Industrial CT image reconstruction method and device, electronic equipment and storage equipment

Abstract

The invention discloses an industrial CT image reconstruction method and device, electronic equipment and storage equipment, and the method comprises the steps: obtaining projection image data, and carrying out the preprocessing of the projection image data, and obtaining target image data; processing the target image data based on a pre-trained U-Net variant lightweight network to obtain fine gradient direction prediction data and structure-level similarity data, the U-Net variant lightweight network being obtained by training based on projection image data and truth value image data, and a loss function of the U-Net variant lightweight network including structure similarity loss; based on the target image data, the subtle gradient direction prediction data and the structure-level similarity data, an alternating direction multiplier method is adopted to carry out iterative updating on a target function to obtain reconstructed image data, and the target function is a high-order regularization function obtained through construction based on second-order total generalized variation of iterative image data and a structure-level non-local mean value. According to the method, stream artifacts and photon noise of the sparse image can be suppressed, so that the requirements of rapid scanning and sparse image reconstruction quality are considered.

Description

Technical Field

[0001] The present invention relates to the field of automated detection technology, and in particular to an industrial CT image reconstruction method, device, electronic equipment and storage device. Background Art

[0002] In industrial nondestructive testing, fast computed tomography (CT) scanning and high-quality image reconstruction are key to improving inspection efficiency. Related technologies achieve fast scanning by reducing projection angles (sparse images) and shortening X-ray exposure times. However, this introduces severe streak artifacts and photon noise, reducing the quality of CT image reconstruction. Summary of the Invention

[0003] The present invention aims to address at least one of the technical problems existing in the prior art. To this end, the present invention provides an industrial CT image reconstruction method, apparatus, electronic device, and storage device capable of suppressing streak artifacts and photon noise in sparse images, thereby achieving both fast scanning and high-quality sparse image reconstruction.

[0004] In a first aspect, an embodiment of the present invention provides an industrial CT image reconstruction method, comprising: Acquiring projection image data, wherein the projection image data is image data obtained using an industrial CT scanning device in a fast scanning mode; Preprocessing the projection image data to obtain target image data; Processing the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, wherein the U-Net variant lightweight network is trained based on the projected image data and the true image data, and the loss function of the U-Net variant lightweight network includes a structural similarity loss; Based on the target image data, the subtle gradient direction prediction data and the structural level similarity data, the alternating direction multiplier method is used to iteratively update the objective function to obtain reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural level non-local mean.

[0005] According to some embodiments of the present invention, the objective function is a high-order regularization function constructed based on the second-order total generalized variation and the structural non-local mean of the iterative image data, including: The objective function is: u = argmin(||u|| TGV +λ∙u * NLM), st||fW∙u||2≤ξ, where u is the iterative image data, ||u|| TGV is the second-order total generalized variation of the iterative image data, u * NLM is the structural non-local mean of the iterative image data, f is the projected image data, W is the system matrix, λ is the regularization coefficient, and ξ is the preset tolerance.

[0006] According to some embodiments of the present invention, the second-order total generalized variation of the iterative image data is: ||u|| TGV = ||D1u||1+α||D2u||1, where D1u is the first-order derivative of the iterative image data, D2u is the second-order derivative of the iterative image data, and α is the weight coefficient.

[0007] According to some embodiments of the present invention, the iterative updating of the objective function using an alternating direction multiplier method based on the target image data, the subtle gradient direction prediction data, and the structural similarity data to obtain the reconstructed image data includes: Using the target image data as the initial value of iterative image data; Based on the iterative image data and the subtle gradient direction prediction data, updating the second-order total generalized variation of the objective function to obtain first updated data; Based on the iterative image data and the structural similarity data, obtaining second updated data by updating the structural non-local mean of the objective function; determining reconstructed image data according to the first update data and the second update data; performing data consistency constraint processing according to the reconstructed image data; An image difference between two adjacent iterative reconstructed image data is determined, and the reconstructed image data is used as new iterative image data to repeat the above iterative steps until the image difference is less than a preset tolerance or the number of iterations reaches a preset number.

[0008] According to some embodiments of the present invention, the obtaining first updated data by updating the second-order total generalized variation of the objective function based on the iterative image data and the subtle gradient direction prediction data includes: generating a direction mask according to the subtle gradient direction prediction data; determining first and second derivatives of the iterative image data; Correcting the first-order derivative and the second-order derivative of the iterative image data according to the directional mask to obtain a directional first-order derivative and a directional second-order derivative; determining a gradient of a second-order total generalized variation based on the directional first-order derivative and the directional second-order derivative; The iterative image data is updated along the reverse direction of the gradient of the second-order total generalized variation to obtain first updated data.

[0009] According to some embodiments of the present invention, the obtaining second updated data by updating the structural non-local mean of the objective function based on the iterative image data and the structural similarity data includes: determining a target pixel of the iterative image data; Determining a search window for the target pixel, and determining pixels within the search window as candidate pixels; Determining a weight value for each candidate pixel according to the structural similarity data; A weighted sum is performed according to the weight value and the grayscale value of each candidate pixel to obtain second update data.

[0010] According to some embodiments of the present invention, performing data consistency constraint processing according to the reconstructed image data includes: generating theoretical simulation projection image data based on the reconstructed image data; Determining image residual data according to the target image data and the theoretically simulated projection image data; The iterative image data is adjusted according to the image residual data.

[0011] In a second aspect, an embodiment of the present invention provides an industrial CT image reconstruction device, comprising: A data acquisition module is used to acquire projection image data, wherein the projection image data is image data obtained by using an industrial CT scanning device in a fast scanning mode; A preprocessing module, configured to preprocess the projection image data to obtain target image data; a first processing module, configured to process the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, wherein the U-Net variant lightweight network is trained based on the projected image data and the true image data, and a loss function of the U-Net variant lightweight network includes a structural similarity loss; The second processing module is used to iteratively update the objective function based on the target image data, the subtle gradient direction prediction data and the structural level similarity data using the alternating direction multiplier method to obtain reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural level non-local mean.

[0012] In a third aspect, an embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is used to implement the above-mentioned industrial CT image reconstruction method when running the computer program.

[0013] In a fourth aspect, an embodiment of the present invention provides a storage device, wherein a computer program is stored in the storage device, and when the computer program is executed, the above-mentioned industrial CT image reconstruction method is implemented.

[0014] The embodiments of the present invention have at least the following beneficial effects: The target image data is processed based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structure-level similarity data. The subtle gradient direction prediction data can assist the second-order total generalized variation of high-order regularization to capture the complex edges of the target image data, and the structure-level similarity data can assist the structure-level non-local mean of the high-order regularization function to achieve local structure matching. The objective function is a high-order regularization function constructed based on the second-order total generalized variation and structure-level non-local mean of iterative image data. The second-order total generalized variation suppresses noise and artifacts by controlling the gradient and curvature of the image while maintaining the edges and details of the image. The structure-level non-local mean uses information from similar areas in the image to reduce noise and suppress blocking effects during image reconstruction. In this way, when reconstructing image data obtained in fast scanning mode, the streak artifacts and photon noise of the sparse image can be suppressed, thereby meeting the requirements of fast scanning and sparse image reconstruction quality.

[0015] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments with reference to the following drawings, in which: Figure 1 This is a flowchart of the steps of the industrial CT image reconstruction method according to an embodiment of the present invention; Figure 2 This is a principle block diagram of an industrial CT image reconstruction device according to an embodiment of the present invention; Figure 3 This is a principle block diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0018] In the description of the present invention, "several" means one or more, "multiple" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, and "above," "below," and "within" are understood to include the number itself. The use of terms such as "first" and "second" is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0019] Please refer to Figure 1 This embodiment discloses an industrial CT image reconstruction method, including steps S100 to S400. It should be noted that the numbering of the steps in this embodiment is only for the convenience of review and understanding, and does not limit the order in which the steps are executed. The content of each step is detailed below: S100, obtaining projection image data, where the projection image data is image data obtained using an industrial CT scanning device in a fast scanning mode; For example, projection image data is obtained by scanning a target object using a CT scanner. The scanning device in this embodiment is an industrial CT scanner, and its scanning object is an industrial product for non-destructive defect detection. During the scanning process, more scanning angles mean more projection angles for the image data, which facilitates three-dimensional image reconstruction. However, more scanning angles also means longer scanning times, which is detrimental to improving industrial production efficiency. To address this issue, related technologies achieve rapid scanning by reducing projection angles and shortening X-ray exposure times to obtain sparse views. While sparse views can shorten scanning times, they can introduce severe streak artifacts and photon noise. In extreme cases, such as when the projection angle is less than 50° and the exposure time is less than 20 milliseconds, structural distortion and blocking effects can occur. To address the drawbacks of sparse views, this embodiment acquires projection image data to facilitate subsequent processing of the projection image data to suppress artifacts and noise. The projection image data in this embodiment is obtained using an industrial CT scanner in rapid scanning mode and contains artifacts and noise.

[0020] S200, pre-processing the projection image data to obtain target image data; Exemplarily, the projection image data is single-channel grayscale image data. The number of pixels in the projection image data is determined by the detector parameters of the industrial CT scanner, for example, 3040 × 2400 pixels. Typically, the projection image data is stored in raw format. During preprocessing, the projection image data undergoes noise reduction using the Penalized Weighted Least Squares (PWLS) method to smooth the image and improve the signal-to-noise ratio. To ensure compatibility with a wider range of image data formats, if the projection image data is not in raw format, it is converted to raw format during preprocessing.

[0021] S300, processing the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, the U-Net variant lightweight network is trained based on the projected image data and the true value image data, and the loss function of the U-Net variant lightweight network includes a structural similarity loss; For example, a lightweight U-Net variant network is based on the symmetric structure of the classic U-Net encoder and decoder, optimized for the rapid reconstruction requirements of industrial CT. Optimizations include simplifying the number of convolutional layers and employing lightweight convolutional layers in the convolutional units during the downsampling and upsampling phases. For example, this involves using depthwise separable convolutions or reducing the number of convolution kernels to reduce computational overhead. By simplifying the number of convolutional layers, employing lightweight convolution operations, and reducing channel expansion, the lightweight U-Net variant network improves computational speed while maintaining feature extraction capabilities, facilitating adaptation to the rapid reconstruction requirements of industrial CT. It should be noted that the purpose and performance of neural networks are not only related to the network architecture but also to the training data. The training process of the lightweight U-Net variant network in this embodiment achieves end-to-end learning based on projected image data and ground-truth image data. It utilizes an industrial CT dataset containing samples of different materials and defects, and the loss function includes a structural similarity loss. The core of the structural similarity loss is to measure the similarity of local image structures (including brightness, contrast, and structural consistency), rather than simply pixel grayscale differences. Therefore, during training, the lightweight U-Net variant learns to capture local structural features in the image, such as defect edges and the morphology and gradient information of material interfaces. The output structure-level similarity data is used to quantify the degree of local structural similarity at different locations in the target image data. The numerical value of the structure-level similarity data directly reflects the degree of structural match between two regions. The subtle gradient direction prediction data has the same dimensionality as the target image data, and each pixel value in the subtle gradient direction prediction data reflects the edge gradient direction at the corresponding location. In industrial CT images, object edges are important features, and accurately capturing complex edges is crucial for image reconstruction quality. The subtle gradient direction prediction data provides edge direction information for the second-order total generalized variation (TGV) term described below, enabling the TGV to more accurately constrain edge regions. For example, when reconstructing CT images of industrial components with complex shapes and multiple materials, the subtle gradient direction prediction data can accurately indicate the edge orientation between different materials, allowing the TGV to optimize edges based on their orientation to avoid edge blurring or distortion.

[0022] S400. Based on the target image data, the subtle gradient direction prediction data and the structural similarity data, the alternating direction multiplier method is used to iteratively update the objective function to obtain the reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural non-local mean.

[0023] For example, image denoising is a crucial issue in image processing, aiming to restore a clear, realistic image from a noise-contaminated image. The presence of noise can severely impact image quality, reducing the performance of subsequent image analysis, recognition, and applications. Among the numerous denoising methods, models based on variational methods have attracted considerable attention due to their excellent denoising performance and interpretability. Total variation (TV) regularization is one of the most widely used regularization terms in variational methods, constraining the smoothness of the solution by penalizing the sum of image gradients. However, TV regularization suffers from the staircase effect, which results in unnatural blocky structures in the restored image. To overcome this drawback, total generalized variation (TGV) regularization has been proposed as a higher-order extension of TV regularization that effectively suppresses the staircase effect, preserves image detail, and achieves smoother denoising results.

[0024] The non-local means (NL-means, abbreviated as NLM) filtering algorithm is a noise reduction technique. Its core is to achieve noise reduction by calculating the grayscale similarity of pixel neighborhoods at different locations in the image and then performing a weighted average of the target pixel. The non-local means similarity metric relies primarily on direct comparison of pixel-level grayscale values, such as calculating the Euclidean distance of pixel grayscale within a neighborhood window. This method focuses solely on the consistency of grayscale distribution in a local region, while ignoring the local structural features of the pixel (such as edge orientation, texture pattern, and defect morphology). In industrial CT images, due to streak artifacts, noise, and material differences, simple pixel-level grayscale matching can misidentify regions with similar structures but slightly different grayscales as dissimilar, or misidentify regions with similar grayscales but different structures due to noise as similar. This can lead to blurring, blocking artifacts, and loss of detail in the denoised image. The non-local mean in this embodiment is a structural non-local mean. Weights are assigned based on the structural similarity data (output by a lightweight U-Net variant) of the local regions where the target pixel and candidate pixels reside. The greater the structural similarity (such as consistent edge orientation or identical texture patterns), the greater the weight, while the greater the structural difference, the smaller the weight. Compared to traditional non-local means that rely on grayscale differences, the similarity judgment criterion for the structural non-local mean in this embodiment has been upgraded from "pixel grayscale similarity" to "local structural matching."

[0025] The high-order regularization function constructed based on the second-order total generalized variation and the structural-level non-local mean is an optimization problem with constraints. The alternating direction method of multipliers (ADMM) can be used to decompose the complex optimization problem into multiple sub-problems and alternately update the variables to gradually approach the optimal solution.

[0026] During the iterative update of the objective function using the alternating direction multiplier method, the target image data is used as the initial value for the iterative image data. The gradient direction information of the subtle gradient direction prediction data provides a clear optimization direction for each iteration, thereby updating the TGV term based on the subtle gradient direction prediction data. When updating the TGV term, the gradient direction of the subtle gradient direction prediction data is used to make targeted adjustments to areas of the target image data that may be subject to streak artifacts and blocking effects. If the subtle gradient direction prediction data indicates complex edge gradient directions in a certain area, the TGV term is constrained in this area to suppress artifacts and blocking effects, making the edges of the intermediate image clearer and closer to the true structure. Through continuous iteration, guided by the subtle gradient direction prediction data, the TGV term is gradually optimized, continuously improving the quality of the reconstructed image. During the structure-level non-local means filtering, dynamic updates are performed based on the current iteration image data and the structure-level similarity data output by a lightweight U-Net variant network, preserving edge continuity during denoising.

[0027] Therefore, the target image data is processed based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data. The subtle gradient direction prediction data can assist the high-order regularized second-order total generalized variation to capture the complex edges of the target image data, and the structural similarity data can assist the structural non-local mean of the high-order regularization function to achieve local structure matching. The objective function is a high-order regularization function constructed based on the second-order total generalized variation and structural non-local mean of the iterative image data. The second-order total generalized variation suppresses noise and artifacts by controlling the gradient and curvature of the image while maintaining the edges and details of the image. The structural non-local mean uses information from similar areas in the image to reduce noise and suppress blocking effects during image reconstruction. In this way, based on deep learning and high-order regularization, when reconstructing image data obtained in fast scanning mode, it is possible to suppress streak artifacts and photon noise of sparse images, thereby meeting the requirements of fast scanning and sparse image reconstruction quality.

[0028] In step S400, the objective function is a high-order regularization function constructed based on the second-order total generalized variation and structural non-local mean of the iterative image data, including: The objective function is: u = argmin(||u|| TGV +λ∙u * NLM ), st||fW∙u||2≤ξ, where u is the iterative image data, ||u|| TGV is the second-order total generalized variation of the iterative image data, u * NLMis the structural non-local mean of the iterative image data, f is the projected image data, W is the system matrix, λ is the regularization coefficient, and ξ is the preset tolerance. It should be noted that argmin is used to represent the value of the independent variable that minimizes the function. st is the abbreviation of such that, which is used to indicate that a variable or condition must meet specific requirements or constraints, and ||∙||2 is used to represent the L2 norm. This embodiment fuses the second-order total generalized variation and the structural non-local mean of the iterative image data through the objective function, and constrains the data consistency through constraints to ensure that the image reconstruction result matches the target image data.

[0029] Among them, the second-order total generalized variation of the iterative image data is: ||u|| TGV = ||D1u||1+α||D2u||1, where D1u is the first derivative of the iterated image data, D2u is the second derivative of the iterated image data, and α is the weight coefficient. It should be noted that ||∙||1 is used to represent the L1 norm.

[0030] To facilitate understanding of the application concept of the second-order total generalized variation in this embodiment, the various parts of the second-order total generalized variation are described below: The first-order derivative describes the local rate of change of pixel grayscale in iterative image data. Its physical meaning is to capture image edge information, such as object outlines and defect boundaries. In regularization constraints, constraining the L1 norm of the first-order derivative achieves the following effects: preserving key edges and suppressing high-frequency noise through the "sparseness assumption." The sparsity assumption assumes that edges in natural images are sparsely distributed in the gradient domain. Preserving key edges emphasizes existing edge structures in the image to avoid edge blurring caused by noise or artifacts. Suppressing high-frequency noise penalizes regions with rapidly fluctuating grayscale values ​​(such as noise) to reduce their interference with the overall image structure. However, relying solely on first-order derivative constraints has limitations: overemphasizing the sparsity of first-order gradients can lead to stair-step distortion (a manifestation of blocking artifacts) in localized image regions. This is especially true in areas with smooth grayscale but complex structures (such as subtle textures within materials). Overly strict gradient constraints can easily lead to loss of detail. Therefore, the second-order derivative is introduced.

[0031] The second-order derivative describes the rate of change of the first-order derivative, reflecting the "curvature" or "smoothness" of image grayscale variations. It is used to characterize structural smoothness and reduce staircase artifacts. Constraining the L1 norm of the second-order derivative is crucial for achieving the following functions: characterizing smooth structural transitions and reducing staircase artifacts by controlling the continuity of gradient changes. Characterizing smooth structural transitions means that in edge regions, the second-order derivative can distinguish between "true edges" (abrupt but continuous gradients) and "false edges" (abrupt but discontinuous gradients, such as noise or blockiness), penalizing the latter. Reducing staircase artifacts means that in regions with flat grayscale, the second-order derivative constraint (||D2u||1) avoids the "piecewise flattening" phenomenon (i.e., forcing adjacent pixels to the same grayscale level, resulting in blocky partitions) caused by the first-order derivative alone, resulting in more natural and continuous grayscale transitions. In short, the second-order derivative constraint addresses the limitation of the first-order derivative, which focuses solely on the existence of gradients, by focusing on the rationality of gradient changes. This provides a higher-dimensional constraint on the rationality of image structure.

[0032] The core function of the weight coefficient α is to dynamically adjust the proportion of the first-order derivative and the second-order derivative in the total constraint, thereby balancing the strength of the two and ultimately reducing the blocking effect. The specific logic is as follows: When α is set to a small value, the first-order derivative constraint (||D1u||1) dominates, effectively preserving image edges but potentially causing blocking artifacts due to oversuppression of high-frequency components. When α is set to a large value, the second-order derivative constraint (||D2u||1) becomes more dominant, enhancing the smoothness of the image structure and effectively mitigating stair-step distortion. However, overemphasizing the second-order constraint can lead to blurred edges. By setting a reasonable α value (which must be optimized experimentally based on the material properties and defect types in the specific industrial CT scenario), the first-order derivative constraint dominates at edges, ensuring sharp edges; while the second-order derivative constraint dominates in flat areas, ensuring continuous grayscale changes and avoiding blocking artifacts. Overall, the image preserves edge information of key structures while ensuring smooth transitions in local areas, thereby reducing structural distortion caused by the effects of either the first-order or second-order constraints alone.

[0033] In high-order regularization functions, u * NLM is the image data after improved non-local mean filtering, u * NLM The structural similarity data of the U-Net variant lightweight network is introduced, where u * NLM The relationship is: Where i and j are used to represent the row index and column index of the target pixel in the image, respectively, to locate the pixel position that needs to be denoised. For example, in an image with a pixel size of N*N, the value range of i and j is 1≤i≤N and 1≤j≤N, corresponding to the coordinates of each specific pixel in the image; m and s are used to represent the row index and column index of the candidate pixel in the image, respectively, and are used to locate other pixel positions for similarity comparison with the target pixels i and j. In the NLM algorithm, candidate pixels usually come from the search window around the target pixel (i.e., a local area within a certain range). By calculating the structural similarity between the candidate pixel and the target pixel, a weighted basis is provided for noise reduction of the target pixel. Similarly, in an image with a pixel size of N*N, the value range of m and s is related to the size of the search window, and m and s take values ​​within a small range around i and j. h represents the filter coefficient, which is used to control the decay speed of the similarity weight and determines the influence of the similarity distance on the weight. The physical meaning of the filter coefficient is that when the structural similarity distance between the candidate pixel and the target pixel is When it is small, the exponential term Larger values ​​of indicate a higher weighted contribution of the candidate pixel to the target pixel. Conversely, when the structural similarity distance is large, the exponential term decays rapidly, significantly reducing its contribution. The filter coefficient should be adjusted based on the image noise level. Larger filter coefficients result in a smoother weight distribution (stronger noise reduction but potentially blurring details), while smaller filter coefficients prioritize preserving local structure.

[0034] Among them, the structural similarity distance The structure-level similarity data output by the U-Net variant lightweight network is used to improve the non-local mean similarity metric, enabling an upgrade from pixel-level matching to local structure-based matching. It is worth noting that the structure-level similarity data output by the U-Net variant lightweight network can be either weight data or distance data. For example, the U-Net variant lightweight network outputs structure-level similarity weights, calculates structure-level similarity distances based on the structure-level similarity weights, or adjusts the structure-level similarity distances based on the structure-level similarity weights.

[0035] Step S400, based on the target image data, the subtle gradient direction prediction data and the structural similarity data, an alternating direction multiplier method is used to iteratively update the objective function to obtain reconstructed image data, including: S410, using the target image data as the initial value of the iterative image data; S420, obtaining first updated data by updating the second-order total generalized variation of the objective function based on the iterative image data and the subtle gradient direction prediction data; S430, obtaining second updated data by updating the structural non-local mean of the objective function based on the iterative image data and the structural similarity data; S440, determining reconstructed image data according to the first updated data and the second updated data; S450, performing data consistency constraint processing according to the reconstructed image data; S460: Determine the image difference between the reconstructed image data of two adjacent iterations, and use the reconstructed image data as new iterative image data to repeat the above iterative steps until the image difference is less than a preset tolerance or the number of iterations reaches a preset number.

[0036] In step S420, based on the iterative image data and the subtle gradient direction prediction data, the second-order total generalized variation of the objective function is updated to obtain the first updated data, including: S421, generating a direction mask based on the subtle gradient direction prediction data; For example, as described above, the U-Net variant lightweight network can output subtle gradient direction prediction data to provide prior edge information. Each pixel value of the subtle gradient direction prediction data corresponds to the edge gradient direction at that location in the image, such as the direction of an object outline or defect boundary. This prediction result is learned based on the projected image data and true image data in the training data, and can capture complex edges that are easily overlooked by traditional gradient calculations, such as subtle structures with large curvature changes or multi-material junctions. Based on the subtle gradient direction prediction data, a direction mask is generated, denoted as M. In the direction mask, the pixel value in the true edge direction area is 1, and the pixel value in the artifact or noise area is 0, thereby providing a directional basis for updating the TGV term.

[0037] S422, determining the first-order derivative and the second-order derivative of the iterative image data; For example, the first-order derivative and the second-order derivative of the current iterative image data are determined by calculation. The method for calculating the derivative is a conventional technical means for those skilled in the art and will not be described in detail in this embodiment. The iterative image data is denoted as u prev , the first-order derivative is recorded as D1u prev , the second-order derivative is denoted as D2u prev .

[0038] S423, respectively correcting the first-order derivative and the second-order derivative of the iterative image data according to the direction mask to obtain a directional first-order derivative and a directional second-order derivative; For example, the directional first-order derivative D1u'=M*D1u prev , only retaining the gradient component of the true edge direction, weakening the false gradient caused by artifacts; directional second-order derivative D2u'=M*D2u prev, smoothing constraints are only imposed on continuous regions of true edges to avoid over-smoothing of non-edge regions.

[0039] S424. Determine the gradient of the second-order total generalized variation based on the directional first-order derivative and the directional second-order derivative; For example, the gradient of the second-order total generalized variation is denoted as ▽||u|| TGV , then ▽||u|| TGV =sign(D1u')+α∙sign(D2u'), where sign represents the sign function. A positive input returns 1, a zero returns 0, and a negative input returns -1. The weight α is determined through experimental optimization and must be adapted to the material properties and defect types of the industrial CT scenario. For example, in edge regions, α is set to a small value, and the first-order derivative constraint dominates to ensure sharp edges. In flat regions, α is set to a large value, and the second-order derivative constraint dominates to ensure continuous grayscale changes and avoid step-like blocking artifacts.

[0040] S425 , iteratively updating the image data along the reverse direction of the gradient of the second-order total generalized variation to obtain first updated data.

[0041] For example, the data is updated in the opposite direction of the gradient, and the updated data is recorded as u TGV , then u TGV= u prev -β∙▽||u|| TGV , where β is the iteration step size. In this way, the image retains clear gradient features in the true edge area and weakens the false gradient features in the artifact or noise area, achieving the dual effects of streak artifact suppression and block effect reduction.

[0042] Step S430: Based on the iterative image data and the structural similarity data, the structural non-local mean of the objective function is updated to obtain second updated data, including: S431, determining a target pixel for iterating image data; S432, determining a search window for the target pixel, and determining pixels within the search window as candidate pixels; For example, for a target pixel (i, j), candidate pixels (m, s) are selected within a preset search window. These candidate pixels are potential regions with similar structures. The size of the search window can be adaptively adjusted based on the actual application, for example, the search window size can be 5×5 or 7×7.

[0043] S433, determining a weight value for each candidate pixel based on the structure-level similarity data; Exemplarily, as described above, the similarity measure of the traditional NLM is based on the gray Euclidean distance of the pixel neighborhood, only focusing on the numerical difference of the pixel gray value, and easily misjudging the gray fluctuation caused by noise as structural difference, or misjudging the region with similar structure but slightly different gray value as dissimilar. The present embodiment upgrades the pixel-level similarity measure of the traditional NLM to the matching based on local structure by introducing the structure-level similarity distance, so as to better preserve the edge and detail structure of the image while suppressing noise. The structure-level similarity distance corresponding to the target pixel (i, j) is denoted as In some application examples, the structure-level similarity distance can be the structural similarity data output by the U-Net variant lightweight network, or, in other application examples, the structural similarity data output by the U-Net variant lightweight network is the structure-level similarity weight data, and the structure-level similarity distance is determined according to the structure-level similarity weight data. The structure-level similarity distance is a distance calculated based on local structural features (such as gradient direction and texture consistency, etc.), and the smaller the value of the structure-level similarity distance, the more similar the local structure of the target pixel and the candidate pixel. In the defect edge region of the industrial CT image, the traditional non-local mean value may determine that the pixels on both sides of the edge are dissimilar due to the gray mutation on both sides of the edge, but the structure-level non-local mean value of the present embodiment can identify that both sides belong to the edge structure through the structure-level similarity distance, and determine that they are highly similar, so as to preserve the continuity of the edge when denoising. According to the u * NLM It can be seen from the relationship that the weight value of the candidate pixel (m, s) is: , Wherein, the filter coefficient h can be dynamically adjusted according to the noise level or optimized by network training.

[0044] S434, weighted sum according to the weight value and the gray value of each candidate pixel to obtain the second update data.

[0045] Exemplarily, the gray values of all candidate pixels in the search window are weighted and summed to obtain the denoised value of the target pixel (i, j), which achieves the effect of "smoothing noise with the pixel value of the structure similar region while preserving the structural features", wherein the gray value of the candidate pixel is .

[0046] It is worth mentioning that the structure-level non-local mean filter of the present embodiment is not independently executed, but is based on the current iteration image data u n-1The structural similarity data output by the lightweight network of the U-Net variant is dynamically updated. For example, the local structural information of the current iteration image data is combined with the structural similarity data to recalculate the similarity weight value of each target pixel to ensure that the weight value adapts to the changes in the image during iteration. Then, the current iteration image data u is updated by weighted summation. n-1 Perform noise reduction to obtain new iterative image data u n . The new iterative image data u n As the input of the data consistency constraint processing, it ensures that the denoised image still matches the projection image data and avoids structural distortion caused by over-smoothing.

[0047] Step S450: performing data consistency constraint processing based on the reconstructed image data, including: S451, generating theoretical simulation projection image data based on the reconstructed image data; For example, during the iterative update process, the reconstructed image data is constantly changing, and the final reconstructed image data is output only after the conditions are met. When performing data consistency constraint processing, the current reconstructed image data is used as input. After the update of step S420 and step S430, the TGV term (||u|| TGV ) and NLM terms (u * NLM ), then the reconstructed image data can be determined according to the objective function above. That is, step S440 is implemented by substituting the first updated data and the second updated data into the objective function to determine the image reconstruction data. Based on the reconstructed image data (denoted as u), the theoretical simulation projection data is generated through the system matrix W, denoted as .

[0048] S452, determining image residual data based on the target image data and the theoretical simulation projection image data; For example, the image residual data is recorded as r, and the target image data is recorded as f, then the image residual data , which is used to measure the deviation between the target image data and the theoretical simulation projection image data.

[0049] S453: Adjust the iterative image data according to the image residual data.

[0050] For example, by back-projection matrix W T The image residual data is mapped back to the image domain through the relationship u n =u n-1 +γW TThe iterative image data is adjusted to reduce the deviation to within a tolerance of ξ, where γ is the step size. It should be noted that the tolerance ξ and step size γ are adjusted based on the projection noise level to avoid excessively loose or tight constraints, which can introduce structural distortion or noise. Data consistency constraint processing works in conjunction with the TGV term regularization in step S420 and the structure-level NLM denoising in step S430 to balance structural rationality and physical realism, ensuring that the reconstruction results suppress artifacts and noise while remaining consistent with the actual data.

[0051] During the iterative update process, a preset number of iterations or the image difference between two iterations can be used as the termination condition. The image difference between two iterations can be calculated by quantifying the pixel-level differences between the image data of two adjacent iterations, for example, using the L2 norm or mean square error. After the iteration is completed, the final reconstructed image is saved in a standard format for subsequent defect analysis or visualization.

[0052] The following describes the architecture of the lightweight U-Net variant network (hereinafter referred to as the network). This network, based on the classic U-Net symmetric "encoder-decoder" structure, is lightweight and optimized for the rapid reconstruction requirements of industrial CT. The overall architecture consists of five parts: input layer, encoder module, bottleneck layer, decoder module, and output layer. Each module uses skip connections to achieve cross-layer feature fusion, resulting in a compact structure and high computational efficiency.

[0053] 1. Input layer Input content: Receive the preprocessed target image data, where the target image data is a single-channel grayscale image in raw format with a pixel size of 3040×2400, and the pixel size corresponds to the CT detector parameters.

[0054] Function: Converts target image data into a tensor format that can be processed by the U-Net variant lightweight network, providing initial input for subsequent feature extraction without changing the spatial size and basic information of the data.

[0055] 2. Encoder module (downsampling stage) The encoder module consists of 3-4 downsampling stages (the number of layers is simplified according to the lightweight requirements), each stage includes a convolution unit and a downsampling layer: Convolutional unit: Each unit contains 1-2 3×3 lightweight convolutional layers (using depthwise separable convolution or reducing the number of convolution kernels to reduce computational complexity), with batch normalization (BN) and ReLU activation function.

[0056] Function: Extract multi-scale features of projection data layer by layer, from low-level features (such as noise texture and local edges) to mid-level features (such as regional structure and defect contours), enhancing feature expression capabilities.

[0057] Down-sampling layer: Down-sampling is achieved by using 2x2 max pooling or convolution layer with stride 2.

[0058] Function: By reducing the spatial size of feature maps (size is halved after each down-sampling), increasing the number of feature channels (moderate increase in lightweight design to avoid redundancy), focusing on global structure information extraction.

[0059] 3. Bottleneck layer Located between the encoder module and the decoder module, it is the deepest feature processing unit in the network: Structure: Contains 1-2 3x3 convolution units (same as the lightweight design of the encoder module), without down-sampling operation.

[0060] Function: Integrates deep abstract features output by the encoder module, captures the most critical global structure information in the target image data (such as the overall shape of the workpiece, large-scale defect distribution), and provides semantic anchor points for feature restoration of the decoder module.

[0061] 4. Decoder module (up-sampling stage) Symmetrical to the encoder module, it contains 3-4 up-sampling stages, each including an up-sampling layer, a feature fusion unit, and a convolution unit: Up-sampling layer: Up-sampling is achieved by using transpose convolution unit (1x1 or 3x3) or interpolation unit (such as bilinear interpolation).

[0062] Function: Gradually restores the spatial size of feature maps (size doubles after each up-sampling), reduces the number of channels, and matches the feature scale of the corresponding stage of the encoder module.

[0063] Feature fusion unit: Shallow features (containing detailed information) from the corresponding stage of the encoder module are spliced or weighted fused with the current deep features (containing semantic information) of the decoder module through jump connection.

[0064] Function: Supplement the detailed information (such as small edges, local texture) lost in the up-sampling process, balance the feature expression of global structure and local details.

[0065] Convolution unit: Same design as the encoder module, each convolution unit contains 1-2 3x3 lightweight convolution layers, with batch normalization (BN) and ReLU activation function.

[0066] Function: Further process the fused features, eliminate redundant information, and strengthen effective features (such as defect edges, structural texture).

[0067] 5. Output layer Structure: Contains 1 1x1 convolution layer (without activation function), with 2 output channels.

[0068] Function: Convert the feature map of the final output of the decoder module into two target outputs: Subtle gradient direction prediction data: Single-channel image, the pixel value reflects the edge gradient direction of the corresponding position, assisting subsequent high-order regularization (TGV term) to accurately capture complex edges and suppress streak artifacts.

[0069] Structural similarity data: single-channel image, pixel value is structural similarity distance or structural similarity weight, used to improve the similarity metric of non-local means (NLM), upgrading from pixel-level matching to local structure-based matching, and reducing block effects.

[0070] Core design features: Lightweight is reflected in simplifying the number of convolution layers, adopting lightweight convolution operations, and reducing the expansion of the number of channels. This improves the computing speed while ensuring feature extraction capabilities, adapting to the rapid reconstruction needs of industrial CT. Through skip connections and dual-output design, it realizes the dual functions of "detail capture and structure modeling", providing key auxiliary information for subsequent iterative reconstruction.

[0071] The network training uses a small-scale industrial CT dataset (containing samples of different materials and defects) and performs end-to-end training on "projection data-true value image" pairs. The loss function is mean square error (MSE) + structural similarity loss (SSIMLoss).

[0072] Please refer to Figure 2 , an embodiment of the present invention provides an industrial CT image reconstruction device, comprising: The data acquisition module 110 is used to acquire projection image data, which is image data obtained by using an industrial CT scanning device in a fast scanning mode; A preprocessing module 120 is used to preprocess the projection image data to obtain target image data; a first processing module 130 for processing the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, wherein the U-Net variant lightweight network is trained based on the projected image data and the true image data, and the loss function of the U-Net variant lightweight network includes a structural similarity loss; The second processing module 140 is used to iteratively update the objective function based on the target image data, subtle gradient direction prediction data and structural level similarity data using the alternating direction multiplier method to obtain reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural level non-local mean.

[0073] The inventive concept of the industrial CT image reconstruction device embodiment is the same as that of the above-mentioned industrial CT image reconstruction method embodiment. The content not involved in the industrial CT image reconstruction device embodiment can refer to the above-mentioned industrial CT image reconstruction method embodiment, which will not be repeated here. The target image data is processed based on the pre-trained U-Net variant lightweight network to obtain fine gradient direction prediction data and structure-level similarity data. The fine gradient direction prediction data can assist the second-order total generalized variation of the high-order regularization to capture the complex edge of the target image data. The structure-level similarity data can assist the structure-level non-local mean of the high-order regularization function to realize the matching of the local structure. The target function is a high-order regularization function constructed based on the second-order total generalized variation and the structure-level non-local mean of the iterative image data. The second-order total generalized variation controls the gradient and curvature of the image to suppress noise and artifacts while maintaining the edge and details of the image. The structure-level non-local mean uses the information of similar regions in the image to reduce noise and suppress the blocking effect in the image reconstruction process. Therefore, when reconstructing the image data obtained in the fast scanning mode, the streak artifacts and photon noise of the sparse image can be suppressed, thereby meeting the requirements of fast scanning and sparse image reconstruction quality.

[0074] Please refer to Figure 3 The embodiment of the present application provides an electronic device, which comprises a processor 210 and a memory 220. The memory 220 stores a computer program. When the processor 210 runs the computer program, it is used to implement the above-mentioned industrial CT image reconstruction method. The details of the industrial CT image reconstruction method can refer to the above, which will not be repeated here. The target image data is processed based on the pre-trained U-Net variant lightweight network to obtain fine gradient direction prediction data and structure-level similarity data. The fine gradient direction prediction data can assist the second-order total generalized variation of the high-order regularization to capture the complex edge of the target image data. The structure-level similarity data can assist the structure-level non-local mean of the high-order regularization function to realize the matching of the local structure. The target function is a high-order regularization function constructed based on the second-order total generalized variation and the structure-level non-local mean of the iterative image data. The second-order total generalized variation controls the gradient and curvature of the image to suppress noise and artifacts while maintaining the edge and details of the image. The structure-level non-local mean uses the information of similar regions in the image to reduce noise and suppress the blocking effect in the image reconstruction process. Therefore, when reconstructing the image data obtained in the fast scanning mode, the streak artifacts and photon noise of the sparse image can be suppressed, thereby meeting the requirements of fast scanning and sparse image reconstruction quality.

[0075] An embodiment of the present invention provides a storage device storing a computer program that, when executed, implements the aforementioned industrial CT image reconstruction method. The details of the industrial CT image reconstruction method can be found above and will not be repeated here. The target image data is processed based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structure-level similarity data. The subtle gradient direction prediction data can assist the second-order total generalized variation of high-order regularization to capture the complex edges of the target image data, and the structure-level similarity data can assist the structure-level non-local mean of the high-order regularization function to achieve local structure matching. The objective function is a high-order regularization function constructed based on the second-order total generalized variation and structure-level non-local mean of iterative image data. The second-order total generalized variation suppresses noise and artifacts by controlling the gradient and curvature of the image while maintaining the edges and details of the image. The structure-level non-local mean uses information from similar areas in the image to reduce noise and suppress blocking effects during image reconstruction. In this way, when reconstructing image data obtained in fast scanning mode, the streak artifacts and photon noise of the sparse image can be suppressed, thereby meeting the requirements of fast scanning and sparse image reconstruction quality.

[0076] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the scope of the present invention.

Claims

1. An industrial CT image reconstruction method, characterized in that: include: Acquiring projection image data, wherein the projection image data is image data obtained using an industrial CT scanning device in a fast scanning mode; Preprocessing the projection image data to obtain target image data; Processing the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, wherein the U-Net variant lightweight network is trained based on the projected image data and the true image data, and the loss function of the U-Net variant lightweight network includes a structural similarity loss; Based on the target image data, the subtle gradient direction prediction data and the structural level similarity data, the alternating direction multiplier method is used to iteratively update the objective function to obtain reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural level non-local mean.

2. The industrial CT image reconstruction method according to claim 1, characterized in that: The objective function is a high-order regularization function constructed based on the second-order total generalized variation and structural non-local mean of iterative image data, including: The objective function is: u = argmin(||u|| TGV +λ∙u * NLM ), st||fW∙u||2≤ξ, where u is the iterative image data, ||u|| TGV is the second-order total generalized variation of the iterative image data, u * NLM is the structural level non-local mean of iterative image data, f is the projected image data, W is the system matrix, λ is the regularization coefficient, and ξ is the preset tolerance.

3. The industrial CT image reconstruction method according to claim 2, characterized in that: The second-order total generalized variation of the iterative image data is: ||u|| TGV = ||D1u||1 +α||D2u||1, where D1u is the first-order derivative of the iterative image data, D2u is the second-order derivative of the iterative image data, and α is the weight coefficient.

4. The industrial CT image reconstruction method according to any one of claims 1 to 3, characterized in that: The step of iteratively updating the objective function based on the target image data, the subtle gradient direction prediction data, and the structural similarity data using an alternating direction multiplier method to obtain reconstructed image data includes: Using the target image data as the initial value of iterative image data; Based on the iterative image data and the subtle gradient direction prediction data, updating the second-order total generalized variation of the objective function to obtain first updated data; Based on the iterative image data and the structural similarity data, obtaining second updated data by updating the structural non-local mean of the objective function; determining reconstructed image data according to the first update data and the second update data; performing data consistency constraint processing according to the reconstructed image data; An image difference between two adjacent iterative reconstructed image data is determined, and the reconstructed image data is used as new iterative image data to repeat the above iterative steps until the image difference is less than a preset tolerance or the number of iterations reaches a preset number.

5. The industrial CT image reconstruction method according to claim 4, characterized in that: The obtaining of first updated data by updating the second-order total generalized variation of the objective function based on the iterative image data and the subtle gradient direction prediction data includes: generating a direction mask according to the subtle gradient direction prediction data; determining first and second derivatives of the iterative image data; Correcting the first-order derivative and the second-order derivative of the iterative image data according to the directional mask to obtain a directional first-order derivative and a directional second-order derivative; determining a gradient of a second-order total generalized variation based on the directional first-order derivative and the directional second-order derivative; The iterative image data is updated along the reverse direction of the gradient of the second-order total generalized variation to obtain first updated data.

6. The industrial CT image reconstruction method according to claim 4, characterized in that: The obtaining of second updated data by updating the structural non-local mean of the objective function based on the iterative image data and the structural similarity data comprises: determining a target pixel of the iterative image data; Determining a search window for the target pixel, and determining pixels within the search window as candidate pixels; Determining a weight value for each candidate pixel according to the structural similarity data; A weighted sum is performed according to the weight value and the grayscale value of each candidate pixel to obtain second update data.

7. The industrial CT image reconstruction method according to claim 4, characterized in that: The performing data consistency constraint processing according to the reconstructed image data includes: generating theoretical simulation projection image data based on the reconstructed image data; Determining image residual data according to the target image data and the theoretically simulated projection image data; The iterative image data is adjusted according to the image residual data.

8. An industrial CT image reconstruction device, characterized in that: include: A data acquisition module is used to acquire projection image data, wherein the projection image data is image data obtained by using an industrial CT scanning device in a fast scanning mode; A preprocessing module, configured to preprocess the projection image data to obtain target image data; a first processing module, configured to process the target image data based on a pre-trained U-Net variant lightweight network to obtain subtle gradient direction prediction data and structural similarity data, wherein the U-Net variant lightweight network is trained based on the projected image data and the true image data, and a loss function of the U-Net variant lightweight network includes a structural similarity loss; The second processing module is used to iteratively update the objective function based on the target image data, the subtle gradient direction prediction data and the structural level similarity data using the alternating direction multiplier method to obtain reconstructed image data. The objective function is a high-order regularization function constructed based on the second-order total generalized variation of the iterative image data and the structural level non-local mean.

9. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, wherein: When the processor runs the computer program, it is used to implement the industrial CT image reconstruction method according to any one of claims 1 to 7.

10. A storage device storing a computer program, wherein: When the computer program is executed, the industrial CT image reconstruction method according to any one of claims 1 to 7 is implemented.

Images