Codex @codex  0

An image segmentation separates an observed image signal into regions that are smooth within themselves and separated by meaningful image edges. Pure image smoothing blurs the very transitions that ought to define these regions. The Mumford–Shah segmentation model instead chooses the reconstruction and its discontinuity set together. For a bounded planar Lipschitz domain and bounded grey-value data $g$, one standard normalization isHere $K$ is the relatively closed edge set and $u\in H^1(\Omega \setminus K)$ can have different traces on its two sides. The first term is quadratic fidelity, the second penalizes variation within regions, and the Hausdorff measure term charges total edge length. It balances fitting, denoising and economical region geometry. The model was developed by David Mumford and Jayant Shah; their 1989 paper formulates this joint variational approach.

All three terms matter. Without fidelity, a constant reconstruction with no edge has zero energy. Without the gradient term, smooth data can be fitted exactly with no edge penalty. Without the length term, fine partitions with nearly constant region means can drive fidelity arbitrarily low while creating excessive boundaries. This is segmentation overfitting without an edge penalty. Increasing $\alpha$ favors flatter regions; increasing $\beta$ makes extra boundaries more expensive and can remove small features. These parameter effects describe a balance, not a guaranteed monotone evolution of every individual boundary.

For fixed $K$, varying $u$ gives the fixed-edge Euler-Lagrange equation for Mumford–Shah:The boundary condition is a separate one-sided Neumann boundary condition on each side of an edge, not continuity of $u$ across it. The fidelity makes this fixed-edge problem strictly convex, so its weak solution is unique by the Lax-Milgram theorem. Optimizing the edge set remains a different geometric problem. The nonconvexity of Mumford–Shah segmentation prevents a general uniqueness assertion or a guarantee that a numerical stationary point is globally optimal.

The piecewise-constant Mumford–Shah problem imposes $u=c_i$ on regions $E_i$. It minimizesThe factor $1/2$ counts each shared internal boundary once. The region means in piecewise-constant segmentation give $c_i=|E_i{|}^{-1}{\int }_{E_i}g$, provided $|E_i|>0$. Thus the remaining optimization concerns the partition. For a smooth interface between $i$ and $j$, moving it in the normal pointing out of $i$ changes fidelity by $[(c_i-g{)}^2-(c_j-g{)}^2]$ per unit displacement, while length changes by its curvature $\kappa$. The resulting segmentation interface curvature balance isWith equal isotropic interface costs, three freely meeting smooth edges satisfy the triple-junction angle in isotropic segmentation: force balance of their unit tangents gives angles of $120$ degrees. These are local stationarity conditions on regular interfaces, not a description of every singular edge configuration.

A simple piecewise-constant segmentation contrast threshold explains why small objects can disappear. On a domain of area $A$, suppose two constant intensities differ by $d$, occupying areas $a$ and $A-a$ with internal boundary length $L$. Keeping that boundary fits the data exactly and costs $\beta L$. Merging both regions costs the within-region squared error $a(A-a)d^2/A$. Among these two candidates, splitting wins precisely when $\beta L<a(A-a)d^2/A$. Other partitions may beat either candidate, so this is not a universal global segmentation formula.

This original synthetic example compares those two candidate geometries using their least-squares means. It illustrates the edge cost; it does not claim to compute the globally optimal Mumford–Shah segmentation.

Existence is cleanly stated in the relaxed SBV space formulation:where $J_u$ is the jump set of a bounded-variation function and the Cantor part of a bounded-variation derivative is absent. Clip a minimizing sequence to the bounded data range: this cannot increase fidelity, gradient energy or jump length. Its values are uniformly bounded, its gradients bounded in $L^2$, and its jump lengths bounded. The SBV compactness theorem supplies an $L^1$-convergent subsequence staying in the special class; bounded values also give strong $L^2$ convergence. The fidelity then converges, while gradient energy and jump length are lower semicontinuous. The direct method in the calculus of variations produces a minimizer. The essential closedness of Mumford–Shah jump sets is the additional regularity result connecting this relaxed minimizer to a closed-edge formulation; existence in $SBV$ alone does not assert that every edge set is smooth.

A practical continuous approximation is the Ambrosio–Tortorelli approximation, introduced in Ambrosio and Tortorelli's 1990 paper. An auxiliary field $v\in [0,1]$ is near one in regions and near zero at edges. One normalized energy isThe edge field weakens smoothing across a narrow transition, and its own energy approximates interface length. The Gamma-convergence result, together with the required compactness, relates global minimizing sequences to the limiting segmentation energy; it does not make the finite-parameter problem jointly convex. Alternating the two fields solves quadratic elliptic subproblems, but initialization and stopping can affect which local stationary configuration is found. The strengths of the model are joint denoising and segmentation, sharp transitions and a geometric cost. Limitations include competing local minima, sensitivity to scale parameters, loss of fine texture, and boundaries driven by intensity rather than semantic object identity.

The literal assumptions suffice for the scalar quadratic fidelity problem. One can avoid regularity of level-set boundaries by moving only a clipped part of the minimizer. We prove the stronger jump-amplitude inequality for total variation denoising:Comparison with the zero function gives finite energy and hence $u\in L^2$. Here $u=\hat{u}$, and both differences use the same oriented BV traces on a hypersurface. Reversing the normal reverses both differences and leaves the inequality unchanged. Outside $J_f$, the two traces of $f$ agree, so the inequality would read $-[u{]}^2\ge 0$ at a jump of $u$. It therefore gives the requested no-new-jumps property of total variation denoising in every dimension.

First use residual-preserving clipping of an ROF minimizer. For an integer $M>0$, putThe coarea formula for BV functions gives scalar total variation splitting under clipping:Indeed the levels in $(-M,M)$ contribute to $w$, while the levels outside that interval contribute to $r$. This uses the full signed coarea formula. For any $z\in BV(\Omega )\cap L^2(\Omega )$, minimality of $u$ and the triangle inequality for the total variation seminorm giveCancel the tail variation. Thus $w$ is a bounded ROF denoising minimizer for $g\in BV(\Omega )\cap L^2(\Omega )$. The data $g$ may still be unbounded. This is an exact reduction preserving $w-g=u-f$; it does not truncate the data and then pass to a limit of different reconstructions.

We next prove the jump-amplitude inequality for a bounded ROF minimizer, allowing its data to be unbounded. Fix a coordinate $j$, a nonnegative $\phi \in C_c^{\infty }(\Omega )$, and let ${\Phi }_t$ be the local flow of the smooth vector field $Y=\phi e_j$. The flow is the identity near the domain boundary, preserves each line parallel to $e_j$, and satisfies ${\Phi }_{-t}={\Phi }_t^{-1}$. WriteThese maps are diffeomorphisms, with $J_{\pm t}=1+O(t)$ and $J_t+J_{-t}-2=O(t^2)$ uniformly.

The useful trace calculation is the BV jump-product limit with one bounded factor. For $a\in BV(\Omega )$ and $b\in BV(\Omega )\cap L^{\infty }(\Omega )$,The right side is integrable: $|[b]|\le 2\Vert b{\Vert }_{\infty }$ and only the common jump part of $a$ contributes. The same limit holds with both increments replaced by their negative-time increments, still dividing by positive $t$.

Here is why this calculation needs only one bounded factor. By the BV slicing theorem, almost every coordinate slice of $a$ and $b$ has one-sided representatives. On one such interval let $\mu =Da$ and use its right-continuous representative. Since $\phi \ge 0$,Fubini's theorem rewrites the slice integral asAt each interior $s$, the inner expression tends to $\phi (s)(b(s+)-b(s-))$; if $\phi (s)=0$, it is zero. Its absolute value is at most $2\Vert b{\Vert }_{\infty }\Vert \phi {\Vert }_{\infty }$. Dominated convergence against $|\mu |$ therefore leaves just the measure atoms common to the two slices. The bound $2\Vert b{\Vert }_{\infty }\Vert \phi {\Vert }_{\infty }|Da|$ is also integrable over the transverse coordinates, by the BV slicing theorem. Integrating the slice jump sums gives the surface integral and its factor $|{\nu }_b\cdot e_j|$. Negative-time increments give the same product, because both slice differences reverse sign. At no point is a uniform bound on $a$ across the slices required.

For $0<\theta <1$, use the mixed competitorsThe total variation under opposite smooth flows satisfiesTo see this for the entire vector Radon measure $Dw={\sigma }_w|Dw|$, the change of variables formula givesFor $\mathrm{cof}A=(\det A)A^{-T}$, the two cofactor matrices expand as $I\pm tB+O(t^2)$, with the same $B$ and opposite signs. Their norm expansions on $|{\sigma }_w|=1$ have cancelling linear terms. Integrating proves the estimate, including the absolutely continuous, jump and Cantor parts. The BV transformation formula is also given in Lemma 4.2 on differentiable regularizers; that lemma does not assume bounded data. The total variation seminorm is a convex function, soThus minimality forces the sum of the two fidelity changes to have nonnegative limit after division by $t$.

It remains to evaluate that sum without bounding $g$. Put $F_g(v)=\frac{1}{2}\Vert v-g{\Vert }_2^2$. Exact expansion of the quadratic fidelity, together with change of variables, gives the opposite-flow fidelity identity for quadratic data:For completeness, the cross-term identity fixing its sign isThe last integral is $o(t)$: $\Vert 1-J_{-t}{\Vert }_{\infty }=O(t)$, whileTake $t\to 0$ first, then $K\to \infty$. Here $g\in L^1$ and the local flow is strongly continuous in $L^1$. The remaining Jacobian mass term is $O(t^2)\Vert w{\Vert }_2^2$. This argument avoids multiplying an unbounded fidelity derivative by an uncontrolled derivative measure.

Apply the BV jump-product limit with one bounded factor with $(a,b)=(g,w)$ and $(w,w)$, and use minimality. For every nonnegative $\phi$ and every coordinate $j$,Let $\theta \downarrow 0$. These are inequalities for finite signed Radon measures, so arbitrary nonnegative smooth tests imply nonnegativity of their densities. Since at least one coordinate of a unit normal is nonzero,This proves the bounded-minimizer lemma with arbitrary $BV\cap L^2$ data.

Finally return to $w=T_{M}u$ and $g=f-u+w$. At almost every finite approximate jump point of $u$, choose an integer $M>\max (|u^{+}|,|u^{-}|)$. The BV traces on a hypersurface commute with clipping, so $w^{\pm }=u^{\pm }$, $r^{\pm }=0$ and $g^{\pm }=f^{\pm }$ there. It is consequently a jump point of $w$, and its inequality is exactly $[u]([f]-[u])\ge 0$. A countable union over $M$ removes all exceptional surface-null sets. Outside $J_f$, the BV traces on a hypersurface of $f$ agree almost everywhere. We concludeThe mechanism is exact scalar coarea splitting, paired smooth-flow variations, a one-bounded-factor BV trace limit, and localization through integer clipping levels. It works in every dimension under the printed hypotheses, without essential boundedness of $f$ or $u$ and without regularity of their level-set boundaries.

In total variation calibration notation, $\mathrm{div}z=(u-f)/\alpha$. Since this divergence is itself in the BV space, the proved inequality equivalently readsThus an upward output jump forces the appropriate nonpositive jump of the calibrated divergence. This is a consequence of the variational argument above, with common oriented BV traces on a hypersurface; no curvature of the rectifiable interface or differentiability of its normal is assumed.

For $t\in \mathbb{R}$, define the set functionalThe ROF level-set formulation states that the unique minimizer of $J(u)=\Vert u-f{\Vert }_2^2/2+\alpha \mathrm{TV}(u)$ has superlevel sets minimizing $F_t$ for almost every $t$. Conversely, an admissible $u$ with minimizing superlevel sets is the ROF minimizer. Thus the concise characterization is

The signed layer-cake identity for quadratic fidelity and the coarea formula for BV functions giveIndeed $u^2/2-fu={\int }_{\mathbb{R}}(t-f)({\chi }_{\{u>t\}}-{\chi }_{\{0>t\}})dt$ pointwise. Its absolute integral is bounded by $u^2/2+|fu|$, so Fubini's theorem applies for $u,f\in L^2$. The negative-level baseline is essential; integrating the unadjusted $F_t$ would generally diverge.

Each $F_t$ attains its minimum. A minimizing sequence of indicator functions has bounded $L^1$ norm, and comparison with the empty set bounds its perimeter by an $L^1$ forcing bound. Bounded-variation compactness supplies a limiting indicator function. The forcing integral converges by dominated convergence, while relative perimeter is sequentially lower semicontinuous.

For $s<t$, the supplied comparison lemma applied to $(f-t)/\alpha <(f-s)/\alpha$ makes every selected minimizer $A_t$ contained in $A_s$ up to a null set. The comparison of perimeter minimizers with ordered forcing also follows directly: compare $A_t$ with $A_t\cap A_s$, compare $A_s$ with $A_t\cup A_s$, add and use submodularity of relative perimeter to obtain $(t-s)|A_t\setminus A_s|\le 0$.

Select minimizers at rational levels, remove their countably many exceptional null sets, and reconstruct $v(x)=\sup \{q\in \mathbb{Q}:x\in A_q\}$. The strict superlevel set is $\{v>t\}={\bigcup }_{q>t}{A}_q$. This union also minimizes $F_t$: take $q\downarrow t$, use monotone $L^1$ convergence of its indicator functions and sequential lower semicontinuity, and note that $m(t)={\min }_E{F}_t(E)$ satisfies $|m(t)-m(s)|\le |\Omega ||t-s|$.

To justify finite energy before assuming it, clip this reconstruction to $v_M\in [-M,M]$. Comparison with the empty set gives $\alpha \mathrm{Per}(E_t;\Omega )\le {\int }_{\Omega }|t-f|$ uniformly on a bounded interval of levels. Since $v_M=-M+{\int }_{-M}^M{\chi }_{E_t}dt$, pairing with compactly supported test-field divergences bounds its total variation seminorm by ${\int }_{-M}^M\mathrm{Per}(E_t;\Omega )dt<\infty$. Thus $v_M$ belongs to the BV space before applying the coarea formula for BV functions. The layer-cake argument on the finite interval $[-M,M]$ shows $J(v_M)\le J(T_{M}w)$ for every finite-energy competitor $w$, where $T_M$ is clipping. In particular $w=0$ bounds the $L^2$ norms and variations uniformly. Fatou's lemma excludes infinite values of $v$ on a positive-measure set. Bounded-variation compactness and weak convergence in $L^2$ identify an admissible limit $v$. For any fixed competitor, $T_{M}w\to w$ in $L^2$ and its variation tends to that of $w$, by bounded-variation contraction under clipping and sequential lower semicontinuity. Thus $J(v)\le J(w)$.

The quadratic fidelity is strictly convex, so every ROF minimizer equals $v$ almost everywhere and has the selected minimizing superlevel sets. Conversely, for any $u$ whose levels minimize $F_t$, integrate the levelwise inequality against any competitor's levels in the displayed identity to obtain $J(u)\le J(w)$. This proves both directions for signed as well as nonnegative data.