Abstract
Local M-smoothers are interesting and important signal and image processing techniques with many connections to other methods. In our paper, we derive a family of partial differential equations (PDEs) that result in one, two, and three dimensions as limiting processes from M-smoothers which are based on local order-p means within a ball the radius of which tends to zero. The order p may take any nonzero value \(>-1\), allowing also negative values. In contrast to results from the literature, we show in the space-continuous case that mode filtering does not arise for \(p \rightarrow 0\), but for \(p \rightarrow -1\). Extending our filter class to p-values smaller than \(-1\) allows to include, e.g. the classical image sharpening flow of Gabor. The PDEs we derive in 1D, 2D, and 3D show large structural similarities. Since our PDE class is highly anisotropic and may contain backward parabolic operators, designing adequate numerical methods is difficult. We present an \(L^\infty \)-stable explicit finite difference scheme that satisfies a discrete maximum–minimum principle, offers excellent rotation invariance, and employs a splitting into four fractional steps to allow larger time step sizes. Although it approximates parabolic PDEs, it consequently benefits from stabilisation concepts from the numerics of hyperbolic PDEs. Our 2D experiments show that the PDEs for \(p<1\) are of specific interest: Their backward parabolic term creates favourable sharpening properties, while they appear to maintain the strong shape simplification properties of mean curvature motion.
Similar content being viewed by others
References
Alvarez, L., Guichard, F., Lions, P.L., Morel, J.M.: Axioms and fundamental equations in image processing. Arch. Ration. Mech. Anal. 123, 199–257 (1993)
Alvarez, L., Lions, P.L., Morel, J.M.: Image selective smoothing and edge detection by nonlinear diffusion, II. SIAM J. Numer. Anal. 29, 845–866 (1992)
Arehart, A.B., Vincent, L., Kimia, B.B.: Mathematical morphology: the Hamilton–Jacobi connection. In: Proceedings of Fourth International Conference on Computer Vision, pp. 215–219. IEEE Computer Society Press, Berlin (1993)
Armatte, M.: Fréchet et la médiane: un moment dans une histoire de la robustesse. J. Soc. Fr. Stat. 147(2), 23–37 (2006)
Aubert, G., Kornprobst, P.: Mathematical problems in image processing: partial differential equations and the calculus of variations, Applied Mathematical Sciences, vol. 147, second edn. Springer, New York (2006)
Barral Souto, J.: El modo y otras medias, casos particulares de una misma expresión matemática. Report No. 3, Cuadernos de Trabajo, Instituto de Biometria, Universidad Nacional de Buenos Aires, Argentina (1938)
Bertero, M., Poggio, T.A., Torre, V.: Ill-posed problems in early vision. Proc. IEEE 76(8), 869–889 (1988)
Brakke, K.A.: The Motion of a Surface by Its Mean Curvature. Princeton University Press, Princeton (1978)
Breuß, M., Weickert, J.: A shock-capturing algorithm for the differential equations of dilation and erosion. J. Math. Imaging Vis. 25(2), 187–201 (2006)
Brockett, R.W., Maragos, P.: Evolution equations for continuous-scale morphological filtering. IEEE Trans. Signal Process. 42, 3377–3386 (1994)
Carmona, R., Zhong, S.: Adaptive smoothing respecting feature directions. IEEE Trans. Image Process. 7(3), 353–358 (1998)
Caselles, V., Morel, J.M., Sbert, C.: An axiomatic approach to image interpolation. IEEE Trans. Image Process. 7(3), 376–386 (1998)
Chan, T.F., Shen, J.: Image Processing and Analysis: Variational, PDE, Wavelet, and Stochastic Methods. SIAM, Philadelphia (2005)
Chu, C.K., Glad, I., Godtliebsen, F., Marron, J.S.: Edge-preserving smoothers for image processing. J. Am. Stat. Assoc. 93(442), 526–556 (1998)
Cohen, I., Falik, A., Gilboa, G.: Stable explicit p-Laplacian flows based on nonlinear eigenvalue anaysis. In: Lellmann, J., Burger, M., Modersitzki, J. (eds.) Scale Space and Variational Methods in Computer Vision. Lecture Notes in Computer Science, vol. 11603, pp. 315–327. Springer, Cham (2019)
Ellis, R.L.: On the foundations of the theories of probablities. Trans. Cambridge Philos. Soc. 8(1), 204–219 (1844)
Fechner, G.T.: Ueber den Ausgangswerth der kleinsten Abweichungssumme, dessen Bestimmung, Verwendung und Verallgemeinerung. Abhandlungen der mathematisch-physischen Classe der Königlich Sächsischen Gesellschaft der Wissenschaften 11, 1–76 (1878)
Fréchet, M.: Sur l’extension de certaines evaluations statistiques au cas de petits echantillons. Revue de l’Institut International de Statistique 11(3), 182–205 (1943)
Fréchet, M.: Nouvelles définitions de la valeur moyenne et des valeurs equiprobables d’un nombre aléatoire. Sciences, pp. 5–26. Annales de l’Université de Lyon, Lyon (1946)
Fréchet, M.: Les élements aléatoires de nature quelconque dans un espace distancié. Annales de l’Institut Henri Poincaré 10, 215–310 (1948)
Fréchet, M.: Les valeurs typiques d’ordre nul ou infini d’un nombre aléatoire. Revue de l’Institut International de Statistique 16(1), 1–22 (1948)
Fréchet, M.: Positions typiques d’un élément aléatoire de nature quelconque. Annales scientifiques de l’É.N.S., 3e série 65, 211–237 (1948)
Gabor, D.: Information theory in electron microscopy. Lab. Investig. 14, 801–807 (1965)
Galić, I., Weickert, J., Welk, M., Bruhn, A., Belyaev, A., Seidel, H.P.: Towards PDE-based image compression. In: Paragios, N., Faugeras, O., Chan, T., Schnörr, C. (eds.) Variational Geometric and Level-Set Methods in Computer Vision. Lecture Notes in Computer Science, vol. 3752, pp. 37–48. Springer, Berlin (2005)
Gauss, C.F.: Theoria Motus Corporum Coelestium in Sectionibus Conicis Solem Ambientium. Perthes & Besser, Hamburg (1809)
Gilboa, G., Sochen, N.A., Zeevi, Y.Y.: Forward-and-backward diffusion processes for adaptive image enhancement and denoising. IEEE Trans. Image Process. 11(7), 689–703 (2002)
Griffin, L.D.: Mean, median and mode filtering of images. Proc. R. Soc. Lond. Ser. A 456(2004), 2995–3004 (2000)
Griffin, L.D., Lillholm, M.: Mode estimation using pessimistic scale space tracking. In: Griffin, L.D., Lillholm, M. (eds.) Scale-Space Methods in Computer Vision. Lecture Notes in Computer Science, vol. 2695, pp. 266–280. Springer, Berlin (2003)
Guichard, F., Morel, J.M.: Partial differential equations and image iterative filtering. In: Duff, I.S., Watson, G.A. (eds.) The State of the Art in Numerical Analysis, no. 63 in IMA Conference Series (New Series), pp. 525–562. Clarendon Press, Oxford (1997)
Huber, P.J.: Robust Statistics. Wiley, New York (1981)
Iijima, T.: Basic equation of figure and observational transformation. Syst. Comput. Controls 2(4), 70–77 (1971)
Jackson, D.: Note on the median of a set of numbers. Bull. Am. Math. Soc. 27, 160–164 (1921)
Jordan, C.: Statistique Mathématique. Gauthier-Villars, Paris (1927)
Kass, M., Solomon, J.: Smoothed local histogram filters. ACM Trans. Graph. 29(4), 100 (2010)
Kramer, H.P., Bruckner, J.B.: Iterations of a non-linear transformation for enhancement of digital images. Pattern Recogn. 7, 53–58 (1975)
Kuijper, A.: Geometrical PDEs based on second-order derivatives of gauge coordinates in image processing. Image Vis. Comput. 29, 1023–1034 (2009)
Laplace, P.S.: Mémoire sur la probabilité des causes par les événements. Mémoires de l’Académie royale des Sciences de Paris (Savants étrangers) VI, 621pp. (1774). In: Oeuvres complètes de Laplace, vol. 8, 25–65, Gauthier-Villars, Paris (1891)
Legendre, A.M.: Nouvelles Méthodes pour la détermination des Orbites des Comètes. Firmin Didot, Paris (1805)
Lindenbaum, M., Fischer, M., Bruckstein, A.: On Gabor’s contribution to image enhancement. Pattern Recogn. 27, 1–8 (1994)
Marchuk, G.I.: Splitting and alternating direction methods. In: Ciarlet, P.G., Lions, J.L. (eds.) Handbook of Numerical Analysis, vol. I, pp. 197–462. North Holland, Amsterdam (1990)
Mitchell, A.R., Griffiths, D.F.: The Finite Difference Method in Partial Differential Equations. Wiley, Chichester (1980)
Mrázek, P., Weickert, J., Bruhn, A.: On robust estimation and smoothing with spatial and tonal kernels. In: Klette, R., Kozera, R., Noakes, L., Weickert, J. (eds.) Geometric Properties from Incomplete Data, Computational Imaging and Vision, vol. 31, pp. 335–352. Springer, Dordrecht (2006)
Osher, S., Paragios, N. (eds.): Geometric Level Set Methods in Imaging Vision and Graphics. Springer, New York (2003)
Osher, S., Rudin, L.: Shocks and other nonlinear filtering applied to image processing. In: Tescher, A.G. (ed.) Applications of Digital Image Processing XIV. Proceedings of SPIE, vol. 1567, pp. 414–431. SPIE Press, Bellingham (1991)
Osher, S., Rudin, L.I.: Feature-oriented image enhancement using shock filters. SIAM J. Numer. Anal. 27, 919–940 (1990)
Pearson, K.: Contributions to the mathematical theory of evolution. II. Skew variation in homogeneous material. Philos. Trans. R. Soc. A 186, 343–414 (1895)
Perona, P., Malik, J.: Scale space and edge detection using anisotropic diffusion. IEEE Trans. Pattern Anal. Mach. Intell. 12, 629–639 (1990)
Rouy, E., Tourin, A.: A viscosity solutions approach to shape-from-shading. SIAM J. Numer. Anal. 29, 867–884 (1992)
Schönlieb, C.B.: Partial Differential Equation Methods for Image Inpainting. Cambridge University Press, New York (2015)
Torroba, P.L., Cap, N.L., Rabal, H.J., Furlan, W.D.: Fractional order mean in image processing. Opt. Eng. 33(2), 528–534 (1994)
Tukey, J.W.: Exploratory Data Analysis. Addison-Wesley, Menlo Park (1971)
van den Boomgaard, R., Smeulders, A.: The morphological structure of images: the differential equations of morphological scale-space. IEEE Trans. Pattern Anal. Mach. Intell. 16, 1101–1113 (1994)
van den Boomgaard, R., van de Weijer, J.: On the equivalence of local-mode finding, robust estimation and mean-shift analysis as used in early vision tasks. In: Proceedings of 16th International Conference on Pattern Recognition, vol. 3, pp. 927–930. Quebec City, Canada (2002)
Vladimirov, V.S.: Generalized Functions in Mathematical Physics. Mir, Moscow (1979)
Weickert, J.: Anisotropic Diffusion in Image Processing. Teubner, Stuttgart (1998)
Weickert, J., Steidl, G., Mrázek, P., Welk, M., Brox, T.: Diffusion filters and wavelets: What can they learn from each other? In: Paragios, N., Chen, Y., Faugeras, O. (eds.) Handbook of Mathematical Models in Computer Vision, pp. 3–16. Springer, New York (2006)
Welk, M.: Multivariate median filters and partial differential equations. J. Math. Imaging Vis. 56, 320–351 (2016)
Welk, M.: Multivariate medians for image and shape analysis. arXiv:1911.00143 (2019)
Welk, M., Breuß, M., Vogel, O.: Morphological amoebas are self-snakes. J. Math. Imaging Vis. 39(2), 87–99 (2011)
Welk, M., Weickert, J.: PDE evolutions for M-smoothers: from common myths to robust numerics. In: Burger, M., Lellmann, J., Modersitzki, J. (eds.) Scale Space and Variational Methods in Computer Vision. Lecture Notes in Computer Science, vol. 11603, pp. 248–263. Springer, Cham (2019)
Welk, M., Weickert, J., Galić, I.: Theoretical foundations for spatially discrete 1-D shock filtering. Image Vis. Comput. 25(4), 455–463 (2007)
Winkler, G., Aurich, V., Hahn, K., Martin, A.: Noise reduction in images: some recent edge-preserving methods. Pattern Recogn. Image Anal. 9(4), 749–766 (1999)
Wladimirow, W.S.: Gleichungen der mathematischen Physik. Deutscher Verlag der Wissenschaften, Berlin (1972). English edition: V. S. Vladimirov, Equations of Mathematical Physics, Marcel Dekker, New York (1971)
Yanenko, N.N.: The Method of Fractional Steps: the Solution of Problems of Mathematical Physics in Several Variables. Springer, New York (1971)
Acknowledgements
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant Agreement No. 741215, ERC Advanced Grant INCOVID). We thank Luis Alvarez (University of Las Palmas de Gran Canaria) for interesting and inspiring discussions on this topic.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Proofs of PDE Approximation Results
1.1 Proof of Proposition 1
1.1.1 Preliminaries: Some Important Integrals
We start by collecting some definite integrals that will be useful in the following. We define for \(\varrho \in (0,1)\) and \(q\in \mathbb {R}\)
With the additional abbreviation
we can derive via substituting \(\xi =\sin \varphi \) and integration by parts (integrating \(\sin ^q\varphi \cos \varphi \) and differentiating \(\cos \varphi \))
for \(q\ne -1\). Moreover, we have by \(1=\sin ^2\varphi +\cos ^2\varphi \)
for \(q\ne -1\) which allows to transform \(S_q\) into \(S_{q+2}\) and vice versa.
From (62) we can obtain thereby
for real p with exception of some odd integers. Note that also for the exceptional values (where some of the denominators become zero) the integrals exist.
1.1.2 Regular Points: Ansatz via Taylor Expansion
Let the image u and mean order p be given as in the proposition. Assume w.l.o.g. that the regular location \(\varvec{x}_0\) is (0, 0) with \(u(0,0)=0\), and that the gradient of u at (0, 0) is in the positive x direction, i.e. \(u_x>0\), \(u_y=0\). Let a neighbourhood radius \(\varrho >0\) be given, and denote the closed (Euclidean) \(\varrho \)-neighbourhood of (0, 0) by \(\mathrm {D}_{\varrho }\).
Using Taylor expansion of u up to third order, we can write for \((x,y)\in \mathrm {D}_\varrho \)
where \(\alpha =u_x\), \(2\beta = u_{xx}/u_x\), \(\gamma = u_{xy}/u_x\), \(2\delta =u_{yy}/u_x\).
We assume that \(\varrho \) is chosen small enough such that \(u_x\) is positive everywhere in \(\mathrm {D}_\varrho \), each level set of u within the disc \(\mathrm {D}_\varrho \) is either a smooth line connecting two points at the circular boundary of the disc, or one of two single points on the boundary of \(\mathrm {D}_\varrho \) where u takes its maximum and minimum on \(\mathrm {D}_\varrho \), respectively.
The order-p mean of u within \(\mathrm {D}_\varrho \) is the minimiser of
By some rough estimates one can conclude that for \(\varrho \rightarrow 0\), \(\mu \sim \varrho ^2\). We substitute therefore
and obtain
In the following we focus therefore on finding the extremum of E (minimum for \(p>0\), maximum for \(p<0\)).
1.1.3 Separation of the Integral
The integral E from (72) can be reorganised into a nested integration where the inner integral integrates along a level line of \(\omega \) going through \((\xi ,0)\), and the outer integral then integrates along the \(\xi \) axis. We have
where \(\tilde{\xi }\) is a function of \(\eta \) that describes the level line of \(\omega \) that goes through \((\xi ,0)\), and reaches the boundary of \(\mathrm {D}_1\) at \(\eta ^*_+>0\) and \(\eta ^*_-<0\). (Note that the fact that \(\omega _\xi \) is positive throughout \(\mathrm {D}_1\) implies that the level line through \((\xi ,0)\) can be described in this way.)
The error term \(\mathcal {O}(\varrho ^3)\) results from the neglection of those level lines near the maximum and minimum of \(\omega \) within \(\mathrm {D}_1\) that do not reach the \(\xi \) axis within \(\mathrm {D}_1\).
In (74), the inner integral
measures the density of the value \(\omega (\xi ,0)\) in the overall distribution of \(\omega \) values within \(\mathrm {D}_1\) by integrating along the level line \(\tilde{\xi }(\eta )\) with \(\eta \) as integration parameter the inverse density of level lines in \(\xi \) direction. It is important here that the inverse density of level lines is measured in a direction perpendicular to that of integration. The density of level lines in \(\xi \) direction is exactly the derivative \(\partial \omega /\partial \xi \) taken at the point \((\tilde{\xi },\eta )\), i.e. the denominator of the integrand.
Integrating the quantity V multiplied with the penaliser \(|\omega -\kappa \varrho |^p\) would directly yield \(E(\kappa )\) if the integration were carried out w.r.t. \(\omega \). We prefer, however, to keep the integration over \(\xi \) in order to avoid plugging in the inverse function of \(\omega (\xi )\equiv \omega (\xi ,0)\) everywhere in the expressions. This is compensated by the factor \((\partial \omega /\partial \xi )(\xi ,0)\) placed at the end of the integrand of (74) that represents just the substitution of \(\omega \) with \(\xi \) (along the \(\xi \) axis \(\eta =0\)) as integration variable.
For ease of evaluation, we combine in the following the substitution factor with the weight \(V(\xi )\) in one single expression:
1.1.4 Evaluation of the Inner (Weight) Integral
To evaluate (76), we determine first the level line function \(\tilde{\xi }(\eta )\) for given \(\xi =\tilde{\xi }(0)\) by using the Taylor expansion (73):
The \(\eta \) coordinates \(\eta ^*_{\pm }\) of the end points of the level line are obtained from the condition \(\tilde{\xi }^2+{\eta ^*}^2=1\) as
where
are polynomials in \(\xi \) the exact coefficients of which are not further needed.
Based on the Taylor expansion (73) we obtain
and with (79)
Combining (84) and (85) we have
and therefore
From (80) one sees that
which allows to continue (87) into
with
1.1.5 Domain Splitting of the Outer Integral
The outer integral of (74), i.e. the integration of \(W(\xi )\) with the penaliser function \(|\omega -\kappa \varrho |^p\), is now split into four parts.
First, we split the integration interval at \(\xi =\nu \varrho \) where \(\omega (\nu \varrho )=\kappa \varrho \) to reduce \(|\omega -\kappa \varrho |\) to either \(\omega -\kappa \varrho \) or \(-\omega +\kappa \varrho \) in each subinterval. By (73) one has \(\nu =\kappa +\mathcal {O}(\varrho ^2)\).
Second, the density term \(W(\xi )\) contains \(\sqrt{1-\xi ^2}\) which is not differentiable at \(\pm 1\), precluding Taylor expansion of this term near the outer interval boundaries. On the other hand, the p-th power penaliser is for \(p\le 1\) not differentiable at 0 and can therefore not be treated by Taylor expansion at the boundary \(\nu \) between the two integration intervals. For this reason, we split each of the two intervals again at \(|\xi |=\sqrt{\varrho }\). This allows to simplify the integrals in later steps by applying Taylor expansion to either \(W(\xi )\) or the penaliser function, safely avoiding the critical regions of each.
As a result, we have
1.1.6 Evaluation of the Outer Integral I
We start by evaluating the integrals \(F_{\mp }\). In the following the upper signs refer to \(F_-\), the lower ones to \(F_+\). In expanding the power \((1+\ldots )^p\) by a Taylor series, it is important to note that \(\varrho /\xi \) is of order \(\mathcal {O}\bigl (\sqrt{\varrho }\bigr )\) due to the lower integral bound.
This gives
In the intermediate step (100) the factors \(p-1\), \(p-3\) occur in the denominators of some terms, which would necessitate the exclusion of \(p=1\) and \(p=3\). However, we see in (99) that the coefficient \(\left( {\begin{array}{c}p\\ 4\end{array}}\right) \) in front of \(I_{p-4}\) vanishes for \(p=1\) and \(p=3\), and similarly \(\left( {\begin{array}{c}p\\ 2\end{array}}\right) \) in front of \(I_{p-2}\) vanishes for \(p=1\), thus sparing the expansion of the respective integrals via (64) and (65). With this consideration, (101) can be obtained also in these cases.
1.1.7 Evaluation of the Outer Integral II
We turn now to evaluating \(G_\mp \). After expanding \(\omega \) in the penaliser function and cancelling terms due to \(\nu =\kappa +\mathcal {O}(\varrho ^2)\) we substitute \(\xi =\sqrt{\varrho }\,\zeta \). Using furthermore the Taylor expansion of \(\omega \) in \(\xi \) direction around \(\nu \varrho \),
we obtain
1.1.8 Extremum of the Combined Integral
Combining (93), (101) and (104), applying (92) and \(\nu =\kappa +\mathcal {O}(\varrho ^2)\) we obtain
It is worth noting that the \(\varrho ^{(p+3)/2}\) contributions cancelling in the last step belong to the integration boundaries of \(F_{\mp }\) and \(G_{\mp }\) at \(\pm \sqrt{\varrho }\).
For \(\varrho \rightarrow 0\), the last expression (105) is a quadratic function of \(\kappa \) with its apex at
Due to the sign of the \(\kappa ^2\) coefficient in (105) the apex is a minimum for \(p>0\) and a maximum for \(p<0\). The \(\mathrm {sgn}\,(p)\) factor in the original energy function \(E_0\) compensates for this such that \(E_0\) is always minimised.
1.1.9 Conclusion for Regular Points
From (106) the claim of the proposition for regular points follows by substituting back \(\kappa \alpha \varrho ^2=\mu \), \(\alpha \beta =u_{xx}/2\), \(\alpha \delta =u_{yy}/2\), and noticing that by our ansatz \(u_x>0\), \(u_y=0\) the coordinates x, y coincide with the geometric coordinates \(\eta \), \(\xi \) as used in the proposition.
1.1.10 Critical Points
The inequalities for local minima (maxima) are obvious consequences of the fact that for any \(\varrho >0\) the mean-p filter value is in the convex hull of values \(u(\varvec{x})\), \(\varvec{x}\in \mathrm {D}_\varrho (\varvec{x}_0)\). \(\square \)
1.2 Proof of Proposition 2
With the same substitutions as in the previous proof, the mode of \(\omega \) is given by the maximiser of \(V(\xi )\). By a slight modification of the calculations of the previous proof one finds
Equating \(V'(\xi )\) to zero yields \(\omega (\xi )=(\delta -2\beta )\varrho +\mathcal {O}(\varrho ^2)\) for the mode. For local minima (maxima), the same reasoning as in the previous proof applies. \(\square \)
1.3 Proof of Proposition 3
We proceed largely analogous to the proof of Proposition 1 in Appendix A.1. However, the integral decomposition gets simpler since no infinite ascents of the weighting at the integral boundaries \(\pm 1\) need to be controlled.
1.3.1 Regular Points: Ansatz via Taylor Expansion
Let the signal u and mean order p be given as in the proposition. Assume w.l.o.g. that the regular location \(x_0\) is 0 with \(u(0)=0\), and that the derivative of u at 0 is positive, \(u_x(0)>0\). Let a neighbourhood radius \(\varrho \) be given.
Using Taylor expansion of u up to third order, we obtain for \(-\varrho \le x\le \varrho \) the following expression:
where \(\alpha =u_x\), \(2\beta = u_{xx}/u_x\).
We assume that \(\varrho \) is chosen small enough so \(u_x\) is positive throughout \([-\varrho ,\varrho ]\), i.e. u is strictly monotonic within this interval. The order-p mean of u within \([-\varrho ,\varrho ]\) is the minimiser of
By rough estimates one can again conclude that for \(\varrho \rightarrow 0\), \(\mu \sim \varrho ^2\). We substitute therefore
and obtain
In the following we focus therefore on finding the extremum of E (minimum for \(p>0\), maximum for \(p<0\)).
1.3.2 Domain Splitting of the Integral
We split the integral (112) into two parts, using again the location \(\xi =\nu \varrho \) where \(\omega (\nu \varrho )=\kappa \varrho \) as splitting point. By (113), one has \(\nu =\kappa +\mathcal {O}(\varrho ^2)\). We have then
By substituting the integration variables, one obtains
where again the upper and lower signs refer to \(F_-\) and \(F_+\), respectively.
1.3.3 Evaluation of the Integrals
The Taylor expansion for \(\omega \) around \(\nu \varrho \) is identical with (102). Inserting this into (117), we have further
from which by evaluating the standard integrals, adding \(F_-\) and \(F_+\) and inserting \(\nu =\kappa +\mathcal {O}(\varrho ^2)\) we reach
The extremum of E is again found as the apex of the quadratic function on the r.h.s., from which the claim for regular points follows.
For critical points, the reasoning from Appendix A.1 applies. \(\square \)
1.4 Proof of Proposition 4
Assuming again that the regular location for the signal u is \(x_0=0\), and u is strictly monotonically increasing and Lipschitz within \([-\varrho ,\varrho ]\), the density of each value u(x) for \(-\varrho \le x\le \varrho \) is proportional to \(1/u'(x)\). The maximum of these values is reached at \(u(-\varrho )\) if u is convex, or \(u(\varrho )\) if u is concave. This proves the claim for regular points. If \(x_0\) is a local extremum, the density has a pole at \(u(x_0)\) and is finite for all other values, making \(u(x_0)\) the mode. \(\Box \)
1.5 Proof of Proposition 5
1.5.1 Regular Points: Ansatz via Taylor Expansion
Let the volume image u and mean order p be given as in the proposition. Assume w.l.o.g. that the regular location \(\varvec{x}_0\) is (0, 0, 0) with \(u(0,0,0)=0\), and that the gradient of u at (0, 0, 0) is in the positive x direction, i.e. \(u_x>0\), \(u_y=u_z=0\). Let a neighbourhood radius \(\varrho >0\) be given, and denote the closed (Euclidean) \(\varrho \)-neighbourhood of (0, 0, 0) by \(\mathrm {B}_{\varrho }\).
Using Taylor expansion of u up to third order, we can write for \((x,y,z)\in \mathrm {B}_\varrho \) the ansatz
We assume that \(\varrho \) is chosen small enough such that \(u_x\) is positive everywhere in \(\mathrm {B}_\varrho \), each level set of u within the ball \(\mathrm {B}_\varrho \) is either a smooth surface patch bounded by a closed regular curve on the boundary of the ball, or one of two single points on the boundary of \(\mathrm {B}_\varrho \) where u takes its maximum and minimum on \(\mathrm {B}_\varrho \), respectively.
The order-p mean of u within \(\mathrm {B}_\varrho \) is the minimiser of
Rough estimates again ensure \(\mu \sim \varrho ^2\) for \(\varrho \rightarrow 0\). Combining an appropriate rescaling with a transition to cylindrical coordinates with the axis in gradient (x) direction, we substitute
and obtain
where the integration in cylindrical coordinates has been written using the disc \(\mathrm {D}_1\) for the \(\xi \), \(\eta \) coordinates. The Taylor expansion of u transfers to
We aim again at finding the extremum of E.
1.5.2 Separation of the Integral
Similar to Appendix A.1.3, the integral E from (125) can be reorganised into a nested integration where the inner double integral (in polar coordinates) integrates over a level surface of \(\omega \) going through \((\xi ,0,0)\), and the outer integral then integrates along the \(\xi \) axis. We have
where \(\tilde{\xi }\) is a function of \(\eta ,\varphi \) that describes the level set of \(\omega \) which goes through \((\xi ,0,0)\), and reaches the boundary of \(\mathrm {B}_1\) at \((\eta ^*(\varphi ),\varphi )\). (Note that our initial assumptions on u ensure that the level set can be described in this way.)
Analogously to Appendix A.1.3 we rewrite (132) as
1.5.3 Evaluation of the Weight Integral
Within any axial plane (\(\varphi =\mathrm {const}\)), (126) is exactly (73). We can therefore transfer verbatim the analysis from Appendix A.1.4, which leads to the expression (79) for \(\tilde{\xi }\), the expression for \(\eta ^*_+\) from (80) for \(\eta ^*(\varphi )\), and (86) for \(\frac{\partial \omega }{\partial \xi }(\xi ,0,0)/\frac{\partial \omega }{\partial \xi } (\tilde{\xi },\eta ,\varphi )\).
Inserting (86) into the inner integral of (134) leads to
To finally obtain \(W(\xi )\), the latter expression needs to be integrated over \(\varphi \). From (80) one obtains by lengthy but straightforward calculation
where \(\varPsi \) is a third-order polynomial in \(\xi \) obtained by integrating (82) w.r.t. \(\varphi \),
Analogously one obtains
Inserting (135)–(142) into (134) yields after sorting terms, similarly to (91), (92),
with
1.5.4 Domain Splitting of the Outer Integral
As the outer integral of (132) has the same structure as in the 2D case, we use the same domain splitting (93).
1.5.5 Evaluation of the Outer Integral I
The first steps in evaluating the integrals \(F_{\mp }\) are as in the 2D case. In (98), the longer expansion (143) has to be used for \(W(\mp )\), which then leads to
which yields
Using the abbreviation
we can sort this into
which by \(J_q= \tfrac{1}{q+1}(1-\varrho ^{(q+1)/2})\) for \(q\ne -1\) (the special case \(J_{-1}\) only occurs as \(J_{p-2}\) for \(p=1\) and has then a vanishing coefficient) yields
1.5.6 Evaluation of the Outer Integral II
Starting with the same substitution \(\xi =\sqrt{\varrho }\,\zeta \) and Taylor expansion of \(\omega \) in \(\xi \) direction as in Appendix A.1.7, we evaluate
1.5.7 Extremum of the Combined Integral
When we finally combine (93), (149) and (151) and apply (144) and \(\nu =\kappa +\mathcal {O}(\varrho ^2)\), we observe as in the 2D case than all terms originating from \(G_-+G_+\) (151) cancel, and it remains
For \(\varrho \rightarrow 0\), the extremum of \(E(\kappa )\) can again be found as the apex of the quadratic function in (152), which yields
1.5.8 Conclusion of the Proof
From (153) the claim of the proposition for regular points follows by substituting back \(\kappa \alpha \varrho ^2=\mu \), \(\alpha \beta =u_{xx}/2\), \(\alpha \delta _0=u_{yy}/2\), \(\alpha \delta _2=u_{zz}/2\) and noticing that by our ansatz \(u_x>0\), \(u_y=u_z=0\) the coordinates x, y, z coincide with the geometric coordinates \(\eta \), \(\xi \), \(\chi \) as used in the proposition.
For critical points, the reasoning from Appendix A.1 applies. \(\square \)
1.6 Proof of Proposition 6
Analogous to Appendix A.2, we calculate
The relevant solution of \(V'(\xi )=0\) yields up to higher order terms \(\omega (\xi )=\tfrac{1}{2}(\delta _0+\delta _2-2\beta )\varrho \). \(\square \)
Continuous Order-p Means and Mode: A Toy Example
To understand the behaviour of order-p mean filters for \(p>-1\), \(p\ne 0\) and their relation to the mode of a continuous density, we consider the following simple example. Let z be a real random variable with (non-normalised) density
Here, \(0<m<<1\) is a fixed parameter, and \(0<\lambda \le (1+m)^{-2}\) to ensure that \(\gamma (z)\ge 0\) for all z. Obviously, the mode of z is the maximum of \(\gamma \), i.e. m.
For any \(p>-1\), \(p\ne 0\) the order-p mean of z is given by the minimiser of \(E(\mu )\) where
from which the minimiser \(\mu ^*\) of \(E(\mu )\) can be read off as the apex of the quadratic function of \(\mu \) as
For any fixed \(\lambda \in \bigl (0,(1+m)^{-2}\bigr )\), the minimiser \(\mu ^*\) goes to m for \(p\rightarrow -1\), but approaches \(2\lambda m / (1+\lambda )<m-m^2/3\) for \(p\rightarrow 0\). Moreover, if we send \(\lambda \) to 0, making the density more and more uniform, for any \(p>-1\) we have \(\mu ^*\rightarrow 0\) which comes as no surprise as for a flattening out density, any penalisation where the penaliser increases with distance will end up in the symmetry centre of the support interval \([-1,1]\).
Rights and permissions
About this article
Cite this article
Welk, M., Weickert, J. PDE Evolutions for M-Smoothers in One, Two, and Three Dimensions. J Math Imaging Vis 63, 157–185 (2021). https://doi.org/10.1007/s10851-020-00986-1
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10851-020-00986-1