Glass substrate hole quality detection method and system based on image feature extraction

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a glass substrate hole quality detection method and system based on image feature extraction, which comprises the following steps: obtaining an ordered discrete pixel point set of a hole and constructing an adaptive window; calculating a discrete curvature eigenvalue reflecting the geometric change of the contour; and adaptively calculating a smoothing weight based on the eigenvalue, and constructing an ideal edge contour without defects through weighted average; and realizing the positioning of defects by calculating the normalized deviation of the actual contour point and the ideal contour point. Through the adaptive window and the normalized saliency measure, the present application avoids the problem that the traditional curvature algorithm is sensitive to image noise and cannot consider different scale defect detection, and improves the accuracy and robustness of detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a glass substrate via hole quality detection method and system based on image feature extraction. BACKGROUND

[0002] In precision electronic manufacturing, the via hole on the glass substrate as a key through structure, its profile integrity directly affects the product performance and yield. Therefore, using industrial cameras to conduct automatic quality detection on the via hole image is a core link to ensure production quality.

[0003] However, in the actual production environment, due to the interference of uneven imaging light, sensor noise and substrate texture and other factors, the edge profile of the via hole in the image often appears as a gray level mutation area with high frequency noise. This makes it a technical problem to distinguish between real profile defects and normal process noise.

[0004] To solve this problem, the prior art usually uses an algorithm based on edge profile curvature analysis. This method first extracts a discrete set of via hole profile points through an edge detection operator, then calculates the local curvature at each profile point, and identifies defects by abnormal changes in curvature. However, in practical applications, this mainstream algorithm has significant defects: first, the traditional curvature calculation relies on the second derivative, which is extremely sensitive to image noise and easily amplifies noise into false defect signals, resulting in a high false positive rate. Second, this type of algorithm usually calculates within a fixed-size analysis window, which cannot adapt to changes in different scale via holes or image resolution. When the window is too large, it will smooth out small defects, and when it is too small, it will be easily disturbed by local noise, with poor robustness. Finally, due to the inherent roughness of the via hole profile at different locations, using a globally uniform fixed threshold to determine defects makes it difficult to balance the high sensitivity of smooth areas and the stability of rough areas, and can easily cause missed detection or misjudgment. SUMMARY

[0005] To solve the technical problem of poor glass substrate via hole quality detection effect, the present application provides solutions in the following aspects.

[0006] In a first aspect, the present application provides a glass substrate via hole quality detection method based on image feature extraction, comprising:

[0007] An image of the glass substrate with plugged holes is acquired, and an ordered set of discrete pixels is extracted. An adaptive window is set along the set of discrete pixels, with each discrete pixel corresponding to an adaptive window. The size of the adaptive window is positively correlated with the number of discrete pixels. The discrete curvature feature value of each discrete pixel is calculated. The discrete curvature feature value is positively correlated with the sum of the absolute values ​​of the outer products of adjacent tangential vectors within the corresponding adaptive window, and inversely proportional to the square of the arc length of the discrete pixel within the adaptive window. An ideal edge contour is constructed, which is composed of ideal coordinates. The ideal coordinates of each discrete pixel are the weighted average of the coordinates of all discrete pixels within the corresponding adaptive window, and the weights are negatively correlated with the corresponding discrete curvature feature value. The defect significance of each discrete pixel is calculated. The defect significance is positively correlated with the distance between the discrete pixel coordinates and the corresponding ideal coordinates, and inversely proportional to the standard deviation of the distance between all discrete pixels within the corresponding adaptive window. Discrete pixels with a defect significance greater than a preset threshold are identified as defect points.

[0008] This invention constructs an adaptive window positively correlated with the total number of contour points, ensuring that the analysis scale dynamically matches the size of the plug or the image resolution. Simultaneously, it employs discrete curvature features based on the cross product of adjacent tangential vectors to suppress the interference of image noise on geometric feature extraction. This invention reconstructs a defect-free ideal contour as a dynamic benchmark through adaptive weighted smoothing, and finally calculates the defect significance of actual contour points deviating from this ideal contour. This transforms the defect measurement into a relative statistic independent of local background roughness, thereby enabling the identification of minute cracks in smooth areas and true edge breakage in rough areas, improving the accuracy and robustness of detection.

[0009] Preferably, the size of the adaptive window is the square root of the number of discrete pixels in the ordered discrete pixel set, rounded down.

[0010] Preferably, the discrete curvature characteristic value satisfies the expression:

[0011] ;

[0012] In the formula, Indicates the first Discrete curvature feature values ​​of discrete pixels; Indicates the first An adaptive window for discrete pixels; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the j-th discrete pixel to the (j+1)-th discrete pixel; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the (j+1)th discrete pixel to the (j+2)th discrete pixel; The symbol representing the magnitude of a vector.

[0013] This invention maps the curvature of the geometry directly to feature values ​​by accumulating the outer product of adjacent tangential vectors within a local window and normalizing it with the square of the total arc length. It is insensitive to pixel-level random noise and can more stably reflect the true geometric abrupt changes of the contour, thereby enhancing the high-frequency features of defects while suppressing the response of normal smooth contours.

[0014] Preferably, the construction of the ideal edge contour includes:

[0015] Calculate the image background noise baseline value, which is the product of the median absolute deviation of all discrete curvature feature values ​​and a preset distribution consistency factor;

[0016] Calculate the smoothing weight for each discrete pixel. The smoothing weight is inversely proportional to the corresponding discrete curvature feature value and directly proportional to the image background noise baseline value.

[0017] The smoothing weights are used as the weights for the weighted average to calculate the ideal coordinates of each discrete pixel.

[0018] This invention robustly estimates the background noise baseline of the current image by calculating the median absolute deviation of all curvature feature values, and calculates the smoothing weight based on this. It can dynamically adjust the filtering intensity to effectively remove real defect signals and maximize the preservation of normal contour details, avoiding missed detections due to over-smoothing or false alarms due to under-smoothing.

[0019] Preferably, the smoothing weights satisfy the expression:

[0020] ;

[0021] In the formula, Indicates the first Smoothing weights for discrete pixels; Indicates the first Discrete curvature feature values ​​of discrete pixels; Indicates the sensitivity coefficient; This represents the baseline value for image background noise. This represents the first minute value.

[0022] This invention employs a negative exponential function to calculate smoothing weights, achieving a non-linear weight mapping relationship. When the curvature feature value of a pixel significantly exceeds the background noise baseline, its weight rapidly decays to near zero, thus being effectively ignored during the reconstruction of the ideal contour. Conversely, the weight of normal contour points is close to 1 and is fully preserved. This more effectively suppresses the interference of defect points, ensuring that the reconstructed ideal contour more closely resembles the real, defect-free state.

[0023] Preferably, the ideal coordinates satisfy the expression:

[0024] ;

[0025] In the formula, Indicates the first The ideal coordinates of a discrete pixel; Indicates the first An adaptive window for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... Smoothing weights for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... The coordinates of a discrete pixel; This represents the second smallest value.

[0026] Preferably, the significance of the defect satisfies the expression:

[0027] ;

[0028] In the formula, Indicates the first The significance of defects in individual discrete pixels; This represents the coordinates of the i-th discrete pixel. Indicates the first The ideal coordinates of a discrete pixel; Indicates the first An adaptive window for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... The coordinates of a discrete pixel; Indicates the first Within the adaptive window of discrete pixels, the first... The ideal coordinates of a discrete pixel; Indicates the first The mean residual within an adaptive window of discrete pixels; Indicates the sensor noise constant; The symbol representing the magnitude of a vector.

[0029] This invention normalizes the Euclidean distance between a contour point and an ideal point by dividing it by the standard deviation of all distances within its local neighborhood. This transforms defect determination from relying on a globally fixed threshold into an adaptive local statistical test. This allows small deviations in smooth regions to achieve high significance, while larger deviations in rough regions may be considered normal, thus avoiding detection problems caused by inherent uneven roughness in different parts of the contour.

[0030] Preferably, the obtaining of the residual mean value in the adaptive window of the first discrete pixel point comprises:

[0031] The obtaining of the distance mean value of all discrete pixel points in the adaptive window of the first discrete pixel point comprises: The distance mean value of all discrete pixel points in the adaptive window of the first discrete pixel point is recorded as the residual mean value in the adaptive window of the first discrete pixel point.

[0032] Preferably, the obtaining of the ordered discrete pixel point set comprises:

[0033] The edge detection of the image is performed using a Canny operator to obtain discrete pixel points.

[0034] The sub-pixel positioning of the initial discrete pixel points is performed using a gray scale moment method to obtain the ordered discrete pixel point set.

[0035] In a second aspect, the present application provides a glass substrate hole quality detection system based on image feature extraction, comprising a processor and a memory, and the memory stores computer program instructions, when the computer program instructions are executed by the processor, the above-mentioned glass substrate hole quality detection method based on image feature extraction is realized.

[0036] By adopting the above technical scheme, the above-mentioned glass substrate hole quality detection method based on image feature extraction is generated into a computer program and stored in the memory to be loaded and executed by the processor, so that a terminal device is manufactured according to the memory and the processor, and the use is convenient.

[0037] The present application has the following advantages:

[0038] (1) The present application solves the adaptability problem of different scale targets by constructing an adaptive analysis window related to the number of contour points.

[0039] (2) The present application adopts a discrete curvature feature based on vector outer product, which overcomes the defect that traditional operators are sensitive to noise.

[0040] (3) The present application converts the absolute deviation into a relative statistical quantity independent of the local background roughness by constructing a dynamic ideal contour and calculating the normalized defect saliency, thereby realizing accurate identification of various defects under different texture backgrounds. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flow chart schematically showing the glass substrate hole quality detection method based on image feature extraction in the present application;

[0042] Figure 2 ​​is a schematic diagram showing a local comparison between an ordered discrete pixel set and an ideal edge profile. DETAILED DESCRIPTION

[0043] The embodiment of the present application discloses a glass substrate hole quality detection method based on image feature extraction, referring to Figure 1 , comprising steps S1-S4:

[0044] S1: based on an edge detection algorithm, an ordered discrete pixel set of a glass substrate hole image is obtained; based on the number of the ordered discrete pixel set, an adaptive window is constructed.

[0045] It should be noted that in the scene of industrial vision detection, the edge of the glass substrate hole is usually shown as a gray mutation area with high-frequency noise in the image, and if a single pixel gray threshold is relied on for segmentation, it is easy to be disturbed by noise and cause false edge positioning. In order to realize high-precision defect detection, the present application converts the edge pixel points into a vectorized discrete point set with topological order. In addition, considering the scale difference of the hole in different product images, a fixed size convolution kernel or filtering window cannot balance noise suppression under large scale and detail preservation under small scale. Therefore, the present application constructs an adaptive window to ensure that the subsequent feature extraction operator has scale invariance under different scales.

[0046] Specifically, based on an edge detection algorithm, an ordered discrete pixel set of a glass substrate hole image is obtained, comprising:

[0047] The glass substrate hole image is acquired by a camera, and edge pixel points are extracted from the glass substrate hole image using an edge detection algorithm and a gray moment sub-pixel positioning method to construct an ordered discrete pixel set , wherein is the coordinate of the i-th discrete pixel point, and the number of discrete pixel points is It should be noted that the edge detection algorithm uses a canny operator, and the edge detection algorithm and the gray moment sub-pixel positioning method are prior art and will not be described here.

[0048] It should be noted that the number of discrete pixels is positively correlated with the physical perimeter of the hole edge profile and the image resolution. When the image resolution is increased while the hole edge profile remains unchanged, the number of discrete pixels increases. When the resolution remains unchanged while the hole edge profile increases, the number of discrete pixels also increases, indicating that the number of discrete pixels is a global geometric feature quantity that can comprehensively reflect the target scale and image resolution. In order to make the subsequent local feature analysis not affected by this global change, the application constructs an adaptive window related to the number of discrete pixels. Further, by selecting a scale related to the number of discrete pixels as the window size based on the law of large numbers in image statistics, it can be ensured that the window always covers a statistically significant edge arc segment.

[0049] Preferably, based on the number of the ordered discrete pixel set, an adaptive window is constructed, comprising:

[0050] The adaptive window size satisfies the expression:

[0051] ;

[0052] In the formula, indicates the adaptive window size; indicates the number of discrete pixels; indicates the down rounding symbol.

[0053] At this point, the ordered discrete pixel set of the glass substrate hole image and the adaptive window are obtained.

[0054] S2: According to the ordered discrete pixel set and the adaptive window, the discrete curvature feature value of each discrete pixel is calculated.

[0055] It should be noted that in the glass substrate hole image, defects and normal processes have great differences in geometric shape. Defects are irregular shapes such as edge collapse and cracks, which show sharp and disordered changes in the direction of the profile tangent in a small neighborhood, while normal processes are regular shapes such as smooth fillets, which show smooth and unidirectional changes. The traditional curvature calculation operator based on the second derivative is extremely sensitive to image noise and cannot effectively distinguish between the two. Therefore, the application maps the degree of geometric shape mutation to a feature response value by calculating the rotation accumulation of the tangent vector in the local window, thereby enhancing the high-frequency defect features and suppressing the low-frequency shape features.

[0056] Specifically, according to the ordered discrete pixel set and the adaptive window, the discrete curvature feature value of each discrete pixel is calculated, comprising:

[0057] It should be noted that in a two-dimensional plane, the absolute value of the cross product of two vectors is proportional to the area of ​​the parallelogram formed by these two vectors, and also reflects the size of the included angle between them. Within a local window, continuously calculating and summing the cross products of adjacent tangential vector pairs is equivalent to integrating the curvature of the local contour. Simultaneously, to eliminate the influence of local arc length variations on the results, normalization using the square of the arc length is required. Based on this, this invention constructs discrete curvature characteristic response values.

[0058] The discrete curvature feature value of any discrete pixel satisfies the expression:

[0059] ;

[0060] In the formula, Indicates the first Discrete curvature feature values ​​of discrete pixels; Indicates the first An adaptive window for discrete pixels; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the j-th discrete pixel to the (j+1)-th discrete pixel; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the (j+1)th discrete pixel to the (j+2)th discrete pixel; The symbol representing the magnitude of a vector.

[0061] In the formula, This represents the cross product of two adjacent tangential vectors, and its value reflects the instantaneous change in the contour direction between two adjacent points. This indicates that the outer product of all adjacent tangential vectors within the adaptive window is accumulated. This value reflects the overall trend of change in the contour direction within the local window, i.e., the cumulative rotation amount. This represents the absolute value of the cumulative rotation; the larger the value, the more severe the curvature of the local contour. This represents the distance between two adjacent discrete pixels; Indicates the first The square of the total arc length within the adaptive window of each discrete pixel is used to normalize the cumulative rotation amount and eliminate the influence caused by the different spacing between pixels. This represents the cumulative rotation over the square of a unit arc length. The larger the value, the more drastic and disordered the change in the direction of the profile is under the same physical scale. Conversely, a smaller value indicates that the change in the profile is gentle or monotonous, thus achieving a high response to defect features.

[0062] At this point, the discrete curvature feature response values ​​of all pixels have been obtained.

[0063] S3: calculate the median absolute deviation of the discrete curvature feature response values, combine the distribution consistency factor to obtain an image background noise reference value; calculate the smoothing weight of any discrete pixel point according to the image background noise reference value; and use the smoothing weight to perform weighted average on the pixel coordinates in the local window to obtain the ideal coordinates of the arbitrary discrete pixel point, and construct an ideal edge contour.

[0064] It should be noted that after obtaining the discrete curvature feature response values, a filter operation is needed to reconstruct a defect-free ideal contour as a comparison reference for subsequent defect judgment. However, in actual production scenarios, the images of different batches of glass substrates will exhibit different background textures due to slight fluctuations in the processing technology, that is, the inherent roughness of the contour, which makes the fixed parameter filter prone to over-smoothing, resulting in missed detection, or under-smoothing, resulting in false positives. Therefore, the present application uses the median absolute deviation of the image data to estimate the background noise level of the current image, and based on this, constructs an adaptive weight nonlinear smoothing filter to automatically remove abnormal defect feature points and edge fitting.

[0065] Specifically, the median absolute deviation of the discrete curvature feature response values is calculated, combined with the distribution consistency factor to obtain an image background noise reference value, including:

[0066] It should be noted that the median absolute deviation is a robust statistic that is not sensitive to outliers, and can reflect the dispersion of the main body distribution of the data. Among all the discrete curvature feature response values, most data points correspond to normal contours, and their response values constitute the main body of the data; while a small number of defect points correspond to outliers. Therefore, the median absolute deviation can accurately depict the roughness of the normal contour.

[0067] The discrete curvature feature values of all pixel points are arranged in sequence according to the pixel point sequence number to form a discrete curvature feature value sequence.

[0068] A preset distribution consistency factor is denoted as , which is used to convert the robust statistic into a standard statistic. Exemplarily, is set to 1.5. If is too large, for example, 2, it will lead to an overestimated image background noise reference value, and the system will misjudge the real small defects as normal texture noise, resulting in missed detection; if is too small, for example, 1, it will lead to an underestimated image background noise reference value, and the system will be too sensitive to normal process textures, resulting in false positives.

[0069] The image background noise reference value satisfies the expression:

[0070] ;

[0071] In the formula, denotes an image background noise reference value; denotes a median filter function; denotes a sequence of discrete curvature feature values; denotes a distribution consistency factor; denotes an absolute value function.

[0072] wherein, denotes an absolute value of a difference between each discrete curvature feature value in the sequence of discrete curvature feature values and a median value in the sequence, reflecting a deviation degree of each point relative to an overall average level; denotes a median value of the deviation degree, that is, a median absolute deviation, which can robustly represent a main fluctuation range of the data sequence, that is, an inherent roughness of a normal profile, without being affected by extreme abnormal values; denotes a conversion of the median absolute deviation into an equivalent standard deviation scale by the distribution consistency factor, so that the image background noise reference value can be used as a universal reference threshold to accurately evaluate a background noise level of a current image.

[0073] It should be noted that, in order to effectively eliminate defects and retain normal details during filtering, a smoothing weight needs to be set, so that a pixel point with a feature response value higher than the background noise obtains a smaller smoothing weight, and a pixel point close to the background noise obtains a larger smoothing weight.

[0074] Preferably, the smoothing weight of an arbitrary discrete pixel point is calculated according to the image background noise reference value, including:

[0075] A sensitivity coefficient is set for controlling an attenuation rate of the smoothing weight to an abnormal response value. Exemplarily, the sensitivity coefficient is set to If the sensitivity coefficient takes a value too large, such as 5, the smoothing weight will be too flat, and a defect point with a high response value can still obtain a large weight, causing the reconstructed profile to collapse to the defect position and the filtering to fail. If the sensitivity coefficient takes a value too small, such as 1, the smoothing weight will be relatively steep, so that normal process texture fluctuations will also be assigned a weight close to zero, causing the reconstruction process to be unstable.

[0076] The smoothing weight of an arbitrary discrete pixel point satisfies an expression:

[0077] ;

[0078] wherein, denotes a smoothing weight of an i-th discrete pixel point; denotes a smoothing weight of an i-th discrete pixel point; denotes a discrete curvature feature value of an i-th discrete pixel point; denotes a discrete curvature feature value of an i-th discrete pixel point; denotes a sensitivity coefficient; denotes an image background noise reference value; This represents the first minute value, used to prevent the denominator from being zero. For example, ; This represents the natural exponential function.

[0079] In the formula, The signal-to-noise ratio (SNR) represents the ratio of the discrete curvature eigenvalue to the background noise baseline. The larger the value, the greater the abrupt change at the current point compared to the background noise, and the higher the probability that it is an outlier. This indicates a nonlinear mapping relationship constructed using a negative exponential function. When the signal-to-noise ratio is high, the smoothing weight approaches 0, thus suppressing defect points; when the signal-to-noise ratio is low, the smoothing weight approaches 1, thus preserving normal contour points.

[0080] It should be noted that the reconstruction of the ideal contour is essentially a process of center-of-gravity correction. Each point within the local window is treated as a point mass, and its mass is determined by adaptive weights. For defective points, since the weights approach zero, their pulling effect on the center of gravity is shielded; for normal points, the original weights are maintained. By calculating the weighted average coordinates, the center point can be automatically shifted to an ideal position that conforms to the normal process trend.

[0081] Preferably, a weighted average of pixel coordinates within a local window is performed using smoothing weights to obtain the ideal coordinates of any discrete pixel point, thereby constructing an ideal edge contour, including:

[0082] The ideal coordinates of any discrete pixel point satisfy the expression:

[0083] ;

[0084] In the formula, Indicates the first The ideal coordinates of a discrete pixel; Indicates the first An adaptive window for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... Smoothing weights for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... The coordinates of a discrete pixel; This represents the second smallest value, used to prevent the denominator from being zero. For example, .

[0085] In the formula, Indicates the first The weighted coordinates of the i-th discrete pixel within an adaptive window of discrete pixels; the larger the weight, the greater the contribution of that point to the centroid position. Indicates the first The sum of the weighted coordinates of all discrete pixel points in the adaptive window of the discrete pixel point reflects the overall position tendency under the action of the smoothing weight; The sum of the smoothing weights of all discrete pixel points in the adaptive window of the discrete pixel point; The sum of the smoothing weights of all discrete pixel points in the adaptive window of the discrete pixel point; The weighted barycenter coordinate is such that the pulling action of the defect point on the barycenter is reduced, and the normal point mainly affects the barycenter position, so that the calculated ideal coordinate corrects the defect deviation in the original coordinate and fits the normal process profile.

[0086] The ideal edge profile is obtained by reordering all the ideal coordinates of the discrete pixel points.

[0087] Thus, the ideal edge profile is obtained.

[0088] S4: Based on the positional relationship between the ordered discrete pixel point set and the ideal edge profile, the defect saliency of each discrete pixel point is calculated; and based on the difference between the defect saliency and the preset segmentation threshold, a defect detection result is obtained.

[0089] It should be noted that after obtaining the original profile and the ideal profile, although the distance between the two can be calculated to preliminarily evaluate the deviation, in the complex scene of industrial images, the inherent roughness of the profile at different positions is not uniform. This physical feature causes the normal residual error in the rough area to be greater in value than the defect residual error in the smooth area. Therefore, instead of directly using the absolute value of the residual error for discrimination, the application constructs a locally adaptive saliency measurement index, and discriminates the defects by calculating the ratio of the residual error to the local background noise variance.

[0090] Specifically, based on the positional relationship between the ordered discrete pixel point set and the ideal edge profile, the defect saliency of each discrete pixel point is calculated, including:

[0091] It should be noted that the essence of local anisotropy is to compare the value of a data point with the mean and standard deviation of the data set in which it is located. In the application, this data set is the residual distance of all pixel points in the local window. By calculating the comparison of the residual error of each pixel point with the residual error distribution in its neighborhood, the deviation degree saliency of the corresponding pixel point can be judged.

[0092] The defect saliency of any discrete pixel point satisfies the expression:

[0093] ;

[0094] ;

[0095] In the formula, The sum of the smoothing weights of all discrete pixel points in the adaptive window of the discrete pixel point; the defect conspicuity of the i-th discrete pixel point; the coordinate of the i-th discrete pixel point; the ideal coordinate of the i-th discrete pixel point; the adaptive window of the i-th discrete pixel point; the coordinate of the i-th discrete pixel point within the adaptive window of the i-th discrete pixel point; the ideal coordinate of the i-th discrete pixel point within the adaptive window of the i-th discrete pixel point; the mean of the residual within the adaptive window of the i-th discrete pixel point; the coordinate of the i-th discrete pixel point within the adaptive window of the i-th discrete pixel point; the ideal coordinate of the i-th discrete pixel point within the adaptive window of the i-th discrete pixel point; the mean of the residual within the adaptive window of the i-th discrete pixel point; the sensor noise constant; the modulus of the vector. It should be noted that, is used to maintain numerical stability in the locally smooth area. For example, the sensor noise constant is 0.1 pixel. It should be noted that if the sensor noise constant is too large, for example, 1 pixel, it will mask the local small changes, resulting in a decrease in the sensitivity to micro-cracks in the smooth area; if the sensor noise constant is too small, it will cause the denominator to tend to zero, so that the inherent thermal noise of the imaging sensor is amplified infinitely, resulting in false high defect conspicuity values. In the formula,

[0096] the Euclidean distance between the coordinate and the ideal coordinate of the i-th discrete pixel point, which directly reflects the degree of deviation of the i-th discrete pixel point from the ideal profile; the difference between the residual of each discrete pixel point within the adaptive window of the i-th discrete pixel point and the mean of the residuals, which reflects the fluctuation of the local residual; the variance of the local window residual, which reflects the dispersion degree of the local background noise;

[0097] the local normalization of the residual by dividing the original residual by the standard deviation of the local background noise, the larger the value, the more significant the deviation of the current point under the consideration of the dispersion degree, thereby realizing the adaptive discrimination of defects in different texture areas. the defect conspicuity of the i-th discrete pixel point; the defect conspicuity of the i-th discrete pixel point; the defect conspicuity of the i-th discrete pixel point; the defect conspicuity of the i-th discrete pixel point; the defect conspicuity of the i-th discrete pixel point; the defect conspicuity of the i-th discrete pixel point;

[0098] Preferably, the defect conspicuity of each pixel point is compared with a preset segmentation threshold, if the defect conspicuity is greater than or equal to the preset segmentation threshold, the pixel point is recorded as a defect point; if the defect conspicuity is less than the preset segmentation threshold, the pixel point is recorded as a normal process point. For example, the preset segmentation threshold is set to 3.​

[0099] It should be noted that, as Figure 2 Fig. 6 is a schematic diagram of a local comparison between an ordered discrete pixel set and an ideal edge profile, showing the comparison effect between part of the discrete pixel set and the ideal edge profile, and the detected defect points.

[0100] Thus far, the glass substrate hole quality detection based on image feature extraction is completed.

[0101] The embodiment of the present application also discloses a glass substrate hole quality detection system based on image feature extraction, comprising a processor and a memory, and the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the glass substrate hole quality detection method based on image feature extraction according to the present application is realized.

[0102] The above system also comprises other components such as a communication bus and a communication interface which are well known to those skilled in the art, and the settings and functions thereof are known in the art, thus will not be described here.

[0103] Although the present application has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made thereto without departing from the spirit and scope of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.

Claims

1. A method for detecting the quality of via plugging in glass substrates based on image feature extraction, characterized in that, include: Obtain images of the glass substrate with plugged holes and extract an ordered set of discrete pixels; An adaptive window is set along the set of discrete pixels, with each discrete pixel corresponding to an adaptive window; the size of the adaptive window is positively correlated with the number of discrete pixels. Calculate the discrete curvature feature value of each discrete pixel. The discrete curvature feature value is positively correlated with the sum of the absolute values ​​of the outer products of adjacent tangential vectors within the corresponding adaptive window, and inversely proportional to the square of the arc length of the discrete pixel within the adaptive window. Constructing an ideal edge contour includes: calculating an image background noise baseline value, wherein the image background noise baseline value is the product of the median absolute deviation of all discrete curvature feature values ​​and a preset distribution consistency factor; calculating a smoothing weight for each discrete pixel, wherein the smoothing weight is inversely proportional to the corresponding discrete curvature feature value and directly proportional to the image background noise baseline value; using the smoothing weight as the weight of a weighted average to calculate the ideal coordinates of each discrete pixel, wherein the ideal edge contour is composed of the ideal coordinates; the ideal coordinates of each discrete pixel are the weighted average of the coordinates of all discrete pixels within the corresponding adaptive window, and the weights are negatively correlated with the corresponding discrete curvature feature values; The defect significance of each discrete pixel is calculated. The defect significance is positively correlated with the distance between the discrete pixel coordinates and the corresponding ideal coordinates, and inversely proportional to the standard deviation of the distance between all discrete pixels in the corresponding adaptive window. Discrete pixels with a defect significance greater than a preset threshold are identified as defect points.

2. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, The size of the adaptive window is the square root of the number of discrete pixels in the ordered discrete pixel set, rounded down.

3. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, The discrete curvature eigenvalues ​​satisfy the expression: ； In the formula, Indicates the first Discrete curvature feature values ​​of discrete pixels; Indicates the first An adaptive window for discrete pixels; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the j-th discrete pixel to the (j+1)-th discrete pixel; Indicates the first Within an adaptive window of discrete pixels, the tangent vector pointing from the (j+1)th discrete pixel to the (j+2)th discrete pixel; The symbol representing the magnitude of a vector.

4. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, The smoothing weights satisfy the expression: ； In the formula, Indicates the first Smoothing weights for discrete pixels; Indicates the first Discrete curvature feature values ​​of discrete pixels; Indicates the sensitivity coefficient; This represents the baseline value for image background noise. This represents the first minute value.

5. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, Ideal coordinates satisfy the expression: ； In the formula, Indicates the first The ideal coordinates of a discrete pixel; Indicates the first An adaptive window for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... Smoothing weights for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... The coordinates of a discrete pixel; This represents the second smallest value.

6. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, The significance of the defect satisfies the expression: ； In the formula, Indicates the first The significance of defects in individual discrete pixels; This represents the coordinates of the i-th discrete pixel. Indicates the first The ideal coordinates of a discrete pixel; Indicates the first An adaptive window for discrete pixels; Indicates the first Within the adaptive window of discrete pixels, the first... The coordinates of a discrete pixel; Indicates the first Within the adaptive window of discrete pixels, the first... The ideal coordinates of a discrete pixel; Indicates the first The mean residual within an adaptive window of discrete pixels; Indicates the sensor noise constant; The symbol representing the magnitude of a vector.

7. The method for detecting the quality of via plugging in a glass substrate based on image feature extraction according to claim 6, characterized in that, The first The acquisition of the mean residual within an adaptive window for discrete pixels includes: Get the The average distance between the coordinates of all discrete pixels and their ideal coordinates within an adaptive window of n discrete pixels is denoted as the nth. The mean residual within an adaptive window of discrete pixels.

8. The method for detecting the quality of hole plugging in a glass substrate based on image feature extraction according to claim 1, characterized in that, The acquisition of the ordered discrete pixel set includes: Edge detection is performed on the image using the Canny operator to obtain discrete pixels; Sub-pixel localization is performed on the initial discrete pixels using the gray-scale moment method to obtain the ordered discrete pixel set.

9. A glass substrate hole plugging quality inspection system based on image feature extraction, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the glass substrate hole-filling quality inspection method based on image feature extraction according to any one of claims 1-8.

Images