Abstract
In this paper, we consider the efficient and robust reconstruction of signals and images via $\ell_{1}-\alpha \ell_{2}~(0<\alpha\leq 1)$ minimization in impulsive noise cases. We introduce two new models: $\ell_1-\alpha\ell_2$ minimization with $\ell_1$ constraint ($\ell_1-\alpha\ell_2$-LAD) and with Dantzig selector constraint ($\ell_1-\alpha\ell_2$-DS). We demonstrate that sparse or nearly sparse signals can be exactly or stably recovered under conditions based on the restricted $1$-isometry property ($\ell_1$-RIP). Furthermore, we propose the $\ell_1-\alpha\ell_2$LA algorithm to solve the unconstrained $\ell_1-\alpha\ell_2$-PLAD model. Numerical experiments show that when the sensing matrix is ill-conditioned (coherence > 0.99), the $\ell_1-\alpha\ell_2$LA method outperforms existing convex and non-convex solvers. The method also shows superior performance in magnetic resonance imaging (MRI) reconstruction under impulsive noise.
Introduction
Compressed sensing (CS) allows for the reconstruction of sparse signals from fewer measurements than traditionally required. However, standard CS methods often struggle with two real-world issues: impulsive noise (outliers) and high coherence in sensing matrices. Traditional $\ell_2$-norm error minimization is highly sensitive to outliers, and high coherence violates the Restricted Isometry Property (RIP) needed for $\ell_1$ minimization to work. This paper addresses these challenges by combining the non-convex $\ell_1 - \alpha \ell_2$ sparsity metric with robust $\ell_1$-norm data fidelity.
Methodology: The $\ell_1 - \alpha \ell_2$LA Algorithm
The authors propose using Least Absolute Deviation (LAD) and the Dantzig Selector (DS) to handle impulsive noise. The core contribution is the $\ell_1 - \alpha \ell_2$LA algorithm, which solves the non-convex $\ell_1 - \alpha \ell_2$ minimization problem. It utilizes the Difference of Convex functions Algorithm (DCA) to handle non-convexity and the Linearized Alternating Direction Method of Multipliers (ADMM) for efficient computation.
Key Results and Insights
The proposed method was tested against standard solvers like YALL1 using synthetic signals and MRI phantoms under heavy-tailed noise models (e.g., Cauchy noise).
Robustness in Ill-Conditioned Settings
A significant advantage of the $\ell_1 - \alpha \ell_2$LA method is its performance when sensing matrices are highly coherent. While traditional solvers fail completely as coherence approaches 1, the proposed method maintains high recovery success rates.
MRI Reconstruction
In MRI simulations using the Shepp-Logan phantom corrupted by Cauchy noise, the $\ell_1 - \alpha \ell_2$LA method achieved a Root-Mean-Square (RMS) error of 0.0762, significantly lower than YALL1 (0.1346) and the non-convex $L_q$ method (0.0985). This demonstrates the method's practical utility in medical imaging where impulsive noise is common.
Conclusion
The $\ell_1 - \alpha \ell_2$LA framework offers a robust alternative for signal recovery in challenging environments. By leveraging non-convex sparsity promotion and $\ell_1$ data fidelity, it overcomes the limitations of standard compressed sensing in the presence of outliers and ill-conditioned matrices. Future research will explore automated parameter selection and clinical validation.
Summary updated