$ell_0$ Minimization for Wavelet Frame Based Image Restoration

The theory of (tight) wavelet frames has been extensively studied in the past twenty years and they are currently widely used for image restoration and other image processing and analysis problems. The success of wavelet frame based models, including balanced approach and analysis based approach, is due to their capability of sparsely approximating piecewise smooth functions like images. Motivated by the balanced approach and analysis based approach, we shall propose a wavelet frame based $\ell_0$ minimization model, where the $\ell_0$ “norm” of the frame coefficients is penalized. We adapt the penalty decomposition (PD) method to solve the proposed optimization problem. Numerical results showed that the proposed model solved by the PD method can generate images with better quality than those obtained by either analysis based approach or balanced approach in terms of restoring sharp features as well as maintaining smoothness of the recovered images. Some convergence analysis of the PD method will also be provided.

💡 Research Summary

The paper introduces a novel image restoration framework that directly penalizes the ℓ₀ “norm” of wavelet frame coefficients, aiming to achieve a better balance between edge sharpness and smoothness than existing ℓ₁‑based methods. After reviewing the limitations of classical least‑squares restoration and the success of variational and wavelet‑frame approaches, the authors motivate the use of ℓ₀ regularization: while ℓ₁ serves as a convex surrogate for sparsity, it may not faithfully capture the true cardinality of non‑zero coefficients, especially when the forward operator A does not satisfy the restricted isometry conditions required by compressed‑sensing theory.

The proposed model (equation 2.1) minimizes a data‑fidelity term ½‖Au − f‖²_D plus a weighted sum of ℓ₀ counts of the frame coefficients α = W u, where W is a tight wavelet frame transform (WᵀW = I). Because the ℓ₀ term is integer‑valued, discontinuous, and non‑convex, the authors adopt the Penalty Decomposition (PD) methodology of Lu and Zhang (2010). They reformulate the problem as a constrained optimization over (u, α) with the linear constraint α = W u, introduce a quadratic penalty with parameter β, and solve the resulting subproblems iteratively.

Each PD iteration consists of two blocks: (i) a quadratic subproblem in u, which reduces to solving a linear system (efficiently handled by conjugate‑gradient or direct methods depending on A), and (ii) an ℓ₀ subproblem in α, solved by a hard‑thresholding operation that retains the largest coefficients according to the current penalty weight. The Block Coordinate Descent (BCD) scheme alternates between these blocks, while a non‑monotone gradient projection step ensures stability when updating α.

A significant theoretical contribution is the convergence analysis for the BCD method applied to a non‑continuous objective. The authors prove that, despite the ℓ₀ term’s discontinuity, the BCD sequence possesses limit points that are stationary for the penalized problem, and that as β → ∞ the PD iterates converge to a feasible point of the original constrained formulation. This fills a gap in the literature where prior PD analyses assumed smooth regularizers.

Experimental validation covers three classic restoration tasks: denoising of blurred noisy images, deconvolution, and inpainting. The ℓ₀‑based model is compared against the ℓ₁‑based balanced approach (which penalizes both sparsity and smoothness) and the ℓ₁‑based analysis approach (which penalizes only the analysis coefficients). Quantitative metrics (PSNR, SSIM) consistently favor the ℓ₀ method, and visual inspection shows sharper edges with fewer ringing artifacts than the balanced model and less blurring than the analysis model. Moreover, the algorithm’s runtime is comparable to, and sometimes faster than, the competing methods because the hard‑thresholding step is cheap and the penalty parameter can be increased adaptively without sacrificing convergence speed.

In summary, the paper makes three core contributions: (1) a wavelet‑frame‑based image restoration model that directly employs ℓ₀ sparsity, (2) an adapted PD algorithm with a rigorously analyzed BCD subroutine for handling the non‑convex, discontinuous regularizer, and (3) comprehensive experiments demonstrating superior visual and quantitative performance over state‑of‑the‑art ℓ₁‑based frame methods. The work opens avenues for further research on ℓ₀ regularization in more complex inverse problems, integration with learned frame dictionaries, and extensions to non‑linear measurement models.