A method, system, medium, and apparatus for nnbi visual tomographic image reconstruction

Abstract

This invention discloses a method, system, medium, and device for NNBI visual tomographic image reconstruction, relating to the field of neutral beam imaging technology using negative ion sources. The NNBI visual tomographic image reconstruction method and system provided by this invention utilizes a tomographic reconstruction algorithm to invert the observed projection map of an NNBI multi-angle beam image, obtaining an initial reconstructed image and constructing its ROI mask. Then, the mask is input into a pre-trained residual denoising network via two channels to acquire the residual image and calculate the denoised image. This suppresses interference from background and noise outside the ROI and improves the stability of denoising, thereby improving image reconstruction efficiency and image quality. Furthermore, based on a forward projection operator constructed using camera calibration parameters and the observed projection vector, the reprojection residual of the denoised image is calculated and a correction value is generated. This correction value is used to perform data consistency correction on the denoised image to obtain the target reconstructed image, maintaining projection consistency and preventing consistency degradation caused by depth denoising. This invention effectively suppresses isolated points and stripe artifacts during image reconstruction, improving the quality of image reconstruction.

Description

Technical Field

[0001] This application relates to the field of neutral beam imaging technology using negative ion sources, and in particular to a method, system, medium, and device for reconstructing NNBI visual tomographic images. Background Technology

[0002] Visual tomography (MTT) of a negative ion source neutral beam injection (NNBI) system acquires radiation signals generated by the interaction between a neutral beam and plasma using a multi-angle optical detector array. Combined with tomographic reconstruction algorithms, it inverts the beam distribution within the three-dimensional space of the tokamak device, thereby achieving high-precision diagnosis of beam focusing efficiency, neutralization efficiency, and spatial uniformity. However, due to device geometry limitations, the available viewing angles are often distributed in a "restricted angle" pattern, leading to ill-conditioning inversion. Furthermore, the actual imaging signal is superimposed with background disturbances, gain fluctuations, and multiplicative noise, making the results obtained using algebraic reconstruction methods prone to problems such as fringe artifacts, blurred interlayer interfaces, and excessively large local extreme pixel errors.

[0003] Currently, while traditional TV regularization methods can suppress noise to some extent, they are prone to causing interface blurring and residual extreme errors under high-noise conditions. Directly applying deep denoising to the reconstruction results can improve image domain metrics, but it may cause the output to deviate from the feasible set of the measurement equation, leading to consistency degradation. Therefore, under conditions of limited angles and strong noise, existing methods often struggle to simultaneously address overall error, extreme isolated point errors, and consistency in the reconstruction results, resulting in reconstructed images with significant artifacts and room for improvement in image quality. Summary of the Invention

[0004] Therefore, it is necessary to provide an NNBI visual tomography image reconstruction method, system, medium, and device to address the aforementioned technical problems.

[0005] The following technical solution is adopted in this specification: This specification provides a method for visual tomography image reconstruction using NNBI, including: Multi-angle beam images of the NNBI system are acquired simultaneously using multiple cameras; The multi-angle beam image is processed to generate an observation projection map, and a forward projection operator and an observation projection vector are constructed based on the camera calibration parameters; the observation projection map is inverted using a tomographic reconstruction algorithm to obtain an initial reconstructed image; A ROI mask for the target region of the image is constructed. The initial reconstructed image and the ROI mask are input as two channels to a pre-trained residual denoising network, which outputs a residual image. A denoised image is obtained based on the residual image, and the reprojection residual of the denoised image is calculated using the forward projection operator and the observation projection vector. The reprojection residual is back-projected to generate a correction amount, and the denoised image is subjected to at least one data consistency correction based on the correction amount to obtain the target reconstructed image.

[0006] Optionally, the residual denoising network includes a multi-layer convolutional residual structure, and an ROI mask gating module and a structural prior module are inserted every few convolutional blocks on the image feature channel.

[0007] Optionally, the weights of the ROI mask gating module are generated by a 1×1 convolution-ReLU-Sigmoid jointly driven by image features and the ROI mask.

[0008] Optionally, the structure prior module predicts the structural parameters of the beam cross section based on feature-adaptive global pooling, and generates a two-dimensional or elliptical Gaussian weight field to weight and modulate the image features.

[0009] Optionally, the ROI mask is a fixed circular ROI mask adaptively generated by thresholding and morphological processing of the initial reconstructed image, wherein the fixed circular ROI mask is configured with a boundary margin ratio or a shrinkage ratio.

[0010] Optionally, the data consistency correction may be applied only within the image ROI, or applied after weighting the correction amount according to the ROI mask weight.

[0011] Optionally, it also includes calculating the image domain error index and the projection domain consistency index based on the target reconstructed image and outputting them to the reconstructed display area for visualization.

[0012] This specification provides an NNBI visual tomography image reconstruction system, including: The acquisition module is used to simultaneously acquire multi-angle beam images of the NNBI system using multiple cameras; The generation and inversion module is used to process the multi-angle beam image to generate an observation projection map, and to construct a forward projection operator and an observation projection vector based on camera calibration parameters; the observation projection map is inverted using a tomographic reconstruction algorithm to obtain an initial reconstructed image; The construction and calculation module is used to construct a ROI mask for the target region of the image, input the initial reconstructed image and the ROI mask as two channels into a pre-trained residual denoising network and output a residual image; obtain a denoised image based on the residual image, and calculate the reprojection residual of the denoised image using the forward projection operator and the observation projection vector; The correction module is used to back-project the reprojection residual to generate a correction amount, and perform data consistency correction on the denoised image at least once based on the correction amount to obtain the target reconstructed image.

[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described NNBI visual tomography image reconstruction method.

[0014] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described NNBI visual tomography image reconstruction method.

[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: The NNBI visual tomographic image reconstruction method and system provided in this specification utilizes a tomographic reconstruction algorithm to invert the observed projection map of the NNBI multi-angle beam image, obtains an initial reconstructed image, constructs its ROI mask, and then inputs it into a pre-trained residual denoising network in two channels to obtain the residual image and calculate the denoised image. This method can suppress error interference caused by background and noise outside the ROI and improve the stability of denoising, thereby improving image reconstruction efficiency and image quality. Furthermore, based on the forward projection operator constructed using camera calibration parameters and the observed projection vector, the reprojection residual of the denoised image is calculated and a correction amount is generated. This correction amount is used to perform data consistency correction on the denoised image to obtain the target reconstructed image. This method can pull the increased reprojection residual after image denoising back to the initial reconstruction level, thereby maintaining projection consistency and preventing consistency degradation caused by depth denoising. This invention can effectively suppress isolated points and stripe artifacts that occur during image reconstruction, further improving the quality of image reconstruction. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 This document provides a schematic flowchart of an NNBI visual tomography image reconstruction method. Figure 2 This is a schematic diagram of the ROI mask gating module (MaskGate) and the structure prior module (Prior) provided in this specification; Figure 3 This diagram illustrates how MAE_proj, provided in this specification, serves as a projection domain consistency index to measure data fit at different noise levels. Figure 4 This is a visual interface diagram of the NNBI system provided in this manual. Figure 5 This is a schematic diagram of an NNBI visual tomography image reconstruction system provided in this specification; Figure 6 This is a schematic diagram of a computer device for implementing the NNBI visual tomographic image reconstruction method provided in this specification. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0019] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0020] This invention provides a method for visual tomography image reconstruction using NNBI. Figure 1 The flowchart of this method is shown, and it specifically includes the following steps: S101: Acquire multi-angle beam images of the NNBI system simultaneously using multiple cameras.

[0021] This invention utilizes high-voltage / injection timing or external trigger signals as a reference to perform global shutter synchronous exposure on multiple industrial cameras circumferentially arranged in the NNBI system, obtaining a multi-view beam image dataset at the same physical moment. Specifically, 3–8 industrial cameras (or positions such as 0 / 90 / 180 / 270°, 0 / 120 / 240°, etc., selected according to engineering conditions) are arranged at equal angles circumferentially around the NNBI beamline, with the cameras operating in global shutter mode. Then, using high-voltage triggering as a time base, a synchronization pulse is output through an isolated timing distributor and used to drive the trigger ports of each camera, achieving multi-angle synchronous acquisition of beam images. The acquired data is aggregated to the main control computer via optical fiber. The main control computer distributes parameters such as exposure gate width, gain, and delay, performs frame drop detection and writes timestamps / trigger counts, and caches the images at high speed to the NVMe array to ensure continuous acquisition. To suppress strong electromagnetic interference, shielded cables are used for signal and power supply, with the shielding layer grounded 360° and single-point grounded. Common-mode suppression and surge protection are configured at the power input if necessary.

[0022] S102: Process the multi-angle beam image to generate the observation projection map, and construct the forward projection operator and the observation projection vector based on the camera calibration parameters; use the tomographic reconstruction algorithm to invert the observation projection map to obtain the initial reconstructed image.

[0023] First, the acquired multi-angle beam images undergo geometric calibration and photometric preprocessing, including flat-field correction, exposure time normalization, filter bandwidth response correction, background subtraction (such as bispectral ratio subtraction), and bad / saturated pixel processing. Specifically, geometric calibration is preferably performed during the maintenance window using a pinhole array or checkerboard calibration board to obtain camera intrinsic, distortion, and extrinsic parameters. The extrinsic parameter angle RMS ≤ 0.2° and reprojection error ≤ 0.3 pixels are used as acceptable thresholds; exceeding these thresholds triggers recalibration. Then, through steps such as acquiring calibration parameters, coordinate transformation, perspective projection, applying distortion correction, and converting to pixel coordinates, a forward projection operator A is constructed. Photometric preprocessing includes dark-field / beamless background acquisition, flat-field correction, exposure time normalization, and bad / saturated pixel processing. To reduce stray light and background drift, dual-spectral channels (e.g., Hα and background bandpass) can be configured and the background channel signal can be subtracted from the main channel using the dual-spectral scaling method. Then, the images from each viewpoint are mapped to observation projection maps and uniformly normalized to a preset dynamic range (e.g., [0,1]) to form the observation projection vector b.

[0024] Then, based on the geometric calibration parameters, the forward projection operator A is established, and the observed projection image is used as input. The initial reconstructed image x0 is obtained by iterative inversion using the Synchronous Algebraic Tomography Reconstruction (SART) algorithm. The SART algorithm is an iterative reconstruction algorithm for tomographic imaging (such as CT, PET, etc.). It requires only a few iterations to achieve good reconstruction quality and accuracy, and has high reconstruction efficiency. In this embodiment, the number of SART iterations is preferably 20–100, and the relaxation factor is preferably 0.1–0.5, which can be determined by the validation set and fixed during operation. To suppress restricted angular fringe artifacts, a round of TV denoising can be embedded after each round of SART update to form a SART+TV baseline, and the stopping criterion is that the relative decrease in projection residual is <1% or the maximum number of iterations is reached.

[0025] S103. Construct a ROI mask for the target region of the image, input the initial reconstructed image and the ROI mask as two channels to the pre-trained residual denoising network and output the residual image; obtain the denoised image based on the residual image, and use the forward projection operator and the observation projection vector to calculate the reprojection residual of the denoised image.

[0026] Among them, the ROI (Region of Interest) mask is one of the core technologies in digital image processing. It is used to specify the region of interest or processing in an image, masking irrelevant background and improving computational efficiency and analysis accuracy. In this embodiment of the invention, the ROI mask is a fixed circular ROI mask m (the physically effective area of ​​the beam section) adaptively generated by thresholding and morphological processing of the target region in the initial reconstructed image. The center of the fixed circular ROI mask m is given by the beamline calibration center or estimated by the intensity centroid of the initial reconstructed image. The fixed circular ROI mask m has a boundary margin ratio (e.g., 0.05) to avoid boundary noise. The ROI mask is used for subsequent image denoising, correction, and to unify the calculation criteria for indicators.

[0027] The aforementioned residual denoising network includes a multi-layer convolutional residual structure, and inserts an ROI mask gating module and a structural prior module every few convolutional blocks in the image feature channels.

[0028] Specifically, the residual denoising network is preferably a 17-layer convolutional residual structure: the first layer is a 3×3 convolution plus ReLU to expand the channels to 64; the middle 15 convolutional blocks consist of 3×3 convolutions, GroupNorm(8,C) and ReLU; the last layer is a 3×3 convolution to regress the single-channel residual. To suppress background interference outside the ROI, an ROI mask gating module MaskGate is inserted every two convolutional blocks on the feature channels. The weight g of the MaskGate module is generated by a 1×1 convolution-ReLU-Sigmoid jointly driven by image features and the ROI mask, and a learnable coefficient β is set and initialized to 0. This causes the network to degenerate into a standard DnCNN in the early stages of training, and then the gating is gradually opened to enhance the representation inside the ROI.

[0029] In addition, to introduce prior beam cross-section structure, after several convolutional blocks, a prior structure module Prior can be connected in series after the ROI mask gating module MaskGate: based on feature-based adaptive global pooling, predict the two-dimensional Gaussian center (c_x,c_y)∈[-1,1] and scale (s_x,s_y)>0, generate a two-dimensional or elliptical Gaussian weight field P and modulate the features in a residual manner; in another implementation, the rotation angle θ can be further predicted to form a rotating elliptical Gaussian. Specifically, the Prior module takes the feature hmask output by the MaskGate module as input and includes a parameter generation branch and a prior enhancement branch. The parameter generation branch is used to construct a two-dimensional Gaussian prior. First, the parameter generation branch performs global average pooling (GAP) on the input feature hmask to extract global statistical information of the input feature. Then, the pooled feature is input into the parameter head (param_head) composed of 1×1 convolutions to generate parameters for describing the two-dimensional Gaussian distribution. Further, the above parameters are subjected to tanh+softplus nonlinear mapping processing, where tanh is used to constrain the range of the center position parameter and softplus is used to ensure that the scale parameter is positive, thereby obtaining a two-dimensional Gaussian prior map g(x,y) that satisfies the distribution constraints.

[0030] The aforementioned prior enhancement branch is used to inject the two-dimensional Gaussian prior map g(x,y) into the input features. Specifically, the input features hmask are first subjected to Dropout2D to obtain the robust feature hdrop; then, the robust feature hdrop is multiplied element-wise with the two-dimensional Gaussian prior map g(x,y) to obtain the feature response modulated by the spatial prior; subsequently, the prior modulation intensity is adaptively controlled by tanh(α) and a scaling factor; finally, the scaled prior modulation result is added element-wise with the robust feature hdrop to obtain the prior-enhanced output feature hprior. Through this structure, the Prior module can adaptively generate a two-dimensional Gaussian spatial prior that matches the current sample based on the input features, and embed the spatial prior into the feature representation in a residual enhancement manner.

[0031] The denoising network training and inference preferably calculates loss and statistical metrics within the Region of Interest (ROI) of the image, and can crop the output range to maintain consistency with the normalization domain. The structures of the ROI mask gating module (MaskGate) and the structure prior module (Prior) are as follows: Figure 2 As shown.

[0032] The ROI mask gating module MaskGate in this embodiment generates gating weights g based on the ROI mask m and intermediate image features. Then, it performs point-by-point or channel-by-channel soft gating on the features, enabling the network to focus on learning the effective region within the ROI, reducing interference from irrelevant backgrounds, suppressing background interference outside the ROI, improving the network's denoising stability under constrained angles and strong noise conditions, and reducing background contamination of gradient updates. The structure prior module Prior predicts the structural parameters of the beam cross-section (such as center, scale, or elliptical axis length) from image features, generates a two-dimensional Gaussian weight field P, and weights the features accordingly. This explicitly introduces a structural prior of approximately radial / elliptical symmetry of the beam cross-section, enhancing the suppression of fringe artifacts and isolated extrema, and reducing the risk of extreme errors.

[0033] It should be noted that the training and testing sample data of the above residual denoising network are generated by simulation using a neutral beam forward model of a negative ion source, or by synchronously acquiring measured projection data using multiple cameras. Geometric parameters are configured according to the actual beam source design values, including beam cross-section size, distance from the source to the observation window, and camera field of view opening position. Beam current intensity is set according to existing data and other requirements. The beam cross-section is discretized into a regular grid within the physical ROI. A line integral of the line of sight matrix A is constructed using geometric ray tracing, and the ideal projection is obtained by performing a line integral on a given emissivity distribution x. The emissivity distribution includes upper and lower layered uniform fields, linearly gradient fields, and non-uniform scenes. Multiplicative noise with intensities α = 0.08, 0.12, 0.15, and 0.20 is superimposed on each scene, and noisy initial reconstructed images are obtained through SART. The total number of samples can be 960, with the training set, validation set, and test set divided into 80%, 15%, and 5% respectively.

[0034] In this implementation, the network can be trained and inferred using the aforementioned sample data within the PyTorch framework. The hardware platform can be an NVIDIA RTX 4090 GPU; the optimizer can be AdamW, with a learning rate of 1×10⁻³, a batch size of 8, and 100 epochs, and AMP mixed precision is enabled. The learning rate scheduling can employ ReduceLROnPlateau, using the validation set PSNR as the monitoring variable; the loss function is preferably calculated within the ROI, using a combination of 0.7×masked MSE + 0.3×Charbonnier loss. The validation phase can employ four test-time augmentations consisting of the original image, horizontal flip, vertical flip, and transpose, and the output is averaged to improve stability.

[0035] Furthermore, the initial reconstructed image x0 and the ROI mask m are fed into the residual denoising network as two-channel inputs. The network outputs a residual image r. The initial reconstructed image x0 is subtracted from the residual r to obtain the denoised image x1 = x0 - r. Then, the forward projection operator A and the observation projection vector b are used to calculate the reprojection residual d = b - Ax1 of the denoised image x1.

[0036] S104. Perform back projection on the reprojection residual to generate a correction amount, and perform data consistency correction on the denoised image at least once based on the correction amount to obtain the target reconstructed image.

[0037] In this embodiment of the invention, to address the potential projection consistency degradation introduced by the aforementioned image domain denoising, after calculating the reprojection residual d = b - Ax1, a data consistency (DC) correction needs to be performed to obtain the final target reconstructed image x2: x2 = x1 + γ·Δx, where Δx is the correction amount obtained by performing backprojection or weighted backprojection on the reprojection residual d, preferably Δx = A^T(b - Ax1) or its normalized or weighted form. To maintain the physical meaning of the beam cross section, the above consistency correction is preferably applied only within the ROI or applied weighted according to the ROI weight; the value of γ ranges from 0.05 to 0.8, and in one set of embodiments, it can be 0.3. The consistency correction can be performed once or multiple times, or the "relative decrease in projection residual <1%" can be used as the stopping criterion, and then the target reconstructed image x2 is output. The above data consistency correction can pull back the increased reprojection residual after pure image domain denoising to a level close to the initial reconstruction baseline, maintaining the physical consistency of the output result with the observed projection while preserving the image domain gains.

[0038] Furthermore, after obtaining the target reconstructed image x2, image domain error metrics and projection domain consistency metrics are calculated based on the target reconstructed image x2 and output to the system reconstruction display area for visualization. Image domain error metrics include MAE, MaxAE, PSNR, SSIM, etc., within the ROI. Projection domain consistency metrics can be MAE_proj or RMSE_proj. By calculating the above image domain error metrics and projection domain consistency metrics and comparing them with corresponding preset thresholds, warnings are issued for metrics exceeding the thresholds, and corresponding suggestions are made. The parameters and structure of the denoising network can be adjusted based on the comparison results to better suppress overall errors and extreme isolated point errors, while maintaining consistency, thereby improving the quality of the reconstructed image. MAE_proj is used as the projection domain consistency metric to measure the fitting performance of the original data with and without data consistency correction (DnCNN) at different noise levels, as well as with the data after data consistency correction (DnCNN+DC). Figure 3 As shown.

[0039] The NNBI system's visual interface provides convenient parameter input areas, running status areas, real-time suggestion areas, and image reconstruction display areas, such as... Figure 4 As shown, when the index or quality threshold exceeds the limit, the interface can prompt suggestions such as triggering recalibration, adjusting the gate width or γ, or checking shading and background subtraction. For online image diagnostic applications, the beam uniformity index, divergence angle (RMS or full width at half maximum), and beam waist position can be further calculated based on the target reconstructed image x2, and a report can be generated on the system interface for easy viewing and analysis by staff.

[0040] based on Figure 1 The NNBI visual tomographic image reconstruction method shown above uses a tomographic reconstruction algorithm to invert the observed projection map of the NNBI multi-angle beam image, obtains the initial reconstructed image, constructs its ROI mask, and then inputs it into a pre-trained residual denoising network in two channels to obtain the residual image and calculate the denoised image. This method can suppress error interference caused by background and noise outside the ROI and improve the stability of denoising, thereby improving image reconstruction efficiency and image quality. Furthermore, based on the forward projection operator constructed using camera calibration parameters and the observed projection vector, the reprojection residual of the denoised image is calculated and a correction amount is generated. This correction amount is used to perform data consistency correction on the denoised image to obtain the target reconstructed image. This method can pull the increased reprojection residual after image denoising back to the initial reconstruction level, thereby maintaining projection consistency and preventing consistency degradation caused by depth denoising, further improving the quality of the reconstructed image.

[0041] When applying the NNBI visual tomography image reconstruction method provided in this manual, it is not necessary to follow the instructions provided. Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0042] The above describes one or more embodiments of the NNBI visual tomography image reconstruction method provided in this specification. Based on the same idea, this specification also provides a corresponding NNBI visual tomography image reconstruction system, such as... Figure 5 As shown.

[0043] Figure 5 The NNBI visual tomography image reconstruction system provided in this specification includes: The acquisition module 401 is used to simultaneously acquire multi-angle beam images of the NNBI system through multiple cameras; The generation and inversion module 402 is used to process the multi-angle beam image to generate an observation projection map, and construct a forward projection operator and an observation projection vector based on camera calibration parameters; and use a tomographic reconstruction algorithm to invert the observation projection map to obtain an initial reconstructed image. The construction and calculation module 403 is used to construct a ROI mask for the target region of the image, input the initial reconstructed image and the ROI mask as two channels to a pre-trained residual denoising network and output a residual image; obtain a denoised image based on the residual image, and calculate the reprojection residual of the denoised image using the forward projection operator and the observation projection vector; The correction module 404 is used to back-project the reprojection residual to generate a correction amount, and perform data consistency correction on the denoised image at least once based on the correction amount to obtain the target reconstructed image.

[0044] Specific limitations regarding the NNBI visual tomography image reconstruction system can be found in the limitations of the NNBI visual tomography image reconstruction method described above, and will not be repeated here. Each module in the aforementioned NNBI visual tomography image reconstruction system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0045] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided NNBI visual tomography image reconstruction method.

[0046] This instruction manual also provides Figure 6 The schematic diagram of the computer device shown is as follows: Figure 6 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The provided NNBI visual tomography image reconstruction method.

[0047] The NNBI visual tomographic image reconstruction method and system provided in this specification utilizes a tomographic reconstruction algorithm to invert the observed projection map of the NNBI multi-angle beam image, obtains an initial reconstructed image, constructs its ROI mask, and then inputs it into a pre-trained residual denoising network in two channels to obtain the residual image and calculate the denoised image. This method can suppress error interference caused by background and noise outside the ROI and improve the stability of denoising, thereby improving image reconstruction efficiency and image quality. Furthermore, based on the forward projection operator constructed using camera calibration parameters and the observed projection vector, the reprojection residual of the denoised image is calculated and a correction amount is generated. This correction amount is used to perform data consistency correction on the denoised image to obtain the target reconstructed image. This method can pull the increased reprojection residual after image denoising back to the initial reconstruction level, thereby maintaining projection consistency and preventing consistency degradation caused by depth denoising. This invention can effectively suppress isolated points and stripe artifacts that occur during image reconstruction, further improving the quality of image reconstruction.

[0048] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0049] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for reconstructing NNBI visual tomographic images, characterized in that, include: Multi-angle beam images of the NNBI system are acquired simultaneously using multiple cameras; The multi-angle beam image is processed to generate an observation projection map, and a forward projection operator and an observation projection vector are constructed based on the camera calibration parameters; the observation projection map is inverted using a tomographic reconstruction algorithm to obtain an initial reconstructed image; A ROI mask for the target region of the image is constructed. The initial reconstructed image and the ROI mask are input as two channels to a pre-trained residual denoising network, which outputs a residual image. A denoised image is obtained based on the residual image, and the reprojection residual of the denoised image is calculated using the forward projection operator and the observation projection vector. The reprojection residual is back-projected to generate a correction amount, and the denoised image is subjected to at least one data consistency correction based on the correction amount to obtain the target reconstructed image.

2. The NNBI visual tomographic image reconstruction method as described in claim 1, characterized in that, The residual denoising network includes a multi-layer convolutional residual structure, and an ROI mask gating module and a structural prior module are inserted every few convolutional blocks in the image feature channel.

3. The NNBI visual tomographic image reconstruction method as described in claim 2, characterized in that, The weights of the ROI mask gating module are generated by a 1×1 convolution-ReLU-Sigmoid jointly driven by image features and the ROI mask.

4. The NNBI visual tomographic image reconstruction method as described in claim 2, characterized in that, The structure prior module predicts the structural parameters of the beam cross section based on feature-based adaptive global pooling, and generates a two-dimensional or elliptical Gaussian weight field to weight and modulate the image features.

5. The NNBI visual tomography image reconstruction method as described in claim 1, characterized in that, The ROI mask is a fixed circular ROI mask adaptively generated by thresholding and morphological processing of the initial reconstructed image. The fixed circular ROI mask has a set boundary margin ratio or shrinkage ratio.

6. The NNBI visual tomographic image reconstruction method as described in claim 5, characterized in that, The data consistency correction is applied only within the image ROI, or applied after weighting the correction amount according to the ROI mask weight.

7. The NNBI visual tomography image reconstruction method as described in claim 1, characterized in that, It also includes calculating the image domain error index and the projection domain consistency index based on the target reconstructed image and outputting them to the reconstructed display area for visualization.

8. An NNBI visual tomographic image reconstruction system, characterized in that, include: The acquisition module is used to simultaneously acquire multi-angle beam images of the NNBI system using multiple cameras; The generation and inversion module is used to process the multi-angle beam image to generate an observation projection map, and to construct a forward projection operator and an observation projection vector based on camera calibration parameters; the observation projection map is inverted using a tomographic reconstruction algorithm to obtain an initial reconstructed image; The construction and calculation module is used to construct the ROI mask of the target region of the image, and input the initial reconstructed image and the ROI mask as two channels into the pre-trained residual denoising network and output the residual image. A denoised image is obtained based on the residual image, and the reprojection residual of the denoised image is calculated using the forward projection operator and the observation projection vector; The correction module is used to back-project the reprojection residual to generate a correction amount, and perform data consistency correction on the denoised image at least once based on the correction amount to obtain the target reconstructed image.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images