A dual-parameter adaptive scatter suppression industrial x-ray inspection method
Patent Information
- Application Number
- CN202610873147.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-17
- Publication Date
- 2026-09-04
AI Technical Summary
[0004]针对成像散射抑制问题，现有技术主要包括以下几类：（1）硬件防散射栅格法，成本高且对小焦点、近距离成像场景适应性差；（2）蒙特卡洛模拟法，精度高但计算量巨大，难以用于实际工程；（3）卷积核估计法，固定尺度核难以适应叠片电池等工件局部厚度与密度的剧烈变化；（4）深度学习法，依赖大量配对训练数据且泛化性受限
[0028] (i) High adaptive accuracy, adaptable to complex workpieces with varying thicknesses
Smart Images
Figure CN122689840A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial X-ray imaging and inspection technology, and in particular to an adaptive scattering suppression industrial X-ray inspection method for digital radiography (DR) and computed tomography (CT) images based on a GPU-accelerated area array detector, which is suitable for industrial non-destructive testing scenarios of complex workpieces with varying thicknesses. Background Technology
[0002] In the field of industrial X-ray inspection, an imaging system typically consists of an X-ray source, the workpiece being inspected, an area array detector, and an industrial X-ray imaging system, such as... Figure 1 As shown, the industrial X-ray imaging system includes a X-ray machine controller, an imaging controller, and an operation monitoring controller. X-ray photons penetrate the workpiece being inspected, are received by a detector, and converted into a projected image. The three-dimensional volume data of the workpiece is then reconstructed using algorithms such as Feldkamp-Davis-Kress (FDK). Ideally, the detector only receives primary photons propagating in a straight line, whose intensity satisfies the Beer-Lambert attenuation law. However, when X-rays pass through matter, they interact with atoms through Compton scattering and Rayleigh scattering, causing photons to change direction and still fall onto the detector pixels, forming scattered signals. Due to their large field of view and two-dimensional flat structure, area array detectors cannot naturally suppress scattering through geometric collimation like linear array detectors. Therefore, the acquired projected image information is actually a superposition of primary and scattered photons, leading to decreased contrast, grayscale distortion, increased artifacts, and obscuring of workpiece details in the image data information of the imaging controller.
[0003] Due to their advantages of fast imaging speed and large field of view, area array detectors are widely used in real-time DR imaging and cone-beam computed tomography (CBCT) 3D reconstruction. One typical application is the stacked power battery of electric vehicles. The article "Application Research of Computed Tomography Technology in Lithium-ion Battery Detection" by Ma Tianyi et al., Journal of Chongqing University of Technology (Natural Science), Vol. 34, No. 2, February 2020, discloses that this type of battery is composed of dozens or even hundreds of layers of alternating positive and negative electrode sheets and separators. Its shape is mostly a large-area flat structure with significant differences in thickness and planar dimensions. Furthermore, it contains various materials with significant density differences, such as aluminum shell, copper foil, aluminum foil, and active materials, making it a typical complex variable-thickness workpiece. At the same time, the interlayer spacing of its electrode sheets is only tens of micrometers, placing extremely high demands on the detection of defects such as electrode alignment, tab welding quality, and internal foreign matter. Any tiny defect can lead to thermal runaway or even a safety accident. These characteristics result in high scattering and prominent artifacts in the DR and CT images of stacked batteries, severely limiting defect detection capabilities.
[0004] To address the problem of scattering suppression in imaging, existing technologies mainly include the following categories: (1) Hardware anti-scattering grid method, which is costly and poorly adaptable to small focal points and close-range imaging scenarios; (2) Monte Carlo simulation method, which is highly accurate but computationally intensive and difficult to apply to practical engineering; (3) Convolution kernel estimation method, which is difficult to adapt to the drastic changes in local thickness and density of workpieces such as stacked batteries with fixed-scale kernels; (4) Deep learning method, which relies on a large amount of paired training data and has limited generalization ability. In addition, for cone-beam CT image reconstruction, the traditional approach is to perform scattering correction on the reconstructed three-dimensional volume data of the workpiece, but once scattering artifacts enter the reconstruction process, they will be amplified and diffused along the ray path, and the post-image correction effect is limited. More importantly, existing scattering estimation methods often adopt pixel-level iterative strategies, and processing a single high-resolution projection takes several minutes. For the hundreds to thousands of projections required for CT images, the time consumption is extremely staggering, which seriously restricts the engineering application of industrial X-ray detection; at the same time, due to the fixed single parameter of the method, the effect is often poor for images at different scales.
[0005] Therefore, there is an urgent need for a scattering suppression method that can adaptively adjust the local scale and has high computational efficiency, and to embed it into the front end of the FDK reconstruction process to achieve integrated scattering correction from DR images to CT images. Therefore, this invention proposes a dual-parameter adaptive scattering suppression industrial X-ray detection method. Summary of the Invention
[0006] In view of the above-mentioned technical problems, this invention proposes a two-parameter adaptive scattering suppression method for industrial X-ray inspection based on GPU-accelerated area array detector DR and CT images, namely the CUDA-GPU method. The CUDA-GPU method introduces two-parameter scattering suppression related to local grayscale scale and uniformity into the DR image, which can adaptively adjust the scattering estimation kernel according to the characteristics of different regions of complex inspected workpieces, overcoming the limitations of fixed kernel methods. Furthermore, with the parallel acceleration of the GPU in the imaging controller, the processing time of a single high-resolution projected image is reduced from minutes to seconds, meeting the engineering processing requirements of batch projection of industrial CT images. Simultaneously, it does not rely on paired training data and has good generalization and engineering applicability for complex variable-thickness workpieces of different models and batches. The CUDA-GPU method of this invention aims to transform scattering estimation from "pixel-by-pixel" to "layer-by-layer full-image" through rapid summation of integral images, geometric series-scale scattering estimation, two-parameter robust uniformity criteria, and logarithmic domain multiplicative scattering suppression processing, achieving a hundredfold speed improvement while maintaining adaptive scale accuracy. This method is also used as a pre-processing step for FDK reconstruction, suppressing scattering artifacts in CT images from the source.
[0007] This invention discloses a dual-parameter adaptive scattering suppression method for industrial X-ray detection, belonging to the field of industrial X-ray imaging and detection technology, and specifically includes the following steps:
[0008] Step 1: Construct the integral map in the GPU;
[0009] The image to be processed is obtained, and a symmetrical inverse filling is performed on the image to be processed to obtain a filled image. The square image is calculated on the filled image by parallel scanning on the GPU to obtain a first-order integral image and a second-order integral image.
[0010] Step 2: Calculation of the two-parameter uniformity threshold;
[0011] A window is selected centered on each pixel of the image to be processed. The first-order integral image and the second-order integral image are used to calculate the baseline mean, standard deviation and local coefficient of variation in the neighborhood of each pixel. The median and the corresponding absolute deviation of the median are calculated. A two-parameter uniform threshold is constructed based on the local coefficient of variation and the absolute deviation of the median.
[0012] Step 3: Multi-scale parallel scattering estimation;
[0013] A geometric series scale set is preset, and the baseline mean, standard deviation and local coefficient of variation at each scale are calculated with the help of integral image. The relative difference energy between adjacent scales is constructed, and the optimal kernel for scattering estimation is determined based on the relative difference energy and the two-parameter uniform threshold to obtain the scattering estimation pixel image.
[0014] Step four: Suppress logarithmic domain scattering;
[0015] A logarithmic domain scattering suppression model is constructed based on the X-ray Beer-Lambert attenuation law. Logarithmic and linear transformations are performed on the image to be processed, and percentage grayscale remapping is performed on the transformed image to obtain a scattering suppression image.
[0016] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0017] After performing steps one to four on all projection sequences, a scattering-suppressed projection sequence is obtained. The scattering-suppressed projection sequence is then weighted, filtered, and back-projected using GPU parallel scanning to reconstruct tomographic image data.
[0018] Further, in step one, the image to be processed is a DR image or a CT projection image, and the maximum fill radius of the symmetrical inverse fill is denoted as... The first-order integral graph and the second-order integral graph are completed by parallel row scanning and parallel column scanning.
[0019] Further, in step two, the two-parameter uniform threshold is constructed as follows: ,in For the coefficient of variation weight, The absolute deviation weight is used. It is the minimum value. The coefficient of variation is the local variation. This represents the median and its corresponding absolute deviation. This is the benchmark mean within the neighborhood.
[0020] Further, in step three, the geometric series scale set contains multiple adaptive kernels of different scales; the relative difference energy between adjacent scales is used to characterize the stability of the scale space, and the scattering kernel expansion stops when the local mean no longer changes significantly; if there is no scale that meets the conditions, the largest kernel is taken as the optimal kernel for scattering estimation of that pixel.
[0021] Further, in step four, the pixels of the image to be processed are reduced to logarithmic transform pixel images after logarithmic domain scattering suppression. The logarithmic transform pixel images are then reduced to linear transform linear-logarithmic transform pixel images after nonlinear compensation coefficients.
[0022] Further in step four, the percentage grayscale remapping is performed by stretching the linear-logarithmic transformed pixel image at the 1% and 99% grayscale quantiles, wherein the grayscale quantiles are implemented using GPU parallel radix sorting.
[0023] Further, in step five, the weighted filtering back projection uses a Ram-Lak ramp filter, and the sampling of the projection coordinates is completed through hardware bilinear interpolation of the GPU texture memory.
[0024] Furthermore, the method of the present invention is applied to an industrial X-ray imaging system, wherein the imaging controller of the system is equipped with GPU hardware, and the method is embedded in the imaging controller as a pre-image processing step for FDK reconstruction.
[0025] Furthermore, the image to be processed in this invention is acquired by an area array detector, and the workpiece to be inspected is a complex workpiece with varying thickness, including stacked power batteries, cylindrical batteries, castings, or workpiece welds.
[0026] Furthermore, the scattering estimation of this invention achieves full-layer parallel computation through fast summation of integral images, reducing the processing time of a single high-resolution projected image from minutes to seconds, and improving the overall efficiency by more than two orders of magnitude compared to traditional pixel-level iterative methods.
[0027] Compared with the prior art, the technical advantages of the method of the present invention are as follows:
[0028] (i) High adaptive accuracy, adaptable to complex workpieces with varying thicknesses
[0029] This invention introduces a two-parameter robust uniformity criterion, simultaneously considering both local grayscale and local uniformity information. It can automatically select the optimal scattering estimation scale based on local image features. Compared to traditional fixed-kernel convolution methods, this mechanism can accurately characterize the differences in scattering distribution at thickness transition zones, electrode edges, and multi-material interfaces in complex workpieces with varying thicknesses. It effectively avoids overcorrection at edges and undercorrection in flat areas, significantly improving the accuracy of scattering estimation and image contrast.
[0030] (ii) High computational efficiency, meeting the needs of industrial batch processing
[0031] This invention employs an integral image fast summation and geometric series scaling mechanism to reduce the computational complexity of scattering estimation from pixel-level iteration to layer-level full-image computation. Furthermore, it leverages the parallel acceleration of a CUDA-based graphics processor (i.e., the CUDA-GPU method), reducing the processing time for a single high-resolution projection from minutes to milliseconds. This overall efficiency is approximately two orders of magnitude higher than traditional methods. This feature enables it to handle batch processing tasks of hundreds to thousands of projections in cone-beam CT images, meeting the stringent real-time and throughput requirements of industrial environments.
[0032] (III) Integrated DR-CT correction to suppress reconstruction artifacts at the source
[0033] This invention places image scattering suppression processing before FDK reconstruction as a preprocessing step in the projection domain. This avoids the problem of scattering errors being amplified and diffused along the ray path by backprojection in traditional volume data post-correction methods, thus eliminating the impact of scattering artifacts on CT image reconstruction at the source. Furthermore, this method is based on a physical and image statistical model, requiring no paired training data, and exhibits good generalization ability for complex workpieces of different models, achieving integrated, high-precision, and high-efficiency scattering correction from DR images to CT images. Attached Figure Description
[0034] Figure 1 This is a system block diagram for industrial X-ray inspection used for non-destructive testing.
[0035] Figure 2 This is a flowchart of the dual-parameter adaptive scattering suppression industrial X-ray detection method of the present invention.
[0036] Figure 3 This is a schematic diagram illustrating the principle of CPU core serial computation and GPU integral graph parallel computation in this invention.
[0037] Figure 4 These are comparison images of stacked batteries before and after processing using the method of this invention.
[0038] Figure 5 These are comparison images of a cylindrical battery before and after treatment using the method of this invention.
[0039] Figure 6 This is a comparison chart of the running time of the layered full-image parallel method in this invention and the traditional pixel-by-pixel method.
[0040] Figure 7 These are comparison images of castings before and after treatment with the method of this invention.
[0041] Figure 8 The scattering suppression method for castings is applied to the center row grayscale curve of the image.
[0042] Figure 9 These are comparison images of the workpiece weld before and after treatment using the method of this invention.
[0043] Figure 10 This is a method for suppressing the scattering of the workpiece weld seam, applied to the center row grayscale curve of the image. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. The examples of the parameters listed are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
[0045] See Figure 1 The industrial X-ray inspection system for non-destructive testing shown in this invention adds a Graphics Processing Unit (GPU) hardware to the imaging controller. The GPU hardware embeds a two-parameter adaptive scattering suppression industrial X-ray inspection method for DR and CT images using an accelerated area array detector; this method is referred to as the CUDA-GPU method. The CUDA-GPU method processes the projected image information before FDK reconstruction, such as... Figure 2 As shown. The CUDA-GPU method of this invention aims to transform the projection image (i.e., the image information to be processed) output by the area array detector into a "layer-by-layer full-image" scattering estimation by rapidly summing integral images, estimating scattering at geometrical scales, applying a two-parameter robust homogeneity criterion, and suppressing logarithmic multiplicative scattering, thereby maintaining adaptive scale accuracy and achieving a hundredfold speed improvement. Furthermore, this method is used as a pre-processing step for FDK reconstruction, suppressing scattering artifacts in CT images from the source.
[0046] See Figure 1 , Figure 2 As shown, this invention proposes a dual-parameter adaptive scattering suppression method for industrial X-ray detection based on GPU-accelerated area array detectors for DR and CT images, namely the CUDA-GPU method; it includes the following steps:
[0047] Step 1: Construct the integral map in the GPU;
[0048] An industrial X-ray inspection system is used to acquire image information of the workpiece under inspection. The imaging controller can obtain the image information to be processed, which can be a DR image or a CT image.
[0049] In the GPU, the image to be processed is defined as... ,in, Where is the X-ray projection angle, and N is the number of projection frames. middle For pixels on the X-axis in the image coordinate system, The pixel point on the Y-axis in the image coordinate system. The image size is H is the total length of the image to be processed, and W is the total width of the image to be processed. Figure 3 As shown.
[0050] The function of applying image fill size Image to be processed Perform symmetrical inverse padding to obtain the filled image, denoted as . The function mentioned above , For the fill radius, In this embodiment, the maximum value of the fill radius is denoted as... ,and In the middle of For pixels on the X-axis in the image coordinate system, is the pixel point on the Y-axis in the image coordinate system.
[0051] Due to the image to be processed Image after image filling Pixels will shift in the image coordinate system. Therefore, the CUDA-GPU method uses parallel scanning to complete parallel row and column scanning, while simultaneously calculating the corresponding integral image. This is applied to the filled image. Performing square graph calculations yields a first-order integral graph. On the other hand, we obtain the second-order integral graph. . In The coordinates of the point on the X-axis in the image coordinate system. Let be the coordinate point on the Y-axis in the image coordinate system. This invention uses square graph calculations to solve the problem of extremely slow calculation times for square-level operations in traditional methods.
[0052] Step 2: Estimate the coefficient of variation and calculate the two-parameter uniformity threshold using the benchmark mean;
[0053] Uniformity threshold calculation based on a two-parameter method using the coefficient of variation and the absolute deviation of the median. (Image to be processed) Corresponding pixels Select the center The window uses the integral image to calculate the local grayscale scale, baseline mean, and standard deviation of each pixel to calculate the local coefficient of variation matrix, as well as the median and corresponding absolute deviation, and uses the above values to construct a two-parameter uniform threshold.
[0054] (1) Calculate the benchmark mean within the neighborhood. ,and . Indicates selection window Lower graph of first-order integral Perform summation.
[0055] (2) Calculate the standard deviation within the neighborhood. ,and . Indicates selection window The graph of the second-order integral below Perform summation.
[0056] (3) Local coefficient of variation ,and , This is a minimum value. The local coefficient of variation can effectively reflect relative fluctuations.
[0057] (4) Median and corresponding absolute deviation ,and , Indicates the image to be processed pixel values, This represents the median. It can effectively detect pulse noise and hot / cold spot interference, improving image clarity while suppressing image noise. Compared with traditional single-parameter methods, it is more suitable for large-scale industrial X-ray images.
[0058] In this invention, a two-parameter uniform threshold is constructed as follows: ,and ;in, For the coefficient of variation weight, The absolute deviation weight is used. It is a local minimum value, and takes the value of In this invention The value ranges from 0.8 to 1.2. The value ranges from 0.4 to 0.7.
[0059] In this step, a two-parameter uniformity threshold construction method is used to construct a uniform region in the image neighborhood by calculating the coefficient of variation and median deviation within the neighborhood. This method has stronger robustness compared to a single criterion.
[0060] Step 3: Geometric series set full-map estimation and multi-scale parallel scattering estimation;
[0061] In multi-scale parallel scattering estimation and stability optimal kernel selection, a pre-defined geometric series scale set is set as follows: , The first adaptive kernel scale, For the second adaptive kernel scale, The third adaptive kernel scale, The fourth adaptive kernel scale, This is the fifth adaptive kernel scale. For ease of explanation, we use... Representing five different scales, namely .
[0062] (1) Construct corresponding parameters for five different scales using integral plots. .
[0063] in This indicates the selection of an adaptive kernel scale at the current time. Calculate the benchmark mean within the neighborhood. This indicates the selection of an adaptive kernel scale at the current time. Select window Lower graph of first-order integral Perform summation. This indicates the selection of an adaptive kernel scale at the current time. Calculate the standard deviation within the neighborhood. This indicates the selection of an adaptive kernel scale at the current time. Select window The graph of the second-order integral below Perform summation. This indicates the selection of an adaptive kernel scale at the current time. Calculate the local coefficient of variation matrix.
[0064] (2) Construct the relative difference energy of the corresponding adjacent scales as ,and .in For located The baseline mean within the previous neighborhood. Calculate. It can effectively characterize the stability of scale space, that is, the expansion of scattering kernels stops when the local mean no longer changes significantly.
[0065] (3) The optimal kernel is denoted as ,and ;in, To ensure stability tolerance, this invention The value ranges from 0.01 to 0.03. It is a two-parameter uniform threshold.
[0066] In this invention, if the image to be processed The pixels do not have an optimal kernel that satisfies the conditions (i.e., none) When ), the fifth adaptive kernel scale will be used. The value is assigned to the image to be processed. This serves as the optimal kernel for its scattering estimation. Fifth adaptive kernel scale. The value of is less than 51, that is .
[0067] (4) Image to be processed Optimal kernel estimated by scattering After calculation, the estimated pixel image of scattering is obtained and denoted as . .
[0068] Step four: Suppress logarithmic domain scattering;
[0069] In this invention, logarithmic domain scattering suppresses image output and grayscale remapping. The image to be processed is constructed based on the X-ray Beer-Lambert attenuation law. pixels The logarithmic domain scattering suppression pixel image is obtained by performing a logarithmic transformation to obtain the logarithmically transformed pixel image. ,and , This is the scattering suppression coefficient. In this invention... The value ranges from 0.4 to 0.7.
[0070] Furthermore, to achieve linear representation, the pixel image after the linear-logarithmic transformation of the linear transformation is: ,and , This is a nonlinear compensation coefficient. In this invention... The value ranges from 0.8 to 1.2.
[0071] In this invention, the A higher value results in a clearer scattering suppression image, but may also lead to excessive sharpness of details on the inspected workpiece; conversely, a lower value better preserves the original contrast information of the image. Furthermore, constructing a suppression model using the X-ray attenuation law yields better physical properties and more closely resembles a true scatter-free image.
[0072] To prevent a decrease in the overall grayscale value of the image due to grayscale subtraction, this invention further modifies the pixel image after linear-logarithmic transformation. Perform percentage grayscale remapping to obtain the remapped pixel image. ,and .
[0073] Alternative express.
[0074] This represents the gray level quantile of 1% of the remapped pixel image.
[0075] This represents the 99th percentile of the grayscale value in the remapped pixel image.
[0076] This represents the gray level quantile of the image to be processed at 1%.
[0077] This represents the 99th percentile of the grayscale value of the image to be processed.
[0078] In this invention, the scattering-suppressed image processed by the GPU is sent back to the imaging controller, thereby completing the scattering suppression of the DR image and obtaining a high-quality X-ray image.
[0079] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0080] like Figure 2 The diagram shows that CT images are processed after the DR images are processed. All images to be processed are shown. After performing steps one through four, the scattering suppression projection sequence was obtained, denoted as... and scattering suppression pixel image .
[0081] The tomographic image data to be reconstructed is loaded from the GPU to the imaging controller for FDK backprojection accumulation, thereby suppressing scattering artifacts in CT images. By first suppressing projection and then reconstructing with FDK, scattering suppression and reconstruction of three-dimensional volume data are achieved, avoiding further amplification of scattering artifacts during the filtered backprojection process.
[0082] Using GPU parallel reduction algorithm to suppress projection sequences Weighted filtering backprojection along the horizontal direction yields tomographic image data. The cone-beam distance at each angle is calculated using cone-beam projection coordinates, resulting in the CT image sequence. ,and ;in, To reconstruct volume voxels, Let π be 3.14, and D be the distance between the ray source and the center of rotation. Let be the projection component of the voxel distance onto the principal optical axis, and h be the filter kernel. These are the pixels on the X and Y axes in the cone-beam projection coordinates of the area array detector. The projection angle.
[0083] In this invention, the GPU parallel reduction algorithm is referenced from "Video Image Processing and Performance Optimization", Liang Jun, Beijing: Machinery Industry Press, June 2017, 1st edition, pp. 102-108.
[0084] This invention employs a Ram-Lak ramp filter. Sampling of the cone-beam projection coordinate system is accomplished through hardware bilinear interpolation of the GPU texture memory, further improving memory access and interpolation efficiency.
[0085] Example 1: X-ray non-destructive testing of stacked solar cells
[0086] The dual-parameter adaptive scattering suppression industrial X-ray inspection method (i.e., the CUDA-GPU method) based on this invention achieves non-destructive X-ray inspection of stacked solar cells, including the following steps:
[0087] Step 1: Construct the integral map in the GPU;
[0088] In this embodiment, the number of projection frames is set to 720, and the resolution of the area array detector is [missing information]. The pixel size is 150 micrometers. The graphics card version used is... 8GB of video memory.
[0089] The selected fill radius is 45 (the same value used for the optimal kernel in scattering estimation). Parallel scanning is performed on the GPU to complete both parallel row and column scans, while simultaneously calculating the corresponding integral image. Zeros are then padded to the top and left of the integral image, in pixels. Centered on, the neighborhood is Construct the pixel summation, where the adaptive kernel scale is used. Each pixel only requires 4 table lookups and 3 addition / subtraction operations to obtain the first-order and second-order integral images of the stacked battery.
[0090] Step 2: Estimate the coefficient of variation and calculate the two-parameter uniformity threshold using the benchmark mean;
[0091] Definition during calculation , and The two-parameter uniform threshold of the stacked battery was obtained. .
[0092] Step 3: Geometric series set full-map estimation and multi-scale parallel scattering estimation;
[0093] In the process of selecting the optimal kernel for scattering estimation, this implementation takes... The scattering estimation kernel is 45, i.e. .
[0094] Multi-scale global parallelism optimizes the original complexity to The complexity is reduced, and the parallel acceleration of the GPU is effectively utilized to improve the computing efficiency, achieving a speedup of hundreds of times compared to the traditional CPU serial operation.
[0095] Step four: Suppress logarithmic domain scattering;
[0096] In the process of suppressing logarithmic domain scattering, this implementation case selects... and The remapped pixel image of the stacked battery was obtained, which suppressed the blurry structure and enhanced the clarity of details.
[0097] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0098] This embodiment employs a Ram-Lak ramp filter. Projection coordinate sampling is performed using hardware bilinear interpolation in the GPU Texture Memory, further improving memory access and interpolation efficiency. CT scans of the stacked cells are conducted, and scattering suppression is applied to the data. The images before and after suppression, as well as slices of the residual map, are shown below. Figure 4 As shown in Table 1, the overall visual effect reveals that the electrode features of the stacked battery sample are more pronounced. In summary, the sliced images processed with scattering suppression significantly improve grayscale uniformity, structural boundary clarity, and detail discernibility, as illustrated in Table 1.
[0099] Table 1 Comparison of image quality indicators before and after scattering correction for different inspected workpieces
[0100]
[0101] Note that the units are dimensionless.
[0102] Example 2: X-ray non-destructive testing of cylindrical batteries
[0103] The dual-parameter adaptive scattering suppression industrial X-ray inspection method (i.e., the CUDA-GPU method) based on this invention achieves non-destructive X-ray inspection of cylindrical batteries, including the following steps:
[0104] Step 1: Construct the integral map in the GPU;
[0105] In this embodiment, the number of projection frames is set to 720, and the resolution of the area array detector is [missing information]. The pixel size is 150 micrometers. The graphics card version used is... 8GB of video memory.
[0106] The selected fill radius is 48 (the same value used for the optimal kernel in scattering estimation). Parallel scanning is performed on the GPU to complete both parallel row and column scans, while simultaneously calculating the corresponding integral image. Zeros are then padded to the top and left of the integral image, in pixels. Centered on, the neighborhood is Construct the pixel summation, where the adaptive kernel scale is used. Each pixel only requires 4 table lookups and 3 addition / subtraction operations to obtain the first-order and second-order integral images of the cylindrical battery.
[0107] Step 2: Estimate the coefficient of variation and calculate the two-parameter uniformity threshold using the benchmark mean;
[0108] Definition during calculation , and The two-parameter uniform threshold of the cylindrical battery was obtained. .
[0109] Step 3: Geometric series set full-map estimation and multi-scale parallel scattering estimation;
[0110] In the process of selecting the optimal kernel for scattering estimation, this implementation takes... The scattering estimation kernel is 48, i.e. .
[0111] Multi-scale global parallelism optimizes the original complexity to maximum kernel complexity. The complexity is reduced, and the parallel acceleration of the GPU is effectively utilized to improve the computing efficiency, achieving a speedup of hundreds of times compared to the traditional CPU serial operation.
[0112] Step four: Suppress logarithmic domain scattering;
[0113] In the process of suppressing logarithmic domain scattering, this implementation case selects... and The remapped pixel image of the cylindrical battery was obtained, which suppressed the blurry structure and enhanced the clarity of details.
[0114] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0115] This embodiment employs a Ram-Lak ramp filter. Projection coordinate sampling is performed using hardware bilinear interpolation in the GPU Texture Memory, further improving memory access and interpolation efficiency. A CT scan of the cylindrical battery is conducted, and scattering suppression is applied to the data. The images before and after suppression, as well as slices of the residual map, are shown below. Figure 5As shown in Table 1, the overall visual effect reveals that the cylindrical battery sample exhibits more pronounced electrode features. The contrast ratio increases from 2.92 before suppression to 3.56 after suppression, representing an improvement of 21.92%. The information entropy increases from 1.49 before suppression to 3.44 after suppression, representing an improvement of 56.49%. In summary, the sliced image after scattering suppression shows significant improvements in grayscale uniformity, structural boundary clarity, and detail discernibility, as illustrated in Table 1.
[0116] Example 3: X-ray non-destructive testing of castings
[0117] The dual-parameter adaptive scattering suppression industrial X-ray inspection method (i.e., the CUDA-GPU method) based on this invention enables non-destructive X-ray inspection of castings, including the following steps:
[0118] Step 1: Construct the integral map in the GPU;
[0119] In this embodiment, the number of projection frames is set to 720, and the resolution of the area array detector is [missing information]. The pixel size is 150 micrometers. The graphics card version used is... 8GB of video memory.
[0120] The selected fill radius is 51 (the same value used for the optimal kernel in scattering estimation). Parallel scanning is performed on the GPU to complete both parallel row and column scans, while simultaneously calculating the corresponding integral image. Zeros are then padded to the top and left of the integral image, in pixels. Centered on, the neighborhood is Construct the pixel summation, where the adaptive kernel scale is used. Each pixel only requires 4 table lookups and 3 addition / subtraction operations to obtain the first-order integral image and second-order integral image of the casting.
[0121] Step 2: Estimate the coefficient of variation and calculate the two-parameter uniformity threshold using the benchmark mean;
[0122] Definition during calculation , and The two-parameter uniform threshold of the casting was obtained. .
[0123] Step 3: Geometric series set full-map estimation and multi-scale parallel scattering estimation;
[0124] In the process of selecting the optimal kernel for scattering estimation, this implementation takes... The scattering estimation kernel is 51, that is .
[0125] Multi-scale global parallelism optimizes the original complexity to maximum kernel complexity. The complexity of the casting, the GPU parallel execution time compared to the traditional CPU serial execution time in this embodiment, is as follows: Figure 6 As shown, the image processing efficiency accelerated by GPU parallel processing using this invention is improved by a factor of 550 compared to CPU serial processing, both significantly reducing algorithm computation time. Figure 6 The maximum kernel usage ratio in the table represents how many pixels in the adaptive algorithm iterate to the maximum kernel number, which, along with the image size, affects the image processing time.
[0126] Step four: Suppress logarithmic domain scattering;
[0127] In the process of suppressing logarithmic domain scattering, this implementation case selects... and The scattering suppression effect on the DR image obtained from the remapped pixel image of the casting is as follows: Figure 7 As shown, this method effectively suppresses blurry structures in images, enhancing detail clarity. For DR images of castings, after applying the scattering suppression algorithm, Region 1 (ROI1) better highlights the letter features on the surface of the casting image, while Region 2 (ROI2) makes the hole structures clearer. In two contrasting DR images, this scattering suppression method effectively suppresses image blur and better highlights image detail structure information.
[0128] At the same time, construct the grayscale curve of the center row of the image as follows: Figure 8 As shown, in this embodiment, the grayscale distribution curve is plotted on row 1536 of the casting. Red represents the grayscale curve before scattering suppression, and blue represents the grayscale curve after scattering suppression, as shown below. Figure 5 The blue image after scattering suppression shows a steeper abrupt change in the center grayscale curve, which reduces the smooth blurring information in the red curve image and better highlights the image details.
[0129] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0130] This invention employs a Ram-Lak ramp filter. Projection coordinate sampling is performed using hardware bilinear interpolation in the GPU texture memory, further improving memory access and interpolation efficiency.
[0131] X-ray inspection of the casting in this example can effectively suppress DR image scattering artifacts, and the GPU can provide sufficient power for the task. On the platform, the processing time for a single DR image is 0.81s, while the traditional CPU processing time is 280.63s, representing a 550-fold improvement over the traditional approach. Furthermore, the tomographic image data of the casting can also be applied to CT reconstruction, effectively achieving scatter suppression in tomographic image reconstruction. Simultaneously, after using GPU parallel acceleration, this solution achieves a tomographic image processing time of 0.81s for a CT scan sequence containing 720 projections. The entire process, including scattering suppression and FDK reconstruction, takes approximately 147 seconds, fully meeting the real-time requirements of industrial online non-destructive testing.
[0132] Example 4: X-ray non-destructive testing of workpiece welds
[0133] The dual-parameter adaptive scattering suppression industrial X-ray inspection method (i.e., the CUDA-GPU method) based on this invention achieves non-destructive X-ray inspection of workpiece welds, including the following steps:
[0134] Step 1: Construct the integral map in the GPU;
[0135] In this embodiment, the number of projection frames is set to 720, and the resolution of the area array detector is [missing information]. The pixel size is 150 micrometers. The graphics card version used is... 8GB of video memory.
[0136] The selected fill radius is 42 (the same value used for the optimal kernel in scattering estimation). Parallel scanning is performed on the GPU to complete both parallel row and column scans, while simultaneously calculating the corresponding integral image. Zeros are then padded to the top and left of the integral image, in pixels. Centered on, the neighborhood is Construct the pixel summation, where the adaptive kernel scale is used. Each pixel only requires 4 table lookups and 3 addition / subtraction operations to obtain the first-order integral image and second-order integral image of the workpiece weld.
[0137] Step 2: Estimate the coefficient of variation and calculate the two-parameter uniformity threshold using the benchmark mean;
[0138] Definition during calculation , and The two-parameter uniform threshold of the workpiece weld was obtained. .
[0139] Step 3: Geometric series set full-map estimation and multi-scale parallel scattering estimation;
[0140] In the process of selecting the optimal kernel for scattering estimation, this implementation takes... The scattering estimation kernel is 42, that is .
[0141] Multi-scale global parallelism optimizes the original complexity to maximum kernel complexity. The complexity of the workpiece weld in this embodiment is compared to the GPU parallel processing and traditional CPU serial processing in terms of execution time. Figure 6 As shown, the image processing efficiency accelerated by GPU parallel processing using this invention is improved by a factor of 346 compared to CPU serial processing, both significantly reducing algorithm computation time. Figure 6 The maximum kernel usage ratio in the table represents how many pixels in the adaptive algorithm iterate to the maximum kernel number, which, along with the image size, affects the image processing time.
[0142] Step four: Suppress logarithmic domain scattering;
[0143] In the process of suppressing logarithmic domain scattering, this implementation case selects... and The scattering suppression effect on the DR image in the remapped pixel image of the workpiece weld is obtained as follows: Figure 9 As shown, this method effectively suppresses blurry structures in images, enhancing detail clarity. For DR images of workpiece welds, after applying the scattering suppression algorithm, Region 1 (ROI1) better highlights the letter features on the surface of the casting image, while Region 2 (ROI2) makes the hole structure clearer. In two contrasting DR images, this scattering suppression method effectively suppresses image blur and better highlights image detail structure information.
[0144] At the same time, construct the grayscale curve of the center row of the image as follows: Figure 10 As shown, in this embodiment, the grayscale distribution curve is plotted on the 1440th row of the workpiece weld. Red represents the grayscale curve before scattering suppression, and blue represents the grayscale curve after scattering suppression, as shown below. Figure 10 The blue image after scattering suppression shows a steeper abrupt change in the center grayscale curve, which reduces the smooth blurring information in the red curve image and better highlights the image details.
[0145] Step 5: Batch processing of CT projection sequences and FDK reconstruction;
[0146] This invention employs a Ram-Lak ramp filter. Projection coordinate sampling is performed using hardware bilinear interpolation in the GPU texture memory, further improving memory access and interpolation efficiency.
[0147] X-ray inspection of the casting in this example can effectively suppress DR image scattering artifacts, and the GPU can provide sufficient power for the task. On the platform, the processing time for a single DR image is 0.81s, while the traditional CPU processing time is 280.63s, representing a 346-fold improvement over the traditional approach. Furthermore, the tomographic image data of the casting can also be applied to CT reconstruction, effectively achieving scatter suppression in tomographic image reconstruction. Simultaneously, with GPU acceleration, this solution achieves a significantly faster tomographic image processing time for a CT scan sequence containing 720 projections. The entire process, including scattering suppression and FDK reconstruction, takes approximately 147 seconds, fully meeting the real-time requirements of industrial online non-destructive testing. As shown in Table 1, the contrast ratio improved from 0.215 before suppression to 0.218 after suppression, an improvement of 1.48%. Sharpness improved from 0.00221 before suppression to 0.00587 after suppression, an improvement of 165.05%. Information entropy improved from 5.92 before suppression to 6.60 after suppression, an improvement of 10.17%.
Claims
1. A two-parameter adaptive scattering suppression method for industrial X-ray detection, characterized in that, Includes the following steps: Step 1: Construct the integral map in the GPU; The image to be processed is obtained, and a symmetrical inverse filling is performed on the image to be processed to obtain a filled image. The square image is calculated on the filled image by parallel scanning on the GPU to obtain a first-order integral image and a second-order integral image. Step 2: Calculation of the two-parameter uniformity threshold; A window is selected centered on each pixel of the image to be processed. The first-order integral image and the second-order integral image are used to calculate the baseline mean, standard deviation and local coefficient of variation in the neighborhood of each pixel. The median and the corresponding absolute deviation of the median are calculated. A two-parameter uniform threshold is constructed based on the local coefficient of variation and the absolute deviation of the median. Step 3: Multi-scale parallel scattering estimation; A geometric series scale set is preset, and the baseline mean, standard deviation and local coefficient of variation at each scale are calculated with the help of integral image. The relative difference energy between adjacent scales is constructed, and the optimal kernel for scattering estimation is determined based on the relative difference energy and the two-parameter uniform threshold to obtain the scattering estimation pixel image. Step four: Suppress logarithmic domain scattering; A logarithmic domain scattering suppression model is constructed based on the X-ray Beer-Lambert attenuation law. Logarithmic and linear transformations are performed on the image to be processed, and percentage grayscale remapping is performed on the transformed image to obtain a scattering suppression image. Step 5: Batch processing of CT projection sequences and FDK reconstruction; After performing steps one to four on all projection sequences, a scattering-suppressed projection sequence is obtained. The scattering-suppressed projection sequence is then weighted, filtered, and back-projected using GPU parallel scanning to reconstruct tomographic image data.
2. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 1, characterized in that, In step one, the image to be processed is a DR image or a CT projection image, and the maximum fill radius of the symmetrical inverse fill is denoted as . The first-order integral graph and the second-order integral graph are completed by parallel row scanning and parallel column scanning.
3. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 1, characterized in that, In step two, the dual-parameter uniform threshold is constructed as follows: ,in For the coefficient of variation weight, The absolute deviation weight is used. It is the minimum value. The coefficient of variation is the local variation. This represents the median and its corresponding absolute deviation. This is the benchmark mean within the neighborhood.
4. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 1, characterized in that, In step three, the geometric series scale set contains multiple adaptive kernels of different scales; the relative difference energy between adjacent scales is used to characterize the stability of the scale space, and the expansion of the scattering kernel stops when the local mean no longer changes significantly. If no scale meets the conditions, the largest kernel is used as the optimal kernel for scattering estimation of that pixel.
5. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 1, characterized in that, In step four, the pixels of the image to be processed are reduced to logarithmic transform pixel images after logarithmic domain scattering suppression. The logarithmic transform pixel images are then reduced to linear transform linear-logarithmic transform pixel images after nonlinear compensation coefficients.
6. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 5, characterized in that, The percentage grayscale remapping is performed by stretching the linear-logarithmic transformed pixel image according to the 1% and 99% grayscale quantiles, wherein the grayscale quantiles are implemented using GPU parallel radix sorting.
7. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to claim 1, characterized in that, In step five, the weighted filtering back projection uses a Ram-Lak ramp filter, and the sampling of the projection coordinates is completed through hardware bilinear interpolation of the GPU texture memory.
8. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to any one of claims 1 to 7, characterized in that, The method is applied to an industrial X-ray imaging system, in which the imaging controller is equipped with GPU hardware, and the method is embedded in the imaging controller as a pre-image processing step for FDK reconstruction.
9. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to any one of claims 1 to 7, characterized in that, The image to be processed is acquired by an area array detector, and the workpiece being inspected is a complex workpiece with varying thickness, including stacked power batteries, cylindrical batteries, castings, or workpiece welds.
10. The dual-parameter adaptive scattering suppression industrial X-ray detection method according to any one of claims 1 to 7, characterized in that, The scattering estimation achieves full-layer parallel computation through fast summation of integral images, reducing the processing time of a single high-resolution projected image from minutes to seconds, and improving the overall efficiency by more than two orders of magnitude compared to traditional pixel-level iterative methods.