Steffen Kieß
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Steffen Kieß has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
Sven Simon
8 shared publicationsUniversity of Stuttgart
6 shared publicationsJajnabalkya Guhathakurta
4 shared publications- Ahmed Baraka
IQS Intelligence in Quality Sensing Laboratory for Machine Tools and Production Engineering WZL RWTH Aachen University Germany
3 shared publications Ct-Lab Ug
3 shared publications- Dominik Wolfschläger
IQS Intelligence in Quality Sensing Laboratory for Machine Tools and Production Engineering WZL RWTH Aachen University Germany
3 shared publications
Latest Publications
- Open Access
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Multi-view Diffusion Model with Gaussian Splatting in Projection Domain for Limited Angle Reconstruction
Industrial computed tomography (CT) plays a critical role in non-destructive testing and quality control; however, limited-angle acquisition often leads to severe artifacts due to restricted scanning rotation. To address this issue, we propose a novel framework that exploits the generative capability of a video diffusion model for multi-view generation in the projection domain and uses a novel reconstruction method. Specifically, a well-trained video diffusion model uses the available projections from limited-angle tomography as conditions to generate the missing-angle projections. To further enhance reconstruction quality and reduce the information loss introduced by traditional methods such as the Feldkamp–Davis–Kress (FDK) algorithm, we integrate a Gaussian splatting–based 3D reconstruction method, enabling volumetric reconstruction from the generated projections. Based on our experimental results, the proposed framework can generate novel projections of size 256 x 256 pixel in approximately 1 minute and complete the subsequent 3D reconstruction in about 5 minutes, offering new opportunities for inline CT applications. Experimental results demonstrate that novel-view generation substantially improves reconstruction performance, achieving significant gains in MSE, PSNR, and SSIM compared with conventional limited-angle tomography.
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Open AccessMulti-view Diffusion Model with Gaussian Splatting in Projection Domain for Limited Angle Reconstruction
Industrial computed tomography (CT) plays a critical role in non-destructive testing and quality control; however, limited-angle acquisition often leads to severe artifacts due to restricted scanning rotation. To address this issue, we propose a novel framework that exploits the generative capability of a video diffusion model for multi-view generation in the projection domain and uses a novel reconstruction method. Specifically, a well-trained video diffusion model uses the available projections from limited-angle tomography as conditions to generate the missing-angle projections. To further enhance reconstruction quality and reduce the information loss introduced by traditional methods such as the Feldkamp–Davis–Kress (FDK) algorithm, we integrate a Gaussian splatting–based 3D reconstruction method, enabling volumetric reconstruction from the generated projections. Based on our experimental results, the proposed framework can generate novel projections of size 256 x 256 pixel in approximately 1 minute and complete the subsequent 3D reconstruction in about 5 minutes, offering new opportunities for inline CT applications. Experimental results demonstrate that novel-view generation substantially improves reconstruction performance, achieving significant gains in MSE, PSNR, and SSIM compared with conventional limited-angle tomography.
- Open Access
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Enhanced Energy Resolution of Synchrotron Multispectral CT based on a Silicon Wafer Edge as Spectral‑Splitting Optics
Multispectral imaging at synchrotrons is of great importance since it enables the simultaneous acquisition of structural and chemical information while reducing implicitly beam hardening artefacts. This allows complex samples from the fields of materials sciences, biology, cultural heritage and others to be analyzed non-destructively. Such a comprehensive approach provides deeper insights into processes and properties that are not accessible with standard integrating detectors alone. At synchrotrons, multispectral imaging is usually based on spectral-dispersive optics for high photon fluxes, as energy-resolved detectors suffer from pile-up under synchrotron radiation. The flagship BM18 beamline at the ESRF, with a beam width of over 300 mm and a spectral width of up to 400 keV, is an ideal source of radiation for such a multispectral imaging approach. In this paper, we propose a 90° cut Si wafer edge as an effective, low-cost refractive optic compared to the Si prism array (100 prisms) inspired by prism-based setups in the literature. For the tungsten K-edge (69.5 keV), the prism array produces a beam shift of 25 pixels, while the wafer edge produces 174 pixels, which corresponds to a 7-fold higher refraction. The wafer edge thus enables shorter distances between the optics and the detector or higher energy resolution.
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Open AccessEnhanced Energy Resolution of Synchrotron Multispectral CT based on a Silicon Wafer Edge as Spectral‑Splitting Optics
Multispectral imaging at synchrotrons is of great importance since it enables the simultaneous acquisition of structural and chemical information while reducing implicitly beam hardening artefacts. This allows complex samples from the fields of materials sciences, biology, cultural heritage and others to be analyzed non-destructively. Such a comprehensive approach provides deeper insights into processes and properties that are not accessible with standard integrating detectors alone. At synchrotrons, multispectral imaging is usually based on spectral-dispersive optics for high photon fluxes, as energy-resolved detectors suffer from pile-up under synchrotron radiation. The flagship BM18 beamline at the ESRF, with a beam width of over 300 mm and a spectral width of up to 400 keV, is an ideal source of radiation for such a multispectral imaging approach. In this paper, we propose a 90° cut Si wafer edge as an effective, low-cost refractive optic compared to the Si prism array (100 prisms) inspired by prism-based setups in the literature. For the tungsten K-edge (69.5 keV), the prism array produces a beam shift of 25 pixels, while the wafer edge produces 174 pixels, which corresponds to a 7-fold higher refraction. The wafer edge thus enables shorter distances between the optics and the detector or higher energy resolution.
- Open Access
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Compensating Projection Interpolation Artifacts in Sparse-View CT using a Two- Stage CNN with learnable Upscaling
X-ray Computed Tomography (CT) is a valuable tool for non-destructive testing, typically requiring thousands of projections to reconstruct high-quality images. For inline CT applications, however, scan time constraints necessitate reducing either the tube current or the number of projections. These limitations introduce significant noise and streak artifacts in reconstructed images. Building on existing approaches for artifact compensation, we propose a novel method to address missing projections through learned angular upscaling of sparse-view sinograms. Our approach modifies the standard U-Net architecture by integrating additional convolution and pixel-shuffle layers to recover missing data. During training, the upscaling kernel parameters are optimized to maximize alignment with ground-truth projections through three training phases. Even at an extreme sparsity factor of 20× fewer projections, our method demonstrates substantial improvement over interpolation-based upscaling deep learning techniques and achieves significant enhancement of sparse-view reconstructions for experimental results.
- 1 March 2026
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 31 (3)
Open AccessCompensating Projection Interpolation Artifacts in Sparse-View CT using a Two- Stage CNN with learnable Upscaling
X-ray Computed Tomography (CT) is a valuable tool for non-destructive testing, typically requiring thousands of projections to reconstruct high-quality images. For inline CT applications, however, scan time constraints necessitate reducing either the tube current or the number of projections. These limitations introduce significant noise and streak artifacts in reconstructed images. Building on existing approaches for artifact compensation, we propose a novel method to address missing projections through learned angular upscaling of sparse-view sinograms. Our approach modifies the standard U-Net architecture by integrating additional convolution and pixel-shuffle layers to recover missing data. During training, the upscaling kernel parameters are optimized to maximize alignment with ground-truth projections through three training phases. Even at an extreme sparsity factor of 20× fewer projections, our method demonstrates substantial improvement over interpolation-based upscaling deep learning techniques and achieves significant enhancement of sparse-view reconstructions for experimental results.
- Open Access
Multispectral CT Denoising via Simulation-Trained Deep Learning: Experimental Results at the ESRF BM18
Multispectral computed tomography (CT) enables advanced material characterization by acquiring energy-resolved projection data. However, since the incoming X-ray flux is be distributed across multiple narrow energy bins, the photon count per bin is greatly reduced compared to standard energy-integrated imaging. This inevitably introduces substantial noise, which can either prolong acquisition times and make scan durations infeasible or degrade image quality with strong noise artifacts. To address this challenge, we present a dedicated neural network-based denoising approach tailored for multispectral CT projections acquired at the BM18 beamline of the ESRF. The method exploits redundancies across angular, spatial, and spectral domains through specialized sub-networks combined via stacked generalization and an attention mechanism. Non-local similarities in the angular-spatial domain are leveraged alongside correlations between adjacent energy bands in the spectral domain, enabling robust noise suppression while preserving fine structural details. Training was performed exclusively on simulated data replicating the physical and noise characteristics of the BM18 setup, with validation conducted on CT scans of custom-designed phantoms containing both high-Z and low-Z materials. The denoised projections and reconstructions demonstrate substantial improvements in image quality compared to classical denoising methods and baseline CNN models. Quantitative evaluations confirm that the proposed method achieves superior performance across a broad spectral range, generalizing effectively to real-world experimental data while significantly reducing noise without compromising structural fidelity.
Open AccessMultispectral CT Denoising via Simulation-Trained Deep Learning: Experimental Results at the ESRF BM18
Multispectral computed tomography (CT) enables advanced material characterization by acquiring energy-resolved projection data. However, since the incoming X-ray flux is be distributed across multiple narrow energy bins, the photon count per bin is greatly reduced compared to standard energy-integrated imaging. This inevitably introduces substantial noise, which can either prolong acquisition times and make scan durations infeasible or degrade image quality with strong noise artifacts. To address this challenge, we present a dedicated neural network-based denoising approach tailored for multispectral CT projections acquired at the BM18 beamline of the ESRF. The method exploits redundancies across angular, spatial, and spectral domains through specialized sub-networks combined via stacked generalization and an attention mechanism. Non-local similarities in the angular-spatial domain are leveraged alongside correlations between adjacent energy bands in the spectral domain, enabling robust noise suppression while preserving fine structural details. Training was performed exclusively on simulated data replicating the physical and noise characteristics of the BM18 setup, with validation conducted on CT scans of custom-designed phantoms containing both high-Z and low-Z materials. The denoised projections and reconstructions demonstrate substantial improvements in image quality compared to classical denoising methods and baseline CNN models. Quantitative evaluations confirm that the proposed method achieves superior performance across a broad spectral range, generalizing effectively to real-world experimental data while significantly reducing noise without compromising structural fidelity.
- Open Access
- 1 August 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (8)
A 3D Denoising Neural Network for Improved Surface Quality in Inline CT Metrology
Inline computed tomography (CT) is becoming increasingly important for 100% inspection in production technology. Rapid scanning cycles, often limited to minutes, require a reduction in the number of X-ray projections, resulting in significantly reduced photon counts compared to laboratory-based systems. This reduction adversely impacts the peak signal-to-noise ratio (PSNR), increasing noise levels and compromising the accuracy of surface measurements essential for dimensional metrology. To address this challenge, we use neural network-based denoising techniques that enhance image quality without the need for clean reference data. Additionally, we introduce a novel surface quality assessment metric based on the Hausdorff distance, enabling real-time monitoring of surface accuracy during CT scans. Our approach uses a three-dimensional U-Net architecture and an innovative data augmentation method to optimize denoising performance. Experimental results demonstrate that the proposed methods significantly improve surface quality and PSNR in CT data within the time constraints necessary for real-time applications.
- 1 August 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (8)
Open AccessA 3D Denoising Neural Network for Improved Surface Quality in Inline CT Metrology
Inline computed tomography (CT) is becoming increasingly important for 100% inspection in production technology. Rapid scanning cycles, often limited to minutes, require a reduction in the number of X-ray projections, resulting in significantly reduced photon counts compared to laboratory-based systems. This reduction adversely impacts the peak signal-to-noise ratio (PSNR), increasing noise levels and compromising the accuracy of surface measurements essential for dimensional metrology. To address this challenge, we use neural network-based denoising techniques that enhance image quality without the need for clean reference data. Additionally, we introduce a novel surface quality assessment metric based on the Hausdorff distance, enabling real-time monitoring of surface accuracy during CT scans. Our approach uses a three-dimensional U-Net architecture and an innovative data augmentation method to optimize denoising performance. Experimental results demonstrate that the proposed methods significantly improve surface quality and PSNR in CT data within the time constraints necessary for real-time applications.
- Open Access
- 1 August 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (8)
A Neural Network for Denoising Multispectral Tomography Data at BM18 of the European Synchrotron Radiation Facility
Multispectral imaging is susceptible to high levels of noise due to the low photon flux observed, particularly in narrow energy bins. To address this issue, multispectral data was acquired at the BM18 beamline of the ESRF using a silicon prism array to refract a polychromatic beam into energy bins, which were then captured by an SCMOS detector. Unlike photon counting detectors, which suffer from pile-up and low energy resolution, the proposed method can handle the high flux of the unfiltered BM18 beam and split the beam into 100 energy bins. The resulting data, which was significantly affected by noise, was denoised using a neural network that had been trained on simulated data. The efficacy of the denoising method was evaluated on computed tomography scans of two phantoms containing high and low atomic number (Z) materials, as well as other samples. Despite differences between the simulation and experimental setups, the network demonstrated a significant reduction in noise, as illustrated in the projections and reconstructed slices. The denoised reconstructions exhibited a notable suppression of noise without compromising image detail or introducing additional artifacts, emphasizing the model's potential for enhancing multispectral imaging in material analysis.
- 1 August 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (8)
Open AccessA Neural Network for Denoising Multispectral Tomography Data at BM18 of the European Synchrotron Radiation Facility
Multispectral imaging is susceptible to high levels of noise due to the low photon flux observed, particularly in narrow energy bins. To address this issue, multispectral data was acquired at the BM18 beamline of the ESRF using a silicon prism array to refract a polychromatic beam into energy bins, which were then captured by an SCMOS detector. Unlike photon counting detectors, which suffer from pile-up and low energy resolution, the proposed method can handle the high flux of the unfiltered BM18 beam and split the beam into 100 energy bins. The resulting data, which was significantly affected by noise, was denoised using a neural network that had been trained on simulated data. The efficacy of the denoising method was evaluated on computed tomography scans of two phantoms containing high and low atomic number (Z) materials, as well as other samples. Despite differences between the simulation and experimental setups, the network demonstrated a significant reduction in noise, as illustrated in the projections and reconstructed slices. The denoised reconstructions exhibited a notable suppression of noise without compromising image detail or introducing additional artifacts, emphasizing the model's potential for enhancing multispectral imaging in material analysis.
- Open Access
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
A Framework for the AI-based visualization and analysis of massive amounts of 4D tomography data for end users of beamlines
The size of 4D tomography datasets acquired at synchrotron or neutron imaging facilities can reach several terabytes, which presents a significant challenge for their evaluation. This paper presents a framework that allows a compressed dataset to be kept in memory and makes it possible to evaluate and manipulate the dataset without requiring enough memory to decompress the entire dataset. The framework enables the compensation of imaging artifacts, including the compression artifacts of the 4D dataset, through the integration of neural networks. The reduction of imaging artifacts can be performed at the imaging facility or at the user's home institution. This framework reduces the computational burden on the computing infrastructure of large synchrotron and neutron facilities by allowing end users to process datasets on their institution's computers. This is made possible by compressing TBs of data to less than 128 GB, allowing powerful PCs to process TBs of 4D tomography data.
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Open AccessA Framework for the AI-based visualization and analysis of massive amounts of 4D tomography data for end users of beamlines
The size of 4D tomography datasets acquired at synchrotron or neutron imaging facilities can reach several terabytes, which presents a significant challenge for their evaluation. This paper presents a framework that allows a compressed dataset to be kept in memory and makes it possible to evaluate and manipulate the dataset without requiring enough memory to decompress the entire dataset. The framework enables the compensation of imaging artifacts, including the compression artifacts of the 4D dataset, through the integration of neural networks. The reduction of imaging artifacts can be performed at the imaging facility or at the user's home institution. This framework reduces the computational burden on the computing infrastructure of large synchrotron and neutron facilities by allowing end users to process datasets on their institution's computers. This is made possible by compressing TBs of data to less than 128 GB, allowing powerful PCs to process TBs of 4D tomography data.
- Open Access
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Surface Quality Monitoring and Improvement for Dimensional Metrology in Inline CT by Denoising with Neural Networks and Fast Surface Quality Metric
Inline computed tomography (CT) is becoming increasingly important for 100% inspection in production technology. However, the short machine cycle times—often just a few minutes—require either fewer X-ray projections or shorter detector exposure times, significantly reducing photon counts by one to two orders of magnitude compared to lab-based CT systems. This reduction leads to a decrease in the signal-to-noise ratio (SNR) by at least an order of magnitude, thereby increasing noise and diminishing the accuracy of surface measurements essential for metrology. To address this challenge, denoising methods based on neural networks have been used, outperforming traditional algorithms such as BM3D. In this work, we investigate the Noise2Noise method, which trains a neural network using pairs of noisy images, eliminating the need for clean ground truth data—typically unavailable in CT. Furthermore, we propose a fast surface quality metric based on the marching cubes algorithm and Hausdorff distance to enable realtime monitoring of voxel dataset quality during continuous CT scans, as required in inline CT. This metric demonstrates a strong correlation with measurement errors associated with surface quality degradation due to noise. Experimental results show that our proposed methods significantly enhance surface quality and SNR of full CT data volumes, achieving denoising within minutes, as required for inline CT applications.
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Open AccessSurface Quality Monitoring and Improvement for Dimensional Metrology in Inline CT by Denoising with Neural Networks and Fast Surface Quality Metric
Inline computed tomography (CT) is becoming increasingly important for 100% inspection in production technology. However, the short machine cycle times—often just a few minutes—require either fewer X-ray projections or shorter detector exposure times, significantly reducing photon counts by one to two orders of magnitude compared to lab-based CT systems. This reduction leads to a decrease in the signal-to-noise ratio (SNR) by at least an order of magnitude, thereby increasing noise and diminishing the accuracy of surface measurements essential for metrology. To address this challenge, denoising methods based on neural networks have been used, outperforming traditional algorithms such as BM3D. In this work, we investigate the Noise2Noise method, which trains a neural network using pairs of noisy images, eliminating the need for clean ground truth data—typically unavailable in CT. Furthermore, we propose a fast surface quality metric based on the marching cubes algorithm and Hausdorff distance to enable realtime monitoring of voxel dataset quality during continuous CT scans, as required in inline CT. This metric demonstrates a strong correlation with measurement errors associated with surface quality degradation due to noise. Experimental results show that our proposed methods significantly enhance surface quality and SNR of full CT data volumes, achieving denoising within minutes, as required for inline CT applications.
- Open Access
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Impact of JPEG compression on the metrological characteristics of industrial CT data
The basis of metrology with computed tomography (CT) are volume datasets with increasing sizes in the two- or three-digit GB range due to the increasing detector resolutions in the double-digit MPixel range. Handling these large amounts of data poses a considerable challenge in metrology in terms of sufficiently fast visualization, analysis and data transfer via networks (local networks or over the Internet) as well as cost- and energy-efficient archiving of TB to PB data volumes. Therefore, this work investigates the impact of lossy compression of volume data on the metrological quality of CT datasets using JPEG, which reduces the amount of data by more than one order of magnitude. The JPEG compression standard is used for compression here because it is available in CT voxel data viewers and its compression artifacts are well known from its widespread use for image compression. It is shown that the impact of JPEG compression artifacts on the metrological quality of CT datasets is limited, how the measurement uncertainty contribution due to compression can be determined and how its acceptability for metrology in general can be evaluated.
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Open AccessImpact of JPEG compression on the metrological characteristics of industrial CT data
The basis of metrology with computed tomography (CT) are volume datasets with increasing sizes in the two- or three-digit GB range due to the increasing detector resolutions in the double-digit MPixel range. Handling these large amounts of data poses a considerable challenge in metrology in terms of sufficiently fast visualization, analysis and data transfer via networks (local networks or over the Internet) as well as cost- and energy-efficient archiving of TB to PB data volumes. Therefore, this work investigates the impact of lossy compression of volume data on the metrological quality of CT datasets using JPEG, which reduces the amount of data by more than one order of magnitude. The JPEG compression standard is used for compression here because it is available in CT voxel data viewers and its compression artifacts are well known from its widespread use for image compression. It is shown that the impact of JPEG compression artifacts on the metrological quality of CT datasets is limited, how the measurement uncertainty contribution due to compression can be determined and how its acceptability for metrology in general can be evaluated.
- Open Access
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Deep Incremental Angle Refinement Model for Limited-Angle CT Reconstruction - A Case Study on Concrete Specimens
Industrial computed tomography (CT) is widely utilized for non-destructive testing and quality control in various industries. However, a common challenge in industrial CT is the presence of artifacts caused by limited angle tomography, where the object cannot be rotated fully due to geometric constraints or time limitations. To eliminate the artifacts, we propose a novel framework based on diffusion model: Deep Incremental Angle Refinement Model (DI-ARM). Our approach leverages the characteristics of CT projection by using reconstructed data of different limited angles as intermediate steps in the training process, replacing the traditional diffusion model of adding random Gaussian noise. This approach ensures data consistency in training process, mitigating the instability caused by sampling randomness of diffusion models. Furthermore, our method requires fewer steps compared to conventional diffusion models, significantly reducing computational resource consumption.
- 1 February 2025
- Published by NDT.net GmbH & Co. KG in e-Journal of Nondestructive Testing
- Vol. 30 (2)
Open AccessDeep Incremental Angle Refinement Model for Limited-Angle CT Reconstruction - A Case Study on Concrete Specimens
Industrial computed tomography (CT) is widely utilized for non-destructive testing and quality control in various industries. However, a common challenge in industrial CT is the presence of artifacts caused by limited angle tomography, where the object cannot be rotated fully due to geometric constraints or time limitations. To eliminate the artifacts, we propose a novel framework based on diffusion model: Deep Incremental Angle Refinement Model (DI-ARM). Our approach leverages the characteristics of CT projection by using reconstructed data of different limited angles as intermediate steps in the training process, replacing the traditional diffusion model of adding random Gaussian noise. This approach ensures data consistency in training process, mitigating the instability caused by sampling randomness of diffusion models. Furthermore, our method requires fewer steps compared to conventional diffusion models, significantly reducing computational resource consumption.