A non-destructive method for focal drift correction in micro-CT combining temporal optimization and Fourier domain translation
By employing a micro-CT focus drift correction method based on temporal optimization and Fourier domain shift, the imaging error problem caused by X-ray focus drift was solved, achieving sub-micron level high-precision detection and improving the signal-to-noise ratio and structural clarity of the micro-CT imaging system.
Patent Information
- Application Number
- CN202611377512.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-07
AI Technical Summary
In micro-CT imaging systems, X-ray focus drift causes inter-frame misalignment of projection sequences, resulting in artifacts in reconstructed slices, edge distortion, and blurring of microscopic details. Traditional focus drift correction techniques cannot meet the requirements for sub-micron level high-precision detection.
A non-destructive correction method for focus drift in micro-CT is proposed, which combines temporal optimization and Fourier domain translation. A low-drift reference projection sequence is established through temporal optimization, and sub-pixel-level accurate correction is performed by feature point matching and Fourier domain translation to construct a high-precision three-dimensional slice model.
It effectively suppresses noise and artifacts introduced by focus drift, improves the signal-to-noise ratio and structural clarity, and meets the sub-micron-level high-precision detection requirements of precision devices such as semiconductor chips.
Smart Images

Figure CN122866271B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computed tomography imaging technology, and relates to a time-optimized micro-CT imaging system and a focus drift correction method. Background Technology
[0002] Microfocus X-ray micro-CT, with its advantages of non-destructive penetration imaging, three-dimensional reconstruction, and layer-by-layer analysis, can achieve sub-micron ultra-high resolution imaging, making it a mainstream core technology for detecting micro- and nano-defects in semiconductor devices. In practical engineering applications, long exposure and high-power imaging conditions are commonly used to improve the imaging signal-to-noise ratio and highlight the details of micro-defects. However, long-term high-load operation of the X-ray source easily leads to continuous heat accumulation and structural thermal deformation. Coupled with interference from cooling system vibration, environmental temperature drift, and mechanical micro-disturbances, this causes micron-level dynamic drift of the X-ray focus, forming an inherent technical contradiction: "high-precision imaging depends on high-load conditions, high load induces focus drift, drift introduces imaging errors, and the upper limit of accuracy is limited." This severely restricts the imaging accuracy and stability of high-end chip detection.
[0003] The main causes of X-ray source focus drift can be divided into two categories: First, the continuous bombardment of the anode target by the electron beam generates a large amount of heat, which causes the target sheet and target base to expand and deform slightly, resulting in a shift in the physical position of the focus. Second, long-term operation causes temperature drift of electron optical components such as cathode, focusing coil, and deflection coil, which induces distortion of electric field and magnetic field, causing the electron beam transmission trajectory to deviate, indirectly causing focus drift.
[0004] Currently, the industry mainly relies on physical-level optimization solutions to suppress focus drift, covering areas such as target-end thermal management, X-ray tube structure optimization, and upgrades to electro-optics and power supply systems. However, these solutions cannot completely eliminate instantaneous dynamic drift and cumulative drift during the scanning process. In sub-micron precision inspection scenarios, residual drift errors can disrupt the system's imaging geometric calibration relationship, causing inter-frame projection offsets. This results in artifacts such as blurred 3D reconstructed slices, ghosting, ring-shaped stripes, and uneven grayscale, significantly reducing the image signal-to-noise ratio and defect identification accuracy. This can easily lead to missed or false detections of micro- and nano-defects, making it unsuitable for the inspection needs of high-end semiconductor manufacturing processes.
[0005] Existing drift correction techniques generally suffer from design flaws, making it difficult to meet the requirements of ultra-high precision imaging. Patent CN109685877B employs a correction sequence of slow scanning followed by fast scanning. During the long slow scanning process, the focus continuously accumulates drift, and the subsequent fast scanning reference image cannot characterize the dynamic drift features throughout the slow scanning process, resulting in an inherent systematic reference mismatch error. Patent CN117075177A uses a real-time correction scheme for the center point of each frame of the image, which highly depends on the clear geometric texture features of the sample. In the detection scenarios of workpieces with complex microstructures and missing features, the drift solution accuracy drops significantly.
[0006] In addition, the mainstream phase correlation method combined with spatial domain affine transformation registration scheme has inherent technical bottlenecks: the phase correlation method is susceptible to motion blur, low signal-to-noise ratio, and spectral leakage, and has low sub-pixel offset solution accuracy, poor adaptability, and is prone to matching failure; the spatial domain affine transformation relies on pixel interpolation calculation, which has inherent truncation error and high-frequency aliasing problem, and the registration accuracy has a theoretical upper limit, which cannot meet the sub-micron level high-precision correction requirements for semiconductor micro-nano defect detection. Summary of the Invention
[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0008] The primary technical problem this invention aims to solve is that, under long-term imaging conditions in micro-CT, the X-ray source is prone to continuous thermal drift, causing inter-frame misalignment in the projection sequence, leading to artifacts in the reconstructed slices, edge distortion, and blurring of microscopic details. Simultaneously, traditional focus drift correction techniques suffer from drawbacks such as unreasonable scanning timing, large reference matching deviations, weak algorithm robustness, limited sub-pixel registration accuracy, and the easy introduction of additional noise during correction, failing to meet the sub-micron level high-precision defect detection requirements of precision devices such as chips.
[0009] Therefore, to solve the above-mentioned technical problems, the present invention provides the following technical solution: a non-destructive correction method for focus drift in micro-CT combining time-series optimization and Fourier domain translation, the correction method comprising the following steps: The micro-CT imaging system is initialized and its geometric parameters are calibrated before the sample to be tested is pre-scanned. A full-view low-drift scan was performed using preset parameters, and the scan parameters were determined after validity verification and parameter optimization. Based on the scanning parameters, the first and second scans are performed at a fixed timing sequence to establish a low-drift reference projection sequence and a second projection sequence containing focus drift error, respectively. Using the projection viewpoint of the low-drift reference projection sequence as an index, extract the second projection image with the same viewpoint from the second projection sequence at a fixed step size to construct the matching dataset corresponding to the viewpoint. Perform normalization preprocessing on the low-drift reference projection sequence and the matching dataset, extract high-contrast feature points and perform feature matching, and filter to obtain effective feature matching pairs; The initial offset is solved based on effective feature matching pairs, and the sub-pixel offsets in the X and Y directions of the second projected image are obtained through refined calculation. Registration is completed based on Fourier domain translation and the single registration residual is calculated. A closed-loop iterative registration and error verification mechanism is introduced to iteratively optimize the sub-pixel offset of the second projection image from each viewpoint. Based on the qualified sub-pixel offset obtained by verification, the continuous drift under the full projection view is obtained by bicubic spline interpolation. Based on the full-view drift, the second projection sequence is registered and corrected frame by frame to construct a complete projection image set after focus drift compensation. Three-dimensional reconstruction is performed based on the complete projected image set after focus drift compensation and geometric parameters to obtain a corrected three-dimensional slice model. This model is then compared with the uncorrected reconstructed model, and the optimization degree and effectiveness of the correction method are evaluated by quantifying image quality indicators.
[0010] In a preferred embodiment of the present invention, the determination of scanning parameters through validity verification and parameter optimization includes calculating the relative offset between the first and last frame projected images and comparing it with a preset stability threshold; if the relative offset is less than the preset stability threshold, the current acquisition is determined to be valid; subsequently, a low-drift scan is performed again with the optimized scanning parameters and verification is implemented to confirm that the focus state remains stable after parameter adjustment, thereby determining the optimal first scanning parameters and its corresponding scanning time; if the relative offset exceeds the preset stability threshold, the current acquisition is determined to be invalid, the pre-exposure time of the X-ray source is extended, and after the micro-CT imaging system is further thermally stabilized, a low-drift scan is performed again to acquire and generate a new low-drift reference projection sequence, and the above verification process is repeated until the focus drift meets the stability condition, thereby determining the optimal first scanning parameters and its corresponding scanning time.
[0011] In a preferred embodiment of the present invention, the step of performing a first scan and a second scan in a fixed timing sequence based on scanning parameters to establish a low-drift reference projection sequence and a second projection sequence containing focus drift error, respectively, includes confirming that the micro-CT imaging system meets the low-drift scanning conditions, and then performing scanning steps in a fixed timing sequence: precisely resetting the sample to be tested to a preset starting angle to ensure that the initial spatial pose of the two scans is consistent and that the sample to be tested can rotate completely twice; performing a first scan with the first scanning parameters to obtain a low-drift reference projection sequence; keeping the order unchanged, and under the premise of maintaining the same scanning starting position and angle coverage, performing a second scan with the second scanning parameters to obtain a second projection sequence containing focus drift error; wherein, the angle sampling step size set by the second scanning parameters is smaller than that of the first scanning parameters; the order of the first scan and the second scan is a predetermined fixed timing sequence.
[0012] In a preferred embodiment of the present invention, the standardization preprocessing of the low-drift reference projection sequence and the matching dataset includes: first, performing global gray-level normalization to linearly map the image gray-level value range to a unified interval, thereby unifying the gray-level reference of each projection image; then, using a contrast-limited adaptive histogram equalization algorithm to enhance the local microstructure information of the sample under test; next, using bilateral filtering to suppress random noise by combining spatial neighborhood and gray-level similarity as dual weights; and finally, performing local zero-mean standardization correction on the projection image.
[0013] As a preferred embodiment of the present invention, the screening to obtain effective feature matching pairs includes: in the selected area of the projected image, using the Sobel operator to solve the horizontal and vertical gray-level gradients and magnitudes pixel by pixel, delineating the effective detection area based on the structural areas with significant and stable gradients in the boundary, defects and interface of the sample to be tested, and eliminating background and low signal-to-noise ratio interference. Within the effective detection area, the scale-invariant feature transformation algorithm is used to locate high-contrast feature points as candidate key points, and local neighborhood information is extracted for the candidate key points to generate corresponding high-dimensional descriptors. Based on the Euclidean distance metric of high-dimensional descriptors, initial matching of feature points is performed to select candidate matching pairs; then, the ratio of the nearest neighbor to the second nearest neighbor distance is introduced as a constraint to eliminate ambiguous mismatches and obtain high-confidence effective feature matching pairs.
[0014] As a preferred embodiment of the present invention, the process of filtering to obtain effective feature matching pairs is further provided with an adaptive feature point quantity control mechanism, including: real-time statistics of the total number of effective feature points in the selected area of the projected image, and determination of whether the current number of feature points meets the requirements of high-precision matching operation. If the current number of feature points is determined to be insufficient, the Lowe threshold is adaptively reduced to relax the constraints of feature selection in order to extract more stable and effective feature points. A non-maximum suppression algorithm is also introduced to eliminate redundant weak feature points caused by the relaxation of the threshold, so as to achieve a spatially uniform distribution of feature points in the first and second projected images.
[0015] As a preferred embodiment of the present invention, the calculation process of the sub-pixel offset includes: based on the effective feature matching pairs obtained by screening under any projection viewpoint, using the random sampling consensus algorithm to robustly fit the global image translation model, and using the median of the point displacements within the model as the initial offset under that viewpoint. Using the initial offset as the initial value for iteration, an enhanced correlation coefficient algorithm is introduced to perform sub-pixel level refinement calculation. By maximizing the inter-frame normalized correlation coefficient, the image grayscale mapping relationship is accurately converged. The bidirectional sub-pixel offset in the X and Y orthogonal directions under each projection viewpoint is obtained by solving the problem from each viewpoint.
[0016] In a preferred embodiment of the present invention, the Fourier domain translation registration includes performing a two-dimensional Fourier transform on the second projected image containing focus drift to convert the spatial domain image signal into a frequency domain spectrum signal; constructing a frequency domain linear phase compensation factor based on the sub-pixel offset of the corresponding viewpoint obtained by solving, performing phase offset compensation on the spectrum signal to offset the residual drift deviation between image frames; and reconstructing the compensated spectrum signal into a spatial domain image through an inverse two-dimensional Fourier transform to obtain the corrected projected image.
[0017] In a preferred embodiment of the present invention, the closed-loop iterative registration and error verification includes: For the second projected image from each projection viewpoint, Fourier domain subpixel translation is performed based on the current subpixel offset, and error verification is performed based on the residual offset between the translated image and the corresponding first projected image. If the residual offset meets the preset registration accuracy condition and the current sub-pixel offset is within the preset physical drift constraint range, then the registration is deemed qualified, and the current offset is used as the final sub-pixel offset for that viewpoint. If the residual offset does not meet the preset registration accuracy condition, the feature points of the first projected image and the corresponding second projected image are re-selected and matched, the sub-pixel offset is re-solved based on the updated effective feature matching pair, and Fourier domain translation and error verification are performed again. The above process is repeated until the residual offset meets the preset registration accuracy condition and the sub-pixel offset is within the preset drift constraint range. Then the iteration is terminated and the current offset is output as the final result. If the residual offset still does not meet the preset registration accuracy condition when the number of iterations reaches the preset maximum number of iterations, then the sub-pixel offsets in the current X and Y directions will be halved as the final result.
[0018] In a preferred embodiment of the present invention, the step of obtaining the continuous drift amount under the full projection view through bicubic spline interpolation includes: The projection viewpoints that have passed iterative verification and whose parameters have been solved are selected as interpolation reference control points, and the number of reference control points is consistent with the total number of frames of the first projection image. Establish a one-to-one mapping relationship between the scanning angle of each reference control point and the corresponding sub-pixel offset in the X and Y directions, and construct an angle-offset discrete dataset; The discrete drift data of the reference control points are nonlinearly fitted based on the bicubic spline interpolation algorithm to construct a global drift function in which the drift amount changes continuously with the scanning angle.
[0019] By using the fitted global drift function, high-density sampling and interpolation are performed on the remaining projection viewpoints that did not participate in the iterative solution to obtain a smooth, continuous drift amount covering all scanning angles.
[0020] As a preferred embodiment of the present invention, the frame-by-frame registration and correction of the second projection sequence based on the full-view drift includes: Based on the sub-pixel offset corresponding to the second projection image of each frame, calculate the actual drift amount of the current frame, and compare the actual drift amount with the preset drift threshold for determination; If the actual drift is lower than the preset drift threshold, the current frame imaging accuracy is determined to meet the requirements, and the original second projection image is directly retained as the standardized projection image. If the actual drift exceeds the preset drift threshold, it is determined that the current frame has excessive drift. Fourier domain translation registration is performed on the image to be corrected in this frame to eliminate the imaging drift deviation and obtain the standardized projection image after drift correction.
[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a time-series optimization architecture of "fast scanning to establish a baseline, followed by slow scanning to acquire data," overcoming the problems of focus drift accumulation and inter-frame baseline inconsistency caused by unreasonable traditional time-series arrangements. A high-precision baseline is established using short-duration fast scanning within a thermally stable window, with a 0.3-pixel threshold set to ensure baseline reliability. In subsequent high-density scanning, a 0.5-pixel correction-free mechanism is introduced; images with drift below this threshold are directly retained, avoiding redundant correction that introduces additional noise. Fourier-domain targeted correction is only performed on drift exceeding the threshold. Because focus drift exhibits a slowly accumulating characteristic, most images are exempt from correction, avoiding high-frequency artifacts and noise accumulation introduced by interpolation resampling from the source. The reconstructed input maintains the original grayscale distribution, effectively suppressing 3D reconstruction noise and significantly improving the signal-to-noise ratio and structural clarity.
[0022] 2. This invention replaces the traditional phase correlation method with feature point matching, overcoming its dependence on global grayscale information and its tendency to fail in low signal-to-noise ratio and weak texture scenarios. It employs Scale-Invariant Feature Transform (SIFT) for local feature extraction, actively locking in significant and stable gradient structures such as sample edges, defects, and interfaces as matching primitives, thus eliminating global grayscale dependence. Highly reliable matching pairs are obtained through Euclidean distance initial screening, distance ratio constraint elimination of ambiguous matches, and robust fitting of outliers. A robust RANSAC fitting global translation model is used, with the median of interior point displacements as the initial offset. This initial value is then used to introduce the ECC (Enhanced Correlation Coefficient) algorithm for sub-pixel-level refinement, outputting bidirectional sub-pixel offsets in the X and Y directions per viewpoint. This multi-level solution process from coarse to fine, combined with a fault-tolerant iteration strategy, can still stably output high-confidence offsets even under complex conditions of sparse features, high noise, and uneven grayscale, significantly improving the algorithm's anti-interference capability and versatility.
[0023] 3. This invention employs a Fourier domain phase compensation lossless registration scheme, overcoming the grayscale distortion and detail blurring caused by traditional spatial domain affine transformations and pixel interpolation. Traditional methods achieve alignment through coordinate remapping and grayscale resampling, directly altering the grayscale and pixel value distribution of the original image. The interpolation smoothing effect leads to high-frequency detail loss and edge blurring. Based on the Fourier translation theorem, registration is transformed into a spectral domain phase compensation problem, only linearly correcting the spectral phase. The original grayscale amplitude and pixel range remain unchanged throughout the process, eliminating the need for spatial domain resampling. Mathematically, this achieves truly lossless sub-pixel-level precise alignment, effectively preserving the original grayscale distribution and high-frequency details. It significantly improves the edge integrity and microstructure fidelity of the reconstructed image, exhibiting excellent artifact suppression, and fully meeting the sub-micron-level high-precision defect detection requirements of precision devices.
[0024] This invention has significant application value in the field of semiconductor chip defect detection and high-precision metrology. It can effectively overcome image distortion caused by focus drift, significantly improve the geometric accuracy and structural fidelity of 3D reconstruction, and provide key quality control means for early identification of minute defects and yield improvement in chip manufacturing. Attached Figure Description
[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1This is a schematic diagram of the micro-CT imaging system on which the present invention, which combines time-series optimization and Fourier domain translation for non-destructive correction of focus drift in micro-CT, is based.
[0026] Figure 2 This is a schematic diagram showing the positional relationship between the X-ray source, the sample to be tested, and the detector in a micro-CT imaging system.
[0027] Figure 3 This is a schematic diagram showing the X and Y direction shift of the projected image caused by focus drift in a micro-CT imaging system.
[0028] Figure 4 This is a schematic diagram of the distribution of feature markers in Maltesers in Embodiment 2 of the present invention.
[0029] Figure 5 This is a dynamic change curve of the X and Y bidirectional offset scanning angle under the "fast scan first, slow scan later" timing architecture in Embodiment 2 of the present invention, using Maltesers as the test sample.
[0030] Figure 6 This is a dynamic curve of the X and Y bidirectional offset scanning angle under the traditional "slow scan first, fast scan later" timing architecture, using Maltesers as the test sample.
[0031] Figure 7 This is a schematic diagram of the projected image before focus drift correction, using Maltesers as the test sample in Embodiment 2 of the present invention.
[0032] Figure 8 This is a schematic diagram of the projected image after focus drift correction, using Maltesers as the test sample in Embodiment 2 of the present invention.
[0033] Figures 9-12 This is the test diagram used in Embodiment 2 of the present invention to evaluate and analyze the corrected projection view of Maltesers as the test sample. Detailed Implementation
[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0035] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0036] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0037] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.
[0038] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0039] Unless otherwise specified, the equipment used in this invention are all commonly used equipment or devices in the art; the methods used in this invention are all commonly used methods in the art.
[0040] Focus drift correction requires the use of a pre-built micro-CT imaging system, such as... Figure 1 As shown, the system includes a radiation source unit 100, a sample clamping unit 200, a displacement control unit 300, an imaging detection unit 400, and a software control and image reconstruction unit 500. Each unit is mounted or fixed on a flat and stable optical platform T, effectively ensuring the relative position verification of each core component, improving the structural rigidity and detection stability of the overall system, and is equipped with a fully enclosed composite shielding and interlocking protection structure to meet radiation safety requirements and ensure stable scanning conditions.
[0041] Specifically, the X-ray source unit 100 includes an X-ray source 101 and a support frame 102 for fixing the X-ray source 101. The sample clamping unit 200 includes a clamping base 201, a support rod 202 mounted on the top of the clamping base 201, and a sample clamp 203 located at the end of the support rod 202. The sample C to be tested is clamped in the sample clamp 203, which can be adapted to conventional samples such as walnuts and malted milk powder, as well as precision micro-nano devices such as chips.
[0042] The displacement control unit 300 is a four-axis linkage stage, including an X-axis displacement stage 301, a Y-axis displacement stage 302 mounted on the X-axis displacement stage 301, a Z-axis lifting stage 303 mounted on the Y-axis displacement stage 302, and a rotary stage 304 mounted on the Z-axis lifting stage 303. The sample clamping unit 200 is mounted on the rotary stage 304. By controlling the displacement control unit 300, the sample can be electrically translated along the X, Y, and Z axes and rotated around the Z-axis. Each motion axis is equipped with a grating ruler and a high-precision encoder to achieve multi-axis closed-loop motion control.
[0043] The imaging detection unit 400 is located on the side of the displacement control unit 300 away from the X-ray source unit 100, and includes a detector 401 and a support platform 402 for fixing the detector. The detector 401 can be a photon counting detector with a resolution of up to 75 μm.
[0044] The software control and image reconstruction unit 500 is a high-performance computer with built-in device control, image reconstruction and artifact suppression software. It can coordinate and regulate various hardware units to complete the entire process of data acquisition, error correction and image reconstruction, and realize high-precision non-destructive imaging detection of samples.
[0045] Example 1, an embodiment of the present invention, describes a non-destructive correction method for micro-CT focus drift that combines temporal optimization and Fourier domain shift, based on the aforementioned micro-CT imaging system after its construction. This correction method includes the following steps: S1. Initialize and preheat the micro-CT imaging system, and use the phantom to calibrate the geometric parameters of the micro-CT imaging system; after calibration, install the sample to be tested and perform a pre-scan on the sample; Specifically, the initialization of the micro-CT imaging system is performed first: the X-ray source is preheated, and the sample is installed after the system reaches a thermally stable state. The four-axis stage is adjusted to place the sample in a suitable exposure position. After confirming that there are no stripe artifacts in the projected image through trial imaging, the initialization is completed.
[0046] After initialization, the phantom is mounted on the rotating stage, and driven to rotate 360° to acquire a full-view projection image, completing the system geometric parameter calibration. Calibration parameters include the distance from the X-ray source to the sample (SOD), the distance from the X-ray source to the detector (SDD), and the relative positions of the sample with respect to the rotation center and detector, etc. (see [reference]). Figure 2 ).
[0047] After calibration, remove the phantom, install the sample C to be tested, and perform continuous pre-scanning for no less than 30 minutes to eliminate the initial focus fluctuations introduced during the heating stage, so that the focus state is fully converged and the thermal conditions tend to be steady.
[0048] S2. After the pre-scan is completed, the sample to be tested is reset to the preset starting angle, and a full-view low-drift scan is performed with the preset scanning parameters. The relative offset between the first frame and the last frame projection image is calculated and verified with the preset stable threshold until the relative offset is lower than the preset stable threshold. Based on this, the first scanning parameters are determined. Specifically, after the pre-scan, the rotary stage is first reset to the preset starting angle; with the X-ray source continuously on, a 0°~360° full-view low-drift scan is performed with preset parameters. By comparing the relative displacement of the 0° (first frame) and 360° (last frame) projected images, it is determined whether the focus has drifted significantly. The specific determination method is as follows:
[0049] The relative offset is compared with a preset (benchmark) stabilization threshold to determine the validity of the current low-drift reference projection sequence. If the relative offset of the first and last frame images is less than the preset benchmark stabilization threshold, the focus is determined to be stable and the acquisition is valid. The scanning step size is appropriately reduced to obtain more reference images, and the optimized parameters are scanned again for verification. Finally, the optimal parameters for the first scan are determined. If the relative offset exceeds the preset benchmark stabilization threshold, the acquisition is determined to be invalid. The pre-exposure time of the X-ray source needs to be extended and the scan needs to be repeated until the stabilization condition is met. Finally, the parameters for the first scan and the corresponding scan time are determined.
[0050] S3. After determining the first scanning parameters, keep the X-ray source on and the sample continuously rotated. With a fixed and irreversible timing sequence, first perform the first scan to establish a low-drift reference projection sequence, and then immediately switch to the second scanning parameters to perform a high-density scan to obtain the second projection sequence containing the focus drift error.
[0051] After determining the first scanning parameters, the rotary stage is reset to the preset starting angle. Strictly following the predetermined fixed timing sequence, the first scan is performed using the first scanning parameters. The first scan uses a scan step size of m and a single-frame exposure time preferably of 0.5 s, acquiring (360+m) / m projection images to construct a low-drift reference projection sequence. After the first scan is completed, maintaining the same timing sequence, the X-ray source remains continuously on, and the sample rotates continuously. Immediately, the system switches to the second scanning parameters to perform a high-density scan, i.e., a scan step size of n and a single-frame exposure time preferably of 2 s, acquiring a total of (360+n) / n projection images to establish a second projection sequence including focus drift error. Here, m is an integer multiple of n and is divisible by 360°, making the sampling density of the second scan significantly higher than that of the first scan. The two scans are performed sequentially, with the X-ray source continuously exposing the image throughout, ensuring that the second scan fully inherits the thermal conditions at the end of the first scan. The second scan cycle is longer, and the heat accumulation from the X-ray source induces a gradual focus drift. The obtained sequence fully records the drift evolution throughout the entire scan cycle, providing data support for subsequent drift parameter fitting and correction. It should be noted that the execution order of the first and second scans is preset and cannot be reversed.
[0052] S4. Using the projection angle of each first projection image in the low-drift reference projection sequence as an index, extract the second projection image under the corresponding angle from the second projection sequence according to a preset fixed step size, and construct a matching dataset corresponding to the angle of the reference projection image. Specifically, based on the principle of consistent projection viewpoints, starting from the first frame (0°) in the second projection sequence, second projection images with the same viewpoint as each reference image in the low-drift reference projection sequence are extracted at a fixed step size to construct a matching dataset corresponding to the viewpoint of the first projection image.
[0053] Using the projection viewpoints of the first scan as spatial indices, the corresponding images are extracted by traversing the second sequence at m / n steps, resulting in (360+m) / m sets of strictly aligned image pairs. After matching, the two images with the same viewpoint are uniformly named and coded to achieve unique identification and precise association. This results in a fully covered matching dataset corresponding to the viewpoints of the reference projection image, eliminating the risk of cross-viewpoint mismatch.
[0054] S5. Perform standardization preprocessing on the matching dataset, use the scale-invariant feature transformation algorithm to extract local neighborhood information, generate corresponding high-dimensional descriptors, filter out candidate matching pairs, and then obtain high-confidence effective feature matching pairs. When performing feature point matching, due to the complex structure of the samples and the presence of areas with smooth grayscale, noise, and repetitive textures within the image, direct matching of the entire image can easily lead to mismatches. Therefore, before feature point matching, a standardization preprocessing is performed on the paired images, specifically including: First, global grayscale normalization is performed to linearly map the image grayscale value range to a unified interval, eliminating grayscale deviations caused by X-ray source fluctuations and inter-frame exposure differences; this unifies the grayscale benchmark of each projected image. Then, a contrast-limited adaptive histogram equalization algorithm is used to enhance the local microstructure information of the sample under test, strengthening key information such as edges and pores. Next, bilateral filtering is used, combining spatial neighborhood and grayscale similarity as dual weights to suppress random noise while preserving edge details. Finally, local zero-mean normalization eliminates interference from uneven illumination and grayscale shift, reducing the grayscale feature differences between the reference and drift images.
[0055] After image preprocessing, feature points within the sample C to be tested need to be further extracted and matched. First, the Sobel operator is used to calculate the gray-level gradients and magnitudes in the horizontal and vertical directions of the image pixel by pixel. Effective detection areas are delineated using structural regions with significant and stable gradients, such as sample boundaries, defects, and interfaces, while background and low signal-to-noise ratio interference are eliminated. Next, within the effective detection area, the SIFT algorithm is used to prioritize the location of high-contrast feature points as candidate keypoints, extracting multi-scale keypoints and their corresponding high-dimensional descriptors, as shown in the following equation: (1) in, For high-dimensional descriptors, For Gaussian kernel function, These are the two-dimensional spatial coordinates of image pixels. As a scale factor, The scale is the magnification. This formula can be used to detect extreme points at different scales, ensuring that feature points have scale invariance to adapt to changes in the scale of local structures under different imaging conditions.
[0056] Finally, initial feature point matching is performed based on the Euclidean distance between descriptors to generate candidate matching pairs. Then, the ratio of the nearest neighbor to the second nearest neighbor distance is used to eliminate ambiguous mismatches, ultimately obtaining highly reliable and effective feature matching pairs for subsequent accurate calculation of inter-frame drift parameters. The ratio of the nearest neighbor to the second nearest neighbor descriptor distance can be expressed as: (2) In the formula, The nearest neighbor matching distance, The next nearest neighbor matching distance. The filtering threshold proposed by Lowe (usually between 0.6 and 0.8). The smaller this ratio, the stronger the uniqueness of the match, thus effectively eliminating ambiguous matches with high confusion and insufficient stability, and significantly improving the matching accuracy.
[0057] In step S5, an adaptive feature point quantity adjustment mechanism is incorporated during the acquisition of effective feature matching pairs to address the problem of insufficient effective feature points caused by sparse local textures or simple structures in the sample. This mechanism counts the number of effective feature points in real time. When the number of detected points falls below a preset lower limit, the Lowe threshold (i.e., the ratio of the nearest neighbor to the second nearest neighbor descriptor distance) is adaptively lowered to appropriately relax the screening constraints, thereby supplementing the extraction of more stable effective feature points and ensuring that subsequent matching has sufficient constraints to prevent matching degradation.
[0058] Furthermore, after lowering the Lowe threshold, a non-maximum suppression strategy is introduced to eliminate redundant weak feature points introduced by the relaxed threshold, ensuring that feature points are uniformly distributed in the image space. This avoids matching bias caused by excessive concentration of features in local regions, thus providing a highly reliable matching basis for solving sub-pixel offsets. Specifically, within a preset local neighborhood window, only the feature points with the strongest response are retained, while other redundant features with weaker responses within that neighborhood are suppressed, completing the simplification and quality improvement of the feature set. Its mathematical expression is as follows: (3) In the formula, Ω represents the response intensity at the key point, and Ω is the local neighborhood window.
[0059] S6. Based on the effective feature matching pairs extracted from each projection viewpoint, solve for the initial offset of the second projection image relative to the corresponding first projection image under that viewpoint; then calculate the sub-pixel offset of the second projection image relative to the corresponding first projection image in the X and Y orthogonal directions using a sub-pixel level refinement algorithm. This step employs a two-stage solution strategy of "first coarse estimation, then refinement".
[0060] In the first stage, based on the effective feature matching pairs extracted from each projection viewpoint, robust fitting is performed using the RANSAC algorithm combined with a translation model to select the interior point matching set. The median of all interior point displacements is used as the initial coarse-grained offset of the image, effectively suppressing outlier matching interference and improving the noise resistance and reliability of coarse registration. The calculation is as follows: (4) in, This is the initial offset. For the first Each interior point is matched with a displacement vector on the image plane.
[0061] The second stage uses a coarse-grained initial offset as the initial value to drive the ECC algorithm. By maximizing the normalized correlation coefficient, it drives the grayscale mapping to converge to the optimal value, achieving a gradual approximation from pixel-level coarse matching to sub-pixel-level precise matching. Finally, it calculates the bidirectional sub-pixel offsets in the X and Y directions of each viewpoint. (See) Figure 3 The objective function of the ECC algorithm is defined as follows: (5) In the formula, For the reference image (i.e., the first scan low-drift reference projection) in pixel coordinates The grayscale value at that location; The grayscale value of the image to be registered (i.e., the current drift projection of the second scan) after translation transformation, i.e., the overall translation. ; and These are the gray-level mean values of the corresponding regions in the reference image and the image to be registered, respectively. The ECC algorithm maximizes... Iterative convergence, accurately solving for the X and Y bidirectional sub-pixel offsets at each viewpoint. .
[0062] S7. The registration process is implemented based on the Fourier domain translation method. The deviation of the single registration is calculated. Then, a closed-loop iterative registration and error verification mechanism is introduced to iteratively optimize the sub-pixel offset of the second projection image under each projection view. The optimization process is then extended to all second projection images corresponding to each first projection image in the reference projection sequence, and the sub-pixel offset of each view is solved one by one. Specifically, single-stage registration is susceptible to local optima due to factors such as noise, repetitive textures, and sparse features, resulting in residual offsets. This step introduces a closed-loop iterative registration and error verification mechanism: First, Fourier domain lossless sub-pixel pre-registration is performed based on the ECC algorithm output to complete preliminary correction; then, the residual offset between the registered image and the reference image is calculated and judged. If the residual offset is within a preset constraint range, the registration is deemed qualified, and the set of parameters is output; if it exceeds the constraint range, stringent feature filtering is used to re-purify the matching pairs, and Fourier domain translation is performed again, iterating until convergence.
[0063] Furthermore, during the iteration process, the second projected image needs to be registered to the first projected image with the same projection viewpoint. A Fourier domain translation method is used to complete the lossless registration of the slow-scan image. According to the Fourier translation theorem, pixel translation in the spatial domain is equivalent to linear phase shift in the frequency domain, and its core transformation relationship is: (6) In the formula, For two-dimensional Fourier transform operators, For frequency domain coordinates, Image size, This is the linear phase compensation factor.
[0064] The specific steps for Fourier domain translation registration are as follows: First, perform a two-dimensional Fourier transform on the original drifted second projection image to obtain the global spectral information; then substitute the sub-pixel offset of the corresponding viewpoint ( , A precise phase compensation factor is constructed to compensate for the phase shift of the spectrum, thus offsetting inter-frame temporal drift. Finally, the spatial image is reconstructed through inverse two-dimensional Fourier transform to obtain a distortion-free corrected projection image free from drift errors. This process preserves the image's grayscale features and microstructure throughout, achieving truly lossless sub-pixel-level accurate registration. The inverse Fourier transform formula is as follows: (7) These are the pixel coordinates of the corrected projected image. These are the pixel coordinates of the original image where the image has drifted. It is an offset with sub-pixel accuracy that needs to be compensated; this method only corrects the phase offset, does not require spatial domain resampling, does not change the grayscale distribution and pixel values of the original image, and can eliminate interpolation truncation error and detail distortion.
[0065] To avoid infinite iterations, the iteration process also incorporates dual termination constraints: a residual threshold and a maximum number of iterations, along with a fallback strategy of halving the offset. The modified formula is as follows: (8) In the formula, , This is the final subpixel offset after fallback correction. This strategy effectively prevents image distortion caused by overcorrection, steadily improves the registration accuracy and inter-frame correlation of the drifted image in the second scan, thus providing a reliable drift-free projection sequence for subsequent metal artifact repair and high-precision reconstruction. After iteration, the optimal result is retained, invalid data is discarded, and the final output is a high-confidence subpixel offset.
[0066] S8. Using the verified sub-pixel offset as the discrete control node, construct a continuous mapping model of drift amount as the angle based on bicubic spline interpolation, accurately fit its nonlinear evolution law, and output the continuous drift amount of the whole view. Iterative registration only yields sparse, discrete viewpoint drift data, which cannot meet the frame-by-frame correction requirements of high-density sequences. Therefore, bicubic spline interpolation is introduced to deduce the continuous drift parameter field across the entire viewpoint using finite discrete control points. The bicubic spline interpolation model uses the scanning rotation angle θ as the independent variable and the real-time focus offset f(θ) as the dependent variable. It constructs a continuous mapping relationship between adjacent viewpoint intervals based on piecewise third-order polynomials. The single-interval fitting polynomial expression is: (9) In the formula: Let be the scanning angle of the i-th control point. ~ These are the fitting coefficients of the third-order polynomial for the corresponding segmented intervals.
[0067] Specifically, the steps of bicubic spline interpolation are as follows: Select verified projection viewpoints as reference control points, with the number consistent with the total number of frames in the first projection; establish a one-to-one mapping between the angle of each control point and the offsets in the X and Y directions to construct a discrete dataset; with the angle as the independent variable and the offset as the dependent variable, constrain the function value, first-order and second-order derivatives to be continuous across the entire domain, and construct a second-order smooth continuous drift function from 0° to 360°; finally, use this function to perform high-density sampling interpolation on the viewpoints that did not participate in the iteration to obtain a full-coverage smooth continuous drift parameter field.
[0068] S9. Frame-by-frame determination based on full-view drift amount: If the drift amount is below the preset drift threshold, the original image is retained; if the drift amount reaches or exceeds the preset drift threshold, Fourier domain translation registration is performed to complete the adaptive correction processing of all projected images and construct a complete set of projected images after focus drift compensation. Traditional global uniform correction is prone to overcorrection of images with slight drift, introducing additional noise. This embodiment constructs a fusion mechanism of viewpoint-differentiated adaptive correction-free and precise correction: when the offset in both X and Y directions of a single viewpoint is less than 0.5 pixels, it is determined to be a small fluctuation in the steady state of the system, and the original image is directly retained without correction; when the offset exceeds 0.5 pixels, it is determined that there is a significant drift deviation, and Fourier domain translation registration needs to be performed on the frame of the image to be corrected to eliminate the imaging drift deviation and obtain a normalized projection image after drift correction.
[0069] S10. Perform 3D reconstruction based on the complete projected image set and geometric parameters after focus drift compensation to obtain a corrected 3D slice model; compare it with the uncorrected reconstructed model, and evaluate the optimization degree and effectiveness of the correction method by quantifying image quality indicators.
[0070] Three-dimensional CT reconstruction was performed based on standardized projection sequences and calibrated geometric parameters to obtain a corrected three-dimensional reconstruction model of the sample. The corrected model was compared with the uncorrected original sequence reconstruction model, and the correction effect was evaluated by quantifying image quality indicators. A highly consistent standardized projection dataset was constructed by integrating the uncorrected original image and the frequency domain corrected image. Viewpoint matching and spatial coordinate unification were performed based on the rotation angle to establish a precise correlation model between the two-dimensional projection and the three-dimensional spatial position.
[0071] Based on the normalized projection sequence, three-dimensional reconstruction is performed in combination with calibrated imaging geometric parameters. This method preferably uses the FDK (Feldkamp-Davis-Kress, filtered backprojection) algorithm adapted to cone-beam CT: first, spatial frequency filtering enhancement is applied to the projected image, and then the filtered projection is backprojected into three-dimensional space in combination with geometric parameters and rotation angle. By using cone-beam weighting and multi-view fusion, focus drift artifacts are effectively suppressed, and a three-dimensional voxel model of the sample with complete structure and clear edges is reconstructed, which significantly improves the fidelity and detection reliability of micro-CT imaging under complex thermal drift conditions.
[0072] Therefore, this embodiment achieves effective suppression of focus drift error during long-term high-resolution scanning of micro-CT through a complete technical chain of "determination of low drift parameters and scanning time - continuous temporal scanning - feature point extraction and matching - sub-pixel offset calculation and iterative optimization - bicubic spline interpolation global drift modeling - Fourier domain non-destructive correction - three-dimensional reconstruction", which significantly improves the reliability and stability of the system for detecting microscale structures.
[0073] Example 2, Reference Figures 4-12A Maltesers sample was selected for focus drift correction, and a "slow scan followed by fast scan" method (first a slow scan, then a fast scan) was used as a control experiment to fully verify the correction accuracy and robustness of the algorithm. The Maltesers sample, being a porous and loose structure without metal components, is highly sensitive to X-ray source focus drift, allowing for precise quantification of the algorithm's dynamic drift compensation performance.
[0074] Reference Figure 4 This is a schematic diagram of the distribution of feature positioning markers on the Maltesers sample in this embodiment of the invention. The size of the natural feature markers extracted in this invention can be manually customized and flexibly adapted to the sample structure scale and imaging resolution. All markers are selected from high-gradient, high-contrast feature regions inside the sample. The points are independent of each other, do not overlap or cluster, and have a uniform and controllable overall distribution, providing a reliable feature constraint basis for subsequent high-precision feature matching and stable sub-pixel offset calculation.
[0075] Reference Figures 5 to 6 The offset change curves directly demonstrate the technical advantages of the timing optimization strategy of this invention. Among them, Figure 5 This invention illustrates the dynamic variation of X and Y bidirectional offsets with scanning angle under the "fast scan first, slow scan later" timing sequence. Figure 6 This describes the dynamic variation of the X and Y bidirectional offsets with the scanning angle under the traditional "slow scan first, fast scan later" timing sequence.
[0076] Reference Figures 7 to 8 This is a schematic diagram of the three-dimensional reconstructed slice images of a Maltesers sample before and after focus drift correction. Figures 9-12 To comprehensively evaluate the practical application effect of focus drift correction, this invention conducts quantitative analysis from four core evaluation dimensions: image edge sharpness comparison (mean ± standard deviation), contrast-to-noise ratio comparison, image clarity comparison, and artifact suppression degree.
[0077] The verification experiment of the focus drift correction method of the present invention was carried out on the established micro-CT imaging system platform. The specific implementation steps are as follows: Before the experiment, hardware stabilization was implemented from multiple dimensions, including thermoelectric, mechanical, and detection, to suppress focus drift at its source. The thermoelectric end employed constant temperature control, high vacuum maintenance, filament preheating, and constant power supply to avoid focus shift caused by electrical participation and temperature fluctuations. The mechanical end relied on a rigid base and damping vibration isolation structure to reduce displacement errors caused by vibration and structural loosening. The detection end completed the calibration of the detector's dark and bright fields, geometric distortion, and bad pixels, eliminating noise and pixel gain unevenness, providing high-quality raw image data for subsequent registration correction and 3D reconstruction.
[0078] Step 1: Initialize and preheat the micro-CT imaging system, and use a phantom to calibrate the geometric parameters of the micro-CT imaging system; after calibration, install the sample to be tested and perform a pre-scan on the sample; When the X-ray source 101 is started, the voltage and current are gradually increased stepwise from the initial standby state of zero voltage and zero current, and the constant temperature is maintained for 30 minutes to allow the core components of the X-ray source 101, such as the electron optical components and the anode target, to fully enter the thermal buffer state and smoothly transition to the preset standard operating parameters, avoiding sudden changes in operating conditions caused by instantaneous start-up and shutdown.
[0079] After system initialization, the Maltesers test sample is clamped on the sample holder 203 of the rotary stage 304. The working parameters of the X-ray source 101 are set as tube voltage 60 kV and tube current 120 μA. The spatial position of the sample is precisely adjusted by the four-axis linkage stage to make the sample be in the optimal imaging exposure field of view. Pre-imaging debugging is carried out to verify that the acquired projection image has no stripe artifacts, noise distortion and other abnormalities, and to confirm that the initial imaging link of the system is working normally.
[0080] After the system initialization and pre-imaging verification are successful, the test sample is removed and a high-precision standard phantom is precisely clamped onto the sample holder 203 of the rotary stage 304. The rotary stage 304 is driven to move the standard phantom to complete a 360° full-range rotation scan, acquiring a phantom projection image dataset covering the entire field of view. Based on the standard phantom imaging data, the high-precision geometric parameter calibration of the micro-CT system is completed.
[0081] The core geometric parameters calibrated in this study include: source-to-detector distance (SDD), source-to-sample working distance (SOD), and reference coordinates of the detector plane corresponding to the central beam. and the detector's three-dimensional deflection attitude angle ,in, Let be the rotation angle of the detector's normal vector about its normal direction. The angle of rotation of the detector about its central column axis. Let be the angle of rotation of the detector about its central axis. Figure 2 The left, middle, and right images correspond to the following: axis, shaft and The viewing angles along the axes (Z' and Y' represent the Z-axis and Y-axis positions after deflection) show the deflection geometry of the detector in the three orthogonal directions mentioned above. After all geometric parameters were calibrated, the accuracy and effectiveness of the calibration parameters were confirmed through system accuracy self-check, and then the standard calibration phantom was removed.
[0082] To completely eliminate initial focus fluctuations caused by unstable operating conditions during the equipment heating phase, the installation position and mechanical attitude of the rotary stage 304 and sample fixture 203 were kept completely constant throughout the calibration process to avoid calibration errors introduced by mechanical displacement. The malted milk sample to be tested was then re-clamped onto the sample fixture 203. The microfocus X-ray source 101 was restarted, maintaining a constant operating parameter of 60kV tube voltage and 120μA tube current, and continuous pre-scanning preprocessing of the malted milk sample for no less than 30 minutes was performed. Through long-term full-load pre-scanning, the thermal conditions of the X-ray source 101 were fully converged, and the dynamic focus drift gradually decreased and tended to a steady state, completely suppressing the initial focus shift caused by temperature rise disturbances. This provided a stable and reliable equipment operating reference for subsequent high-precision projection acquisition, drift correction, and 3D reconstruction.
[0083] Step 2: After the pre-scan is completed, the sample to be tested is reset to the preset starting angle, and a full-view low-drift scan is performed with the preset scanning parameters. The relative offset between the first frame and the last frame projection image is calculated and verified with the preset stable threshold until the relative offset is lower than the preset stable threshold. Based on this, the first scanning parameters are determined. After completing the thermal stabilization pre-scanning of the Maltesers sample, the rotary stage 304 was initialized and reset to return the sample to the preset scanning starting angle. Throughout the scan, the X-ray source 101 remained continuously on, ensuring constant exposure and locking the sample clamping posture, spatial position, system geometric parameters, and optical path transmission relationship to remain constant. After system pre-preparation, a rapid, low-drift full-domain scan of 0°–360° was performed on the sample. Initial scanning parameters were set to a single-frame exposure time of 0.5 s and an angle scan step size of 18°. A total of 21 reference projection images were acquired across the entire domain, with the timestamp of each image acquisition recorded synchronously frame-by-frame to construct a complete time-series dataset. This provides accurate temporal support for subsequent thermal drift time-series analysis, error tracing, and continuous drift modeling.
[0084] This example uses 0.3 pixels as the focus stability threshold: when the two-dimensional relative displacement between the first and last frames is less than this threshold, the X-ray source thermal condition is considered stable, the thermal drift amplitude is controllable throughout the scan, and the reference data meets the requirements for high-precision imaging and correction modeling. Under this steady-state premise, an adaptive parameter optimization mechanism is activated, refining the initial 18° scan step size to 15°, and increasing the number of reference projection acquisition frames from 21 to 25. This effectively improves the spatial angle sampling density of the reference projection sequence, enriches the global perspective sampling information, significantly suppresses the fitting residuals and waveform oscillations in the subsequent bicubic spline interpolation modeling process, and improves the fitting accuracy of the global dynamic drift field. After optimization and comparison, the 18° scan step size has the best overall effect in this example, and the first scan parameters are finally determined as: scan step size 18°, exposure time 0.5 s.
[0085] Step 3: After determining the first scanning parameters, keep the X-ray source on and the sample to be tested continuously rotated. With a fixed timing sequence, first perform the first scan to establish a low-drift reference projection sequence, and then switch to the second scanning parameters to perform a high-density scan to establish a second projection sequence containing focus drift error. After the validity verification in Step 2 confirms that the global pixel drift of the first scan is stably below the qualified threshold of 0.3 pixels and the device is in a low thermal drift steady state, it enters the dual-sequence continuous scanning stage with a fixed and irreversible timing.
[0086] First, the rotary stage was initialized to precisely reset the Maltesers sample to the preset starting angle. Then, the first scan was performed using the first scan parameters. Specifically, while maintaining the operating conditions of the micro-CT imaging system, a scan step size of 18° and a single-frame exposure time of 0.5 s were used to acquire a total of 21 projection images, establishing a low-drift reference projection sequence. During the scanning process, each frame of projection image was standardized and named according to the rotation angle, and the motion coordinate parameters of each axis of the device were simultaneously stored, constructing a low-drift reference projection sequence with complete temporal sequence, traceable parameters, and high stability.
[0087] After the first scan is completed, the sequence is strictly connected and cannot be changed. The X-ray source is kept on and the sample is rotated continuously. The second scan parameters (scan step size 0.5°, single frame exposure time 2 s) are immediately and without interruption switched to carry out a high-density and fine second scan. 721 projection images are acquired simultaneously to obtain the second projection sequence containing focus drift error.
[0088] Step 4: Using the projection angle of each first projection image in the low drift reference projection sequence as an index, extract the second projection image under the corresponding angle from the second projection sequence according to a preset fixed step size, and construct a matching dataset corresponding to the angle of the reference projection image. Based on the principle of consistent projection viewpoints, in the second projection sequence containing focus drift errors, starting from the first frame (0°), second projection images with the same viewpoint as each reference image in the low-drift reference projection sequence are sequentially extracted in 36° steps until a full 360° viewpoint is covered. A total of 21 matching images are selected to construct a matching dataset corresponding to the viewpoints of the reference projection images in the second scan, with its number strictly consistent with that of the low-drift reference projection sequence. After completing frame extraction and matching, the two images with the same viewpoint are uniformly named and coded to achieve unique identification and accurate association. After traversing all viewpoints and completing pairing, a matching dataset corresponding to the viewpoints of the reference projection images with full coverage is formed, and it is standardized and archived for storage, providing reliable data support for subsequent feature matching, offset calculation, and high-precision correction.
[0089] Step 5: Perform standardization preprocessing on the matching dataset, use the scale-invariant feature transformation algorithm to extract local neighborhood information, generate corresponding high-dimensional descriptors, filter out candidate matching pairs, and then obtain high-confidence effective feature matching pairs; Step 51: Standardize and preprocess the projected image
[0090] First, global grayscale normalization is performed on the paired projection images to linearly map the image grayscale value range to a unified interval, eliminating grayscale deviations caused by ray intensity fluctuations and exposure differences, and unifying the grayscale benchmark of fast and slow scan images. On this basis, the CLAHE algorithm (Contrast Limited Adaptive Histogram Equalization) is used to enhance local microstructure and strengthen key features such as edges and pores. Then, bilateral filtering is introduced, which relies on the dual weights of spatial neighborhood and grayscale similarity to suppress random noise while preserving fine structures, resulting in a high-definition and low-noise preprocessed image.
[0091] Maltesers samples have a complex internal structure, including various microstructural features such as particle interfaces, pores, and sugar shell-chocolate filling boundaries. Direct feature extraction is easily affected by uneven grayscale and noise interference. After the above preprocessing, the edge contours of Maltesers samples are clearer, the contrast of internal pores and particle interfaces is significantly improved, and the noise in the background area is effectively suppressed, providing a high-quality image foundation for the accurate extraction and stable matching of subsequent feature points.
[0092] Step 52: Sample Feature Point Extraction and Matching
[0093] Local zero-mean normalization correction is applied to the preprocessed image to eliminate interference from uneven illumination and regional gray-level shifts, thereby improving the robustness of the descriptor. The Sobel operator is used to calculate the gray-level gradient and magnitude pixel by pixel to capture gray-level abrupt changes such as edges and defects. Combined with adaptive thresholding, flat and low signal-to-noise ratio regions are removed to construct an initial set of candidate feature regions.
[0094] Based on this, SIFT algorithm keypoint detection and descriptor extraction are performed in the effective regions of the two sets of projected images. First, scale-space extremum detection is used to screen initial keypoints, and high-contrast feature points are located as candidate keypoints based on the strength of contrast response, while low-contrast and unstable edge response inferior points are eliminated. Then, for the retained keypoints, the gradient direction and magnitude distribution information in their local neighborhood are extracted to generate high-dimensional descriptor vectors with rotation and scale invariance.
[0095] Subsequently, based on the Euclidean distance metric of high-dimensional descriptors, initial matching of feature points is performed to generate candidate matching pairs; further, the Lowe distance ratio (the ratio of the nearest neighbor to the second nearest neighbor) constraint is introduced to eliminate ambiguous mismatches, and finally obtain high-confidence effective feature matching pairs.
[0096] In this example, refer to Figure 4 The image shows a CT image after feature points were extracted from a Maltesers sample using the aforementioned strategy. As can be seen, after extraction using the SIFT algorithm and subsequent spatial uniformity constraint and gray-level variance quality screening, feature points are concentrated in areas with significant gray-level gradients, such as the sample's edge contours, internal pores, and particle interfaces. No obvious feature points are found in homogeneous areas with gentle gray levels or in the background. Furthermore, all effective feature points are located within the sample. This result demonstrates that the feature extraction and matching strategy employed in this step can effectively lock in high-value structural information within the sample, accurately remove background and redundant interference features, and provide reliable and spatially uniformly distributed marker points for subsequent cross-frame feature matching and sub-pixel offset calculation.
[0097] Step 53: Automatically activate the adaptive feature point quantity adjustment mechanism when feature points are sparse.
[0098] An adaptive adjustment mechanism is implemented: when feature points are sparse or insufficient, the Lowe threshold is automatically lowered to supplement stable feature points. In this example, the Maltesers sample has rich texture and no sparsity is observed; the Lowe threshold remains at its default value, and the number of feature points meets the accuracy requirements. Simultaneously, the neighborhood gray-level variance and gradient direction consistency criteria are comprehensively applied, combined with non-maximum suppression to eliminate redundant weak feature points, retaining only the feature points with the highest response intensity in the neighborhood. This achieves feature sparsity and high-quality selection, ensuring a uniform distribution of feature points across images and providing reliable support for subsequent matching and displacement calculations.
[0099] Step 6: Based on the effective feature matching pairs extracted from each projection viewpoint, calculate the initial offset of the second projection image relative to the corresponding first projection image under that viewpoint; then calculate the sub-pixel offset of the second projection image relative to the corresponding first projection image in the X and Y orthogonal directions using a sub-pixel level refinement algorithm. This step employs a two-stage strategy of "coarse estimation + fine registration". First, the initial offset is obtained through robust fitting, and then it is refined to sub-pixel accuracy through the ECC algorithm, providing accurate displacement parameters for subsequent global drift interpolation and image correction.
[0100] Coarse offset estimation: Based on the effective feature matching pairs extracted from each projection viewpoint, the RANSAC algorithm combined with the translation model is used for robust fitting, the interior point matching set is screened, and the median of all interior point displacements is used as the image coarse offset, which effectively suppresses outlier matching interference and improves the noise resistance and reliability of coarse registration.
[0101] Subpixel fine registration: Using the coarse offset result as the initial value for iteration, the ECC algorithm is driven to maximize the normalized cross-correlation function between the reference image and the image to be registered. Through iterative optimization, it converges to subpixel-level displacement, achieving a gradual approximation from pixel-level coarse matching to subpixel-level precise matching. The final output is the bidirectional subpixel offset in the X and Y directions for each viewpoint. , This provides accurate displacement parameters to support subsequent focus drift modeling and image correction.
[0102] Step 7: Implement the registration process based on the Fourier domain translation method, calculate the deviation of the single registration, and then introduce a closed-loop iterative registration and error verification mechanism to iteratively optimize the sub-pixel offset of the second projection image under each projection viewpoint; and extend this optimization process to all second projection images corresponding to each first projection image in the reference projection sequence, and solve the sub-pixel offset of each viewpoint one by one.
[0103] Because the Maltesers sample contains various structures such as the sugar shell-chocolate filling interface, pores, and particles, there are continuous temporal focus drift, weak thermal deformation disturbances, and projection image noise interference during slow scanning of micro-CT. Single-round coarse and fine registration solutions are prone to residual sub-pixel deviations. At the same time, feature matching is prone to small inter-frame grayscale fluctuations in local sparse regions. Relying solely on a single offset solution cannot meet the requirements for high-precision registration. Therefore, an adaptive iterative mechanism needs to be introduced to continuously optimize the offset parameters and compress residual errors.
[0104] The X and Y bidirectional sub-pixel offsets under a single viewpoint were calculated. , Afterwards, it is necessary to re-examine the residual offset of feature matching. First, based on the sub-pixel offset parameters of each viewpoint, lossless registration of the slow-scan image is completed in the Fourier domain. When performing translation registration of the second scan image using the Fourier domain method, a two-dimensional Fourier transform is first performed on the original drift image to obtain the full-domain spectrum. According to the Fourier translation theorem, a linear phase compensation factor is superimposed to accurately cancel the inter-frame drift deviation. Then, the corrected image is reconstructed through inverse Fourier transform, preserving the image's grayscale features and microstructure throughout the process.
[0105] The frequency domain registration process is as follows: First, perform a two-dimensional Fourier transform on the original drifted second projection image to obtain the full-domain spectrum information; substitute the corresponding sub-pixel offset of the viewpoint to construct an accurate phase compensation factor to offset the inter-frame temporal drift deviation; finally, reconstruct the image through inverse two-dimensional Fourier transform to obtain a distortion-free corrected projection image that eliminates drift error, thus achieving lossless registration.
[0106] After Fourier domain translation and registration of the second scanned projection image from this perspective, an iterative mechanism is triggered. Taking the Maltesers sample as an example, its sugar shell-chocolate filling interface region has significant features, but the internal homogeneous region features are sparse, and residual offsets may still exist after a single registration. During iteration, a registration quality check is performed with a precision threshold of 1 pixel. For cases where residual offsets exceed the standard or image correlation decreases, adaptive iterative optimization is implemented based on the number of effective interior points: when interior points are sparse, the Lowe threshold is relaxed to expand the matching set; when interior points are sufficient, dense cross-correlation is used for fine correction. At the same time, a fault-tolerant constraint mechanism is added to reset the initial value of abnormal offsets in the hyperphysical region and re-drive the ECC algorithm to converge, effectively suppressing noise interference and solution divergence in sparse feature scenarios, and ensuring high-precision and robust convergence of the full-view drift solution.
[0107] To prevent the algorithm from getting stuck in infinite iteration and to improve the stability of engineering deployment, a maximum number of iterations is set. When the number of iterations is ≥10, the iteration operation is forcibly terminated. This strategy can effectively prevent image distortion caused by overcorrection, and steadily improve the registration accuracy and inter-frame correlation of the drifted images of the second scan of the Maltesers sample, thereby providing a reliable drift-free projection sequence for subsequent metal artifact repair and high-precision reconstruction.
[0108] Step 8: Using the verified sub-pixel offset as the discrete control node, construct a continuous mapping model of drift amount as the angle changes based on bicubic spline interpolation, accurately fit its nonlinear evolution law, and output the continuous drift amount of the whole view. The Maltesers sample has a complex structure, with significant differences in the projection characteristics of its sugar shell, chocolate layer, and internal pores at different angles. The focus drift exhibits a continuous and gradual nonlinear evolution with the scanning angle, without any abrupt step shifts. In this example, independent bicubic spline interpolation models are constructed for the X and Y directions respectively. Piecewise third-order polynomials are used to fit the mapping relationship between the drift amount in each dimension and the angle, effectively avoiding bidirectional coupling interference.
[0109] Using the sub-pixel offsets in the X and Y directions of each discrete viewpoint obtained in Step 6 as reference control points, bicubic spline interpolation models are established to perform batch simulations for all unsampled viewpoints across the entire viewing range, quickly obtaining the continuous sub-pixel drift parameters corresponding to all projected viewpoints in the slow-scan sequence. The modeling process constrains the function values, first and second derivatives of all control points to be continuous throughout the entire domain. Optimal coefficients are solved by combining boundary conditions to ensure a smooth transition in drift rate and curvature, eliminating numerical abrupt changes and oscillatory distortions caused by sparse sampling, and constructing a smooth, continuous second-order drift function from 0° to 360°.
[0110] Reference Figure 5Under the "fast scan to establish benchmark, slow scan to collect data" timing of this invention, the bidirectional offset of the X and Y axes of the Maltesers sample remained at a low level throughout the 0°~360° scanning range. The curve trend was gentle, without obvious fluctuations or abrupt changes, indicating that the X-ray source focus was in a good thermal stability state during the benchmark scanning stage, and the drift amplitude was effectively suppressed within an acceptable range.
[0111] Reference Figure 6 In the traditional "slow scan first, then fast scan" timing sequence, the offset continues to accumulate with the angle during the long slow scan process, and the amplitude in both X and Y directions increases significantly. The curve shows obvious fluctuations and undulations, reflecting that the focus drift continues to intensify as the scanning process progresses.
[0112] The comparison of the two figures provides intuitive evidence that the timing optimization strategy of this invention can effectively suppress the cumulative effect of focus drift with angle, ensuring that the reference projection sequence is acquired under low drift steady-state conditions, and providing a reliable reference benchmark for subsequent high-precision drift modeling and sub-pixel correction.
[0113] Step 9: Frame-by-frame determination based on the full-view drift amount: If the drift is below the preset drift threshold, the original image is retained; if the drift reaches or exceeds the preset drift threshold, Fourier domain translation registration is performed to complete the adaptive correction processing of all projected images and construct a complete set of projected images after focus drift compensation. For Maltesers samples, the interface between the sugar shell and the chocolate layer, as well as the fine structures such as internal pores, are sensitive to grayscale changes. Excessive image translation correction can easily lead to significant noise in the reconstructed 3D slice images. Therefore, a correction-free mechanism is set up: when the X and Y bidirectional offsets are less than 0.5 pixels, the drift is considered negligible, and the original image is used directly to avoid the additional noise introduced by correction; images exceeding the standard are included in the correction range. Images requiring correction are subjected to Fourier domain lossless registration based on the interpolation drift parameters of each viewpoint: a two-dimensional Fourier transform is performed on a single frame image to obtain the spectrum, and linear phase compensation is applied in combination with the corresponding viewpoint offset to cancel the spatial shift caused by focus drift. Then, the aligned corrected image is reconstructed through inverse transformation, preserving the original grayscale and fine structural features of the Maltesers sample throughout the process, effectively ensuring the fidelity of key information such as pores and interfaces in subsequent 3D reconstruction.
[0114] Step 10: Perform 3D reconstruction based on the complete projected image set after focus drift compensation and geometric parameters to obtain the corrected 3D slice model; compare it with the uncorrected reconstructed model, and evaluate the optimization degree and effectiveness of the correction method by quantifying image quality indicators.
[0115] Using the drift-corrected full-view accurate projection sequence as input, and combining calibration geometric parameters such as SOD, SDD, and detector attitude, the FDK cone-beam filter backprojection algorithm is employed to perform 3D reconstruction. Through ramp filtering, cone-beam cosine weighting, and voxel accumulation backprojection optimization, edge blurring, structural distortion, and ring artifacts caused by focus drift are effectively suppressed, resulting in a high-fidelity 3D voxel model.
[0116] To objectively quantify the correction effect, this invention conducts quantitative evaluation from four indicators: edge gradient, contrast-to-noise ratio, image information entropy, and gray-level fluctuation in artifact regions, which respectively characterize image sharpness, noise suppression capability, detail preservation capability, and artifact suppression level.
[0117] Reference Figures 9-12 Quantitative testing results show that, compared with the uncorrected image, the corrected image obtained by this method has improved edge sharpness by 66.7%, contrast-to-noise ratio by 40.5%, image clarity by 3.59%, and artifact suppression by 64.3%, with all core indicators showing significant optimization.
[0118] In summary, this invention employs a temporal strategy of "first establishing a baseline with fast scanning, then acquiring data with slow scanning," using low-drift fast-scan images as a baseline to calibrate the full-temporal focus drift of slow-scan images. It combines SIFT algorithm feature matching with ECC algorithm iterative solution of sub-pixel offsets at each viewpoint, achieving lossless registration through Fourier domain phase shift. By relying on bicubic spline interpolation to deduce unsampled viewpoint drift parameters, it completes continuous drift compensation across the entire domain. Ultimately, it forms a complete closed-loop system of "temporally optimized acquisition—global drift modeling—lossless sub-pixel registration—high-precision 3D reconstruction," effectively improving the imaging accuracy and detail restoration capabilities of micro-CT for precision samples.
[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A non-destructive method for correcting focus drift in micro-CT by combining temporal optimization and Fourier domain translation, characterized in that: include The micro-CT imaging system is initialized and its geometric parameters are calibrated before the sample to be tested is pre-scanned. A full-view low-drift scan was performed using preset parameters, and the scan parameters were determined after validity verification and parameter optimization. Based on the scanning parameters, the first and second scans are performed at a fixed timing sequence to establish a low-drift reference projection sequence and a second projection sequence containing focus drift error, respectively. Using the projection viewpoint of the low-drift reference projection sequence as an index, extract the second projection image with the same viewpoint from the second projection sequence at a fixed step size to construct the matching dataset corresponding to the viewpoint. Perform normalization preprocessing on the low-drift reference projection sequence and the matching dataset, extract high-contrast feature points and perform feature matching, and filter to obtain effective feature matching pairs; The initial offset is solved based on effective feature matching pairs, and the sub-pixel offsets of the second projected image relative to the first projected image in the X and Y directions are obtained through refined calculation. Based on Fourier domain translation registration and calculation of registration residuals, combined with closed-loop iterative registration and error verification, the sub-pixel offset of the second projection image relative to the first projection image from each viewpoint is iteratively optimized. Based on the qualified sub-pixel offset obtained by verification, the continuous drift under the full projection view is obtained by bicubic spline interpolation. Based on the full-view drift, the second projection sequence is registered and corrected frame by frame to construct a complete projection image set after focus drift compensation. Three-dimensional reconstruction is performed based on the complete projected image set after focus drift compensation and geometric parameters to obtain a corrected three-dimensional slice model.
2. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain shift as described in claim 1, characterized in that: The scanning parameters are determined through validity verification and parameter optimization, including... Calculate the relative offset between the projected images of the first and last frames and compare it with a preset stability threshold. If the relative offset is less than the preset stability threshold, the current acquisition is determined to be valid. Then, a low-drift scan is performed again with the optimized scanning parameters and a verification is implemented to confirm that the focus state remains stable after the parameter adjustment. Based on this, the optimal first scanning parameters and its corresponding scanning time are determined. If the relative offset exceeds the preset stability threshold, the current acquisition is deemed invalid. The pre-exposure time of the X-ray source is extended, and after the micro-CT imaging system has been further thermally stabilized, a low-drift scan is re-executed to acquire and generate a new low-drift reference projection sequence. The above verification process is repeated until the focus drift meets the stability condition, and the optimal first scan parameters and their corresponding scan time are determined.
3. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 2, is characterized in that... Based on the scanning parameters, the first and second scans are performed at a fixed timing sequence to establish a low-drift reference projection sequence and a second projection sequence including focus drift error, respectively. After confirming that the micro-CT imaging system met the low-drift scanning conditions, the scanning steps were then executed according to a fixed timing sequence: The sample to be tested is precisely reset to the preset starting angle to ensure that the initial spatial pose of the two scans is consistent and that the sample to be tested can rotate completely twice. The first scan is performed using the first scan parameters to obtain a low-drift reference projection sequence; Keeping the order unchanged, and under the premise of maintaining the same scanning start position and angle coverage, a second scan is performed with the second scanning parameters to obtain a second projection sequence containing focus drift error; Wherein, the angle sampling step size set by the second scanning parameter is smaller than that of the first scanning parameter; The order of the first scan and the second scan is a predetermined fixed timing sequence.
4. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 3, is characterized in that... Perform normalization preprocessing on the low-drift reference projection sequence and the matching dataset, including... First, global grayscale normalization is performed to linearly map the image grayscale value range to a unified interval, thus unifying the grayscale benchmark of each projected image. Then, a contrast-limited adaptive histogram equalization algorithm is used to enhance the local microstructure information of the sample under test. Next, bilateral filtering is used to suppress random noise by combining spatial neighborhood and grayscale similarity as dual weights. Finally, local zero-mean standardization correction is performed on the projected image.
5. The non-destructive correction method for focus drift in micro-CT combining time-series optimization and Fourier domain translation as described in claim 4, characterized in that: The filtering process yields valid feature matching pairs, including Within the selected area of the projected image, the Sobel operator is used to solve the horizontal and vertical gray-level gradients and magnitudes pixel by pixel. The effective detection area is delineated based on the boundary, defects and interface of the sample to be tested, and the structural regions with significant and stable gradients, and the background and low signal-to-noise ratio interference are eliminated. Within the effective detection area, the scale-invariant feature transformation algorithm is used to locate high-contrast feature points as candidate key points, and local neighborhood information is extracted for the candidate key points to generate corresponding high-dimensional descriptors. Based on the Euclidean distance metric of high-dimensional descriptors, initial matching of feature points is performed to filter out candidate matching pairs; Then, the distance ratio between the nearest neighbor and the second nearest neighbor is introduced as a constraint to eliminate ambiguous mismatches and obtain high-confidence effective feature matching pairs.
6. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 5, is characterized in that... The process of obtaining effective feature matching pairs through screening also includes an adaptive feature point quantity control mechanism, including... The total number of valid feature points in the selected area of the projected image is counted in real time, and it is determined whether the current number of feature points meets the requirements of high-precision matching operation. If the current number of feature points is determined to be insufficient, the Lowe threshold is adaptively reduced to relax the constraints of feature selection in order to extract more stable and effective feature points. A non-maximum suppression algorithm is also introduced to eliminate redundant weak feature points caused by the relaxation of the threshold, so as to achieve a spatially uniform distribution of feature points in the first and second projected images.
7. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 6, characterized in that, The calculation process for the sub-pixel offset includes: Based on the effective feature matching pairs obtained from screening under any projection viewpoint, the random sampling consensus algorithm is used to robustly fit the global image translation model, and the median of the point displacements within the model is used as the initial offset under that viewpoint. Using the initial offset as the initial value for iteration, an enhanced correlation coefficient algorithm is introduced to perform sub-pixel level refinement calculation. By maximizing the inter-frame normalized correlation coefficient, the image grayscale mapping relationship is accurately converged. The bidirectional sub-pixel offset in the X and Y orthogonal directions under each projection viewpoint is obtained by solving the problem from each viewpoint.
8. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 7, characterized in that, The Fourier domain translation registration includes Perform a two-dimensional Fourier transform on the second projection image containing focus drift to convert the spatial domain image signal into a frequency domain spectrum signal; Based on the sub-pixel offset obtained from the corresponding viewpoint, a frequency domain linear phase compensation factor is constructed to perform phase offset compensation on the spectral signal, thereby offsetting the residual drift deviation between image frames. The compensated spectral signal is reconstructed into a spatial image by inverse two-dimensional Fourier transform, and the corrected projection image is obtained.
9. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation as described in claim 8, characterized in that, The closed-loop iterative registration and error verification include: For the second projected image from each projection viewpoint, Fourier domain subpixel translation is performed based on the current subpixel offset, and error verification is performed based on the residual offset between the translated image and the corresponding first projected image. If the residual offset meets the preset registration accuracy condition and the current sub-pixel offset is within the preset physical drift constraint range, then the registration is deemed qualified, and the current offset is used as the final sub-pixel offset for that viewpoint. If the residual offset does not meet the preset registration accuracy condition, the feature points of the first projected image and the corresponding second projected image are re-selected and matched, the sub-pixel offset is re-solved based on the updated effective feature matching pair, and Fourier domain translation and error verification are performed again. The above process is repeated until the residual offset meets the preset registration accuracy condition and the sub-pixel offset is within the preset drift constraint range. Then the iteration is terminated and the current offset is output as the final result. If the residual offset still does not meet the preset registration accuracy condition when the number of iterations reaches the preset maximum number of iterations, then the sub-pixel offsets in the current X and Y directions will be halved as the final result.
10. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation according to claim 9, characterized in that: The continuous drift under the full projection view obtained by bicubic spline interpolation includes... The projection viewpoints that have passed iterative verification and whose parameters have been solved are selected as interpolation reference control points, and the number of reference control points is consistent with the total number of frames of the first projection image. Establish a one-to-one mapping relationship between the scanning angle of each reference control point and the corresponding sub-pixel offset in the X and Y directions, and construct an angle-offset discrete dataset; The discrete drift data of the reference control points are nonlinearly fitted based on the bicubic spline interpolation algorithm to construct a global drift function in which the drift amount changes continuously with the scanning angle. By using the fitted global drift function, high-density sampling and interpolation are performed on the remaining projection viewpoints that did not participate in the iterative solution to obtain a smooth, continuous drift amount covering all scanning angles.
11. The non-destructive correction method for focus drift in micro-CT combining temporal optimization and Fourier domain translation according to claim 10, characterized in that: The frame-by-frame registration and correction of the second projection sequence based on the full-view drift includes Based on the sub-pixel offset corresponding to the second projected image of each frame, calculate the actual drift amount of the current frame, and compare the actual drift amount with the preset drift threshold for determination; If the actual drift is lower than the preset drift threshold, the current frame imaging accuracy is determined to meet the requirements, and the original second projection image is directly retained as the standardized projection image. If the actual drift exceeds the preset drift threshold, it is determined that the current frame has excessive drift. Fourier domain translation registration is performed on the image to be corrected in this frame to eliminate the imaging drift deviation and obtain the standardized projection image after drift correction.
Citation Information
Patent Citations
A Micro / Nano CT Focus Drift Correction Method Based on Adaptive Projection Image Feature Region Matching
CN109685877B
A micro-nano CT focus drift correction method based on adaptive projection image feature region matching
CN109685877A
Automated application of drift correction to sample studied under electron microscope
CN114041200A
