The Mumford-Shah functional is a cornerstone of mathematical image processing, providing a framework for segmenting images into smooth regions separated by sharp edges. While its properties are well-studied for direct observations of bounded data, many real-world inverse problems involve singular operators (like blur or Radon transforms) and potentially unbounded data (such as point sources in astronomy). This paper establishes existence results for minimizers in these more challenging scenarios.
The Challenge of Inverse Problems
The classical Mumford-Shah functional seeks to minimize an energy involving a piecewise smooth function $u$ and its jump set $K$. Standard existence proofs typically rely on the $L^\infty$ boundedness of the input data. However, in applications like inverse source problems or image deblurring, the data may be unbounded or observed through a singular operator $T$ that loses information.
Without the $L^\infty$ constraint, minimizing sequences can become unstable, and jump sets might exhibit fractal-like behavior that prevents the existence of a well-defined minimizer. The authors investigate the conditions under which a stable solution can be guaranteed for these ill-posed problems.
Geometric Regularization and Rondi’s Class
To overcome the inherent ill-posedness, the authors move beyond the standard "Special Functions of Bounded Variation" (SBV) framework. They introduce geometric constraints by restricting the jump set $K$ to Rondi’s class. This class consists of sets formed by a finite number of Lipschitz arcs with specific stability properties.
By restricting $K$, the authors ensure that the domain $\Omega \setminus K$ satisfies the cone property. This technical requirement prevents the formation of infinitely sharp cusps, allowing for the application of Sobolev space compactness arguments and interpolation inequalities. This approach effectively uses the topology of the edges as a regularizer, bypassing the need for $L^\infty$ data bounds.
Key Existence Results
The paper provides existence theorems for two primary classes of operators:
- Nonlocal Compact Injective Operators: These represent global transformations like convolutions (blurring) that do not entirely destroy the signal.
- Local Differential Operators: These are relevant for physical models involving linear conservation laws.
The authors demonstrate that for these operators, a minimizer $(u, K)$ exists within the restricted geometric class. This result is significant because it confirms that the Mumford-Shah framework remains valid for physics-based inverse problems where data intensity might be extreme.
Discussion and Limitations
While providing a rigorous foundation for singular Mumford-Shah problems, the study highlights several limitations:
- Dimensionality: The results are currently limited to two-dimensional domains ($d=2$). Extending these proofs to 3D surfaces is mathematically complex due to the nature of the geometric constraints required.
- Geometric Stiffness: Rondi's class, while mathematically convenient for proving existence, may be too restrictive for images with high-frequency textures or very complex edge structures.
- Computational Hurdles: The paper focuses on theoretical existence. In practice, finding these minimizers remains an NP-hard problem, and the convergence of discrete numerical approximations to these continuous minimizers is still an open question.
Despite these challenges, the work reinforces the superiority of Mumford-Shah regularization over simpler methods like Total Variation (TV) for preserving edge sharpness in deblurring tasks.
Summary updated