A glass-based circuit board defect detection method and system

Abstract

The application discloses a kind of glass-based circuit board defect detection method and system, it is related to glass substrate defect detection technical field, including using the tomogram of target area of optical coherence tomography system with scanning resolution fixed as 320×320 is collected, and it is three-dimensional reconstruction, obtains three-dimensional body data model.Extract defect depth, transverse distribution profile and defect internal refractive index inhomogeneity etc. from model Multiple three-dimensional morphological characteristics.Classifier based on three-dimensional convolutional neural network is identified by inputting characteristics, and spatial position verification is carried out in combination with known design layout, and finally the detection report is generated.The method realizes the standardization of detection process by fixing the scanning resolution, improves the efficiency and result consistency;By quantifying the internal refractive index inhomogeneity of the defect, the material internal optical property variation type defect is effectively identified, and the detection depth and accuracy are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of glass substrate defect detection technology, specifically a method and system for detecting defects in glass-based circuit boards. Background Technology

[0002] In the field of glass-based circuit board manufacturing, optical coherence scanning microscopy (OCSMS) has been used for non-destructive testing of their internal structures. Conventional methods typically set an adjustable or non-specific resolution parameter during scanning imaging, based on equipment capabilities or general standards. This non-standardized parameter selection leads to differences in detail, data volume, and acquisition time among different batches and equipment, affecting the stability and comparability of production line testing and making it difficult to establish a stable and optimal balance between testing efficiency and information sufficiency.

[0003] From a detection and analysis perspective, existing methods primarily rely on the geometric features of defects extracted from images, such as length, width, depth, and volume. While these features effectively describe physical anomalies, they cannot reveal changes in the optical properties of the material within the defect area. For glass substrates, their dielectric properties and signal transmission quality are highly dependent on the material's optical homogeneity. Changes in refractive index distribution caused by internal stress, impurities, or microstructural variations are key factors affecting the product's electrical performance and long-term reliability. Due to the lack of quantitative indicators for the intrinsic properties of materials, conventional techniques struggle to effectively identify and assess these non-geometric, yet significantly impactful, material defects, resulting in detection blind spots. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a defect detection method for glass-based circuit boards, comprising:

[0006] The tomographic image data of the target area of ​​the glass substrate circuit board under test was acquired using an optical coherence scanning microscopy system. The scanning resolution of the optical coherence scanning microscopy system during scanning imaging was set to 320×320.

[0007] The tomographic image data is subjected to three-dimensional reconstruction processing to generate a three-dimensional volume data model that reflects the internal structural features of the target region;

[0008] Multiple three-dimensional morphological feature parameters related to line defects are extracted from the three-dimensional volume data model. These three-dimensional morphological feature parameters include defect depth, defect lateral distribution profile, and defect internal refractive index non-uniformity.

[0009] The multiple three-dimensional morphological feature parameters are input into a defect classifier built on a three-dimensional convolutional neural network to identify the defect type and obtain preliminary defect identification results.

[0010] Based on the known design layout data of the glass-based circuit board under test, the spatial location of the preliminary defect identification results is verified, and a final defect detection report is generated.

[0011] Further, the step of performing three-dimensional reconstruction processing on the tomographic image data to generate a three-dimensional volumetric data model reflecting the internal structural features of the target region includes:

[0012] Each two-dimensional tomographic image in the tomographic image data undergoes denoising and edge enhancement preprocessing;

[0013] The image registration algorithm is used to precisely align all the preprocessed two-dimensional tomographic images in three-dimensional space.

[0014] A three-dimensional interpolation algorithm is used to spatially interpolate the precisely aligned two-dimensional tomographic image data to fill the three-dimensional data gaps formed by the scanning interval;

[0015] Based on the interpolated complete 3D dataset, a 3D volumetric data model capable of characterizing the distribution of optical scattering intensity inside the material is constructed using voxels as the basic unit.

[0016] Furthermore, the extraction of multiple three-dimensional morphological feature parameters related to the line defect from the three-dimensional volumetric data model, including defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity of the defect, includes:

[0017] In the three-dimensional volume data model, the optical scattering intensity profile of each pixel column is analyzed along the depth direction, the boundary depth of the defect is determined based on the abrupt change points of the profile, and the depth of the defect is calculated.

[0018] In the three-dimensional volume data model, the defect region is sliced ​​along the horizontal plane, the binary contour of the defect region in each slice is extracted, and the contours of all slices are superimposed to generate the horizontal distribution contour of the defect.

[0019] The optical scattering intensity value of each voxel within the defect region in the three-dimensional volumetric data model is calculated relative to the dispersion of its neighboring voxels, thereby quantifying the refractive index inhomogeneity within the defect.

[0020] Further, the step of inputting the multiple three-dimensional morphological feature parameters into a defect classifier constructed based on a three-dimensional convolutional neural network for defect type identification to obtain preliminary defect identification results includes:

[0021] The defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity of the defect are converted into a three-dimensional feature tensor.

[0022] The three-dimensional feature tensor is input into the input layer of the defect classifier constructed based on the three-dimensional convolutional neural network. The defect classifier includes multiple cascaded three-dimensional convolutional layers, three-dimensional pooling layers, and fully connected layers.

[0023] The high-level abstract features of the three-dimensional feature tensor are extracted step by step through the three-dimensional convolutional layer and the three-dimensional pooling layer.

[0024] The extracted high-level abstract features are mapped to probability scores corresponding to different defect types through the fully connected layer, and the defect type with the highest probability score is output as the preliminary defect identification result.

[0025] Furthermore, the process of combining the known design layout data of the glass-based circuit board under test to perform spatial location verification on the preliminary defect identification results and generate a final defect detection report includes:

[0026] The detected defects are mapped to their three-dimensional coordinates in the three-dimensional volume data model and then to the known design layout coordinate system of the glass-based circuit board under test.

[0027] In the known design layout data, identify the structural elements corresponding to the mapped coordinate points, the structural elements including conductor lines, insulating dielectric layers and vias;

[0028] Based on the expected function of the structural elements, determine whether the defect type of the preliminary defect identification result is reasonable in the spatial location; when it is determined to be reasonable, confirm that the preliminary defect identification result is valid, and integrate it with the corresponding three-dimensional coordinates and structural element information into the final defect detection report.

[0029] Furthermore, after acquiring tomographic image data of the target area of ​​the glass-based circuit board under test using an optical coherence scanning microscopy system, the method further includes:

[0030] The quality of the acquired tomographic image data is evaluated, and the signal-to-noise ratio and contrast of each tomographic image are calculated.

[0031] Tomographic images with a signal-to-noise ratio and contrast below a preset threshold are identified as invalid images and are discarded.

[0032] Record the scanning position coordinates corresponding to the invalid images that have been rejected, and control the optical coherence scanning microscope system to reacquire the tomographic image data of the area corresponding to the scanning position coordinates at a scanning resolution of 320×320.

[0033] The newly acquired tomographic image data is added to the original dataset to form the complete tomographic image data.

[0034] Further, the calculation of the signal-to-noise ratio and contrast of each tomographic image includes:

[0035] Select the region of interest containing the circuit structure and the background region on each tomographic image;

[0036] The average value of the pixel grayscale values ​​in the region of interest is calculated as the signal strength, and the standard deviation of the pixel grayscale values ​​in the background region is calculated as the background noise.

[0037] The signal-to-noise ratio of the tomographic image is obtained by dividing the signal intensity by the background noise.

[0038] The difference between the maximum and minimum gray values ​​of pixels within the region of interest is calculated and used as the contrast of the tomographic image.

[0039] Furthermore, the optical coherent scanning microscopy system uses a broadband light source with a center wavelength in the visible light band. The acquisition process of the tomographic image data follows the steps of scanning layer by layer along the depth direction, starting from the surface of the glass substrate circuit board to be tested. After scanning a horizontal 320×320 pixel area, a set of tomographic image data of the pixel area is obtained.

[0040] Further, the step of converting the defect depth, the lateral distribution profile of the defect, and the internal refractive index inhomogeneity of the defect into a three-dimensional feature tensor includes:

[0041] Based on the defect depth, a preset number of intervals are divided at equal intervals along the depth direction, and a one-dimensional depth feature vector is generated by quantizing and encoding the interval into which the actual depth value of the defect falls in the three-dimensional volume data model.

[0042] Based on the lateral distribution contour of the defects, a two-dimensional contour feature matrix is ​​generated on the lateral coordinate grid of the three-dimensional volume data model by assigning binary values ​​according to whether each coordinate point is identified as a defect area.

[0043] Based on the internal refractive index inhomogeneity of the defect, the optical scattering intensity values ​​of each voxel in the defect region of the three-dimensional volume data model are normalized, and a three-dimensional inhomogeneity feature cube is directly formed by the three-dimensional spatial arrangement of the voxels.

[0044] The one-dimensional depth feature vector and the two-dimensional contour feature matrix are multiplied by an outer product in the horizontal plane to obtain a three-dimensional basic feature tensor.

[0045] The three-dimensional basic feature tensor and the three-dimensional non-uniformity feature cube are concatenated along the channel dimension to form the final three-dimensional feature tensor for input to the defect classifier.

[0046] Furthermore, the present invention also includes a glass-based circuit board defect detection system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the glass-based circuit board defect detection method described above.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The scanning resolution of the optical coherent scanning system is fixed at 320×320. This value is a specifically optimized nominal value, achieving a definite balance between scanning rate, single-frame image information, and 3D data volume. This setting parameterizes and standardizes the data acquisition process, eliminating imaging performance differences caused by resolution variations. While ensuring sufficient capture of the basic morphological information of the target defect, this fixed resolution allows the system's imaging efficiency to reach a stable and predictable state, improving the consistency of the inspection process and overall throughput.

[0049] The "internal refractive index inhomogeneity of defects" is extracted as the core three-dimensional feature parameter. This parameter is obtained by calculating the statistical dispersion of the estimated refractive index values ​​of each pixel within the three-dimensional region of the defect; its physical essence is to quantify the spatial distribution uniformity of the material's optical properties. This technique extends defect analysis from external geometric morphology to the level of internal material properties. For regions where geometric deformation is not significant but the material's optical homogeneity has changed, this feature parameter can produce a significant anomalous response, thus effectively detecting internal material variations that traditional geometric feature methods cannot respond to, achieving a deeper characterization and identification of the nature of defects. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the steps of the glass-based circuit board defect detection method described in this invention.

[0051] Figure 2 A flowchart for 3D reconstruction processing;

[0052] Figure 3 A flowchart for extracting three-dimensional morphological feature parameters;

[0053] Figure 4 A visualization of the coordinate mapping relationship in the defect detection process of glass-based circuit boards;

[0054] Figure 5 This is a comparison of the reflection intensity of glass-based circuit boards as a function of depth under an optical coherence scanning microscope system. Detailed Implementation

[0055] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] See Figure 1 This glass-based circuit board defect detection method utilizes an optical coherence scanning microscopy system to acquire tomographic image data of the target area of ​​the glass-based circuit board under test. The scanning resolution of the optical coherence scanning microscopy system during scanning imaging is set to 320×320. The acquired tomographic image data is processed to generate a three-dimensional volumetric data model reflecting the internal structural features of the target area. Multiple three-dimensional morphological feature parameters related to circuit defects are extracted from the three-dimensional volumetric data model, including defect depth, defect lateral distribution contour, and internal refractive index inhomogeneity of the defect. These multiple three-dimensional morphological feature parameters are input into a defect classifier based on a three-dimensional convolutional neural network to identify the defect type and obtain preliminary defect identification results. The preliminary defect identification results are then spatially verified using the known design layout data of the glass-based circuit board under test to generate a final defect detection report.

[0057] See Figure 2 In one embodiment of the present invention, an optical coherence scanning microscopy system acquires a series of two-dimensional tomographic image data of the target area of ​​a glass-based circuit board under test at a scanning resolution of 320×320. These two-dimensional tomographic image data constitute the basic raw dataset for three-dimensional reconstruction. Each two-dimensional tomographic image in the tomographic image data undergoes denoising and edge enhancement preprocessing. The denoising operation aims to suppress randomly distributed noise points in the image, while the edge enhancement operation aims to enhance the signal transition characteristics at the interfaces of different materials in the image. The preprocessed two-dimensional tomographic images retain key structural information while exhibiting superior image quality.

[0058] In some embodiments, an image registration algorithm is used to precisely align all preprocessed two-dimensional tomographic images in three-dimensional space. Since the optical coherence scanning microscope system may introduce slight positional shifts during mechanical scanning, the image registration algorithm calculates the translation and rotation parameters of feature regions between adjacent tomographic images to uniformly correct the coordinates of all images in three-dimensional space, ensuring that each two-dimensional tomographic image is located at its theoretically designed three-dimensional spatial position. A three-dimensional interpolation algorithm is then used to spatially interpolate the precisely aligned two-dimensional tomographic image data. The data points in the two-dimensional tomographic image data are discretely sampled along the depth scanning direction. The three-dimensional interpolation algorithm calculates the data values ​​at the missing positions in the scanning interval based on the image data at known depth positions, thereby filling the three-dimensional data gaps formed by the scanning interval and generating a complete three-dimensional dataset with continuous data points in each upward direction in three-dimensional space.

[0059] It is understandable that, based on the interpolated complete 3D dataset, a 3D volumetric data model capable of characterizing the distribution of optical scattering intensity within a material is constructed using voxels as basic units. Each voxel corresponds to a tiny cubic unit in 3D space, and the stored value represents the intensity of backscattered light from the probe at that spatial location. The 3D volumetric data model completely records the optical properties of every point within the target area, providing a data foundation for subsequent feature extraction. In an example scenario, a region of a glass substrate with internal microcracks is scanned, and original tomographic image data of 320×320×200 (depth direction) is obtained. After the aforementioned 3D reconstruction processing, a 3D volumetric data model containing 320×320×200 voxels is obtained. This model clearly distinguishes the glass matrix, internal cracks, and scattering enhancement regions at the crack edges. Compared to the unprocessed original image stack, the 3D volumetric data model provides a continuous and clear view of the internal structure on any depth section and any direction profile.

[0060] Optionally, in the preprocessing steps, denoising and edge enhancement operations can be performed sequentially or integrated into a single filtering operator. One implementation involves first applying a low-pass filter to smooth the 2D tomographic image to suppress high-frequency noise, and then applying a high-pass filter or gradient operator to highlight edge details. Image registration algorithms can employ registration methods based on image intensity mutual information or feature point matching. 3D interpolation algorithms can be 3D linear interpolation, 3D spline interpolation, or other spatial interpolation methods; the choice of interpolation algorithm affects the smoothness and accuracy of the final 3D volumetric data model. The construction of the 3D volumetric data model involves arranging and storing the 3D dataset according to a regular 3D grid, where the value of each grid node corresponds to the optical scattering intensity value of the voxel.

[0061] In practical implementation, after constructing a three-dimensional volumetric data model, its data format can be a three-dimensional matrix, which is convenient for subsequent image processing algorithms to directly access and manipulate. The three-dimensional volumetric data model not only contains the geometric structure information of the material, but also contains the optical property information of the material. For example, defect areas usually exhibit different optical scattering intensity distribution patterns than normal areas. By comparing the display effect of the same microcrack in the same area before and after three-dimensional reconstruction processing, the unreconstructed original image stack cannot display the structure at non-scanning depth positions, while the three-dimensional volumetric data model can display the continuous structure at any depth position through resampling, so that the morphology, orientation and spatial position of the defect can be presented completely and continuously. This directly supports the accurate extraction of three-dimensional morphological feature parameters from the model.

[0062] See Figure 3In one embodiment of the present invention, the three-dimensional volumetric data model is a regular three-dimensional mesh with voxels as basic units. Each voxel stores the optical scattering intensity value corresponding to its spatial location. The extraction process is performed on regions in the model that have been initially marked as candidate defects. In some embodiments, the optical scattering intensity profile of each pixel column is analyzed along the depth direction in the three-dimensional volumetric data model. The optical scattering intensity profile refers to a curve formed by the optical scattering intensity values ​​of a series of voxels along a direction perpendicular to the surface of the glass substrate (i.e., the depth direction). The boundary depth of the defect is determined based on the abrupt change points in the profile. These abrupt change points are usually manifested as sharp increases or decreases in the optical scattering intensity value, corresponding to the internal interface of the material or the defect boundary. These abrupt change points can be located by detecting the zero-crossing point of the first derivative of the profile curve or by setting an intensity threshold. Calculating the defect depth requires determining the top boundary depth and the bottom boundary depth of the defect. The defect depth value is the coordinate difference between the bottom boundary depth and the top boundary depth in the depth direction.

[0063] In practice, the defect region is sliced ​​along a horizontal plane in the 3D volumetric data model. This horizontal plane is parallel to the surface of the glass substrate. The slicing operation involves acquiring 2D image data of this plane at different depth positions. The binary contour of the defect region in each slice is extracted. An image segmentation algorithm (such as thresholding) is applied to each horizontal slice image to separate pixels belonging to the defect region from the background, resulting in a binary image where pixel values ​​in the defect region are 1 and pixel values ​​in the background region are 0. The outer boundary of the connected components of this binary image represents the defect contour at that depth. The contours of all slices are superimposed to generate the lateral distribution contour of the defect. This superposition operation projects the binary contours from slices at different depths onto the same 2D horizontal reference plane, forming a 2D contour map that integrates the projections of the defect at all depth levels. This contour map represents the overall distribution range and shape of the defect in the lateral direction.

[0064] It can be understood that calculating the dispersion of the optical scattering intensity value of each voxel within a defect region in a three-dimensional volumetric data model relative to its neighboring voxels quantifies the refractive index inhomogeneity within the defect. For each voxel within the defect region, the optical scattering intensity distribution within a small three-dimensional neighborhood is examined; the dispersion characterizes the variation in the difference between the voxel value and the values ​​of other voxels in the neighborhood. One formula for quantifying the dispersion is:

[0065]

[0066] in: This represents the local inhomogeneity value at the current voxel v. This represents a three-dimensional neighborhood window centered on voxel v. Representing the neighborhood The total number of voxels contained therein. Representing the neighborhood The optical scattering intensity value of the j-th voxel. Representing the neighborhood The arithmetic mean of the optical scattering intensity values ​​of all voxels within the defect region. Calculate the optical scattering intensity of all voxels within the defect region. After setting the value, the internal refractive index inhomogeneity of the defect can be defined as these. The statistical characteristics of the value, such as its mean, maximum value, or standard deviation.

[0067] In an example scenario, an analysis is performed on a defect region identified as a bubble in a 3D volumetric data model. The optical scattering intensity profile of the pixel column passing through the bubble's center is analyzed along the depth direction. Within the profile, a point where the intensity drops sharply from high to the background value (upper boundary of the bubble) and a point where the intensity rises sharply from the background value (lower boundary of the bubble) are observed. The depth coordinate difference between these two points is 85 micrometers, representing the defect depth. Within the depth range containing this bubble, a horizontal slice is taken every 5 micrometers, and a binarized contour is extracted. The resulting horizontal defect distribution contour, obtained by superimposing all contours, appears as a near-circular shape with an equivalent diameter of approximately 12 micrometers. The local inhomogeneity of all voxels within this bubble region is calculated. Its average value is 15.3 (relative intensity unit), while a defect-free area of ​​glass... The average value is only 2.1, and the high average value quantifies the non-uniformity of the refractive index inside the bubble.

[0068] Optionally, the defect boundary depth can be determined using an adaptive thresholding method, where the threshold is dynamically calculated based on the background noise level of the optical scattering intensity profile for each pixel column. The superposition of the defect's lateral distribution contours does not have to be a simple Boolean OR operation; depth-related weights can be assigned to the contours of slices at different depths before fusion. Quantifying the refractive index inhomogeneity within the defect can also employ other mathematical descriptors, such as calculating the entropy of the voxel optical scattering intensity values ​​within the defect region, or calculating the contrast characteristics of its gray-level co-occurrence matrix.

[0069] In practice, the three three-dimensional morphological feature parameters extracted through the above steps describe the attributes of the defect from different dimensions: defect depth describes the vertical extension scale of the defect, defect lateral distribution contour describes the geometric shape and size of the defect on the plane, and internal refractive index inhomogeneity describes the disorder or uniformity of the optical properties of the material inside the defect. These parameters together constitute a quantitative description of the three-dimensional morphology of the defect. Compared with using only two-dimensional image features, the three-dimensional morphological feature parameters contain richer spatial structural information. For example, for two defects with similar areas on a two-dimensional projection map, their defect depths may differ significantly; for two defects with similar depths, their internal refractive index inhomogeneity may reveal that one is a material inclusion and the other is a microcrack. These differences can be clearly distinguished through data comparison.

[0070] In one embodiment of the present invention, defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity are three feature parameters with different dimensions and physical meanings, which need to be converted into a unified three-dimensional data structure that can be processed by a three-dimensional convolutional neural network. In some embodiments, the process of converting defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity into a three-dimensional feature tensor is performed step by step. Based on the defect depth, a preset number of intervals are divided along the depth direction at equal intervals, and the interval into which the actual depth value of the defect in the three-dimensional volume data model falls is quantized and encoded to generate a one-dimensional depth feature vector. For example, if the entire possible depth range is preset to be divided into 10 intervals, and a defect with a measured depth of 85 micrometers falls into the 7th interval, then a one-dimensional vector of length 10 is generated, where only the 7th element is 1 and the rest are 0. This vector expresses the quantized position information of the defect in the depth direction. Based on the lateral distribution contour of the defect, a two-dimensional contour feature matrix is ​​generated by assigning a binary value to each coordinate point on the lateral coordinate grid of the three-dimensional volume data model according to whether each coordinate point is identified as a defect area. For a 320×320 lateral grid, the contour feature matrix is ​​also a 320×320 matrix, where the coordinate points belonging to the defect projection area are assigned a value of 1, and otherwise 0. This matrix accurately describes the projection shape and position of the defect on the lateral plane.

[0071] In practical implementation, based on the refractive index inhomogeneity within the defect, the optical scattering intensity values ​​of each voxel within the defect region in the three-dimensional volumetric data model are normalized and directly arranged in three-dimensional space to form a three-dimensional inhomogeneity feature cube. The normalization process linearly scales the optical scattering intensity values ​​of all voxels within the defect region to a range of 0 to 1. The inhomogeneity feature cube spatially extracts a three-dimensional sub-block encompassing the entire defect region. Assuming the size of this sub-block is H (height) × W (width) × D (depth), then this feature cube is a three-dimensional array of size H × W × D. The value at each position in the array is the normalized optical scattering intensity value of the corresponding voxel, directly encoding the three-dimensional spatial distribution variation of the material's optical properties within the defect. It can be understood that performing an outer product operation on the horizontal plane between a one-dimensional depth feature vector and a two-dimensional contour feature matrix yields a three-dimensional basic feature tensor. The outer product operation combines the depth information and the horizontal contour information in three-dimensional space, generating a three-dimensional tensor with a size of 10×320×320. Each "depth layer" of this basic feature tensor is the product of the corresponding elements in the contour feature matrix and the depth feature vector, thus roughly representing the possible distribution space of defects in three-dimensional space. The three-dimensional basic feature tensor and the three-dimensional non-uniformity feature cube are concatenated along the channel dimension to form the final three-dimensional feature tensor for the defect classifier. The concatenation operation connects the two three-dimensional data volumes along the channel dimension. Assuming the size of the non-uniformity feature cube is 30×30×20, the size of the final three-dimensional feature tensor is 10×320×320×(1+1). One channel is the basic feature tensor, and the other channel is the data obtained by spatial interpolation or clipping of the non-uniformity feature cube to the same spatial size (10×320×320) as the basic feature tensor. This makes the three-dimensional feature tensor contain both prior information about the geometric spatial distribution of the defect and detailed information about the internal material.

[0072] In practice, a three-dimensional feature tensor is input into the input layer of a defect classifier built on a three-dimensional convolutional neural network. This classifier comprises multiple cascaded three-dimensional convolutional layers, three-dimensional pooling layers, and fully connected layers. The input three-dimensional feature tensor flows sequentially through these hierarchical structures. The three-dimensional convolutional and pooling layers progressively extract high-level abstract features from the three-dimensional feature tensor. The three-dimensional convolutional layers use small three-dimensional convolutional kernels to perform sliding calculations across the three dimensions of the input tensor, automatically learning and extracting local spatial feature patterns. The three-dimensional pooling layers (such as max pooling) downsample the feature map, reducing the amount of data and increasing the translation invariance of the features while preserving the main features. The fully connected layers map the extracted high-level abstract features into probability scores corresponding to different defect types. The fully connected layers then perform weighted summation and nonlinear transformation on the one-dimensional feature vector expanded after the preceding convolutional and pooling layers, ultimately outputting a vector where each element represents the confidence probability that the input sample belongs to a specific defect type (such as bubbles, cracks, inclusions, etc.). The defect type with the highest probability score is output as the initial defect identification result, that is, the category label corresponding to the maximum value in the output probability vector is selected as the classification result.

[0073] In an example scenario, the defect depth extracted from a detected region is quantized and encoded into a vector of length 10 [0,0,0,0,0,0,1,0,0,0]. The lateral distribution contour of the defect is a 320×320 binary matrix. The outer product of the two matrices generates a 10×320×320 basic feature tensor. The non-uniformity feature cube of this region originally has a size of 25×25×15. After interpolation adjustment to 10×320×320, it is concatenated with the basic feature tensor in the channel dimension to form a final three-dimensional feature tensor with a size of 10×320×320×2. This tensor is input into a defect classifier with 4 three-dimensional convolutional layers, 3 three-dimensional max pooling layers, and 2 fully connected layers. The network finally outputs a probability vector containing 5 elements, such as [0.02, 0.87, 0.05, 0.03, 0.03]. The second element, 0.87, is the largest and corresponds to the "bubble" defect category. Therefore, the initial defect identification result is "bubble".

[0074] Optionally, the depth feature vector can be encoded using a normalized scalar based on the actual depth value instead of one-hot encoding. The generation of the contour feature matrix is ​​not limited to binary; gradient values ​​can be assigned to the contour edges. The generation of the non-uniformity feature cube can be performed without normalization, directly using the original optical scattering intensity values, but batch normalization is required before inputting it into the network. The specific architecture of the 3D convolutional neural network, including the number of layers, kernel size, and number of channels, can be adjusted according to the actual task complexity and computational resources.

[0075] It is understandable that the operations of 3D convolutional layers and 3D pooling layers are performed in 3D space rather than 2D images. This allows them to effectively capture the correlation of features across the three dimensions of depth, width, and height, which is crucial for analyzing defects with complex 3D morphologies. The final fully connected layer maps the learned distributed feature representations to a specific defect category space, and probability scores provide a measure of the reliability of the classification results. The entire processing flow integrates 3D morphological feature parameters from different sources and dimensions into a unified 3D data structure and utilizes the powerful spatial feature learning capabilities of 3D convolutional neural networks for automatic classification and recognition, avoiding the difficulty of manually designing complex classification rules.

[0076] In one embodiment of the present invention, the known design layout data is a digital file describing the geometry, material properties, and spatial relationships of each layer of a glass-based circuit board. It typically contains precise coordinates and dimensions of structural elements such as conductors, insulating dielectric layers, and vias. In some embodiments, the three-dimensional coordinates of detected defects in a three-dimensional volumetric data model are mapped to the known design layout coordinate system of the glass-based circuit board under test. The coordinate system of the three-dimensional volumetric data model has its origin at the scan start point, while the known design layout data has an independent global coordinate system. Coordinate mapping requires a rigid body transformation. This transformation relationship is determined through a calibration process, which uses the coordinates of several known marker points on the glass-based circuit board in both coordinate systems to solve for the rotation matrix and translation vector. The mapping relationship can be expressed by the following formula:

[0077]

[0078] in: This represents the coordinates of the defect in the coordinate system of the three-dimensional volumetric data model. This represents the mapped coordinates in the known design layout coordinate system. It is a 4x4 homogeneous coordinate transformation matrix that includes rotation and translation parameters. After mapping, each defect has a corresponding three-dimensional spatial location in the layout data.

[0079] In practice, identifying the structural elements corresponding to mapped coordinate points in known design layout data includes conductor lines, insulating dielectric layers, and vias. Known design layout data is typically stored in a layered format, with each layer containing a set of polygons or geometric shapes representing specific types of structural elements. The identification process involves performing spatial queries within each layer of the layout data to determine the mapped coordinate points. Within the geometric boundaries of which structural element it is located. For example, for a mapped coordinate point, the system first determines its location based on its... The coordinates determine which dielectric or circuit layer it is located in, and then in the plane of that layer, according to... The coordinates determine which specific conductor circuit pattern, dielectric region, or through-hole structure the point falls within.

[0080] It is understandable that the spatial location of the defect type identified in the preliminary defect identification is judged based on the expected function of the structural element. Different structural elements, due to differences in their materials, functions, and manufacturing processes, may produce or tolerate different types of defects. The judgment logic is based on a pre-defined set of rules. When the judgment is deemed reasonable, the preliminary defect identification result is confirmed as valid, and it is integrated with the corresponding three-dimensional coordinates and structural element information into a final defect detection report. The final defect detection report is a structured document or data file that systematically lists the attributes of each identified defect, including the defect type, its precise location in the layout coordinate system, the structural element it is located in, and its three-dimensional morphological parameter.

[0081] In an example scenario, the defect classifier initially identifies a region as "material missing," with coordinates (155.2 μm, 203.7 μm, 68.5 μm) in the calibrated 3D volumetric data model. This defect is then identified by applying a transformation matrix. The point was mapped to coordinates (2010.5μm, 1755.3μm, 65.8μm) in the known design layout coordinate system. A search in the known design layout database revealed that this coordinate point is located within the sidewall region of a 50μm diameter metallized via structure. The sidewall region of the metallized via is intended to provide vertical electrical connections, and its fabrication involves a deposition process. The presence of "material missing" at this location is a potential defect in the process (such as incomplete plating). Therefore, the initial identification result of "material missing" at this spatial location was deemed reasonable, and the defect was confirmed and recorded. See Table 1 for an example containing multiple defect verification results.

[0082] Table 1: Defect Spatial Location Verification Results

[0083] Defect ID Model In-Field Coordinates Preliminary identification type Mapped coordinates within the map Corresponding map structural elements Factor Expectation Function Verification judgment D001 (155,204,68) Material missing (2010,1755,66) Φ50μm through-hole sidewall Vertical conduction Reasonable / Confirmed D002 (320,110,12) bubble (2175,1662,10) Surface signal line Transmitted signals Reasonable / Confirmed D003 (450,80,95) crack (2305,1632,92) bottom ground layer Provide grounding Reasonable / Confirmed D004 (120,300,25) Metal residue (1995,1851,23) Media isolation area Electrical insulation Unreasonable / Review

[0084] Optional, coordinate transformation matrix The solution can be obtained by fitting multiple corresponding point pairs using the least squares method. When identifying structural elements in known design layout data, if the mapped coordinate point happens to be located on the boundary between two structural elements, it can be associated with both elements and noted in the report. The rules for judging reasonableness can be based on a knowledge base, which stores the associations of common defect types on different structural elements.

[0085] In practical implementation, for example D004 in Table 1, it was initially identified as "metal residue." However, the mapped coordinates point is located in the "medium isolation zone," which is expected to be a pure insulating medium. The intended function of this area is electrical insulation, and the presence of conductive "metal residue" is a serious problem. Although the location relationship is reasonable, it may trigger a more rigorous review or classification adjustment. The spatial location verification step acts as a logical filter. It uses prior knowledge of the design layout to verify the contextual consistency of classification results based solely on image features. This helps to discover misjudgments that the classifier may make due to feature ambiguity, improving the reliability of the final defect detection report. The final defect detection report integrates defect features and types from image analysis with structural background information from design information, providing more comprehensive information for process diagnosis and repair decisions.

[0086] See Figure 4 This is a visualization of the coordinate mapping relationships in the defect detection process of glass-based circuit boards. It intuitively presents the rigid body transformation mapping structure between the coordinate system of the 3D volumetric data model and the global coordinate system of the design layout. This diagram is a key verification tool in the "coordinate mapping calculation stage" of the defect detection process. By visualizing the mapping relationship of the three types of coordinates, it intuitively verifies the accuracy of the coordinate transformation matrix, ensuring that the mapping of the defect location in the two coordinate systems is linear and stable. This provides a reliable positional basis for subsequent "structural element matching" and "defect rationality verification," avoiding misjudgments due to coordinate deviations. As can be seen from the diagram, the distribution trend of all mapping points is consistent, indicating that the transformation matrix obtained in this calibration is reliable. The distribution of mapping points in different dimensions shows no obvious dispersion or offset, indicating that the coordinate transformation accuracy from the 3D volumetric data model to the design layout meets the detection requirements, providing a core basis for the accurate matching of subsequent defect locations.

[0087] In one embodiment of the present invention, after acquiring tomographic image data using an optical coherence scanning microscopy system, the data undergoes quality assessment, invalid data removal, and re-acquisition, and the specific operating parameters of the system are defined. The acquired tomographic image data is the foundation for all subsequent processing steps, and its quality directly affects the accuracy of 3D reconstruction and defect identification; therefore, quality assessment and purification are necessary before executing the core processes. In some embodiments, the acquired tomographic image data undergoes quality assessment, and the signal-to-noise ratio (SNR) and contrast ratio of each tomographic image are calculated. A region of interest (ROI) containing circuit structures and a background region are selected on each tomographic image. The ROI is a local area in the image where circuits, vias, or other functional structures are clearly displayed, and the background region is a smooth area in the image that is expected to be a uniform substrate or unstructured. The average pixel grayscale value within the ROI is calculated as the signal intensity, and the standard deviation of the pixel grayscale value in the background region is calculated as the background noise. The signal intensity is divided by the background noise to obtain the SNR of the tomographic image. The difference between the maximum and minimum pixel grayscale values ​​within the ROI is calculated as the contrast ratio of the tomographic image. Signal-to-noise ratio (SNR) and contrast ratio are two independent quantitative metrics. SNR reflects the relative strength of the useful signal and random noise in an image, while contrast ratio reflects the significance of the grayscale difference between the target and the background. It can be understood that tomographic images with SNR and contrast ratios below preset thresholds are considered invalid and discarded. These preset thresholds are determined based on the statistical distribution of SNR and contrast ratios from a large number of qualified sample images. For example, a separate SNR threshold could be set. and contrast threshold For a tomographic image, if its calculated signal-to-noise ratio or contrast If the image is invalid, it will be removed from the original dataset.

[0088] In practice, the scanning position coordinates corresponding to the rejected invalid images are recorded, and the optical coherence scanning microscopy system is controlled to reacquire tomographic image data of the corresponding area at a scanning resolution of 320×320. The scanning position coordinates uniquely identify the three-dimensional spatial position of each tomographic image on the glass substrate circuit board under test. The system drives the scanning device to reposition itself to this position based on the recorded coordinate information. The reacquired tomographic image data is then added to the original dataset to form complete tomographic image data. This process ensures that the dataset used for 3D reconstruction is spatially continuous and complete, and that data from all positions meets minimum quality requirements. The optical coherence scanning microscopy system uses a broadband light source with a center wavelength in the visible light band, such as a superluminescent diode light source with a center wavelength of 850 nm and a bandwidth of 100 nm. The tomographic image data acquisition process follows a step of scanning layer by layer along the depth direction, starting from the surface of the glass substrate circuit board under test. After scanning a horizontal 320×320 pixel area, a set of tomographic image data for that pixel area is obtained. The broadband light source provides a short coherence length, thereby achieving high resolution in the depth direction.

[0089] In one example scenario, a target area of ​​a glass-based circuit board was scanned, yielding 512 tomographic images. Quality assessment determined that the signal-to-noise ratios (SNRs) of images 127, 308, and 419 were 8.5, 7.9, and 9.1, respectively, below the preset SNR threshold of 10.0. These three images were deemed invalid and removed from the dataset. The system recorded the 3D scan coordinates of these invalid images and drove the scanning head to return to these coordinate positions sequentially, re-acquiring images at the same 320×320 scan resolution. After re-acquisition, the SNRs of the new images improved to 12.7, 11.8, and 13.5, respectively, all exceeding the threshold. These new, valid images were inserted into their corresponding positions in the original dataset, replacing the gaps left by the invalid images, thus forming a complete dataset of 512 valid tomographic images for subsequent processing.

[0090] Optionally, the region of interest and background region can be selected manually or automatically by an algorithm to identify textured and smooth areas in the image. The preset thresholds for signal-to-noise ratio and contrast can be dynamically adjusted based on different sample types, scanning parameters, or detection standards. An upper limit can be set for the number of re-acquisitions; if the image quality after re-acquisition is still unsatisfactory, the location can be marked as an "undetectable area" and noted in the report. The visible light broadband light source for the optical coherence scanning microscopy system can specifically use light sources with center wavelengths of 530 nm, 780 nm, etc., to adapt to glass substrates with different material properties.

[0091] In some embodiments, the step of layer-by-layer scanning in an optical coherence scanning microscope system is as follows: First, the lateral position is fixed, and an A-scan is performed along the depth direction (Z-axis) to obtain the reflected light intensity information at different depths of the point, forming an A-scan line; then, a raster scan is performed in the X and Y directions, covering a region of 320 pixels × 320 pixels; after the A-scan is completed at each lateral pixel, the depth reflection information of all points within the 320 × 320 pixel region is obtained, constituting a complete set of three-dimensional tomographic image data. It can be understood that setting the scanning resolution to 320 × 320 is a balance between lateral resolution and the area of ​​a single scan. A higher number of lateral pixels provides finer lateral details but increases the acquisition and processing time of a single frame of data. The 320 × 320 setting maintains reasonable detection efficiency while ensuring sufficient image detail.

[0092] See Figure 5 This is a comparison of the reflection intensity versus depth curves of a glass-based circuit board under an optical coherence scanning microscopy system, which has significant technical value in the defect detection process. The peak range of the curve directly corresponds to the depth range of the defect, which perfectly matches the technical steps in the patent: "analyzing the optical scattering intensity profile along the depth direction and determining the defect boundary depth based on the abrupt change point." The difference in reflection intensity between the normal area and the defect area can be directly quantified as "refractive index inhomogeneity within the defect," which is one of the core input features of the defect classifier. The overall reflection intensity in the defect area is low and the peak is flat, which conforms to the optical characteristics of "material missing" defects, providing crucial optical evidence for preliminary defect identification. This figure is one of the core outputs of the optical coherence scanning microscopy system. It transforms the depth information in the three-dimensional volumetric data model into intuitive curve features, serving not only as the basis for extracting parameters such as defect depth and refractive index inhomogeneity, but also providing underlying data support for subsequent defect classification by a three-dimensional convolutional neural network.

[0093] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for detecting defects in glass-based circuit boards, characterized in that, include: The tomographic image data of the target area of ​​the glass substrate circuit board under test was acquired using an optical coherence scanning microscopy system. The scanning resolution of the optical coherence scanning microscopy system during scanning imaging was set to 320×320. The tomographic image data is subjected to three-dimensional reconstruction processing to generate a three-dimensional volume data model that reflects the internal structural features of the target region; Multiple three-dimensional morphological feature parameters related to line defects are extracted from the three-dimensional volume data model. These three-dimensional morphological feature parameters include defect depth, defect lateral distribution profile, and defect internal refractive index non-uniformity. The multiple three-dimensional morphological feature parameters are input into a defect classifier built on a three-dimensional convolutional neural network to identify the defect type and obtain preliminary defect identification results. Based on the known design layout data of the glass-based circuit board under test, the spatial location of the preliminary defect identification results is verified, and a final defect detection report is generated.

2. The method for detecting defects in a glass-based circuit board according to claim 1, characterized in that, The step of performing three-dimensional reconstruction processing on the tomographic image data to generate a three-dimensional volumetric data model reflecting the internal structural features of the target region includes: Each two-dimensional tomographic image in the tomographic image data undergoes denoising and edge enhancement preprocessing; The image registration algorithm is used to precisely align all the preprocessed two-dimensional tomographic images in three-dimensional space. A three-dimensional interpolation algorithm is used to spatially interpolate the precisely aligned two-dimensional tomographic image data to fill the three-dimensional data gaps formed by the scanning interval; Based on the interpolated complete 3D dataset, a 3D volumetric data model capable of characterizing the distribution of optical scattering intensity inside the material is constructed using voxels as the basic unit.

3. The method for detecting defects in a glass-based circuit board according to claim 2, characterized in that, The extraction of multiple three-dimensional morphological feature parameters related to line defects from the three-dimensional volumetric data model includes defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity of the defect, including: In the three-dimensional volume data model, the optical scattering intensity profile of each pixel column is analyzed along the depth direction, the boundary depth of the defect is determined based on the abrupt change points in the profile, and the depth of the defect is calculated. In the three-dimensional volume data model, the defect region is sliced ​​along the horizontal plane, the binary contour of the defect region in each slice is extracted, and the contours of all slices are superimposed to generate the horizontal distribution contour of the defect. The optical scattering intensity value of each voxel within the defect region in the three-dimensional volumetric data model is calculated relative to the dispersion of its neighboring voxels, thereby quantifying the refractive index inhomogeneity within the defect.

4. The method for detecting defects in a glass-based circuit board according to claim 3, characterized in that, The step of inputting the multiple three-dimensional morphological feature parameters into a defect classifier constructed based on a three-dimensional convolutional neural network for defect type identification and obtaining preliminary defect identification results includes: The defect depth, defect lateral distribution profile, and internal refractive index inhomogeneity of the defect are converted into a three-dimensional feature tensor. The three-dimensional feature tensor is input into the input layer of the defect classifier constructed based on the three-dimensional convolutional neural network. The defect classifier includes multiple cascaded three-dimensional convolutional layers, three-dimensional pooling layers, and fully connected layers. The high-level abstract features of the three-dimensional feature tensor are extracted step by step through the three-dimensional convolutional layer and the three-dimensional pooling layer. The extracted high-level abstract features are mapped to probability scores corresponding to different defect types through the fully connected layer, and the defect type with the highest probability score is output as the preliminary defect identification result.

5. The method for detecting defects in a glass-based circuit board according to claim 4, characterized in that, The process involves combining the known design layout data of the glass-based circuit board under test, performing spatial location verification on the preliminary defect identification results, and generating a final defect detection report, including: The detected defects are mapped to their three-dimensional coordinates in the three-dimensional volume data model and then to the known design layout coordinate system of the glass-based circuit board under test. In the known design layout data, identify the structural elements corresponding to the mapped coordinate points, the structural elements including conductor lines, insulating dielectric layers and vias; Based on the expected function of the structural elements, determine whether the defect type of the preliminary defect identification result is reasonable in the spatial location; when it is determined to be reasonable, confirm that the preliminary defect identification result is valid, and integrate it with the corresponding three-dimensional coordinates and structural element information into the final defect detection report.

6. The method for detecting defects in a glass-based circuit board according to claim 5, characterized in that, After acquiring tomographic image data of the target area of ​​the glass-based circuit board under test using an optical coherence scanning microscopy system, the method further includes: The quality of the acquired tomographic image data is evaluated, and the signal-to-noise ratio and contrast of each tomographic image are calculated. Tomographic images with a signal-to-noise ratio and contrast below a preset threshold are identified as invalid images and discarded. Record the scanning position coordinates corresponding to the invalid images that have been rejected, and control the optical coherence scanning microscope system to reacquire the tomographic image data of the area corresponding to the scanning position coordinates at a scanning resolution of 320×320. The newly acquired tomographic image data is added to the original dataset to form the complete tomographic image data.

7. The method for detecting defects in a glass-based circuit board according to claim 6, characterized in that, The calculation of the signal-to-noise ratio and contrast of each tomographic image includes: Select the region of interest containing the circuit structure and the background region on each tomographic image; The average value of the pixel grayscale values ​​in the region of interest is calculated as the signal strength, and the standard deviation of the pixel grayscale values ​​in the background region is calculated as the background noise. The signal-to-noise ratio of the tomographic image is obtained by dividing the signal intensity by the background noise. The difference between the maximum and minimum gray values ​​of pixels within the region of interest is calculated and used as the contrast of the tomographic image.

8. The method for detecting defects in a glass-based circuit board according to claim 1, characterized in that, The optical coherent scanning microscopy system uses a broadband light source with a center wavelength in the visible light band. The acquisition process of the tomographic image data follows the steps of scanning layer by layer along the depth direction, starting from the surface of the glass substrate circuit board under test. After scanning a horizontal 320×320 pixel area, a set of tomographic image data of the pixel area is obtained.

9. The method for detecting defects in a glass-based circuit board according to claim 4, characterized in that, The process of converting the defect depth, the lateral distribution profile of the defect, and the internal refractive index inhomogeneity of the defect into a three-dimensional feature tensor includes: Based on the defect depth, a preset number of intervals are divided at equal intervals along the depth direction, and a one-dimensional depth feature vector is generated by quantizing and encoding the interval into which the actual depth value of the defect falls in the three-dimensional volume data model. Based on the lateral distribution contour of the defects, a two-dimensional contour feature matrix is ​​generated on the lateral coordinate grid of the three-dimensional volume data model by assigning binary values ​​according to whether each coordinate point is identified as a defect area. Based on the internal refractive index inhomogeneity of the defect, the optical scattering intensity values ​​of each voxel in the defect region of the three-dimensional volume data model are normalized, and a three-dimensional inhomogeneity feature cube is directly formed by the three-dimensional spatial arrangement of the voxels. The one-dimensional depth feature vector and the two-dimensional contour feature matrix are multiplied by an outer product in the horizontal plane to obtain a three-dimensional basic feature tensor. The three-dimensional basic feature tensor and the three-dimensional non-uniformity feature cube are concatenated along the channel dimension to form the final three-dimensional feature tensor for input to the defect classifier.

10. A glass-based circuit board defect detection system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the glass-based circuit board defect detection method as described in any one of claims 1 to 9.

Images