A chip edge chipping defect detection system

Abstract

This invention relates to the field of defect detection and discloses a chip edge chipping defect detection system for accurately distinguishing between genuine edge chipping and normal edge disturbances. It includes a support device for supporting and fixing the chip to be inspected; a drive control device connected to the support device and / or an imaging device for driving relative movement between the chip and the imaging device; a light source device for providing illumination to the edge region of the chip; an imaging device disposed on one side of the support device for acquiring an edge image of the chip under the illumination of the light source device; and a data processing device communicatively connected to the imaging device and the drive control device for receiving and processing the edge image. This invention improves the accuracy and robustness of detecting minute edge chipping defects in semiconductor chips by comprehensively analyzing the physical geometric shape deviations and local optical gradient variations of the chip edge.

Description

Technical Field

[0001] This invention relates to the field of defect detection, and more particularly to a chip edge chipping defect detection system. Background Technology

[0002] As semiconductor packaging technology continues to advance towards higher density and finer pitch, chip edges are highly susceptible to chipping defects due to mechanical or thermal stress during processes such as dicing, thinning, scribing, and handling. These chipping defects not only reduce the chip's mechanical strength and yield but can also trigger crack propagation during subsequent packaging and soldering, severely impacting product reliability. Therefore, online, high-precision, and high-reliability chipping defect detection has become an indispensable quality control method in the semiconductor back-end packaging and testing process.

[0003] Currently, chip edge chipping detection primarily relies on machine vision technology. Typical solutions include geometric detection methods based on edge contour fitting residuals, texture analysis methods based on grayscale gradient changes, and classification methods based on deep learning. Among these, traditional machine vision-based methods still dominate in industrial settings due to their advantages such as strong interpretability, low deployment cost, and fast inference speed. However, chipping defects themselves are characterized by narrow width (typically only a few micrometers to tens of micrometers), irregular shape, and low contrast with normal edge transition areas, placing extremely high demands on the sensitivity and robustness of detection algorithms.

[0004] Most existing technologies directly use pixel-level edge coordinates as the analysis object, ignoring the sub-pixel-level edge position information contained in the continuous grayscale changes between pixels. The scale of edge chipping defects is often at the sub-pixel level. Relying solely on pixel-level edge coordinates for geometric analysis will cause minute distortions in the edge morphology to be masked by quantization errors, thereby reducing detection sensitivity.

[0005] In existing methods, edge fitting residuals are often directly used as the basis for defect identification. However, the absolute value of this index is greatly affected by factors such as the range of edge point coordinates, window size, and overall edge tilt angle. The absolute values ​​of the residuals are not comparable between windows at different locations and of different sizes, making it difficult for fixed threshold strategies to adapt to global detection needs and easily generating a large number of missed detections or false alarms.

[0006] When calculating the gradient magnitude in edge regions, non-edge-breaking factors such as minute textures and scratches on the chip surface can also generate gradient responses, forming false edge interference. Existing methods do not perform effective sparsification of the gradient sequence, resulting in the gradient variation coefficient being contaminated by false edge signals, further reducing the accuracy of discrimination.

[0007] Therefore, we propose a chip edge chipping defect detection system to solve the above problems. Summary of the Invention

[0008] This invention provides a chip edge chipping defect detection system for accurately distinguishing between actual edge chipping and normal edge disturbances.

[0009] The first aspect of the present invention provides a chip edge chipping defect detection system, the chip edge chipping defect detection system comprising: a carrier device for carrying and fixing a chip to be inspected; a drive control device connected to the carrier device and / or an imaging device for driving relative movement between the chip to be inspected and the imaging device; a light source device for providing illumination to the edge region of the chip to be inspected; an imaging device disposed on one side of the carrier device for acquiring an edge image of the chip to be inspected under the illumination of the light source device; and a data processing device communicatively connected to the imaging device and the drive control device for receiving and processing the edge image, the data processing device comprising: an acquisition module for acquiring the edge image of the chip to be inspected and extracting the edge... The edge points in the edge image are used to obtain a sequence of edge points arranged along the edge direction of the chip to be detected; a processing module is used to move a window of a preset size along the edge point sequence, perform line fitting based on the edge points within the window to obtain a reference benchmark, and calculate a first feature value based on the deviation of the edge points within the window relative to the reference benchmark; an extraction module is used to obtain pixel gradient information in the edge image region corresponding to the window, and calculate a second feature value based on the distribution difference of the pixel gradient information; a fusion module is used to fuse the first feature value and the second feature value to obtain a defect risk score for the window; a judgment module is used to determine whether there is a chipping defect at the position corresponding to the window based on the defect risk score and preset judgment conditions.

[0010] Optionally, in a first implementation of the first aspect of the present invention, it further includes: Calculate the absolute value of the coordinate difference between two adjacent edge points within the window in the direction perpendicular to the edge; Edge points whose absolute value of coordinate difference exceeds a preset difference threshold are removed, and the remaining edge points in the window are taken as the valid edge point sequence. The method of performing line fitting based on edge points within the window includes: performing line fitting on the valid edge points in the valid edge point sequence.

[0011] Optionally, in a second implementation of the first aspect of the present invention, the first feature value is specifically a normalized local residual ratio, and the calculation of the first feature value based on the deviation of the edge points within the window relative to the reference benchmark includes: The reference benchmark is obtained by performing a linear fit on the effective edge point sequence using the least squares method. Calculate the vertical coordinate difference between each valid edge point and its corresponding position on the reference datum to obtain the residual; Calculate the residual standard deviation for all residuals within the window; Obtain the maximum and minimum values ​​of the vertical coordinates in the effective edge point sequence to calculate the range; The normalized local residual ratio is calculated based on the ratio between the residual standard deviation and the range.

[0012] Optionally, in a third implementation of the first aspect of the present invention, when calculating the normalized local residual ratio based on the ratio of the residual standard deviation to the range, a physical precision limit factor is introduced, and the formula for calculating the normalized local residual ratio is: ; In the formula, To normalize the local residual ratio, This represents the number of valid edge points in the valid edge point sequence. Let be the vertical coordinate of the i-th valid edge point in the valid edge point sequence. The vertical coordinates of the position corresponding to the i-th valid edge point on the reference datum. and These represent the maximum and minimum vertical coordinates of the valid edge points in the valid edge point sequence, respectively. This is the physical precision limit factor.

[0013] Optionally, in a fourth implementation of the first aspect of the present invention, the second feature value is the gradient variation coefficient, and the step of obtaining pixel gradient information in the edge image region corresponding to the window and calculating the second feature value based on the distribution difference of the pixel gradient information includes: The gradient operator is used to calculate the gradient magnitude of pixels in the edge image region in the direction perpendicular to the edge, forming a gradient magnitude sequence; Non-maximum suppression processing is applied to the gradient magnitude sequence to retain pixels with local maxima along the edge direction, thus forming an updated gradient magnitude sequence. Calculate the standard deviation and mean of the gradient magnitudes in the updated gradient magnitude sequence; The ratio of the standard deviation to the mean is used as the gradient coefficient of variation.

[0014] Optionally, in a fifth implementation of the first aspect of the present invention, the step of fusing the first feature value and the second feature value to obtain the defect risk score of the window includes: The gradient variation coefficient is added to a preset constant to obtain the weight sum value; Multiply the first feature value by the weight sum value, and use the resulting product as the defect risk score.

[0015] Optionally, in a sixth implementation of the first aspect of the present invention, a step of correcting the defect risk score is further included: Calculate the variance of the vertical coordinates of all edge points within the window; If the variance is less than the first preset variance threshold, the defect risk score is multiplied by a first correction coefficient less than 1 for weight reduction. If the variance is greater than the second preset variance threshold, the defect risk score is multiplied by a second correction coefficient greater than 1 for weighting. The revised defect risk score will be used as the basis for the final determination of whether there is edge chipping defect at the location corresponding to the window.

[0016] Optionally, in a seventh implementation of the first aspect of the present invention, the method includes: A baseline risk score is obtained for multiple sample windows on the edge region of a defect-free chip. A preset threshold is set based on the mean and standard deviation of the baseline risk score as the preset judgment condition. If the defect risk score of the window is greater than the preset threshold, then the window is marked as a candidate defect location; According to the positional order along the edge direction of the chip to be tested, the consecutive adjacent candidate defect positions are merged into a target defect region, and the maximum defect risk score in the target defect region is used as the severity index of the target defect region.

[0017] Optionally, in an eighth implementation of the first aspect of the present invention, a size calibration module is further included: Obtain a first sample chip containing known defects and a second sample chip without defects; For multiple candidate sizes, calculate the first risk score set corresponding to the first sample chip and the second risk score set corresponding to the second sample chip respectively; Calculate the first mean and first standard deviation of the first risk score set, and the second mean and second standard deviation of the second risk score set; Calculate the difference between the first mean and the second mean, and divide the difference by the sum of the first standard deviation and the second standard deviation to obtain the separation evaluation index corresponding to each candidate size; The candidate size corresponding to the largest separation evaluation index is determined as the preset size.

[0018] The mechanism of this invention is as follows: Sub-pixel edge point sequences are extracted along the chip edge through a sliding window. Within each window, the local linear deviation of the edge contour and the local variation of the gray-level gradient are evaluated simultaneously. The former uses the ratio of the standard deviation of the edge point fitting residual to the vertical coordinate range to eliminate the scale effect of normal edge undulations and obtain the normalized local residual ratio, which reflects the geometric abrupt change caused by edge collapse. The latter uses the ratio of the standard deviation of the gradient magnitude to the mean to obtain the gradient variation coefficient, which reflects the gray-level disorder caused by scattering or breakage in the edge collapse area. Beneficial effects: The risk score of edge collapse is constructed by fusing the product of the normalized local residual ratio and the gradient coefficient of variation. The normalized local residual ratio eliminates the influence of coordinate range and window scale by dividing the residual standard deviation by the range, making the geometric deviation at different locations globally comparable. The gradient coefficient of variation, which is the ratio of the standard deviation to the mean, represents the degree of dispersion of the gradient distribution and effectively distinguishes between normal edges and severe gradient fluctuations at the edge collapse. It achieves nonlinear synergistic enhancement of geometric and texture information, and the score only increases significantly when both types of anomalies exist at the same time, which greatly improves the clarity of classification boundaries. It introduces data-driven window length calibration, anomaly point removal based on the difference between adjacent points, and pseudo-edge filtering based on non-maximum suppression, which systematically eliminates the interference caused by noise, pseudo-edges and empirical parameter settings, making the detection scheme have strong robustness and adaptability. At the defect judgment and output level, the threshold is objectively calibrated based on the statistical distribution of the score of the edge-free region, and a variance adaptive correction mechanism is introduced. When the edge is stable, the score is reduced to suppress false alarms, and when the edge fluctuates, the score is increased to reduce false alarms. Combined with the intelligent merging of candidate defect regions and the quantitative output strategy that uses the maximum score to represent the severity, the detection result is no longer a simple binary judgment, but a quantifiable defect report with severity classification. From feature representation, parameter optimization, noise suppression to result output, a complete closed loop is formed. This system systematically solves the shortcomings of existing technologies, such as insufficient positioning accuracy, single feature, poor scale adaptability, subjective threshold, and non-quantifiable output. It achieves highly sensitive, robust, quantifiable, and adaptive online chip edge damage detection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of one embodiment of the chip edge chipping defect detection system of the present invention; Figure 2 This is a schematic diagram of another embodiment of the chip edge chipping defect detection system of the present invention; Figure 3 This diagram illustrates the grayscale image of the chip edge acquired by a line scan camera, and how to accurately extract the chip edge point sequence using sub-pixel processing technology. Figure 4This is a schematic diagram of one embodiment of the chip chipping defect detection device in this invention. Detailed Implementation

[0020] This invention provides a chip edge chipping defect detection system for accurately distinguishing between actual edge chipping and normal edge disturbances. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 An embodiment of the chip edge chipping defect detection system of the present invention includes a carrier device for carrying and fixing the chip to be inspected; a drive control device connected to the carrier device and / or an imaging device for driving relative movement between the chip to be inspected and the imaging device; a light source device for providing illumination to the edge region of the chip to be inspected; an imaging device disposed on one side of the carrier device for acquiring an edge image of the chip to be inspected under the illumination of the light source device; and a data processing device communicatively connected to the imaging device and the drive control device for receiving and processing the edge image, wherein the data processing device includes: 101. Acquisition module, used to acquire grayscale images of the chip edge, extract sub-pixel edge points from each row, and obtain a sequence of sub-pixel edge points arranged along the chip edge direction.

[0022] It is understood that the executing entity of this invention can be a chip chipping defect detection device, a terminal, or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0023] It should be noted that during the hardware image acquisition phase, the system utilizes an industrial area scan camera with a resolution of 4096×3000px, along with a telecentric lens and a coaxial incident light source, to acquire images. The optical system resolution is calibrated to 0.5μm / px. The camera is aimed at the edge of the chip on the stage and exposed to obtain a black and white grayscale image. In the image, the grayscale values ​​of the bright areas of the chip body are distributed between 200 and 240, while the grayscale values ​​of the dark areas of the external background are distributed between 30 and 60.

[0024] A rectangular region of interest (ROI) of 1000 rows × 200 columns was defined in the full-frame image. This region covers a physical length of 500 μm vertically, with row numbers from 1 to 1000 and column numbers from 1 to 200. The chip edge penetrates vertically through this region.

[0025] In the edge localization stage, the system processes 1000 rows of pixel data line by line. Taking row 500 as an example, grayscale values ​​are read from column 1 to column 200, and the grayscale difference between adjacent pixels is calculated to obtain the gradient sequence. The system locates the point where the grayscale gradient jump is the largest in this row between columns 98 and 99, and marks column 98 as a coarse edge point.

[0026] To achieve sub-pixel accuracy, the system extracts the discrete gradient values ​​from columns 97, 98, and 99, and uses these three values ​​for one-dimensional parabolic space interpolation. The extreme coordinates of the fitted parabola's vertex are calculated, deriving the actual horizontal coordinate of the physical edge as 98.35px. The resulting two-dimensional sub-pixel coordinates are (500, 98.35).

[0027] The system iterates through rows 1 to 1000 from top to bottom, outputting a sub-pixel coordinate for each row through spatial interpolation. The 1000 coordinate points are combined in ascending order of row number to form a one-dimensional sub-pixel edge point sequence describing the edge contour undulations within a 500μm length range of the chip, providing underlying data for local geometric feature analysis.

[0028] 102. Processing module, used to slide a window of preset length along the sub-pixel edge point sequence, perform straight line fitting on all edge points in each window, obtain the residual of each edge point relative to the fitted straight line, calculate the standard deviation of all residuals in the window, and calculate the range of the vertical coordinates of the edge points in the window. Divide the standard deviation of the residuals by the range to obtain the normalized local residual ratio.

[0029] It should be noted that, on a sequence of 1000 subpixel 2D coordinate points, the sliding window length is set to 50 data points, and the sliding step size is 1 data point.

[0030] A specific window covering rows 401 to 450 was selected for processing. This window contains 50 consecutive edge points. There is a local gap at the physical edge of the chip, causing the horizontal coordinates of 5 points (rows 423 to 427) to be concave inward, while the remaining 45 points are in normal edge positions.

[0031] The least squares method was applied to fit a straight line to these 50 points to obtain a baseline line. Based on the mathematical property that the algebraic sum of the least squares method approaches zero, the system calculated that the mean residual of the 45 normal edge points relative to the fitted line is +0.15px; while the mean residual of the 5 notched edge points deviates in the opposite direction and to a greater extent, with a mean residual of -1.35px. The sum of the residuals within this window is 45 × 0.15 + 5 × (-1.35) = 6.75 - 6.75 = 0, which conforms to the fitting logic.

[0032] Calculate the standard deviation of these 50 residual data. The system calculates the sum of squared residuals: 45 × (0.15) is generated from the 45 normal points. 2 =1.0125 sum of squares, 5 gap points produce 5×(-1.35) 2 =9.1125. The total sum of squares is 10.125. Dividing by the sample size of 50, the variance is 0.2025. Taking the square root of 0.2025, the standard deviation of the residuals for this window is calculated to be 0.45px.

[0033] To eliminate the influence of window size under different tilt angles, the maximum and minimum values ​​of the vertical coordinates of the effective edge points within the window were extracted to calculate the range. The calculated maximum vertical coordinate (Xmax) of the normal edge points is approximately 98.50 pixels, and the minimum vertical coordinate (Xmin) of the recessed notch is approximately 96.50 pixels. The range of the window's vertical coordinates is 2.00 pixels. Dividing the residual standard deviation (0.45) by the range (2.00) yields a quotient of 0.2250.

[0034] 103. Extraction module, used to calculate the gradient magnitude of pixels within the original grayscale image region corresponding to each window, obtain the gradient magnitude sequence, and then calculate the ratio of the standard deviation of the gradient magnitude to the mean to obtain the gradient variation coefficient.

[0035] It should be noted that the original grayscale image region covering rows 401 to 450 is processed simultaneously. Grayscale information of 50 rows of pixels within this image region is extracted to analyze optical transition characteristics.

[0036] Within each row's pixel column interval, the absolute value of the grayscale difference is calculated, and the maximum value is taken as the gradient magnitude for that row. In this window, the normal edge areas in rows 45 have smooth cross-sections and uniform reflection, maintaining a high gradient magnitude with a mean of 170. However, the five rows from rows 423 to 427, with their notched areas, have rough cross-sections that cause diffuse reflection, resulting in blurred light and dark boundaries. The gradient magnitudes in these five rows decrease, with a mean of 88. Arranging these data points by row number creates a gradient magnitude sequence of 50 data points.

[0037] Statistical calculations are performed on this sequence. First, the arithmetic mean is calculated, and the sum is... Dividing by 50 gives the mean of 161.8.

[0038] The deviation of the 45 normal values ​​from the mean is 170 - 161.8 = 8.2, which squares to 67.24. The deviation of the 5 missing values ​​from the mean is 88 - 161.8 = -73.8, which squares to 5446.44. The total sum of squares is calculated as 45 × 67.24 + 5 × 5446.44 = 3025.8 + 27232.2 = 30258. Dividing 30258 by 50 yields a variance of 605.16. Taking the square root of 605.16 gives a standard deviation of 24.6 for the gradient magnitude.

[0039] Dividing the calculated gradient magnitude standard deviation of 24.6 by the mean of 161.8 yields a quotient of 0.152, which is the gradient variation coefficient for this window, quantifying the dispersion of local edge grayscale changes.

[0040] 104. The fusion module is used to multiply the normalized local residual ratio by one and the sum of the gradient variation coefficient to obtain the edge collapse risk score for each window.

[0041] It should be noted that geometric feature indicators and optical feature indicators are fused. For the gap window containing rows 401 to 450, the system calls the normalized local residual ratio of 0.2250 calculated in step 102 and the gradient coefficient of variation of 0.152 obtained in step 103.

[0042] The calculation process adds the gradient variation coefficient 0.152 to a constant 1, resulting in a sum of 1.152. This result serves as an amplification factor for the optical anomaly feature's geometric deviation. The normalized local residual ratio 0.2250 is multiplied by this amplification factor 1.152, yielding a value of 0.25920. The system rounds the result to four decimal places, resulting in 0.2592. This value of 0.2592 is the edge collapse risk score for this defect window.

[0043] For comparative verification, the system performed the same operation on windows with intact edges covering rows 101 to 150. The physical edges of this normal window are straight and the grayscale transition is consistent. Its vertical coordinate range is measured to be 0.50px, the residual standard deviation is 0.05px, the normalized local residual ratio is calculated to be 0.1000, the gradient coefficient of variation is 0.020, and the resulting edge collapse risk score for the normal window is 0.10200 (rounded to 0.1020).

[0044] Two sets of data verified the effectiveness of feature fusion: the score of normal areas with intact edges was 0.1020, while the score of defective areas with actual damage was amplified to 0.2592. By traversing the one-dimensional sequence formed by 1000 nodes, the score of each window was calculated one by one to form a global edge collapse risk score array.

[0045] 105. Judgment module, used to compare the edge collapse risk score with a preset threshold. If the score is greater than the threshold, it is determined that there is an edge collapse defect at the window position.

[0046] It should be noted that preliminary calculations and measurements on 500 normal chips and 100 defective chips showed that the sliding window risk score for areas with intact edges remained consistently below 0.1500, while the sliding window score for areas with physical damage was concentrated above 0.2000. To ensure that good chips are not mistakenly rejected and defects are not missed, the midpoint between the upper limit of normal and the lower limit of defects, 0.1750, was set as the preset threshold.

[0047] Read the global edge collapse risk score array obtained in step 104, and perform binary classification judgment logic for each window against the preset threshold of 0.1750.

[0048] For the sliding window from row 101 to row 150, the score is 0.1020. Since 0.1020 is less than the preset threshold of 0.1750, the system determines that the area meets the quality requirements and records the status as normal.

[0049] For the sliding window from row 401 to row 450, the score is 0.2592. 0.2592 is greater than the preset threshold of 0.1750, triggering the defect condition. The system determines that there is contour deformation and optical anomaly at this physical location, and marks the area as having a chipped edge defect.

[0050] The data comparison and judgment results for some windows are shown in Table 1 below: Table 1 101 to 150 0.1020 0.1750 Less than the threshold normal 401 to 450 0.2592 0.1750 Greater than the threshold There is a chipping defect. 701 to 750 0.0918 0.1750 Less than the threshold normal 801 to 850 0.2365 0.1750 Greater than the threshold There is a chipping defect. After the full image comparison is completed, if any chipping defect marker is found in the output record, the system will determine that the overall quality of the chip is unqualified and output the physical coordinates corresponding to the defect window.

[0051] Please see Figure 2 Another embodiment of the chip edge chipping defect detection system in this invention includes a carrier device for carrying and fixing the chip to be inspected; a drive control device connected to the carrier device and / or an imaging device for driving relative movement between the chip to be inspected and the imaging device; a light source device for providing illumination to the edge region of the chip to be inspected; an imaging device disposed on one side of the carrier device for acquiring an edge image of the chip to be inspected under the illumination of the light source device; and a data processing device communicatively connected to the imaging device and the drive control device for receiving and processing the edge image, wherein the data processing device includes: 201. Acquisition module, used to acquire grayscale images of the chip edge, extract sub-pixel edge points from each row, and obtain a sequence of sub-pixel edge points arranged along the chip edge direction.

[0052] Specifically, a line scan camera is used to acquire grayscale images of the chip edge. For each row of images, the Otsu method is used to determine the grayscale threshold of the edge region. A pixel band of fixed width is taken near the threshold, and the grayscale values ​​of each pixel in the pixel band are weighted and averaged to calculate the sub-pixel coordinates, thereby obtaining the sub-pixel edge point sequence.

[0053] It should be noted that, taking the detection of a straight edge of a semiconductor chip as an example, the specific process of image acquisition and obtaining sub-pixel edge point sequence is explained.

[0054] The line scan camera is activated to scan along the physical edge of the chip, acquiring a 2D grayscale image of the chip's edge. The image width is set to 1000 pixels, and the height along the chip edge is set to 2000 pixels. The grayscale value of each pixel in the image ranges from 0 to 255. The grayscale value of the background area is concentrated around 30, while the grayscale value of the chip body area is concentrated around 210. There is a grayscale gradient transition area at the boundary between the background and the chip.

[0055] The system extracts all pixel grayscale data from the entire 2000-row by 1000-column image and applies Otsu's method for global analysis. Otsu's method searches for the grayscale level that maximizes the inter-class variance between the background and foreground pixel sets by iterating through all grayscale levels from 0 to 255. The system determines that the grayscale value corresponding to the maximum inter-class variance is 115. This value is used as the globally fixed grayscale threshold for the edge regions of the entire image to avoid threshold fluctuations caused by noise interference during row-by-row calculations.

[0056] Starting from row 1 of the image, the system processes each row sequentially up to row 2000. Taking row 100 as an example, the system searches from left to right for the pixel position where the grayscale value first crosses the global threshold of 115. The grayscale value of the pixel in column 500 is set to 110, and the grayscale value of the pixel in column 501 is set to 155. The threshold of 115 lies between these two columns. Using this position as a reference, the system extracts a pixel band with a fixed width of 5 pixels, covering columns 499 to 503. The specific grayscale values ​​of these five consecutive pixels are set as follows: column 499 is 85, column 500 is 110, column 501 is 155, column 502 is 180, and column 503 is 210.

[0057] Calculate the weighted average of the gray values ​​of the five pixels within the pixel band. Multiply the column coordinate value of each pixel by its corresponding gray value to obtain the product. Add these five products together to get the sum, and then divide the sum by the sum of the gray values ​​of the five pixels.

[0058] In the calculation for row 100, the sum of the products of each column coordinate and its grayscale value is 371060; the sum of the grayscale values ​​of the five pixels is 740. Dividing 371060 by 740 yields the result 501.432. The value 501.432 is the sub-pixel coordinate of the edge of row 100 in the horizontal direction.

[0059] For rows 1 to 2000 of the image, a process of pixel band selection and weighted average calculation is performed one by one based on a globally fixed threshold, calculating a corresponding one-dimensional sub-pixel coordinate for each row. These 2000 coordinate points are arranged in spatial order to form a sub-pixel edge point sequence describing the chip edge direction.

[0060] 202. Processing module, used to slide a window of preset length along the sub-pixel edge point sequence, perform straight line fitting on all edge points in each window, obtain the residual of each edge point relative to the fitted straight line, calculate the standard deviation of all residuals in the window, and calculate the range of the vertical coordinates of the edge points in the window. Divide the standard deviation of the residuals by the range to obtain the normalized local residual ratio.

[0061] Specifically, the length of the window is preset to an odd number of pixels based on the range of edge width. Within each window, the least squares method is used to fit a straight line to all edge points to obtain the parameters of the fitted straight line. Then, the difference between the vertical coordinate of each edge point and the vertical coordinate of the corresponding position on the fitted straight line is calculated as the residual. The square root of the sum of the squares of all residuals within the window is calculated by dividing by the window length to obtain the standard deviation of the residuals. At the same time, the maximum and minimum values ​​of the vertical coordinates of the edge points within the window are found, and the difference between the two is used to obtain the range. The ratio of the standard deviation of the residuals to the range is the normalized local residual ratio.

[0062] Furthermore, before using the least squares method for line fitting, the absolute value of the vertical coordinate difference between each edge point and its adjacent edge points within the window is calculated. When the absolute value exceeds a preset threshold, the edge point is marked as an outlier and removed. The remaining edge points form a valid edge point sequence for subsequent line fitting, residual calculation, and acquisition of the normalized local residual ratio.

[0063] It should be noted that the calculation process of the normalized local residual ratio is explained based on the edge point sequence obtained in the previous step. The chip edge baseline is set to be approximately vertical, and the vertical coordinates refer to the column coordinates perpendicular to the edge direction.

[0064] Set the sliding window length to 5 pixels. Slide the window to rows 101 to 105, and use the local vertical index y of the 5 edge points. i Labeled as 1 to 5, the corresponding sub-pixel vertical coordinates x i The numbers are, in order: 500.0, 500.1, 501.5, 501.6, and 500.2.

[0065] Calculate the absolute value of the vertical coordinate difference between adjacent edge points one by one: |500.1-500.0|=0.1; |501.5-500.1|=1.4; |501.6-501.5|=0.1; |500.2-501.6|=1.4; Given that the maximum allowable physical depth of edge chipping for this type of chip is 1.8 pixels, to prevent real defects from being filtered out, the preset threshold for eliminating isolated impulse noise is strictly set to 2.5 pixels. Since none of the calculated results exceed 2.5, the system determines that there are no impulse noise points within the window, and all five points are retained as valid edge points.

[0066] Linear equations were constructed by fitting lines to the five edge points. .

[0067] The slope k = 0.19 and the intercept b = 500.11 were obtained by least squares method.

[0068] Calculate the coordinates of the corresponding position on the fitted line. : hour, ; hour, ; hour, ; hour, ; hour, ; Calculate the actual vertical coordinates perpendicular coordinates to the fitted coordinates residuals: Residual 1: -0.30; Residual 2: -0.39; Residual 3: 0.82; Residual 4: 0.73; Residual 5: -0.86; The sum of squares of these 5 residuals is 2.187. Dividing the sum of squares by the window length of 5 gives 0.4374, and taking the square root gives the standard deviation of the residuals as 0.661 pixels.

[0069] Maximum vertical coordinate minimum value The range is 1.6 pixels.

[0070] Introducing a calculation formula that includes the physical precision limit factor: ; In the formula, NLRR is the normalized local residual ratio. Let be the optical resolution limiting constant (taken as 0.01 pixels). Substitute the data: The normalized local residual ratio output is 0.411.

[0071] 203. Extraction module, used to calculate the gradient magnitude of pixels within the original grayscale image region corresponding to each window, obtain the gradient magnitude sequence, and then calculate the ratio of the standard deviation of the gradient magnitude to the mean to obtain the gradient variation coefficient.

[0072] Specifically, for the original grayscale image area covered by each window, the Sobel operator is used to calculate the gradient magnitude of each pixel in the area in the direction perpendicular to the chip edge. The gradient magnitudes of all pixels are arranged according to their positions to form a gradient magnitude sequence. The ratio of the standard deviation of the gradient magnitude to the mean of the gradient magnitude is calculated, and the resulting ratio is the gradient variation coefficient.

[0073] Furthermore, after obtaining the gradient magnitude sequence, non-maximum suppression processing is performed on the sequence, retaining only the pixels with local maximum gradient magnitudes along the chip edge direction. The retained pixels form a new gradient magnitude sequence, and the ratio of the standard deviation to the mean of the gradient magnitude is calculated based on the new gradient magnitude sequence to obtain the gradient variation coefficient.

[0074] It should be noted that the current window is located on lines 101 to 105.

[0075] The system retrieves the original grayscale image data covering rows 101 to 105, with column coordinates ranging from 498 to 504. The Sobel operator is used to calculate the horizontal gradient magnitude of each pixel within this region. The system utilizes a horizontal Sobel convolution kernel with weighted coefficients to perform a weighted summation of the grayscale values ​​of the target pixel and its eight neighboring pixels, taking the absolute value. After pixel-by-pixel calculation, the system constructs an initial gradient magnitude sequence.

[0076] The initial gradient magnitude sequence is subjected to non-maximum suppression row by row. Taking row 101 as an example, a horizontal comparison reveals that the gradient magnitude of 120 in column 500 is a local maximum for that row. The system retains the value 120 and sets the remaining gradient magnitudes in that row to zero. The same filtering is performed on rows 102 to 105.

[0077] The five extracted values ​​are as follows: The local maximum value in row 101 is 120; the local maximum value in row 102 is 118; the local maximum value in row 103 is 75; the local maximum value in row 104 is 82; and the local maximum value in row 105 is 115. The five gradient magnitudes in the sequence are added together, and the sum is 510. Dividing the sum by 5, the mean of the gradient magnitudes is 102.

[0078] Calculate the differences between the five gradient magnitudes and the mean of 102, obtaining +18, +16, -27, -20, and +13 respectively; square the differences to get 324, 256, 729, 400, and 169; sum them up to get 1878; divide by 5 to get the variance of 375.6; take the square root to get the standard deviation of 19.38.

[0079] Dividing the standard deviation of 19.38 by the mean of 102 yields a gradient coefficient of variation of 0.19.

[0080] 204. The fusion module is used to multiply the normalized local residual ratio by one and the sum of the gradient variation coefficient to obtain the edge collapse risk score for each window.

[0081] Specifically, the gradient variation coefficient is first added to the value 1 to obtain a sum, and then the normalized local residual ratio is multiplied by this sum. The resulting product is the edge collapse risk score.

[0082] Furthermore, after obtaining the edge collapse risk score, the variance of the vertical coordinates of all edge points within the window is calculated. If the variance is less than the first preset variance threshold, the edge collapse risk score is multiplied by a first correction coefficient less than 1. If the variance is greater than the second preset variance threshold, the edge collapse risk score is multiplied by a second correction coefficient greater than 1. The corrected edge collapse risk score is used as the object for comparison with the preset threshold.

[0083] It should be noted that the calculation and correction process for edge collapse risk scoring based on residual data is to eliminate the absolute coordinate system deviation caused by the tilted placement of the chip.

[0084] The system retrieves the parameters calculated in the previous steps: the window normalized local residual ratio is 0.411, and the gradient coefficient of variation is 0.19. The system adds the gradient coefficient of variation (0.19) to a constant value of 1, resulting in a sum of 1.19. Then, it multiplies the normalized local residual ratio (0.411) by the sum of 1.19. The calculated value is 0.489, which is used as the initial, uncorrected edge collapse risk score for the current window.

[0085] To evaluate the dispersion of the local contour after removing the linear trend, the system extracts five effective residuals obtained from the straight-line fitting in step 202, which are -0.30, -0.39, 0.82, 0.73, and -0.86, respectively. According to the mathematical properties of the least squares method, the mean of the fitted residuals must approach 0. The system squares these five residual values ​​respectively, obtaining square values ​​of 0.0900, 0.1521, 0.6724, 0.5329, and 0.7396. The sum of these five square values ​​is 2.187. Finally, dividing the sum of squares by the number of data points (5) yields the variance of the residuals within this window, which is 0.4374. Calculating the variance based on the residuals completely eliminates the pseudo-high dispersion problem that might be caused by the overall chip placement tilt.

[0086] A first preset variance threshold of 0.10 and a second preset variance threshold of 0.40 are preset to evaluate the degree of residual fluctuation; the first correction coefficient is 0.8 and the second correction coefficient is 1.2.

[0087] The residual variance of the current window, 0.4374, is compared with the threshold. Because the actual variance of 0.4374 is greater than the second preset variance threshold of 0.40, the system determines that the local area still has significant positional changes after eliminating the linear trend. The system extracts the second correction coefficient of 1.2, multiplies it by the initial edge collapse risk score of 0.489, and obtains the final result of 0.587.

[0088] 205. Judgment module, used to compare the edge collapse risk score with a preset threshold. If the score is greater than the threshold, it is determined that there is an edge collapse defect at the window position.

[0089] Specifically, the preset threshold is obtained through calibration: the edge collapse risk scores of multiple windows are collected on the edge of the non-collapse area, and their mean and standard deviation are calculated. The mean plus a preset multiple of the standard deviation is used as the threshold. During online detection, the edge collapse risk score of each window is compared with the threshold. If the score is greater than the threshold, the window is marked as a candidate defect location. For all candidate defect locations, adjacent candidate defect locations are merged into a defect region according to the position order along the chip edge direction, and the maximum edge collapse risk score of each window in the region is used as the severity of the defect region.

[0090] It should be noted that this embodiment illustrates the process of threshold calculation, defect determination, and region merging.

[0091] Data from 1000 sliding windows were collected on the normal edges of the non-collapsing regions. Statistical analysis showed that the mean score for normal windows was 0.150, and the standard deviation was 0.050. A preset multiplier of 3 was set for the standard deviation, and the preset threshold of 0.300 was calculated by adding three times the standard deviation to the mean, and this threshold was fixed as the judgment baseline.

[0092] Slide the window along the edge in 1-pixel increments. Assume we are currently processing windows 99 to 103, where the risk score for window 101 (corresponding to rows 101 to 105) has been calculated to be 0.587.

[0093] The scores for these five consecutive windows are compared one by one with a preset threshold of 0.300. Scores greater than the threshold are marked as candidate defect locations; scores less than or equal to the threshold are marked as normal. The data and marking results are shown in Table 2 below: Table 2 99 Lines 99 to 103 0.210 0.300 normal 100 Lines 100 to 104 0.450 0.300 Candidate defects 101 Lines 101 to 105 0.587 0.300 Candidate defects 102 Lines 102 to 106 0.512 0.300 Candidate defects 103 Lines 103 to 107 0.280 0.300 normal The system performs positional sequence analysis on the windows marked as candidate defects. Windows 100, 101, and 102 are identified as spatially adjacent. The system merges these three consecutive windows, defining them as a single edge-break defect region.

[0094] After merging, extract all risk scores (0.450, 0.587, 0.512) within the region, filter out the maximum value of 0.587, record the location range of the edge chipping defect area, and output 0.587 as the severity index of the defect area.

[0095] 206. Size calibration module: Using at least one sample chip containing known edge chipping defects and at least one normal chip without edge chipping, the module is executed under multiple candidate window lengths. The edge chipping risk score of all windows under each candidate window length is calculated. The score distribution of the edge chipping samples and the score distribution of the normal samples are statistically analyzed. The candidate window length with the greatest separation between the two is determined as the preset length, which is used as the preset length of the sliding window in subsequent online detection.

[0096] Furthermore, for each candidate window length, the mean of the edge collapse risk score for the edge collapse sample and the mean of the edge collapse risk score for the normal sample are calculated, and the standard deviation of the edge collapse sample score and the standard deviation of the normal sample score are calculated. The difference between the mean of the edge collapse sample and the mean of the normal sample is divided by the sum of the standard deviations of the edge collapse sample and the standard deviations of the normal sample. The candidate window length with the largest value is the preset length.

[0097] It should be noted that two chips were selected: a sample chip containing known edge chipping defects and a normal sample chip with smooth edges. Four candidate window lengths were set: 3, 5, 7, and 9 pixels.

[0098] For each candidate window length, scores are collected at the edges of both collapsed and normal samples.

[0099] Taking a length of 5 pixels as an example: the mean score for normal samples is 0.150, and the standard deviation is 0.050; the mean score for samples with edge damage is 0.580, and the standard deviation is 0.080. The system also performed statistical calculations for lengths of 3, 7, and 9 pixels.

[0100] Based on the obtained statistical parameters, the degree of separation between the score distributions of normal samples and samples with edge collapse is calculated. The calculation logic is as follows: first, calculate the difference between the mean of the samples with edge collapse and the mean of the normal samples; then, calculate the sum of the standard deviations of the samples with edge collapse and the standard deviations of the normal samples; finally, divide the difference by the sum.

[0101] Taking a length of 5 pixels as an example, the mean difference is 0.580 minus 0.150, which gives 0.430; the sum of the standard deviations is 0.080 plus 0.050, which gives 0.130; dividing 0.430 by 0.130 gives a separation degree of approximately 3.31.

[0102] The statistical indicators and separation degree of candidate length are summarized in Table 3 below: Table 3 3 0.180 0.080 0.400 0.120 1.10 5 0.150 0.050 0.580 0.080 3.31 7 0.140 0.040 0.450 0.110 2.07 9 0.130 0.035 0.350 0.105 1.57 The comparison revealed that a window size of 3 pixels is susceptible to random noise, resulting in low separation; a window size of 9 pixels smooths out minor defects, further reducing separation. The separation score reaches its maximum at a length of 5 pixels, at 3.31. The system has determined 5 as the optimal window length and fixed it as the preset length for online detection.

[0103] Figure 4 This is a schematic diagram of a chip chipping defect detection device according to an embodiment of the present invention. The device 300 can vary considerably due to differences in configuration or performance. The device 300 includes a transmitter 301, a receiver 302, and a processor 303. The processor 303 can also be a controller. Figure 4 The device is designated as "controller / processor 303". Optionally, the device 300 may also include a modem processor 305, which may include an encoder 306, a modulator 307, a decoder 308, and a demodulator 309.

[0104] In one example, transmitter 301 modulates (e.g., analog-to-analog conversion, filtering, amplification, and up-conversion, etc.) the output sample to obtain an uplink signal, which is transmitted via an antenna to an access network device. On the downlink, the antenna receives the downlink signal transmitted by the access network device. Receiver 302 modulates (e.g., filtering, amplification, down-conversion, and digitization, etc.) the signal received from the antenna and provides an input sample. In modem processor 305, encoder 306 receives service data and signaling messages to be transmitted on the uplink and processes (e.g., formatting, encoding, and interleaving) the service data and signaling messages. Modulator 307 further processes (e.g., symbol mapping and modulation) the encoded service data and signaling messages and provides an output sample. Demodulator 309 processes (e.g., demodulates) the input sample and provides a symbol estimate. Decoder 308 processes (e.g., deinterleaving and decoding) the symbol estimate and provides decoded data and signaling messages to device 300. Encoder 306, modulator 307, demodulator 309, and decoder 308 can be implemented by a combined modem processor 305. These units perform processing according to the radio access technology adopted by the radio access network (e.g., LTE and other evolved systems access technologies). It should be noted that when device 300 does not include modem processor 305, the above-mentioned functions of modem processor 305 can also be performed by processor 303.

[0105] The processor 303 controls and manages the operation of the device 300, and is used to execute the processing procedures performed by the device 300 in the above embodiments of this disclosure. For example, the processor 303 is also used to execute various steps of the transmitting or receiving device in the above method embodiments, and / or other steps of the technical solutions described in the embodiments of this disclosure.

[0106] Furthermore, the device 300 may also include a memory 304 for storing program code and data for the device 300.

[0107] Understandable, Figure 4 Only a simplified design of device 300 is shown. In practical applications, device 300 can include any number of transmitters, receivers, processors, modem processors, memory, etc., and all devices that can implement the embodiments of this disclosure are within the protection scope of the embodiments of this disclosure.

[0108] The present invention also provides a chip edge chipping defect detection device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the chip edge chipping defect detection system in the above embodiments.

[0109] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the chip edge chipping defect detection system.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0111] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A chip edge chipping defect detection system, characterized in that, Includes the following steps: A support device, used to support and fix the chip to be tested; A drive control device, connected to the carrier device and / or imaging device, is used to drive the chip under test to move relative to the imaging device; a light source device is used to provide illumination for the edge area of ​​the chip under test; and an imaging device is disposed on one side of the carrier device to acquire an edge image of the chip under test under the illumination of the light source device. A data processing device, communicatively connected to the imaging device and the drive control device, is used to receive and process the edge image. The data processing device includes: The system comprises: an acquisition module for acquiring an edge image of the chip to be inspected and extracting edge points from the edge image to obtain a sequence of edge points arranged along the edge direction of the chip to be inspected; a processing module for moving a window of a preset size along the edge point sequence, performing line fitting based on the edge points within the window to obtain a reference benchmark, and calculating a first feature value based on the deviation of the edge points within the window relative to the reference benchmark; an extraction module for acquiring pixel gradient information within the edge image region corresponding to the window and calculating a second feature value based on the distribution difference of the pixel gradient information; a fusion module for fusing the first feature value and the second feature value to obtain a defect risk score for the window; and a judgment module for determining whether a chipping defect exists at the location corresponding to the window based on the defect risk score and preset judgment conditions.

2. The chip chipping defect detection system according to claim 1, characterized in that, Also includes: Calculate the absolute value of the coordinate difference between two adjacent edge points within the window in the direction perpendicular to the edge; Edge points whose absolute value of coordinate difference exceeds a preset difference threshold are removed, and the remaining edge points in the window are taken as the valid edge point sequence. The method of performing line fitting based on edge points within the window includes: performing line fitting on the valid edge points in the valid edge point sequence.

3. The chip chipping defect detection system according to claim 2, characterized in that, The first feature value is specifically the normalized local residual ratio. The calculation of the first feature value based on the deviation of the edge points within the window relative to the reference benchmark includes: The reference benchmark is obtained by performing a linear fit on the effective edge point sequence using the least squares method. Calculate the vertical coordinate difference between each valid edge point and its corresponding position on the reference datum to obtain the residual; Calculate the residual standard deviation for all residuals within the window; Obtain the maximum and minimum values ​​of the vertical coordinates in the effective edge point sequence to calculate the range; The normalized local residual ratio is calculated based on the ratio between the residual standard deviation and the range.

4. The chip chipping defect detection system according to claim 3, characterized in that, When calculating the normalized local residual ratio based on the ratio of the residual standard deviation and the range, a physical precision limit factor is introduced, and the formula for calculating the normalized local residual ratio is as follows: ； In the formula, To normalize the local residual ratio, This represents the number of valid edge points in the valid edge point sequence. Let be the vertical coordinate of the i-th valid edge point in the valid edge point sequence. The vertical coordinates of the position corresponding to the i-th valid edge point on the reference datum. and These represent the maximum and minimum vertical coordinates of the valid edge points in the valid edge point sequence, respectively. This is the physical precision limit factor.

5. The chip chipping defect detection system according to claim 1, characterized in that, The second feature value is the gradient variation coefficient. The step of obtaining pixel gradient information within the edge image region corresponding to the window and calculating the second feature value based on the distribution difference of the pixel gradient information includes: The gradient operator is used to calculate the gradient magnitude of pixels in the edge image region in the direction perpendicular to the edge, forming a gradient magnitude sequence; Non-maximum suppression processing is applied to the gradient magnitude sequence to retain pixels with local maxima along the edge direction, thus forming an updated gradient magnitude sequence. Calculate the standard deviation and mean of the gradient magnitudes in the updated gradient magnitude sequence; The ratio of the standard deviation to the mean is used as the gradient coefficient of variation.

6. The chip chipping defect detection system according to claim 5, characterized in that, The process of fusing the first feature value and the second feature value to obtain the defect risk score of the window includes: The gradient variation coefficient is added to a preset constant to obtain the weight sum value; Multiply the first feature value by the weight sum value, and use the resulting product as the defect risk score.

7. The chip chipping defect detection system according to claim 6, characterized in that, This also includes revising the defect risk score: Calculate the variance of the vertical coordinates of all edge points within the window; If the variance is less than the first preset variance threshold, the defect risk score is multiplied by a first correction coefficient less than 1 for weight reduction. If the variance is greater than the second preset variance threshold, the defect risk score is multiplied by a second correction coefficient greater than 1 for weighting. The revised defect risk score will be used as the basis for the final determination of whether there is edge chipping defect at the location corresponding to the window.

8. The chip chipping defect detection system according to claim 1, characterized in that, include: A baseline risk score is obtained for multiple sample windows on the edge region of a defect-free chip. A preset threshold is set based on the mean and standard deviation of the baseline risk score as the preset judgment condition. If the defect risk score of the window is greater than the preset threshold, then the window is marked as a candidate defect location; According to the positional order along the edge direction of the chip to be tested, the consecutive adjacent candidate defect positions are merged into a target defect region, and the maximum defect risk score in the target defect region is used as the severity index of the target defect region.

9. The chip chipping defect detection system according to claim 1, characterized in that, It also includes a size calibration module: Obtain a first sample chip containing known defects and a second sample chip without defects; For multiple candidate sizes, calculate the first risk score set corresponding to the first sample chip and the second risk score set corresponding to the second sample chip respectively; Calculate the first mean and first standard deviation of the first risk score set, and the second mean and second standard deviation of the second risk score set; Calculate the difference between the first mean and the second mean, and divide the difference by the sum of the first standard deviation and the second standard deviation to obtain the separation evaluation index corresponding to each candidate size; The candidate size corresponding to the largest separation evaluation index is determined as the preset size.

Images