Abstract
Although difference of convex model has attracted many research efforts due to its superior performance for image processing, no attention has focused on robust data fidelity for this model. In this paper, we propose a novel model, which combines the $l_1$ and $l_2$ fidelity terms with a weighted difference of anisotropic and isotropic total variation (TV). Since our model takes a new difference form of convex terms, we employ difference of convex algorithm (DCA). In this paper, we adopt split Bregman iteration (SBI) to solve each DCA subproblem of the proposed model. Image denoising verifies the convergence of optimal solution and monotone decreasing of objective function. Experimental results on image denoising demonstrate that the proposed methods outperform other competing methods in terms of quantitative criteria and perceptual quality.
Introduction
Image restoration is a fundamental inverse problem in computer vision, typically modeled as $f = Ku + n$, where $f$ is the observed noisy image, $u$ is the latent clean image, $K$ is a blurring operator, and $n$ represents noise. Traditional methods often struggle with two primary issues: noise diversity and the "blocky" artifacts associated with standard Total Variation (TV) regularization.
Standard $l_2$ fidelity terms are optimized for Gaussian noise but perform poorly against impulsive "salt-and-pepper" noise. Conversely, $l_1$ fidelity is robust against impulsive noise but can sacrifice fine details. Furthermore, while TV regularization is excellent for edge preservation, it often produces cartoon-like artifacts in smooth regions. This paper addresses these challenges by proposing a hybrid model that integrates multiple fidelity and regularization strategies.
Methodology
The proposed approach introduces a novel objective function designed for robustness and detail preservation.
Hybrid Data Fidelity
The model blends $l_1$ and $l_2$ norms within the data fidelity term. This hybridization allows the algorithm to simultaneously handle Gaussian and impulsive noise, making it more versatile for real-world applications where noise characteristics may be mixed or unknown.
Weighted Difference of Total Variation
To mitigate the blocky artifacts of standard TV, the authors utilize a weighted difference between anisotropic TV (sum of absolute gradients) and isotropic TV (gradient magnitude). Specifically, the regularization term takes the form of $L_1 - \alpha L_2$ on the image gradient. This "difference of convex" (DC) approach provides a better approximation of the heavy-tailed gradient distributions found in natural images compared to traditional convex models.
The DCASBI Solver
The resulting objective function is non-convex due to the difference of convex terms. To solve this, the authors employ the Difference of Convex Algorithm (DCA), which decomposes the problem into a sequence of convex sub-problems. Each sub-problem is then efficiently solved using Split Bregman Iteration (SBI), a technique renowned for its speed and stability in $l_1$-minimization tasks.
Experimental Evaluation
The model was rigorously tested on 14 standard grayscale images, including natural scenes like "Lena" and "Barbara" as well as synthetic textured images.
Quantitative Results
The proposed DCASBI method was compared against standard $l_1$ TV, $l_1 - l_2$ TV, and weighted $l_1 - 0.5l_2$ TV variants. In nearly all test scenarios, DCASBI achieved superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) values. For instance, on the "Barbara" image, DCASBI reached a PSNR of 29.8230, significantly higher than competing methods.
Qualitative Assessment
Visually, the restored images exhibited fewer artifacts and sharper boundaries. The hybrid model successfully preserved fine textures—such as the patterns in Barbara's shawl—while avoiding the over-smoothing typical of pure TV-based methods.
Conclusion
This research demonstrates that combining hybrid fidelity terms with a weighted difference of TV regularization significantly improves image denoising performance. The synergy between DCA and Split Bregman Iteration provides a robust framework for solving complex, non-convex restoration problems. Future work will focus on automated hyperparameter tuning and extending the model to handle multi-channel color images.
Summary updated