Method and system for detecting air bubbles in semiconductor adhesive tape based on optical coherence tomography

Abstract

The present application relates to the technical field of bubble detection, in particular to a semiconductor adhesive tape internal bubble detection method and system based on optical coherence tomography. The method comprises: obtaining adhesive tape OCT data and preliminary reconstruction; calculating voxel sub-block local structure entropy, and supplementally scanning high-entropy areas; segmenting signals along depth, independently calculating dispersion compensation coefficient correction, and reconstructing high-precision three-dimensional data; calculating gradient amplitude field and Hessian matrix, combining spherical structure response to generate a saliency map to segment bubble candidate areas; identifying upper and lower surfaces of the candidate areas, and calculating a matching confirmation factor to determine internal bubbles. The scheme of the present application balances detection efficiency and accuracy, overcomes complex dispersion effects, effectively suppresses interference, and significantly reduces the misjudgment rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bubble detection. More particularly, the present application relates to a method and system for detecting bubbles inside semiconductor adhesive tape based on optical coherence tomography. BACKGROUND

[0002] Semiconductor adhesive tapes, such as wafer dicing tapes, backside thinning tapes, etc., are important materials in the process of semiconductor chip manufacturing and packaging. Micro-bubbles often occur inside the semiconductor adhesive tapes due to material, process or environmental factors. As a typical internal defect, bubbles can affect the adhesion uniformity and mechanical properties of the adhesive tape. Therefore, rapid and accurate non-destructive testing of bubbles inside the adhesive tape is a key link to ensure the quality of semiconductor production. Currently, the detection methods for internal defects of materials mainly include X-ray detection and ultrasonic detection. However, these traditional techniques have obvious limitations: X-ray detection equipment is expensive, and the imaging contrast of micro-bubbles in polymer materials is low; while ultrasonic detection usually requires the use of coupling agents, resulting in a complex operation process, and its detection resolution is difficult to meet the precision detection requirements of micron-level bubbles.

[0003] To solve the above problems, the industry has introduced optical coherence tomography (OCT) technology. As a new type of non-contact, high-resolution optical detection technology, OCT technology can realize three-dimensional tomographic imaging of samples, providing a more potential solution for the detection of internal micro-defects of semiconductor adhesive tapes.

[0004] However, when applying OCT technology to semiconductor adhesive tape bubble detection, challenges still exist in data acquisition, image reconstruction and bubble recognition. First, in data acquisition, there is a contradiction between detection efficiency and data quality. Although full-field high-density scanning can obtain detailed structural information, it is time-consuming and difficult to adapt to the rhythm of industrial production. While fast sparse scanning improves efficiency, it is easy to miss bubbles that are small in size or scattered in distribution. Second, in image reconstruction, since semiconductor adhesive tapes are usually multi-layer composite structures, the dispersion effect of the light beam during propagation in the structure will reduce the axial resolution of the OCT image, resulting in blurred image details and affecting the accurate positioning of the bubble profile, especially its upper and lower surfaces. Finally, in bubble recognition and confirmation, the inherent speckle noise of the OCT image, as well as the interference of other impurities, interfaces and other structures inside the adhesive tape, make it difficult to accurately segment bubbles from the complex background and exclude false alarms and missed detections by relying solely on signal intensity threshold segmentation or simple edge detection algorithms. SUMMARY

[0005] The application aims to provide a semiconductor adhesive tape internal bubble detection method and system based on optical coherence tomography, so as to solve the problem of high error rate and poor reliability of the detection result in the prior art.

[0006] In the first aspect, the application provides a semiconductor adhesive tape internal bubble detection method based on optical coherence tomography, comprising the following steps:

[0007] The three-dimensional optical coherence tomography original interference data of the semiconductor adhesive tape are acquired, and preliminary three-dimensional reconstruction is performed to obtain preliminary three-dimensional volume data; the preliminary three-dimensional volume data are divided into a plurality of voxel sub-blocks, and a local structure entropy value is calculated for each voxel sub-block; when the local structure entropy value of any voxel sub-block is greater than a preset entropy threshold value, high-density supplementary scanning is performed on the physical region corresponding to the voxel sub-block, and the supplementary scanning data are merged with the original data; each A-scan signal in the merged original interference data is divided into a plurality of segmented intervals according to the local signal characteristics of the signal along the depth direction; for each segmented interval, a dispersion compensation coefficient is independently calculated based on the second-order and third-order phase terms of the signal; after the signal in each corresponding segmented interval is corrected by using the dispersion compensation coefficient, Fourier transform is performed to reconstruct high-precision three-dimensional volume data; based on the high-precision three-dimensional volume data, a three-dimensional gradient amplitude field and a three-dimensional Hessian matrix of the high-precision three-dimensional volume data are calculated; the spherical structure response degree of each voxel is calculated by using the eigenvalue of the Hessian matrix; the gradient amplitude field and the spherical structure response degree are multiplied point by point at the voxel level to generate a bubble candidate region saliency map, and a plurality of bubble candidate regions are segmented based on the saliency map; for each bubble candidate region, the upper surface and the lower surface of the candidate region are identified along the optical axis propagation direction; a pairing confirmation factor is calculated, and the bubble candidate region is determined as an internal bubble according to the pairing confirmation factor.

[0008] Preferably, the preliminary three-dimensional volume data are divided into a plurality of voxel sub-blocks, and a local structure entropy value is calculated for each voxel sub-block, comprising: the preliminary three-dimensional volume data are divided into a plurality of voxel sub-blocks with a size of 32*32*16 voxels and mutually non-overlapping in space; for each voxel sub-block, the intensity values of all voxels inside the voxel sub-block are counted to generate a gray level histogram with 256 levels; based on the gray level histogram, the Shannon entropy of the sub-block is calculated as the local structure entropy value.

[0009] Preferably, for each A-scan signal in the merged original interferometric data, the signal is divided into multiple segmented intervals according to the local signal characteristics along the depth direction, including: using a sliding window with a width of 64 sampling points to calculate the local average intensity of the A-scan signal along the depth direction; when the gradient of the local average intensity exceeds a preset gradient threshold, the sampling point is recorded as a segmentation point; and the A-scan signal is divided into multiple segmented intervals using all segmentation points.

[0010] Preferably, for each segmented interval, a dispersion compensation coefficient is independently calculated based on the second-order and third-order phase terms of the signal, including: the optimized search range for the second-order dispersion coefficient is [-4000, 4000]. The optimization search range for the third-order dispersion coefficient is [-25000, 25000]. Within a two-dimensional parameter space, a particle swarm optimization algorithm is employed, with the objective function being to maximize the sum of squared intensities of the reconstructed image within the segmented intervals. The optimal combination of second- and third-order dispersion coefficients is iteratively optimized to obtain the dispersion compensation coefficients. It is a femtosecond.

[0011] Preferably, the step of calculating the spherical structure responsivity of each voxel using the eigenvalues ​​of the Hessian matrix includes: for each voxel, calculating the three eigenvalues ​​of the Hessian matrix. And satisfy ;when When all values ​​are greater than a preset positive threshold, the voxel is determined to be a candidate point for a dark-colored spherical structure, and the spherical structure responsivity... Calculate using the following formula: In other cases, the spherical structure response... It is 0.

[0012] Preferably, the step of multiplying the gradient magnitude field with the spherical structure responsivity at the voxel level to generate a bubble candidate region saliency map, and segmenting multiple bubble candidate regions based on the saliency map, includes: multiplying the three-dimensional gradient magnitude of each voxel with the spherical structure responsivity of the voxel to obtain a saliency value; performing binarization on the generated three-dimensional saliency map using a fixed threshold of 25% of the global maximum saliency value; and using a three-dimensional connected component analysis algorithm to extract each independent connected region as a bubble candidate region.

[0013] Preferably, identifying the upper and lower surfaces of each bubble candidate region along the optical axis propagation direction includes: within the bubble candidate region, for each lateral position... Along the direction of optical propagation, i.e. the z-axis, find the first local maximum point of the gradient magnitude field as the upper surface point, and find the last local maximum point as the lower surface point; gather all the upper surface points in the lateral position to form the upper surface, and gather all the lower surface points to form the lower surface.

[0014] Preferably, calculating a pairing confirmation factor includes calculating the following three indicators: at the centroid of the bubble candidate region, calculating the dot product of the upper surface normal vector and the lower surface normal vector. ; Calculate the average axial spacing between all paired upper and lower surface points within the bubble candidate region. ; Calculate the area 30 directly below the lower surface Optical attenuation coefficient of signal strength within a depth range of m .

[0015] Preferably, the step of determining the bubble candidate region as an internal bubble based on the pairing confirmation factor includes: when the condition is met... , ,and If three conditions are met, the pairing confirmation factor is assigned a value of 1; otherwise, it is assigned a value of 0. The dot product of the normal vectors of the upper and lower surfaces is given. The average axial distance between all paired upper and lower surface points within the bubble candidate region. This is the optical attenuation coefficient.

[0016] In the second aspect, a semiconductor tape internal bubble detection system based on optical coherence tomography includes:

[0017] The processor; the memory storing computer instructions for detecting internal bubbles in semiconductor tape based on optical coherence tomography, which, when executed by the processor, cause the system to perform the aforementioned method for detecting internal bubbles in semiconductor tape based on optical coherence tomography.

[0018] The beneficial effects of this invention are as follows: First, by employing a supplementary scanning strategy based on local structural entropy, detailed investigation of suspicious areas is achieved, significantly improving the overall efficiency of data acquisition without sacrificing the detection accuracy of key areas, thus resolving the contradiction between detection speed and data quality. Second, by segmenting the signal along the depth direction and independently performing dispersion compensation, the complex dispersion effect caused by the multi-layer structure of the tape is effectively overcome, resulting in higher axial resolution and laying the foundation for accurate bubble contour identification. Third, by fusing gradient edge information with spherical morphology information based on the Hessian matrix to segment candidate targets, the unique three-dimensional morphology of the bubble can be highlighted more accurately, effectively suppressing interference from background noise and interlayer interfaces. Finally, by establishing a confirmation factor that integrates geometric pairing relationships and optical attenuation characteristics for final judgment, the candidate targets undergo rigorous physical characteristic verification, greatly reducing the false judgment rate and ensuring the final reliability and stability of the detection results. Attached Figure Description

[0019] Figure 1 The flowchart illustrating the steps of the method for detecting internal bubbles in semiconductor tape based on optical coherence tomography in this embodiment is shown in the schematic diagram.

[0020] Figure 2 The schematic diagram illustrates the structural block diagram of the semiconductor tape internal bubble detection system based on optical coherence tomography in this embodiment. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] like Figure 1 As shown, the method for detecting internal bubbles in semiconductor tape based on optical coherence tomography in this embodiment includes the following steps:

[0023] Step S1: Obtain the original three-dimensional optical coherence tomography interference data of the semiconductor tape and perform preliminary three-dimensional reconstruction to obtain preliminary three-dimensional volume data; divide the preliminary three-dimensional volume data into multiple voxel sub-blocks, and calculate a local structural entropy value for each voxel sub-block; when the local structural entropy value of any voxel sub-block is greater than a preset entropy threshold, perform a high-density supplementary scan on the physical region corresponding to the voxel sub-block, and merge the supplementary scan data with the original data.

[0024] Specifically, a swept-frequency optical coherence tomography system is used to perform an initial sparse scan on the semiconductor tape sample. This process involves controlling a galvanometer to achieve rapid scanning along the X-axis to form a B-scan, followed by a line-by-line scan along the Y-axis using a motor platform to form a C-scan, thereby obtaining three-dimensional raw interferometric data covering the entire test area. This raw data is then reconstructed into preliminary three-dimensional volumetric data after standard k-space resampling, fast Fourier transform, and global dispersion compensation. To locate potential defects, the aforementioned three-dimensional volumetric data is spatially divided into multiple non-overlapping voxel sub-blocks (e.g., 32×32×32 voxels).

[0025] Subsequently, for each sub-block, the local structural entropy value is obtained by calculating the information entropy of each sub-block based on its internal gray-level histogram. This value quantifies the signal complexity of the local region. For example, sub-blocks containing bubble boundaries have high entropy values ​​due to their drastic gray-level changes. An empirical entropy threshold is set. When the calculated local structural entropy value exceeds the threshold, the spatial coordinates of the sub-block are recorded, and the scanning system is driven back to the physical region corresponding to those coordinates. In this physical region, the scan line spacing of the B-scan and the step distance of the C-scan are both reduced to half of their original values ​​to perform high-density scanning. Finally, the newly acquired high-density raw interferometric data replaces the sparse data corresponding to the physical region in the original data, forming a locally high-precision merged dataset for subsequent analysis.

[0026] In an optional embodiment, the preliminary three-dimensional volume data is divided into multiple voxel sub-blocks, and a local structural entropy value is calculated for each voxel sub-block, including:

[0027] The preliminary three-dimensional volume data is spatially divided into multiple non-overlapping voxel sub-blocks with a size of 32×32×16 voxels. For each voxel sub-block, the intensity values ​​of all voxels within it are counted to generate a grayscale histogram with 256 levels. Based on the grayscale histogram, the Shannon entropy of the sub-block is calculated as the local structural entropy value.

[0028] Specifically, the entire preliminary 3D volume data (e.g., a total size of 512×512×256 voxels) is meshed and divided into multiple independent cuboid sub-blocks of size 32×32×16 voxels along the x, y, and z axes. For any sub-block, all voxels contained within it are traversed, the intensity value of each voxel is read, and the frequency of each intensity value at each gray level from 0 to 255 is counted to construct a gray-level histogram for that sub-block. Finally, based on the probability distribution of this gray-level histogram, the Shannon entropy formula is used to calculate the local structural entropy value of the aforementioned voxel sub-block, which is used to quantify the complexity and uncertainty of the signal within that local region.

[0029] Step S2: For each A-scan signal in the merged original interferometric data, divide the signal into multiple segment intervals according to the local signal characteristics along the depth direction; for each segment interval, calculate a dispersion compensation coefficient independently based on the second and third phase terms of the signal; after correcting the signal in each corresponding segment interval using the dispersion compensation coefficient, perform Fourier transform to reconstruct high-precision three-dimensional volume data.

[0030] Specifically, for each A-scan interference signal in the merged data, its signal envelope is analyzed along the depth direction (wavenumber k). By identifying peak points, the upper and lower surfaces of the tape, as well as strong reflective interfaces such as the internal bubble interface, are located. These strong reflective interfaces divide the entire A-scan signal into multiple segmented intervals. For example, the interval from air to the upper surface of the tape is the first interval, and the interval from the upper surface of the tape to the upper surface of the bubble is the second interval.

[0031] Subsequently, for each segmented interval, the strongest reflected signal peak within that interval is selected, and a short-time Fourier transform is applied to it to extract the phase information that varies with wavenumber k. This phase information is then fitted to a third-order polynomial in k, where the coefficients of the quadratic and cubic terms are the second- and third-order dispersion compensation coefficients specific to each segmented interval.

[0032] Then, a phase correction function is constructed using the aforementioned dispersion compensation coefficients, and multiplied in the frequency domain by the original interference signal corresponding to the segmented interval, thereby canceling the dispersion effect accumulated by the light along the propagation path.

[0033] Finally, after performing the above independent correction operations on all segments of an A-scan, the corrected segments are stitched together into a complete A-scan signal. A Fast Fourier Transform is then performed on the entire stitched A-scan signal to obtain a depth profile with high axial resolution. This process is repeated for all A-scans to reconstruct high-precision 3D volume data.

[0034] In an optional embodiment, each A-scan signal in the merged original interferometric data is divided into multiple segmented intervals based on the local signal characteristics of the signal along the depth direction, including:

[0035] A sliding window with a width of 64 sampling points is used to calculate the local average intensity of the A-scan signal along the depth direction; when the gradient of the local average intensity exceeds a preset gradient threshold, the sampling point is recorded as a segmentation point; the A-scan signal is divided into multiple segmented intervals using all segmentation points.

[0036] Specifically, an A-scan signal (a one-dimensional signal sequence along the depth direction) is smoothed. A sliding window with a width of 64 data points is used, moving point by point along the depth direction from the signal's starting point. At each window position, the average intensity value of all data points within the window is calculated. This process continues until the window has traversed the entire signal sequence, ultimately generating a local average intensity curve that reflects the signal's macroscopic trend. Subsequently, the gradient at each point on the local average intensity curve is calculated, i.e., the rate of change of intensity between that point and its neighboring points is calculated. The absolute value of the gradient at each point is checked to see if it exceeds a preset gradient threshold. If the gradient value at a certain depth z is greater than the preset gradient threshold, that depth z is marked as a segmentation point. After traversing the entire A-scan signal and recording all segmentation points that meet the conditions, these points are used as dividing boundaries, precisely dividing the original A-scan signal into several continuous segments with distinct signal characteristics for subsequent processing.

[0037] In an optional embodiment, for each segmented interval, a dispersion compensation coefficient is independently calculated based on the second-order and third-order phase terms of the signal, including:

[0038] The optimal search range for the second-order dispersion coefficient is [-4000, 4000]. The optimization search range for the third-order dispersion coefficient is [-25000, 25000]. Within a two-dimensional parameter space, a particle swarm optimization algorithm is employed, with the objective function being to maximize the sum of squared intensities of the reconstructed image within the segmented intervals. The optimal combination of second- and third-order dispersion coefficients is iteratively optimized to obtain the dispersion compensation coefficients. It is a femtosecond.

[0039] Specifically, a two-dimensional parameter search space is constructed for any segmented interval from the A-scan signal. The horizontal axis of the search space represents the second-order dispersion coefficient, with a value range of [-4000, 4000]. The vertical axis represents the third-order dispersion coefficient, with a value range of [-25000, 25000]. .

[0040] Within the aforementioned two-dimensional parameter search space, a group of particles is randomly initialized, each representing a set of candidate combinations of second- and third-order dispersion coefficients. For each coefficient combination represented by a particle, a trial dispersion compensation is performed, and the corresponding image fragment is reconstructed. The sum of squares of the intensities of all pixels in the image fragment is calculated, and the calculated sum of squares is used as the fitness function value to evaluate the quality of the coefficient combination. A higher fitness value indicates a better compensation effect and a clearer image.

[0041] In each iteration, each particle updates its flight speed and position based on its own and other particles' historical best positions. This process causes the entire particle swarm to gradually converge towards the region that maximizes the sum of squared image intensities. When the iteration meets the termination condition, the algorithm outputs the combination of dispersion coefficients corresponding to the global historical best position (e.g., second-order coefficients are 2100, and third-order coefficients are -12000). This combination of dispersion coefficients is then determined as the final dispersion compensation coefficient for the current segment interval.

[0042] Step S3: Based on the high-precision three-dimensional volume data, calculate the three-dimensional gradient magnitude field and the three-dimensional Hessian matrix of the high-precision three-dimensional volume data; use the eigenvalues ​​of the Hessian matrix to calculate the spherical structure responsivity of each voxel; multiply the gradient magnitude field and the spherical structure responsivity point by point at the voxel level to generate a bubble candidate region saliency map, and segment multiple bubble candidate regions based on the saliency map.

[0043] Specifically, for each voxel in the reconstructed high-precision 3D volume data, the partial derivatives of the voxel in the X, Y, and Z directions are calculated using the 3D Sobel operator, and the gradient magnitude of the voxel is calculated accordingly. This process is repeated for all voxels to obtain the 3D gradient magnitude field. Simultaneously, for each voxel, a 3×3 3D Hessian matrix is ​​constructed by calculating the second-order partial derivatives of the voxel with its neighboring voxels, and the three eigenvalues ​​of the Hessian matrix are calculated. Bubbles appear as dark regions in OCT images, representing local minima in the 3D intensity distribution. Therefore, as a spherical dark structure, its voxel's three Hessian eigenvalues ​​should theoretically all be positive. To identify this structure, a spherical structure responsivity function can be constructed; for example, when all three eigenvalues ​​are positive, the responsivity is the product of the three eigenvalues; otherwise, the responsivity is zero. The calculated gradient magnitude field is multiplied point-by-point with the spherical structure response field to obtain a saliency map. In this saliency map, only structures possessing both strong edge characteristics and a spherical shape are enhanced. A global threshold is applied to the saliency map for binarization, and the segmented connected regions are the bubble candidate regions.

[0044] In an optional embodiment, the spherical structure responsivity of each voxel is calculated using the eigenvalues ​​of the Hessian matrix, including:

[0045] For each voxel, calculate the three eigenvalues ​​of the Hessian matrix. And satisfy ;when When all values ​​are greater than a preset positive threshold, the voxel is determined to be a candidate point for a dark-colored spherical structure, and the spherical structure responsivity... Calculate using the following formula:

[0046] ,

[0047] In other cases, the spherical structure response It is 0.

[0048] Specifically, in optical coherence tomography (OCT) three-dimensional volumetric data, solid semiconductor tapes generate strong signals in OCT images due to their scattering properties, corresponding to high values; conversely, the interior of a bubble contains almost no scattering material, resulting in extremely weak signals, corresponding to low values. Therefore, the center of the bubble represents a local minimum in the signal intensity distribution. Mathematically, a local minimum has positive eigenvalues ​​in all eigenvalues ​​of the Hessian matrix.

[0049] For each voxel in the 3D volumetric data, its Hessian matrix is ​​constructed by calculating its second-order partial derivatives with neighboring voxels, and then its three eigenvalues ​​are solved. These eigenvalues ​​reflect the second-order rate of change of image intensity within the voxel's neighborhood. For ease of analysis, the three eigenvalues ​​are sorted by their absolute values ​​after solving, resulting in... , , ,in The absolute value is the largest. A filtering condition is set to check if all three feature values ​​are greater than a preset minimum positive threshold (e.g., 1e-5). This condition is used to filter voxels that exhibit local minima in all directions; these voxels appear as dark, spherical, or speckled structures in the image. If a voxel's feature value, for example... If the value is not greater than 1e-5, it is not considered a candidate point for a dark sphere.

[0050] Only when a voxel passes the above screening criteria will its spherical structure responsivity be calculated. For example, if the eigenvalue of a voxel is... =0.8, =0.7, =0.6, all greater than 1e-5, then the voxel responsiveness Substitute the values ​​into the above formula for calculation. For voxels that fail the screening, their voxel responsiveness... The value is assigned to 0, thus effectively suppressing it in subsequent analysis.

[0051] In an optional embodiment, the gradient magnitude field is multiplied point-by-point at the voxel level with the responsivity of the spherical structure to generate a saliency map of bubble candidate regions, and multiple bubble candidate regions are segmented based on the saliency map, including:

[0052] The saliency value is obtained by multiplying the three-dimensional gradient magnitude of each voxel by the spherical structure responsivity of the voxel; the generated three-dimensional saliency map is binarized by applying a fixed threshold of 25% of the global maximum saliency value; and a three-dimensional connected component analysis algorithm is used to extract each independent connected region as a bubble candidate region.

[0053] Specifically, the algorithm aligns the 3D gradient magnitude map with the 3D spherical structure responsivity map. At each identical 3D coordinate location, both the gradient magnitude and the spherical structure responsivity are simultaneously acquired. The saliency value for that point is obtained by multiplying the gradient magnitude by the spherical structure responsivity. This process is repeated for each voxel in the 3D volume data to generate a complete 3D saliency map, where the value of each voxel represents its overall probability of being a bubble candidate point.

[0054] To transform the saliency map into a binary image that facilitates segmentation, an adaptive global thresholding method is employed. First, the entire 3D saliency map is traversed to determine its global maximum saliency value. Then, a preset percentage (e.g., 25%) of this maximum saliency value is used as the global threshold. The saliency map is traversed again, marking all voxels with values ​​greater than 30 as 1 and all voxels with values ​​less than or equal to 30 as 0, thereby transforming the saliency map into a binary 3D mask.

[0055] The binarized 3D mask generated in the previous step is applied to a 3D connected component analysis algorithm to identify all independent regions formed by interconnected voxels marked as 1. Each of these independent connected regions is treated as a potential bubble and segmented, each marked as an independent bubble candidate region for subsequent verification.

[0056] Step S4: For each bubble candidate region, identify the upper and lower surfaces of the candidate region along the optical axis propagation direction; calculate a pairing confirmation factor, and determine the bubble candidate region as an internal bubble based on the pairing confirmation factor.

[0057] Specifically, within the bubble candidate region, for each lateral position Along the direction of optical propagation, i.e. the z-axis, find the first local maximum point of the gradient magnitude field as the upper surface point, and find the last local maximum point as the lower surface point; gather all the upper surface points in the lateral position to form the upper surface, and gather all the lower surface points to form the lower surface.

[0058] For example, firstly, the algorithm identifies a segmented candidate bubble region and then selects the region covered by the bubble. Starting from the first horizontal coordinate point on the plane, for example... , Examine the gradient magnitudes of all voxels belonging to the candidate region along the vertical downward direction (i.e., the z-axis direction) from the coordinate points mentioned above.

[0059] In the column In the gradient value sequence along the axis, find all local maxima, i.e., points where the gradient value is greater than the gradient values ​​of the two immediately above and below voxels. Assuming local maxima are found at z=10, z=25, and z=40, select... The point with the smallest coordinates, i.e., the point at z=10, is taken as the horizontal position. , The point on the upper surface is selected. Simultaneously, the point with the largest z-coordinate, i.e., z=40, is selected as the point on the lower surface.

[0060] The above process will cover each lateral position of the bubble candidate region. Repeat this process to find a pair of upper and lower surface points for each lateral position. Once all lateral positions have been processed, combine all the found upper surface points to form the complete upper surface of the bubble; similarly, combine all the lower surface points to form the complete lower surface.

[0061] The calculation of the pair confirmation factor includes the following three indicators:

[0062] At the centroid of the bubble candidate region, calculate the dot product of the upper surface normal vector and the lower surface normal vector. ; Calculate the average axial spacing between all paired upper and lower surface points within the bubble candidate region. ; Calculate the area 30 directly below the lower surface Optical attenuation coefficient of signal strength within a depth range of m .

[0063] The step of determining the candidate bubble region as an internal bubble based on the pairing confirmation factor includes:

[0064] When satisfied , ,and If all three conditions are met, the pairing confirmation factor is assigned a value of 1; otherwise, it is assigned a value of 0.

[0065] in, The dot product of the normal vectors of the upper and lower surfaces is given. The average axial distance between all paired upper and lower surface points within the bubble candidate region. This is the optical attenuation coefficient.

[0066] Specifically, three key metrics are calculated for each bubble candidate region. At the geometric center of the region, the normal vectors of the upper and lower surfaces are calculated respectively, and then obtained through vector dot product operation. This value reflects whether the two surfaces are approximately parallel and in opposite directions. For each pair of upper and lower surface points at a lateral position within the region, the distance between the pair of upper and lower surface points on the z-axis is calculated, and all these distances are averaged to obtain the average axial spacing. Within a depth of 30 μm immediately below the lower surface, the intensity attenuation of the OCT signal was analyzed, and the optical attenuation coefficient was calculated. , used to represent the shadow effect caused by bubbles.

[0067] Then, a multi-condition joint judgment is performed to check whether the three calculated indicators simultaneously meet the preset physical constraints. For example, for a candidate region, if the calculated region... =-0.9, =35μm, μ=0.01. , Since the value of μ meets the preset conditions, the candidate region has passed the verification.

[0068] Based on the judgment results, a pairing confirmation factor is assigned to the candidate bubble region. For the candidate regions that pass the verification, the pairing confirmation factor is set to 1, indicating that it is confirmed as a real bubble. If any condition is not met, for example, another candidate region... The calculated value is 5μm, which does not meet the requirement of being greater than 10μm. Therefore, its pairing confirmation factor is set to 0, and the candidate region is excluded.

[0069] This invention also provides a system for detecting internal bubbles in semiconductor tapes based on optical coherence tomography. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the above-described method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to the present invention.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0071] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0072] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0073] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for detecting internal bubbles in semiconductor tape based on optical coherence tomography, characterized in that, Includes the following steps: The raw three-dimensional optical coherence tomography (OCT) data of the semiconductor tape is acquired and preliminary three-dimensional reconstruction is performed to obtain preliminary three-dimensional volume data. The preliminary three-dimensional volume data is divided into multiple voxel sub-blocks, and a local structural entropy value is calculated for each voxel sub-block. When the local structural entropy value of any voxel sub-block is greater than a preset entropy threshold, a high-density supplementary scan is performed on the physical region corresponding to the voxel sub-block, and the supplementary scan data is merged with the original data. For each A-scan signal in the merged original interferometric data, the signal is divided into multiple segmented intervals based on the local signal characteristics along the depth direction; For each segmented interval, a dispersion compensation coefficient is independently calculated based on the second-order and third-order phase terms of the signal, including: The optimal search range for the second-order dispersion coefficient is [-4000, 4000]. The optimization search range for the third-order dispersion coefficient is [-25000, 25000]. Within a two-dimensional parameter space, a particle swarm optimization algorithm is employed, with the objective function being to maximize the sum of squared intensities of the reconstructed image within the segmented intervals. The optimal combination of second- and third-order dispersion coefficients is iteratively optimized to obtain the dispersion compensation coefficients. For femtosecond; After correcting the signals in each corresponding segmented interval using the dispersion compensation coefficient, Fourier transform is then performed to reconstruct high-precision three-dimensional volume data. Based on the high-precision three-dimensional volume data, the three-dimensional gradient magnitude field and the three-dimensional Hessian matrix of the high-precision three-dimensional volume data are calculated; the spherical structure responsivity of each voxel is calculated using the eigenvalues ​​of the Hessian matrix; the gradient magnitude field and the spherical structure responsivity are multiplied point-by-point at the voxel level to generate a bubble candidate region saliency map, and multiple bubble candidate regions are segmented based on the saliency map; For each bubble candidate region, the upper and lower surfaces of the candidate region are identified along the optical axis propagation direction; a pairing confirmation factor is calculated, and the bubble candidate region is determined to be an internal bubble based on the pairing confirmation factor.

2. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, The step of dividing the preliminary three-dimensional volume data into multiple voxel sub-blocks and calculating a local structural entropy value for each voxel sub-block includes: The preliminary three-dimensional volume data is spatially divided into multiple non-overlapping voxel sub-blocks with a size of 32×32×16 voxels. For each voxel sub-block, the intensity values ​​of all voxels within it are counted to generate a grayscale histogram with 256 levels. Based on the grayscale histogram, the Shannon entropy of the sub-block is calculated as the local structural entropy value.

3. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, Each A-scan signal in the merged original interferometric data is divided into multiple segmented intervals based on the local signal characteristics along the depth direction, including: A sliding window with a width of 64 sampling points is used to calculate the local average intensity of the A-scan signal along the depth direction; when the gradient of the local average intensity exceeds a preset gradient threshold, the sampling point is recorded as a segmentation point; the A-scan signal is divided into multiple segmented intervals using all segmentation points.

4. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, The calculation of the spherical structure responsivity of each voxel using the eigenvalues ​​of the Hessian matrix includes: For each voxel, calculate the three eigenvalues ​​of the Hessian matrix. And satisfy ;when When all values ​​are greater than a preset positive threshold, the voxel is determined to be a candidate point for a dark-colored spherical structure, and the spherical structure responsivity... Calculate using the following formula: ， In other cases, the spherical structure response It is 0.

5. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, The process involves performing a voxel-level pointwise multiplication of the gradient magnitude field with the responsivity of the spherical structure to generate a saliency map of bubble candidate regions, and segmenting multiple bubble candidate regions based on the saliency map, including: The saliency value is obtained by multiplying the three-dimensional gradient magnitude of each voxel by the spherical structure responsivity of the voxel; the generated three-dimensional saliency map is binarized by applying a fixed threshold of 25% of the global maximum saliency value; and a three-dimensional connected component analysis algorithm is used to extract each independent connected region as a bubble candidate region.

6. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, For each bubble candidate region, identifying the upper and lower surfaces of the candidate region along the optical axis propagation direction includes: Within the bubble candidate region, for each lateral position Along the direction of optical propagation, i.e. the z-axis, find the first local maximum point of the gradient magnitude field as the upper surface point, and find the last local maximum point as the lower surface point. The upper surface is formed by gathering all the points on the upper surface in the horizontal direction, and the lower surface is formed by gathering all the points on the lower surface.

7. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 1, characterized in that, The calculation of a pair confirmation factor includes calculating the following three indicators: At the centroid of the bubble candidate region, calculate the dot product of the upper surface normal vector and the lower surface normal vector. ; Calculate the average axial spacing between all paired upper and lower surface points within the bubble candidate region. ; Calculate 30 directly below the lower surface Optical attenuation coefficient of signal strength within a depth range of m .

8. The method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to claim 7, characterized in that, The step of determining the bubble candidate region as an internal bubble based on the pairing confirmation factor includes: When satisfied , ,and If all three conditions are met, the pairing confirmation factor is assigned a value of 1; otherwise, it is assigned a value of 0. in, The dot product of the normal vectors of the upper and lower surfaces is given. The average axial distance between all paired upper and lower surface points within the bubble candidate region. This is the optical attenuation coefficient.

9. A system for detecting internal bubbles in semiconductor tape based on optical coherence tomography, characterized in that, include: processor; A memory storing computer instructions for detecting internal bubbles in semiconductor tape based on optical coherence tomography, wherein when the computer instructions are executed by the processor, the system performs the method for detecting internal bubbles in semiconductor tape based on optical coherence tomography according to any one of claims 1-8.

Images