Abstract
We consider the total variation (TV) minimization problem used for compressive sensing and solve it using the generalized alternating projection (GAP) algorithm. Extensive results demonstrate the high performance of proposed algorithm on compressive sensing, including two dimensional images, hyperspectral images and videos. We further derive the Alternating Direction Method of Multipliers (ADMM) framework with TV minimization for video and hyperspectral image compressive sensing under the CACTI and CASSI framework, respectively. Connections between GAP and ADMM are also provided.
Introduction
Compressive Sensing (CS) enables the acquisition of high-resolution signals from a sub-Nyquist number of measurements by exploiting the inherent sparsity of natural signals. Two primary priors are commonly utilized: transform-domain sparsity (e.g., wavelets or DCT) and Total Variation (TV) minimization, which assumes piecewise smoothness. While the Generalized Alternating Projection (GAP) algorithm has proven effective for transform-domain sparsity, TV minimization has traditionally relied on the Alternating Direction Method of Multipliers (ADMM). However, ADMM's performance is often sensitive to the penalty parameter ($\eta$), necessitating tedious tuning. This work introduces GAP-TV, a novel approach that adapts the GAP framework to TV minimization, offering a more robust and nearly parameter-free alternative.
Methodology
The GAP-TV algorithm reformulates the reconstruction task as a series of alternating projections between two sets:
- Euclidean Projection: The current estimate is projected onto a linear manifold defined by the measurement constraint $\Phi x = y$, ensuring data consistency.
- TV Denoising: A TV-regularized denoising problem is solved, typically using an iterative clipping algorithm.
The authors also propose Acc-GAP-TV, an accelerated variant that adaptively modifies the linear manifold during iterations to enhance convergence. A significant theoretical contribution of this paper is the derivation of the relationship between GAP and ADMM. The authors demonstrate that GAP can be interpreted as a specific instance of ADMM where the manifold update is simplified, effectively removing the sensitivity to the penalty parameter found in standard ADMM implementations. The framework is specifically derived for practical CS architectures, including Coded Aperture Compressive Temporal Imaging (CACTI) for video and Coded Aperture Snapshot Spectral Imaging (CASSI) for hyperspectral data.
Experimental Results
The performance of GAP-TV was evaluated across 2D images, video sequences, and hyperspectral datasets.
2D Image Reconstruction
Acc-GAP-TV consistently outperforms existing solvers such as TVAL3, TwIST, and standard wavelet-based GAP across a wide range of Compressive Sensing ratios (CSr). For instance, on the foreman benchmark at a 10% sampling rate (CSr 0.1), Acc-GAP-TV achieves a PSNR of 31.68 dB, significantly surpassing ADMM-TV (28.20 dB) and TVAL3 (27.63 dB).
Video and Hyperspectral Imaging
In video compressive sensing, GAP-TV demonstrates superior stability. Analysis of the "traffic" dataset shows that the algorithm maintains higher and more consistent PSNR values across frames compared to methods like TwIST.
For hyperspectral imaging, both GAP-TV and the derived ADMM-TV framework significantly outperform wavelet-based methods, highlighting the efficacy of the TV prior in capturing spectral-spatial correlations.
Conclusion
This paper successfully integrates TV minimization into the GAP framework, providing a high-performance, robust solver for various compressive sensing modalities. By bridging the gap between GAP and ADMM, the authors provide both a practical tool for hardware-specific reconstruction and a theoretical foundation for parameter-free manifold updates. Future work may address the computational complexity of the internal TV denoising step and explore automated weight selection for group-based sparsity.
Summary updated