Revisit Dictionary Learning for Video Compressive Sensing under the Plug-and-Play Framework
This work addresses the problem of efficient and effective reconstruction for video compressive sensing, offering an incremental improvement in denoising performance for researchers and practitioners in computational imaging.
The paper tackles the challenge of denoising in the Plug-and-Play framework for video snapshot compressive imaging reconstruction by proposing a shallow-learning-based algorithm using kernel singular value decomposition, achieving around 2 dB improvement in PSNR and 0.2 in SSIM compared to a total variation baseline.
Aiming at high-dimensional (HD) data acquisition and analysis, snapshot compressive imaging (SCI) obtains the 2D compressed measurement of HD data with optical imaging systems and reconstructs HD data using compressive sensing algorithms. While the Plug-and-Play (PnP) framework offers an emerging solution to SCI reconstruction, its intrinsic denoising process is still a challenging problem. Unfortunately, existing denoisers in the PnP framework either suffer limited performance or require extensive training data. In this paper, we propose an efficient and effective shallow-learning-based algorithm for video SCI reconstruction. Revisiting dictionary learning methods, we empower the PnP framework with a new denoiser, the kernel singular value decomposition (KSVD). Benefited from the advent of KSVD, our algorithm retains a good trade-off among quality, speed, and training difficulty. On a variety of datasets, both quantitative and qualitative evaluations of our simulation results demonstrate the effectiveness of our proposed method. In comparison to a typical baseline using total variation, our method achieves around $2$ dB improvement in PSNR and 0.2 in SSIM. We expect that our proposed PnP-KSVD algorithm can serve as a new baseline for video SCI reconstruction.