed by the Gaussian mask is determined; it is directly proportional to the smoothness of the image but inversely proportional to the accuracy of line positioning, which may cause inaccurate line positioning during extraction. The extracted skeleton lines can be broken down into a set of points, providing coordinate points for data reconstruction.

[0091] Formula 7;

[0092] Specifically, Gaussian detection is used to extract the skeleton while simultaneously extracting the line width of the XLD contour. Since the image edges are quite sharp, the LineMode is set to parabolic. Finally, the subpixel-level skeleton lines are merged into a single continuous skeleton line, followed by smoothing.

[0093] Output coordinates:

[0094] Since the data reading of the sensing fiber 5 corresponds to the fiber length information, the corresponding coordinates can be found on the extracted skeleton line based on the measurement point position (actual measurement position) determined by the actual system spatial resolution of the sensing fiber 5, and then output. This achieves a one-to-one correspondence between the measurement data of the sensing fiber 5 and the specific position of the silicon wafer 6.

[0095] Specifically, the sub-pixel-level skeleton coordinates extracted by Gaussian detection can be obtained through the get_contour_xld operator. Then, based on the interval d of the measurement points determined by the actual system spatial resolution of the optical fiber, the coordinates of interval d on the skeleton are filtered out and finally input.

[0096] In summary, this invention achieves sub-pixel-level high-precision path detection and fiber measurement point coordinate extraction for sensing fibers laid on silicon wafers. Linewidth analysis of the extracted sensing fibers reveals uniform single-sided linewidth distribution and high system stability. Calculations based on the actual diameter of the sensing fiber show that the detection method achieves an accuracy of 10 micrometers, further demonstrating the high stability and accuracy of the Gaussian skeleton path detection method. The technical solution of this invention exhibits high robustness. Experimental verification shows that even when the actual laid sensing fiber 5 is poorly fitted or has similar edge information interference around it, this method can still accurately identify the skeleton path of the actual sensing fiber 5, unaffected by defects or unexpected interference.

[0097] System Implementation Examples

[0098] According to an embodiment of the present invention, a sub-pixel-level sensing fiber optic path Gaussian extraction system is provided, characterized in that, for the method described in the above method embodiments, the sub-pixel-level sensing fiber optic path Gaussian extraction system according to the embodiment of the present invention specifically includes:

[0099] An industrial camera, connected to a computer, is used to acquire and transmit images of inspected silicon wafers to the computer; the industrial camera is a five-megapixel camera with a resolution of 3856×2764 and a lens focal length of 12mm.

[0100] A computer is used to determine a visual illumination scheme based on material properties, acquire the image, reduce noise in the image through bilateral filtering, extract edge information of the distributed sensing fiber using sub-pixel edge detection technology based on the Canny algorithm, close edge pairs based on the edge information, and extract skeleton information using Gaussian line detection method to obtain the path of the distributed sensing fiber.

[0101] A worktable is used to fix the industrial camera above the silicon wafer to be inspected;

[0102] A square shadowless light source is used to provide visual illumination for a silicon wafer to be inspected, positioned at its center, based on a defined visual illumination scheme. Specifically, the visual illumination scheme involves using the square shadowless light source to provide side illumination from all four sides of the wafer in a dark environment, and acquiring an image with sufficient resolution that is non-reflective and has high distinction between the optical fiber and the silicon wafer surface.

[0103] The computer is specifically used for:

[0104] First, open the acquired image at its original size without stretching. Convert the color image to a grayscale image. Set the region to be processed within the grayscale image, and draw a rectangular Region of Interest (ROI) to be processed, denoted as region1. Draw multiple regions for other unused fiber optic segments, and then merge the drawn regions, denoted as region2. Obtain the image processing area. On the grayscale image, [the image processing area is then defined]. The region requiring fiber optic path identification is obtained, and the preprocessed image is obtained.

[0105] According to Formulas 1 and 2, noise is removed from the image using bilateral filtering:

[0106] Formula 1;

[0107] Formula 2;

[0108] Among them, parameters and To smooth the parameters, and These are pixels ( )and( The grayscale values ​​of () are normalized after the weights are calculated. For pixels ( The grayscale value after noise reduction;

[0109] The image is smoothed by convolving it with a Gaussian filter, with a size of [value missing]. The generation equation for the Gaussian filter kernel is shown in Equation 3:

[0110] Formula 3;

[0111] According to Equation 4, the gradient images in the x and y directions are obtained using a Sobel filter, and then the gradient intensity is calculated. and gradient direction :

[0112] Formula 4;

[0113] in, and Sobel operators and In the image Convolution of window A;

[0114] Based on Equation 5, an algorithm that calculates non-maximum suppression and hysteresis thresholding is used to link edge points into edges, thereby detecting points in the previous image. amplitude If greater than If it is an edge point, it will be immediately accepted as an edge point, and the output image point will be... The grayscale value is set to 255, while the amplitude is less than... Points that are not accepted are rejected, while other points that are connected to accepted edge points are also accepted as edges.

[0115] Formula 5;

[0116] The subpixel edge coordinates are finally obtained by fitting the quadratic polynomial shown in Formula 6. ;

[0117] Formula 6;

[0118] in, and The x and y coordinates of the current integer edge points. and These are the gradient values ​​to the left and right of the edge point. The gradient value at the edge point. The distance from adjacent pixels to the edge point;

[0119] Regions are selected based on area characteristics. For each input region, the area characteristics are calculated. If the calculated characteristics of each region are within the limits (6000, 2e+006), the region will be output.

[0120] Close the XLD outline and then fill it into a region, denoted as region3.

[0121] An opening operation is performed on region3, and the fiber edge is extracted by subtracting the eroded region3 from the preprocessed image.

[0122] The region is formed by merging sub-pixel level edges into edge pairs and then closing them. Image interference is reduced by threshold selection and image opening operation (erosion processing). Finally, the skeleton path is extracted using a Gaussian line detection method. Specifically, the Gaussian line detection method includes:

[0123] The partial derivative of the convolution of the image with a Gaussian mask is used to determine the position of each pixel in the image. and The Taylor quadratic polynomial parameters in the direction are used to calculate the line direction of each pixel.

[0124] The second partial derivative perpendicular to the direction of the line In this process, pixels exhibiting local maxima are marked as skeleton points, and a hysteresis thresholding operation is performed, accepting values ​​where the second derivative is greater than 1. The line point, rejecting the second derivative being less than All other line points that are adjacent to an accepted point are marked as skeleton points, and the discovered line points are eventually connected to form skeleton lines.

[0125] Based on Formula 7, the contrast of grayscale values ​​of the lines to be extracted is calculated. and With the selected Value calculation parameters and :

[0126] Formula 7;

[0127] Among them, parameters The amount of smoothing to be performed by the Gaussian mask is determined, which is directly proportional to the smoothness of the image but inversely proportional to the accuracy of line positioning. The extraction result of the skeleton line breaks down its path into a set of points, providing coordinate points for data reconstruction.

[0128] While extracting the skeleton using Gaussian detection, the line width of the XLD contour line is also extracted. The Parabolic mode is selected using LineMode to merge the subpixel-level skeleton lines into a continuous skeleton line, and finally, smoothing is performed.

[0129] The computer is further used for:

[0130] The sub-pixel-level skeleton coordinates extracted by Gaussian detection are obtained by the get_contour_xld operator. Then, based on the data acquisition interval d of the actual optical fiber, the coordinates at interval d on the skeleton are filtered out to achieve a one-to-one correspondence between the optical fiber detection data and the actual position of the silicon wafer.

[0131] In other words, the system hardware mainly consists of a computer, an industrial camera, a workbench, and a square shadowless light source. The industrial camera is a 5-megapixel camera with a resolution of 3856×2764 and a lens focal length of 12mm, connected to the computer via a gigabit Ethernet cable. The silicon wafer to be inspected is placed in the center of the square shadowless light source, and the industrial camera is mounted above the silicon wafer on the workbench to take pictures from below. To improve the discrimination of the distributed sensing fiber optic path on the surface of the silicon wafer, a suitable visual illumination scheme needs to be designed based on the material properties of the silicon wafer and the optical fiber, so as to perform image preprocessing processes such as grayscale conversion and noise reduction after image acquisition.

[0132] Optical fibers are typically made of highly transparent silica and are extremely small in size. Therefore, when laid on a silicon wafer surface, they exhibit low visual resolution. Furthermore, the extremely smooth surface of a silicon wafer results in significant specular reflection, making it susceptible to interference from ambient light. This necessitates careful selection of the light source. To prevent the camera from failing to capture images with sufficient information, a suitable lighting scheme must be designed.

[0133] The lighting solution of this invention is as follows: a square shadowless light source is used to illuminate the wafer from all sides in a dark environment. By setting up a dark environment, the problem of ambient light interference is solved. Furthermore, the height difference formed by the sensing optical fiber laid on the wafer surface is cleverly utilized. The side lighting method solves the problem of low resolution of the optical fiber on the wafer surface, while avoiding the specular reflection effect produced by the light source illuminating the wafer from the front, and finally obtaining the original image to be detected.

[0134] like Figure 2-3 As shown, the system architecture of the method in this embodiment of the invention consists of a computer (not shown), an industrial camera 1, a 12mm focal length lens 2, a worktable 3, a shadowless light source 4, a sensing fiber optic cable 5 for the object being detected, and a silicon wafer 6. The industrial camera 1 is mounted on the worktable and can move in the x, y, and z directions to adjust the imaging area and distance. The object being detected is the sensing fiber optic cable 5 laid on the silicon wafer 6, located below the lens. Multiple LEDs are distributed around the shadowless light source 4 to illuminate the silicon wafer 6 from all sides. Simultaneously, the entire system operates in a dark environment to avoid issues such as wafer reflection caused by surrounding light sources. This is because the sensing fiber optic cable 5 is made of highly transparent silicon dioxide material with an extremely small diameter, and the silicon wafer surface is smooth with significant specular reflection. Therefore, side lighting can fully utilize the height difference between the sensing fiber optic cable 5 and the silicon wafer 6, ensuring that the side lighting only reflects light through the sensing fiber optic cable 5 into the lens. Due to the smooth surface of the silicon wafer 6, almost no light is reflected into the lens from areas where the sensing fiber optic cable 5 is not laid.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sub-pixel level sensing fiber path Gaussian extraction method, characterized in that, The method for path detection of sensing optical fibers laid on the surface of a silicon wafer includes: A visual lighting scheme is determined based on the material properties, and an image is acquired based on the visual lighting scheme. The image is denoised by bilateral filtering, and the edge information of the distributed sensing fiber is extracted by subpixel edge detection technology based on the Canny algorithm. Based on the edge information, the edge pairs are closed, and the skeleton information is extracted using the Gaussian line detection method to obtain the path of the distributed sensing fiber; specifically including: Regions are formed by merging sub-pixel level edges into edge pairs and then closing them. Image interference is reduced through threshold selection and image opening operations. Finally, the skeleton path is extracted using a Gaussian line detection method. Specifically, the Gaussian line detection method includes: The partial derivative of the convolution of the image with a Gaussian mask is used to determine the position of each pixel in the image. and The Taylor quadratic polynomial parameters in the direction are used to calculate the line direction of each pixel. The second partial derivative perpendicular to the direction of the line In this process, pixels exhibiting local maxima are marked as skeleton points, and a hysteresis thresholding operation is performed, accepting values ​​where the second derivative is greater than 1. The line point, rejecting the second derivative being less than All other line points that are adjacent to an accepted point are marked as skeleton points, and the discovered line points are eventually connected to form skeleton lines. Based on Formula 7, the contrast of grayscale values ​​of the lines to be extracted is calculated. and With the selected Value calculation parameters and : Equation 7; Among them, parameters The amount of smoothing to be performed by the Gaussian mask is determined, which is directly proportional to the smoothness of the image but inversely proportional to the accuracy of line positioning. The extraction result of the skeleton line breaks down its path into a set of points, providing coordinate points for data reconstruction. While extracting the skeleton using Gaussian detection, the line width of the XLD contour line is also extracted. The Parabolic mode is selected using LineMode to merge the subpixel-level skeleton lines into a continuous skeleton line, and finally, smoothing is performed.

2. The method of claim 1, wherein, Determining a visual illumination scheme based on material properties, and acquiring an image based on the visual illumination scheme specifically includes: In a dark environment, a square shadowless light source is used to illuminate the wafer from all sides, and a camera with sufficient resolution is used to capture an image that is non-reflective and has high distinction between the optical fiber and the silicon wafer surface.

3. The method according to claim 1, characterized in that, The image is denoised using bilateral filtering, and the edge information of the distributed sensing fiber is extracted using sub-pixel edge detection technology based on the Canny algorithm. Specifically, this includes: First, open the acquired image at its original size without stretching. Convert the color image to a grayscale image. Set the region to be processed within the grayscale image, and draw a rectangular Region of Interest (ROI) to be processed, denoted as region1. Draw multiple regions for other unused fiber optic segments, and then merge the drawn regions, denoted as region2. Obtain the image processing area. On the grayscale image, [the image processing area is then defined]. The region requiring fiber optic path identification is obtained, and the preprocessed image is obtained. According to Formulas 1 and 2, the image is denoised using bilateral filtering: Formula 1 ; Official 2; Among them, parameters and To smooth the parameters, and These are pixels ( )and( The grayscale values ​​of () are normalized after the weights are calculated. (i, j) represents the pixel point ( The grayscale value after noise reduction; The image is smoothed by convolving it with a Gaussian filter, with a size of [value missing]. The generation equation for the Gaussian filter kernel is shown in Equation 3: Equation 3; According to Equation 4, the gradient images in the x and y directions are obtained using a Sobel filter, and then the gradient intensity is calculated. and gradient direction : Official 4; in, and Sobel operators and In the image Convolution of window A; Based on Equation 5, an algorithm that calculates non-maximum suppression and hysteresis thresholding is used to link edge points into edges, thereby detecting points in the previous image. amplitude If greater than If it is an edge point, it will be immediately accepted as an edge point, and the output image point will be... The grayscale value is set to 255, while the amplitude is less than... Points that are not accepted are rejected, while other points that are connected to accepted edge points are also accepted as edges. Official 5; The sub-pixel edge coordinates are finally obtained by quadratic polynomial fitting shown in formula 6 ; Official 6; in, and The x and y coordinates of the current integer coordinate edge point. and These are the gradient values ​​to the left and right of the edge point. The gradient value at the edge point. The distance from adjacent pixels to the edge point; According to the area feature selection region, for each input region, the area feature is calculated, and if the calculated feature of each region is within the limit (6000, 2 e+006 ) the region will be output; Close the XLD outline and then fill it into a region, denoted as region3; An opening operation is performed on region3, and the fiber edge is extracted by subtracting the eroded region3 from the preprocessed image.

4. The method of claim 1, wherein, The method further includes: The sub-pixel-level skeleton coordinates extracted by Gaussian detection are obtained by the get_contour_xld operator. Then, based on the interval d of fiber data acquisition determined by the spatial resolution of the actual fiber demodulator, the coordinates of interval d on the skeleton are filtered out to achieve a one-to-one correspondence between the fiber detection data and the actual position of the silicon wafer.

5. A sub-pixel-level sensing fiber optic path Gaussian extraction system, characterized in that, The system for the method according to any one of claims 1 to 4 comprises: An industrial camera, connected to a computer, is used to acquire images of inspected silicon wafers and transmit them to the computer. A computer is used to determine a visual illumination scheme based on material properties, acquire the image, reduce noise in the image through bilateral filtering, extract edge information of the distributed sensing fiber using sub-pixel edge detection technology based on the Canny algorithm, close edge pairs based on the edge information, and extract skeleton information using Gaussian line detection method to obtain the path of the distributed sensing fiber. A worktable is used to fix the industrial camera above the silicon wafer to be inspected; A square shadowless light source is used to provide visual illumination for a silicon wafer to be inspected, positioned at its center, based on a defined visual illumination scheme.

6. The system of claim 5, wherein, The industrial camera is a five-megapixel camera with a resolution of 3856×2764 and a lens focal length of 12mm.

7. The system of claim 5, wherein, The visual illumination scheme is as follows: in a dark environment, a square shadowless light source is used to illuminate the wafer from all sides, and a camera with sufficient resolution is used to acquire an image that is non-reflective and has a high degree of distinction between the optical fiber and the silicon wafer surface.

8. The system of claim 5, wherein, The computer is specifically used for: First, open the acquired image at its original size without stretching. Convert the color image to a grayscale image. Set the region to be processed within the grayscale image, and draw a rectangular Region of Interest (ROI) to be processed, denoted as region1. Draw multiple regions for other unused fiber optic segments, and then merge the drawn regions, denoted as region2. Obtain the image processing area. On the grayscale image, [the image processing area is then defined]. The region requiring fiber optic path identification is obtained, and the preprocessed image is obtained. According to Formulas 1 and 2, the image is denoised using bilateral filtering: Official 1; Formula 2; Among them, parameters and To smooth the parameters, and These are pixels ( )and( The grayscale values ​​of () are normalized after the weights are calculated. For pixels ( The grayscale value after noise reduction; The image is smoothed by convolving it with a Gaussian filter, with a size of [value missing]. The generation equation for the Gaussian filter kernel is shown in Equation 3: Equation 3; According to Equation 4, the gradient images in the x and y directions are obtained using a Sobel filter, and then the gradient intensity is calculated. and gradient direction : Formula 4; in, and Sobel operators and In the image Convolution of window A; Based on Equation 5, an algorithm that calculates non-maximum suppression and hysteresis thresholding is used to link edge points into edges, thereby detecting points in the previous image. amplitude If greater than If it is an edge point, it will be immediately accepted as an edge point, and the output image point will be... The grayscale value is set to 255, while the amplitude is less than... Points that are not accepted are rejected, while other points that are connected to accepted edge points are also accepted as edges. Formula 5; The sub-pixel edge coordinates are finally obtained by quadratic polynomial fitting shown in formula 6 ; Official 6; in, and The x and y coordinates of the current integer coordinate edge point. and These are the gradient values ​​to the left and right of the edge point. The gradient value at the edge point. The distance from adjacent pixels to the edge point; According to the area feature selection region, for each input region, the area feature is calculated, and if the calculated feature of each region is within the limit (6000, 2 e+006 ) the region will be output; Close the XLD outline and then fill it into a region, denoted as region3; An opening operation is performed on region3, and the fiber edge is extracted by subtracting the eroded region3 from the preprocessed image. Regions are formed by merging sub-pixel level edges into edge pairs and then closing them. Image interference is reduced through threshold selection and image opening operations. Finally, the skeleton path is extracted using a Gaussian line detection method. Specifically, the Gaussian line detection method includes: The partial derivative of the convolution of the image with a Gaussian mask is used to determine the position of each pixel in the image. and The Taylor quadratic polynomial parameters in the direction are used to calculate the line direction of each pixel. The second partial derivative perpendicular to the direction of the line In this process, pixels exhibiting local maxima are marked as skeleton points, and a hysteresis thresholding operation is performed, accepting values ​​where the second derivative is greater than 1. The line point, rejecting the second derivative being less than All other line points that are adjacent to an accepted point are marked as skeleton points, and the discovered line points are eventually connected to form skeleton lines. Based on Formula 7, the contrast of grayscale values ​​of the lines to be extracted is calculated. and With the selected Value calculation parameters and : Official 7; Among them, parameter s determines the amount of smoothing to be performed by the Gaussian mask, which is directly proportional to the smoothness of the image but inversely proportional to the accuracy of line positioning. The extraction result of the skeleton line splits its path into a set of points, providing coordinate points for data reconstruction. While extracting the skeleton using Gaussian detection, the line width of the XLD contour line is also extracted. The Parabolic mode is selected using LineMode to merge the subpixel-level skeleton lines into a continuous skeleton line, and finally, smoothing is performed.

9. The system of claim 5, wherein, The computer is further used for: The sub-pixel-level skeleton coordinates extracted by Gaussian detection are obtained by the get_contour_xld operator. Then, based on the interval d of fiber optic data acquisition determined by the actual spatial resolution of the demodulator, the coordinates of interval d on the skeleton are filtered out to achieve a one-to-one correspondence between the fiber optic detection data and the actual position of the silicon wafer.

Images