An X-ray-based method for detecting blind via defects in PCB boards

Abstract

This invention relates to the field of PCB blind via defect detection technology and discloses an X-ray-based method for detecting PCB blind via defects. The method includes the following steps: first, the PCB is irradiated with X-rays; then, data is acquired and an initial image is established using CT projection; subsequently, the initial image is reconstructed using filtered back projection to obtain a reconstructed image; then, the blind via region is extracted using the reconstructed image; next, a two-stream neural network architecture is constructed based on a deep residual network; finally, the blind via defect is detected and classified using the two-stream neural network architecture. This invention, through a two-stream neural network architecture constructed based on a deep residual network, comprehensively utilizes geometric and material features, enabling comprehensive and accurate detection and classification of various defects in PCB blind vias, such as cracks, wrinkles, dents, and voids. This improves the efficiency and accuracy of defect detection, helps to promptly identify quality problems in the PCB manufacturing process, and ensures product quality.

Description

Technical Field

[0001] This invention relates to the field of PCB blind hole defect detection technology, specifically to an X-ray-based method for detecting PCB blind hole defects. Background Technology

[0002] PCB blind via defect detection has long faced two core challenges: the invisibility of hidden defects and the contradiction between detection efficiency. Traditional methods such as automated optical inspection (AOI) and ultrasonic testing are limited by physical principles, only able to identify surface or near-surface defects (such as scratches and short circuits), and are powerless against hidden defects such as voids, cracks, and interlayer alignment misalignments in blind vias within multilayer boards. For example, AOI relies on visible light reflection and cannot penetrate copper foil layers to observe internal structures; while ultrasonic testing can penetrate materials, its resolution for micron-level defects is insufficient and it is easily affected by material anisotropy. In addition, manual visual inspection suffers from high subjectivity and low efficiency, making it difficult to meet the needs of large-scale industrial production. Against this backdrop, X-ray nondestructive testing technology, with its high penetration and high resolution, has become a key direction for breaking through the bottlenecks of traditional detection methods. Its ability to image through material density differences can directly display the internal structure of blind vias, providing a reliable basis for defect location and classification.

[0003] Despite the physical advantages of X-ray inspection, the raw images still suffer from noise interference and blurred defect features, especially in complex multilayer PCBs where the contrast between blind vias and surrounding structures is low, making it difficult for traditional image processing algorithms (such as thresholding and edge detection) to accurately extract defects. To address this, researchers have proposed two technical approaches: image reconstruction optimization based on physical models and intelligent feature parsing based on deep learning. The former enhances the signal intensity of the target area by combining a filtered back projection (FBP) algorithm with a blind via priority weight design; simultaneously, it utilizes Hessian matrix eigenvalue analysis to quantify the differences between tubular structures (blind vias) and speckled structures (impurities), improving defect extraction accuracy. The latter introduces 3D ResNet as the backbone network, using residual connections to address the degradation problem of deep networks, and constructs a two-stream neural network to fuse geometric features (curvature tensor) and material features (multi-energy X-ray energy spectrum decomposition), overcoming the limitations of insufficient single-modal information. Furthermore, a weighted cross-entropy loss function further enhances the model's ability to identify minute defects by dynamically balancing the weights of class samples. These algorithmic innovations have enabled X-ray inspection to move from "visible" to "precise and identifiable," driving the evolution of PCB quality control towards intelligence and automation. Summary of the Invention

[0004] The purpose of this invention is to solve the problem that the original images of X-ray inspection still have noise interference and blurred defect features, especially in complex multilayer PCBs, where the contrast between blind holes and the surrounding structure is low, and traditional image processing algorithms (such as threshold segmentation and edge detection) are difficult to accurately extract defects. Therefore, an X-ray-based method for detecting blind holes in PCBs is proposed.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] An X-ray-based method for detecting blind via defects in PCB boards includes the following steps:

[0007] S1. First, X-rays are used to irradiate the PCB board;

[0008] S2. Data acquisition and initial image creation are performed using CT projection;

[0009] S3. The initial image is reconstructed using filtered back projection to obtain the reconstructed image;

[0010] S4. Extract the blind hole region using the reconstructed image;

[0011] S5. Construct a dual-stream neural network architecture based on deep residual networks;

[0012] S6. Detect and classify blind hole defects using a two-stream neural network architecture.

[0013] Based on the above technical solution, the present invention can be further improved as follows.

[0014] Preferably, the projection formula for the CT projection is:

[0015] ;

[0016] in, For projection data, For spatial ray intensity, The strength after penetrating the PCB.

[0017] Preferably, the filtered backprojection reconstruction of the initial image requires the introduction of blind aperture optimization weights:

[0018] ;

[0019] in, This represents the reconstructed 3D image. This represents the distance from the X-ray source to the center of rotation. , These represent the horizontal and vertical coordinates, respectively. It is a blind hole optimization weight. It is a convolution filtering operation;

[0020] ;

[0021] in, It is the depth attenuation coefficient, and z is the blind hole depth. It is the depth of the target blind hole.

[0022] Preferably, the process of extracting the blind hole region is as follows:

[0023] Tubular structure detection based on Hessian matrix is ​​employed at each body point. At this point, calculate the second derivative of the image and construct the Hessian matrix:

[0024] ;

[0025] Eigenvalue decomposition yields ,in The local tubular structure representing the feature point corresponds to a blind hole. The mottled structure representing the feature points corresponds to impurities. A planar structure representing feature points, corresponding to the hole wall;

[0026] The probability of each point belonging to a blind hole is calculated using the blind hole similarity response function. The blind hole similarity response function is as follows:

[0027] ;

[0028] in, It is the similarity response value. , ; , , , These are the first empirical parameter, the second empirical parameter, and the third empirical parameter.

[0029] Preferably, the process of constructing a two-stream neural network architecture based on a deep residual network is as follows:

[0030] Using 3DResNet as the backbone network, geometric feature branches and material feature branches are constructed. A two-stream feature fusion mechanism is then introduced, and a weighted cross-entropy loss function is used to optimize and adjust the two-stream neural network architecture.

[0031] Preferably, the process of constructing the geometric feature branches is as follows:

[0032] The 3D data block of the blind hole region is standardized and used as input. A 3D convolution kernel is used for initial convolution with a stride of 2 to quickly reduce the data size. Simultaneously, basic edge and corner features are extracted. To construct a four-channel curvature feature tensor The curvature feature tensor With the main feature map The splicing formula is as follows:

[0033] ;

[0034] ;

[0035] in, This indicates that the local surface shape at that point resembles a local ellipsoid. This indicates that the local surface shape at that point is a developable surface, similar to a local part of a cylinder or plane. This indicates that the local surface shape at this point resembles a saddle-shaped surface, where K is the Gaussian curvature. The formula for calculating Gaussian curvature is as follows:

[0036] ;

[0037] The final output is a feature tensor that integrates general features and precise set information. This is used to quantify the curvature of the blind hole surface.

[0038] Preferably, the process of constructing the material feature branch is as follows:

[0039] The input is projection data acquired at different transistor voltages. According to the Beer-Lambert law, the attenuation can be expressed as:

[0040] ;

[0041] in, It is the linear decay coefficient of the material at energy E, representing the decay capability of an object located at path s at energy E;

[0042] Energy spectral decomposition transforms this into a linear problem; first, the projected values ​​are calculated:

[0043] ;

[0044] For the known C types of materials It can be decomposed into:

[0045] ;

[0046] in, It is the density distribution of material c along path s. Since the attenuation coefficient of material c is known, the projection equation becomes:

[0047] ;

[0048] For each detector pixel The projection values ​​under multiple energies can be written in matrix form:

[0049]

[0050] Where M is The material attenuation coefficient matrix, where a is... The material density vector; by solving using the least squares method, the material density map of each base material can be obtained. ;

[0051] Material density map As input and expanded using dilated convolution, we obtain

[0052] ;

[0053] Where d is the expansion rate;

[0054] Then, convolutional kernels with d=3, 6, and 9 are used in parallel to convolve the features, and the results are then combined.

[0055] ;

[0056] This enables the network to capture material regions of different scales simultaneously, whether it's a large area of ​​copper foil or a tiny hole; ultimately, the output of the material stream is a feature tensor containing multi-scale material information. .

[0057] Preferably, the dual-stream feature fusion process is as follows:

[0058] First, the feature map size of the material flow is adjusted to be exactly the same as that of the geometry flow feature map through a 3D interpolation algorithm. Then, the feature maps of the two flows are stitched together through a gated fusion mechanism, so that the network can adaptively choose to trust the geometry flow or the material flow, thereby achieving accurate feature fusion.

[0059] The splicing formula is as follows:

[0060] ;

[0061] in, This indicates a splicing operation. It's the Sigmoid function, which compresses the output to between 0 and 1. and These are learnable parameters, and the features that are ultimately fused. It is generated by weighting the gated graph, and the generation formula is as follows:

[0062] ;

[0063] in, This indicates element-wise multiplication.

[0064] Preferably, the weighted cross-entropy loss function is:

[0065] ;

[0066] in, This is the total loss value. , , These are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. It is cross-entropy loss. It is Dice's loss. It is an L2 regularization term;

[0067] The formula for calculating cross-entropy loss is as follows:

[0068] ;

[0069] in, It's a real label. It is a predicted probability;

[0070] The formula for calculating Dice loss is as follows:

[0071] ;

[0072] Where Y is the true label region and P is the predicted region. This is the number of true positive pixels. This is the actual number of pixels in the target object. It predicts the number of target pixels.

[0073] Preferably, the The value range is 0.1-1.0. The value range is 0.1-1.0. The value range is 0.0001-0.01.

[0074] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0075] 1. This invention introduces blind hole optimization weights when reconstructing the initial image through filtered back projection. It comprehensively considers factors such as the distance from the X-ray source to the rotation center, coordinate information, blind hole priority weights, and convolution filtering operations. Furthermore, the depth attenuation coefficient is related to the depth of the blind hole, making the reconstructed three-dimensional image more consistent with the actual situation and able to more accurately present the detailed information of the blind holes on the PCB board, providing a high-quality image foundation for subsequent defect detection.

[0076] 2. This invention employs a tubular structure detection method based on the Hessian matrix. By calculating the second derivative of the image to construct the Hessian matrix and performing eigenvalue decomposition, it can accurately distinguish between blind holes, impurities, and hole walls. The probability of each point belonging to a blind hole is calculated using the blind hole similarity response function, which further improves the accuracy of blind hole region extraction and reduces the interference of impurities and hole walls on blind hole detection.

[0077] 3. This invention inputs the three-dimensional data blocks of the blind hole region after standardization, uses a large three-dimensional convolutional kernel to quickly reduce the data size and extract basic edge and corner features, constructs a curvature feature tensor and splices it with the backbone feature map. This branch can quantify the curvature of the blind hole surface, which helps to distinguish convex, concave and saddle-shaped regions. It is crucial for key defect detection such as crack recognition, enabling the network to have both general visual features and precise geometric shape features.

[0078] 4. This invention utilizes the advantages of multi-energy X-ray scanning and, based on Beer-Lambert's law and the principle of energy spectrum decomposition, transforms projection data into a material density map. Through dilated convolution and spatial pyramid pooling modules, it captures material region information at different scales, effectively identifying both large copper foils and tiny voids, providing rich material information for defect detection.

[0079] 5. This invention adjusts the size of the material flow feature map using a three-dimensional interpolation algorithm, making it consistent with the geometric flow feature map. Figure 1 Furthermore, a gating fusion mechanism is used to stitch the feature maps of the two streams together. This mechanism enables the network to adaptively choose to trust either the geometry stream or the material stream, and to accurately fuse features according to the characteristics of different defect types, thereby improving the effectiveness and reliability of the features.

[0080] 6. This invention uses a weighted cross-entropy loss function to optimize the two-stream neural network architecture. This function combines cross-entropy loss, Dice loss, and L2 regularization. Cross-entropy loss focuses on the difference between the predicted probability and the true label, Dice loss focuses on the degree of overlap between the predicted region and the true label region, and L2 regularization prevents the model from overfitting. By reasonably setting the weight coefficients, the model pays more attention to samples that are difficult to classify or are more important during training, thereby improving the model's performance on samples of different categories and enhancing the model's generalization ability.

[0081] 7. This invention utilizes a dual-stream neural network architecture based on a deep residual network to comprehensively and accurately detect and classify various defects in PCB blind vias, such as cracks, wrinkles, dents, and voids. This improves the efficiency and accuracy of defect detection, helps to promptly identify quality problems in the PCB manufacturing process, and ensures product quality. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the process of the present invention;

[0083] Figure 2 This is a schematic diagram illustrating the construction process of the geometric feature branches of the present invention;

[0084] Figure 3 This is a schematic diagram illustrating the construction process of the material feature branch of the present invention. Detailed Implementation

[0085] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some 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.

[0086] An X-ray-based method for detecting blind via defects in PCB boards includes the following steps:

[0087] S1. First, X-rays are used to irradiate the PCB board;

[0088] S2. Data acquisition and initial image creation are performed using CT projection;

[0089] S3. The initial image is reconstructed using filtered back projection to obtain the reconstructed image;

[0090] S4. Extract the blind hole region using the reconstructed image;

[0091] S5. Construct a dual-stream neural network architecture based on deep residual networks;

[0092] S6. Detect and classify blind hole defects using a two-stream neural network architecture.

[0093] The projection formula for CT projection is:

[0094] ;

[0095] in, For projection data, For spatial ray intensity, The strength after penetrating the PCB.

[0096] Filtered backprojection reconstruction of the initial image requires the introduction of blind aperture optimization weights:

[0097] ;

[0098] in, This represents the reconstructed 3D image. This represents the distance from the X-ray source to the center of rotation. , These represent the horizontal and vertical coordinates, respectively. It is a blind hole optimization weight. It is a convolution filtering operation;

[0099] ;

[0100] in, It is the depth attenuation coefficient, and z is the blind hole depth. This is the target blind via depth, which can be obtained from the PCB design file.

[0101] The process of extracting the blind hole region is as follows:

[0102] Tubular structure detection based on Hessian matrix is ​​employed at each body point. At this point, calculate the second derivative of the image and construct the Hessian matrix:

[0103] ;

[0104] Eigenvalue decomposition yields ,in The local tubular structure representing the feature point corresponds to a blind hole. The mottled structure representing the feature points corresponds to impurities. A planar structure representing feature points, corresponding to the hole wall;

[0105] The probability of each point belonging to a blind hole is calculated using the blind hole similarity response function. The blind hole similarity response function is as follows:

[0106] ;

[0107] in, It is the similarity response value. , , It is the gradient magnitude of the image at point x, representing the intensity of the grayscale change at that point. , , These are the first empirical parameter, the second empirical parameter, and the third empirical parameter, which are generally set to 0.5, 0.5, and 5 respectively.

[0108] The process of constructing a two-stream neural network architecture based on a deep residual network is as follows:

[0109] Using 3DResNet as the backbone network, geometric feature branches and material feature branches are constructed. A two-stream feature fusion mechanism is then introduced, and a weighted cross-entropy loss function is used to optimize and adjust the two-stream neural network architecture (by assigning different weights to samples of different categories, the model pays more attention to samples that are difficult to classify or are more important during training, thereby improving the model's performance in these categories).

[0110] The process of constructing geometric feature branches is as follows:

[0111] The 3D data block of the blind hole region is standardized and used as input. A 3D convolutional kernel (7x7x7) is used for initial convolution with a stride of 2 to quickly reduce the data size. Simultaneously, basic edge and corner features are extracted. To construct a four-channel curvature feature tensor The curvature feature tensor With the main feature map The splicing formula is as follows:

[0112] ;

[0113] ;

[0114] in, This indicates that the local surface shape at that point resembles an ellipsoid, representing a raised structure such as a tiny bump on a hole wall, a solder bump, or a transversely cut spherical impurity. This indicates that the local surface shape at that point is a developable surface, similar to a local part of a cylinder or plane. It can be understood as a perfect tubular or planar structure. For the wall of a blind hole, this is usually the expected normal state, a smooth cylindrical surface. This indicates that the local surface shape at this point resembles a saddle-shaped surface, such as a crack, wrinkle, or depression on the wall of a hole, or the edge area where two surfaces intersect. K is the Gaussian curvature, and the formula for calculating the Gaussian curvature is as follows:

[0115] ;

[0116] The final output is a feature tensor that integrates general features and precise set information. This is used to quantify the curvature of the blind hole surface.

[0117] This helps distinguish between convex, concave, and saddle-shaped regions, which is crucial for crack identification. The module first calculates the Hessian matrix (a matrix composed of the second derivatives of the 3D image in various directions, describing the local curvature of that point) for each point in the image. Then, it performs eigenvalue decomposition on this matrix to obtain three eigenvalues ​​(…). The magnitude and relationship of these eigenvalues ​​reveal whether the local structure at a point is point-like, tubular (like a hole wall), or sheet-like. Finally, these calculated curvature indices (such as the absolute value and sign of the eigenvalues) are used as new feature channels and merged with the features extracted from the backbone network before being output to subsequent layers. In this way, the network possesses both general visual features and precise geometric shape features.

[0118] The process of constructing material feature branches is as follows:

[0119] The input is projection data acquired at different tube voltages (e.g., 80kV, 100kV, 120kV). According to the Beer-Lambert law, the attenuation can be expressed as:

[0120] ;

[0121] in, It is the linear decay coefficient of the material at energy E, representing the decay capability of an object located at path s at energy E;

[0122] Energy spectral decomposition transforms this into a linear problem; first, the projected values ​​are calculated:

[0123] ;

[0124] For known materials C (such as copper, FR4, and air). It can be decomposed into:

[0125] ;

[0126] in, It is the density distribution of material c along path s. Since the attenuation coefficient of material c is known, the projection equation becomes:

[0127] ;

[0128] For each detector pixel The projection values ​​under multiple energies can be written in matrix form:

[0129] ;

[0130] Where M is The material attenuation coefficient matrix, where a is... The material density vector; by solving using the least squares method, the material density map of each base material can be obtained. ;

[0131] Material density map As input and expanded using dilated convolution, we obtain

[0132] ;

[0133] Where d is the dilation rate, which allows the network to consider the contextual information of a larger area around a point when analyzing the material of the point, and to determine the consistency of the material.

[0134] Then, convolutional kernels with d=3, 6, and 9 are used in parallel to convolve the features, and the results are then combined.

[0135] ;

[0136] This enables the network to capture material regions of different scales simultaneously, whether it's a large area of ​​copper foil or a tiny hole; ultimately, the output of the material stream is a feature tensor containing multi-scale material information. .

[0137] The input is 3D data from the same location as the geometric flow, but it leverages the advantages of multi-energy X-ray scanning. The system scans at multiple voltages, including 80kV, 100kV, and 120kV, because different materials exhibit different X-ray attenuation characteristics at different energies. A pre-calibrated transformation matrix converts the CT values ​​at different energies into a set of feature maps representing the density of the underlying material (e.g., pure copper, pure epoxy resin, air). This process is called "energy spectral decomposition" or "material decomposition." A technique called dilated convolution is used to construct the network. This convolution expands the receptive field of each convolution point without increasing the number of parameters, allowing it to see a wider range of contextual information, which is very useful for determining material consistency. Subsequently, a spatial pyramid pooling module is used. This module uses multiple convolutional kernels of different sizes to convolve and pool the features in parallel, thereby capturing material regions of different scales (whether large copper foil areas or tiny impurities), ensuring effective identification regardless of defect size.

[0138] The two-stream feature fusion process is as follows:

[0139] First, the feature map size of the material flow is adjusted to be exactly the same as that of the geometry flow feature map through a 3D interpolation algorithm. Then, the feature maps of the two flows are stitched together through a gated fusion mechanism, so that the network can adaptively choose to trust the geometry flow or the material flow, thereby achieving accurate feature fusion.

[0140] The splicing formula is as follows:

[0141] ;

[0142] in, This indicates a splicing operation. It's the Sigmoid function, which compresses the output to between 0 and 1. and These are learnable parameters, and the features that are ultimately fused. It is generated by weighting the gated graph, and the generation formula is as follows:

[0143] ;

[0144] in, This indicates element-wise multiplication.

[0145] Gated fusion mechanism: Simply concatenating or adding two feature maps is not the optimal solution. This architecture designs an intelligent "gating" unit. The specific process is as follows: First, the feature maps of the two streams are concatenated. Then, this concatenated feature map is passed through a small neural network (usually several convolutional layers and a sigmoid activation function). This network generates a gating map with weights ranging from 0 to 1. Each point in this gating map represents a decision: at the corresponding location in the image, whether to believe the geometric flow (weight close to 1) or the material flow (weight close to 0).

[0146] The final fused feature = weighted graph × geometric flow feature + (1 - weighted graph) × material flow feature. This mechanism allows the network to adaptively select the most reliable information source. For example, when judging a fine crack, the curvature change information of the geometric shape may be more reliable, so the gating mechanism will give higher weight to the geometric flow; while when judging a cavity filling, the material information may be more direct, so the gating will give higher weight to the material flow.

[0147] The fused features are further refined through attention mechanisms and other modules to enhance important features and suppress irrelevant features. Finally, a 1x1x1 3D convolutional layer (equivalent to a fully connected layer) maps the high-dimensional features to the classification output. The output is a classification score map with the same size as the input data, where each voxel contains a 4-dimensional vector (assuming four classes: background, normal pore wall, pore wall rupture, and insufficient filling) representing the probability of it belonging to each class. Ultimately, the class with the highest probability is taken as the classification result for that voxel, generating the final 3D defect classification map.

[0148] The weighted cross-entropy loss function is:

[0149] ;

[0150] in, This is the total loss value. , , These are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. It is cross-entropy loss. It is Dice's loss. It is the L2 regularization term, also known as weight decay, which is used to reduce the complexity of multiplicative models and prevent overfitting.

[0151] The formula for calculating cross-entropy loss is as follows:

[0152] ;

[0153] in, It's a real label. It is a predicted probability;

[0154] The formula for calculating Dice loss is as follows:

[0155]

[0156] Where Y is the true label region and P is the predicted region. This is the number of true positive pixels. This is the actual number of pixels in the target object. It predicts the number of target pixels.

[0157] The value range is 0.1-1.0. The value range is 0.1-1.0. The value range is 0.0001-0.01.

[0158] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0159] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting blind via defects on PCB boards based on X-rays, characterized in that, Includes the following steps: S1. First, X-rays are used to irradiate the PCB board; S2. Data acquisition and initial image creation are performed using CT projection; S3. The initial image is reconstructed using filtered back projection to obtain the reconstructed image; S4. Extract the blind hole region using the reconstructed image; S5. Construct a dual-stream neural network architecture based on deep residual networks; S6. Detect and classify blind hole defects using a two-stream neural network architecture; The initial image reconstruction via filtered backprojection requires the introduction of blind hole optimization weights: ； in, This represents the reconstructed 3D image. This represents the distance from the X-ray source to the center of rotation. , These represent the horizontal and vertical coordinates, respectively. It is a blind hole optimization weight. It is a convolution filtering operation; ； in, It is the depth attenuation coefficient, and z is the blind hole depth. It is the depth of the target blind hole.

2. The X-ray-based method for detecting blind via defects in PCB boards according to claim 1, characterized in that, The projection formula for the CT projection is: ； in, For projection data, For spatial ray intensity, The strength after penetrating the PCB.

3. The X-ray-based method for detecting blind via defects in PCB boards according to claim 1, characterized in that, The process of extracting the blind hole region is as follows: Tubular structure detection based on Hessian matrix is ​​employed at each body point. At this point, calculate the second derivative of the image and construct the Hessian matrix: ； Eigenvalue decomposition yields ,in The local tubular structure representing the feature point corresponds to a blind hole. The patchy structure representing the feature points corresponds to impurities. A planar structure representing feature points, corresponding to the hole wall; The probability of each point belonging to a blind hole is calculated using the blind hole similarity response function. The blind hole similarity response function is as follows: ； in, It is the similarity response value. , ; , , , These are the first empirical parameter, the second empirical parameter, and the third empirical parameter.

4. The X-ray-based method for detecting blind via defects in PCB boards according to claim 3, characterized in that, The process of constructing a two-stream neural network architecture based on a deep residual network is as follows: Using 3DResNet as the backbone network, geometric feature branches and material feature branches are constructed. A two-stream feature fusion mechanism is then introduced, and a weighted cross-entropy loss function is used to optimize and adjust the two-stream neural network architecture.

5. The X-ray-based method for detecting blind via defects in PCB boards according to claim 4, characterized in that, The process of constructing geometric feature branches is as follows: The 3D data block of the blind hole region is standardized and used as input. A 3D convolution kernel is used for initial convolution with a stride of 2 to quickly reduce the data size. Simultaneously, basic edge and corner features are extracted. To construct a four-channel curvature feature tensor The curvature feature tensor With the main feature map The splicing formula is as follows: ； ； in, This indicates that the local surface shape at that point resembles a local ellipsoid. This indicates that the local surface shape at that point is a developable surface, similar to a local part of a cylinder or plane. This indicates that the local surface shape at this point resembles a saddle-shaped surface, where K is the Gaussian curvature. The formula for calculating Gaussian curvature is as follows: ； The final output is a feature tensor that integrates general features and precise geometric information. This is used to quantify the curvature of the blind hole surface.

6. The X-ray-based method for detecting blind via defects in PCB boards according to claim 5, characterized in that, The process of constructing material feature branches is as follows: The input is projection data acquired at different transistor voltages. According to the Beer-Lambert law, the attenuation can be expressed as: ； in, It is the linear decay coefficient of the material at energy E, representing the decay capability of an object located at path s at energy E; Energy spectral decomposition transforms this into a linear problem; first, the projected values ​​are calculated: ； For the known C types of materials It can be decomposed into: ； in, It is the density distribution of material c along path s. Since the attenuation coefficient of material c is known, the projection equation becomes: ； For each detector pixel The projection values ​​under multiple energies can be written in matrix form: ； Where M is The material attenuation coefficient matrix, where a is... The material density vector; by solving using the least squares method, the material density map of each base material can be obtained. ; Material density map As input and expanded using dilated convolution, we obtain ； Where d is the expansion rate; Then, multiple convolutional kernels with d=3, 6, and 9 are used in parallel to convolve the features, and the results are then combined: ； This enables the network to capture material regions of different scales simultaneously, whether it's a large area of ​​copper foil or a tiny hole; ultimately, the output of the material stream is a feature tensor containing multi-scale material information. .

7. The X-ray-based method for detecting blind via defects in PCB boards according to claim 6, characterized in that, The dual-stream feature fusion process is as follows: First, the feature map size of the material flow is adjusted to be exactly the same as that of the geometry flow feature map through a 3D interpolation algorithm. Then, the feature maps of the two flows are stitched together through a gated fusion mechanism, so that the network can adaptively choose to trust the geometry flow or the material flow, thereby achieving accurate feature fusion. The splicing formula is as follows: ； in, This indicates a splicing operation. It's the Sigmoid function, which compresses the output to between 0 and 1. and These are learnable parameters, and the features that are ultimately fused. It is generated by weighting the gated graph, and the generation formula is as follows: ； in, This indicates element-wise multiplication.

8. The X-ray-based method for detecting blind via defects in PCB boards according to claim 4, characterized in that, The weighted cross-entropy loss function is: ； in, It is the total loss value. , , These are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. It is cross-entropy loss. It is Dice's loss. It is an L2 regularization term; The formula for calculating cross-entropy loss is as follows: ； in, It's a real label. It is a predicted probability; The formula for calculating Dice loss is as follows: ； Where Y is the true label region and P is the predicted region. This is the number of true positive pixels. This is the actual number of pixels in the target object. It predicts the number of target pixels.

9. The X-ray-based method for detecting blind via defects in PCB boards according to claim 8, characterized in that, The The value range is 0.1-1.

0. The value range is 0.1-1.

0. The value range is 0.0001-0.01.

Images