Kedi Lyu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kedi Lyu has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Juncheng Hu
College of Computer Science and Technology, Jilin University, Changchun, China
2 shared publications - Zeyu Wang
College of Computer Science and Engineering, Dalian Minzu University, Dalian, China
2 shared publications - Haiyu Song
Dalian Minzu University
1 shared publication Honghao Huang
Beijing Qingzhiyuanshi Technology Company, Ltd., Beijing, China
1 shared publication- Hongwei Chen
Tsinghua University
1 shared publication - Jiayu Wang1 shared publication
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 15407-15415
LSGSNet: Local-Sensitive and Globally-Stable Network for 3D Human Motion Prediction
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 15407-15415
Open AccessLSGSNet: Local-Sensitive and Globally-Stable Network for 3D Human Motion Prediction
- 19 December 2025
- Published by Elsevier in Information Fusion
- Vol. 129
SD-Fuse: An image structure-driven model for multi-focus image fusion
- 19 December 2025
- Published by Elsevier in Information Fusion
- Vol. 129
SD-Fuse: An image structure-driven model for multi-focus image fusion
- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Signal Processing
- Vol. 20 (3) , 407-418
Efficient Video Representation via Hybrid Key Frame Reconstruction: A Test-Time Data Augmentation Approach
Aiming at the quality and credibility of video recon struction in dynamic coded time-domain compressed imaging, we propose a Key Frames Assisted Hybrid Compressive Video Reconstruction with Test-time Data Augmentation (KH-TTA). Our method integrates a dual-mode sampling architecture com bining long-exposure compressed frames with short-exposure key frames, coupled with a depth-expanded reconstruction network enhanced by spatial-temporal transformation strategies (including axis flipping and rotational operations across spatial-temporal dimensions). Through Monte Carlo approximation of posterior distribution sampling, the framework achieves statistical fusion of reconstruction outcomes with simultaneous uncertainty quantifi cation. Comprehensive experiments demonstrate that KH-TTA establishes new state-of-the-art performance in both simulated and hardware-validated systems, showing significant improve ments over existing methods with PSNR, SSIM, and LPIPS metrics improved by up to 1.01 dB (PSNR), 0.0074 (SSIM), and 12.9% (LPIPS). The proposed uncertainty estimation mechanism exhibits exceptional reliability in complex texture regions (e.g., edge transitions and textual elements), achieving an error-standard deviation correlation coefficient of 0.85. Our method does not need to modify the model structure and optimizes the com putational efficiency through parallelization, which provides a new idea for trustworthy video reconstruction in security-critical scenarios.
- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Signal Processing
- Vol. 20 (3) , 407-418
Efficient Video Representation via Hybrid Key Frame Reconstruction: A Test-Time Data Augmentation Approach
Aiming at the quality and credibility of video recon struction in dynamic coded time-domain compressed imaging, we propose a Key Frames Assisted Hybrid Compressive Video Reconstruction with Test-time Data Augmentation (KH-TTA). Our method integrates a dual-mode sampling architecture com bining long-exposure compressed frames with short-exposure key frames, coupled with a depth-expanded reconstruction network enhanced by spatial-temporal transformation strategies (including axis flipping and rotational operations across spatial-temporal dimensions). Through Monte Carlo approximation of posterior distribution sampling, the framework achieves statistical fusion of reconstruction outcomes with simultaneous uncertainty quantifi cation. Comprehensive experiments demonstrate that KH-TTA establishes new state-of-the-art performance in both simulated and hardware-validated systems, showing significant improve ments over existing methods with PSNR, SSIM, and LPIPS metrics improved by up to 1.01 dB (PSNR), 0.0074 (SSIM), and 12.9% (LPIPS). The proposed uncertainty estimation mechanism exhibits exceptional reliability in complex texture regions (e.g., edge transitions and textual elements), achieving an error-standard deviation correlation coefficient of 0.85. Our method does not need to modify the model structure and optimizes the com putational efficiency through parallelization, which provides a new idea for trustworthy video reconstruction in security-critical scenarios.