Kuan Shen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kuan Shen has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Fuquan Wang
Key Lab of Optoelectronic Technology and Systems, and Engineering Research Center of Industrial Computed Tomography Nondestructive Testing, Ministry of Education, Chongqing University, Chongqing, China
2 shared publications - Ao Wang
Chongqing University
1 shared publication Chao Long
1 shared publicationChuandong Tan
1 shared publicationDingmei Jing
College of Optoelectronic Engineering, Chongqing University, Chongqing, China
1 shared publicationEnxuan Zhao
1 shared publication
Latest Publications
- Open Access
- 5 August 2026
- Published by Elsevier in SSRN Electronic Journal
Sparse-Point Initialization and Decaying Opacity Perturbation for Sparse-View CT Reconstruction with 3D Gaussian Splatting
Sparse-view computed tomography (CT) reduces radiation dose and acquisition time, but the limited number of projections makes the reconstruction problem highly underdetermined, leading to streak artifacts and structural degradation. Existing 3D Gaussian Splatting (3DGS)-based CT reconstruction methods typically initialize Gaussian primitives from full FDK point clouds, causing streak artifacts and beam-hardening responses in sparse-view FDK reconstructions to be inherited by the initial Gaussian distribution. Moreover, limited projection supervision tends to drive Gaussian primitives into fixed co-adaptation relationships, resulting in needle-like artifacts during optimization. To address these issues, we propose a sparse-view CT reconstruction method that combines Sparse-Point Initialization (SPI) with Decaying Opacity Perturbation (DOP). SPI selects representative structural points from the FDK reconstruction through joint density-gradient scoring and improves spatial coverage using Adaptive Boundary Expansion (ABE), thereby reducing artifact contamination in the initial Gaussian distribution. DOP applies iteration-decaying multiplicative Laplace noise to Gaussian opacities during training, continuously modulating Gaussian contributions to suppress fixed co-adaptation. Experiments on the public FIPS dataset and self-built industrial CT datasets demonstrate that the proposed method improves PSNR by up to 1.132~dB and SSIM by up to 0.023 over the R2-Gaussian baseline under 25-view reconstruction, while effectively reducing the Co-Adaptation Score (CA).
- 5 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSparse-Point Initialization and Decaying Opacity Perturbation for Sparse-View CT Reconstruction with 3D Gaussian Splatting
Sparse-view computed tomography (CT) reduces radiation dose and acquisition time, but the limited number of projections makes the reconstruction problem highly underdetermined, leading to streak artifacts and structural degradation. Existing 3D Gaussian Splatting (3DGS)-based CT reconstruction methods typically initialize Gaussian primitives from full FDK point clouds, causing streak artifacts and beam-hardening responses in sparse-view FDK reconstructions to be inherited by the initial Gaussian distribution. Moreover, limited projection supervision tends to drive Gaussian primitives into fixed co-adaptation relationships, resulting in needle-like artifacts during optimization. To address these issues, we propose a sparse-view CT reconstruction method that combines Sparse-Point Initialization (SPI) with Decaying Opacity Perturbation (DOP). SPI selects representative structural points from the FDK reconstruction through joint density-gradient scoring and improves spatial coverage using Adaptive Boundary Expansion (ABE), thereby reducing artifact contamination in the initial Gaussian distribution. DOP applies iteration-decaying multiplicative Laplace noise to Gaussian opacities during training, continuously modulating Gaussian contributions to suppress fixed co-adaptation. Experiments on the public FIPS dataset and self-built industrial CT datasets demonstrate that the proposed method improves PSNR by up to 1.132~dB and SSIM by up to 0.023 over the R2-Gaussian baseline under 25-view reconstruction, while effectively reducing the Co-Adaptation Score (CA).
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 12, 1241-1253
GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View Tomographic Reconstruction
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 12, 1241-1253
GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View Tomographic Reconstruction
- Open Access
- 26 May 2025
- Published by Springer Nature in EURASIP Journal on Image and Video Processing
- Vol. 2025 (1) , 1-17
A novel digital radiography image fusion enhancement algorithm based on NSST
Due to the limitations of X-ray digital imaging in handling workpieces with a large thickness variation range, Moreover, digital radiography (DR) itself is prone to noise, uneven grayscale distribution, and it is difficult for the human eye to obtain useful information from it. In this study, we propose an enhanced visual sensitivity fusion method based on the non-subsampled shearlet transform (NSST), leveraging the inherent characteristics of the human visual system. Consequently, the fused DR image exhibits heightened sensitivity to human perception and enhances discernibility. This facilitates easier observation of distinctive features in DR images and enables more effective extraction of crucial information. The NSST is used to decompose the source image into high- and low-frequency coefficients through multi-scale decomposition. The proposed fusion rule, which enhances visual sensitivity, is applied to combine the high-frequency coefficients. The fused image is then reconstructed using the inverse NSST transform by incorporating both high- and low-frequency coefficients. Experimental results demonstrate that our method outperforms traditional fusion methods and other fusion rules in terms of objective evaluation indicators, such as IE, PSNR, and SSIM. Moreover, our method achieves superior fusion results for workpieces with large thickness variations. Compared to deep learning methods, our approach exhibits wider adaptability. Compared with deep learning methods, it is not limited by memory. The fusion rules also enhance details, enabling clear representation of complex internal structures and defects within workpieces. Notably, utilizing GPU parallel computing significantly reduces the algorithm’s running time.
- 26 May 2025
- Published by Springer Nature in EURASIP Journal on Image and Video Processing
- Vol. 2025 (1) , 1-17
Open AccessA novel digital radiography image fusion enhancement algorithm based on NSST
Due to the limitations of X-ray digital imaging in handling workpieces with a large thickness variation range, Moreover, digital radiography (DR) itself is prone to noise, uneven grayscale distribution, and it is difficult for the human eye to obtain useful information from it. In this study, we propose an enhanced visual sensitivity fusion method based on the non-subsampled shearlet transform (NSST), leveraging the inherent characteristics of the human visual system. Consequently, the fused DR image exhibits heightened sensitivity to human perception and enhances discernibility. This facilitates easier observation of distinctive features in DR images and enables more effective extraction of crucial information. The NSST is used to decompose the source image into high- and low-frequency coefficients through multi-scale decomposition. The proposed fusion rule, which enhances visual sensitivity, is applied to combine the high-frequency coefficients. The fused image is then reconstructed using the inverse NSST transform by incorporating both high- and low-frequency coefficients. Experimental results demonstrate that our method outperforms traditional fusion methods and other fusion rules in terms of objective evaluation indicators, such as IE, PSNR, and SSIM. Moreover, our method achieves superior fusion results for workpieces with large thickness variations. Compared to deep learning methods, our approach exhibits wider adaptability. Compared with deep learning methods, it is not limited by memory. The fusion rules also enhance details, enabling clear representation of complex internal structures and defects within workpieces. Notably, utilizing GPU parallel computing significantly reduces the algorithm’s running time.
- 16 May 2025
- Published by Elsevier in Measurement
- Vol. 253
Defect recognition of printed circuit board assemblies via enhanced Swin-Condinst network
- 16 May 2025
- Published by Elsevier in Measurement
- Vol. 253
Defect recognition of printed circuit board assemblies via enhanced Swin-Condinst network
- 12 March 2025
- Published by Springer Nature in Circuits, Systems, and Signal Processing
- Vol. 44 (7) , 5279-5304
Anomalous Sound Detection Using Self-Supervised Classification Deep Hierarchical Reconstruction Network with Symmetric Fusion Attention
- 12 March 2025
- Published by Springer Nature in Circuits, Systems, and Signal Processing
- Vol. 44 (7) , 5279-5304
Anomalous Sound Detection Using Self-Supervised Classification Deep Hierarchical Reconstruction Network with Symmetric Fusion Attention
- 18 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
SCFA-Net: Fast and High-Quality Network for Large-Size DR Image Enhancement
Digital radiography (DR) technology is characterized by high spatial resolution, substantial information capacity, and a wide dynamic range, making it widely applicable in the field of industrial nondestructive testing. However, in practical applications, the diverse types of workpieces often lead to many components with uneven thickness and structural anomalies. This results in challenges such as overexposure in thinner areas and underexposure in larger regions during DR inspections. In addition, as detector technology has advanced, the pixel matrices of detectors have continually increased, with common models now reaching 4k $\times 4$ k, capable of producing large-sized DR images. Nevertheless, existing methods often struggle to efficiently process or even infer 4k images. To successfully infer 4k images while maintaining the quality of enhanced results, we propose a scale-aware channel Fourier transform attention network (SCFA-Net) based on semantic alignment. Semantic-aligned scale-aware module (SAM) is designed to adapt to the feature extraction requirements of large-sized images. Furthermore, we innovatively introduce an attention mechanism to capture the long-range global dependencies inherent in 4k images. The proposed Fourier transform module addresses the issue of exposure inconsistency in DR efficiently. Moreover, to further achieve high-quality inference of large DR images, we ultimately modified the inference network by replacing local information aggregation (LRG) with global information aggregation. Results indicate that our method can process 4k resolution DR images at a speed of 48 frames/s, achieving state-of-the-art (SOTA) performance on two public datasets and a DR dataset.
- 18 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
SCFA-Net: Fast and High-Quality Network for Large-Size DR Image Enhancement
Digital radiography (DR) technology is characterized by high spatial resolution, substantial information capacity, and a wide dynamic range, making it widely applicable in the field of industrial nondestructive testing. However, in practical applications, the diverse types of workpieces often lead to many components with uneven thickness and structural anomalies. This results in challenges such as overexposure in thinner areas and underexposure in larger regions during DR inspections. In addition, as detector technology has advanced, the pixel matrices of detectors have continually increased, with common models now reaching 4k $\times 4$ k, capable of producing large-sized DR images. Nevertheless, existing methods often struggle to efficiently process or even infer 4k images. To successfully infer 4k images while maintaining the quality of enhanced results, we propose a scale-aware channel Fourier transform attention network (SCFA-Net) based on semantic alignment. Semantic-aligned scale-aware module (SAM) is designed to adapt to the feature extraction requirements of large-sized images. Furthermore, we innovatively introduce an attention mechanism to capture the long-range global dependencies inherent in 4k images. The proposed Fourier transform module addresses the issue of exposure inconsistency in DR efficiently. Moreover, to further achieve high-quality inference of large DR images, we ultimately modified the inference network by replacing local information aggregation (LRG) with global information aggregation. Results indicate that our method can process 4k resolution DR images at a speed of 48 frames/s, achieving state-of-the-art (SOTA) performance on two public datasets and a DR dataset.
- 1 January 2009
- Published by SAGE Publications in Journal of X-Ray Science and Technology
- Vol. 17 (3) , 207-220
Efficient x-ray image enhancement algorithm using image fusion
Multiresolution Analysis (MRA) plays an important role in image and signal processing fields, and it can extract information at different scales. Image fusion is a process of combining two or more images into an image, which extracts features from so
- 1 January 2009
- Published by SAGE Publications in Journal of X-Ray Science and Technology
- Vol. 17 (3) , 207-220
Efficient x-ray image enhancement algorithm using image fusion
Multiresolution Analysis (MRA) plays an important role in image and signal processing fields, and it can extract information at different scales. Image fusion is a process of combining two or more images into an image, which extracts features from so