Awesome

HOMS_SignalProcessing

HOMS_SignalProcessing is a C++ toolbox with MATLAB wrapper for joint smoothing and partitioning of one-dimensional signals and time series using higher order Mumford-Shah and Potts models.

Fast and numerically stable algorithms for solving univariate Mumford-Shah problems of any order
The used dynamic programming scheme computes a global minimum
Theoretical quadratic runtime guarantee
Practical linear runtime by using pruning strategies

This is an implementation of the algorithms described in the paper

M. Storath, L. Kiefer, A. Weinmann Smoothing for signals with discontinuities using higher order Mumford-Shah models. Numerische Mathematik, 2019.

Joint Smoothing and Partitioning of 1D Data and Time Series

Piecewise smooth models

Recovering a piecewise smooth signal from noisy data: classical spline approximation smooths out the discontinuities; the classical first order Mumford–Shah model allows for discontinuities, but the estimate misses most of them; higher order Mumford–Shah models provide improved smoothing and segmentation.

titleimageA

Piecewise polynomial models

Recovering a piecewise constant signal from noisy data with the piecewise constant Mumford-Shah model (aka Potts model). titleimageB

Recovering a piecewise affine-linear signal from noisy data with the piecewise affine-linear Mumford-Shah model. titleimageC

Recovering a piecewise quadratic signal from noisy data with the piecewise quadratic Mumford-Shah model. titleimageD

Installation

Compiling

The algorithm depends on a mex script that needs to be compiled before execution. For compilation inside MATLAB, cd into the 'src/cpp' folder and run build.m

Requires the Armadillo library

Tested with Armadillo 8.400 https://launchpad.net/ubuntu/+source/armadillo/1:8.400.0+dfsg-2

On Linux, just use your package manager to install it:

sudo apt-get install libarmadillo-dev

Running

For a test run, run demo.m. demo.m calls the wrapper function higherOrderMumShah1D.m which calls the necessary C++ sources

Arguments of higherOrderMumShah1D are:

data: input data array (double)
gamma: complexity penalty (larger choice -> fewer segments)
beta: smoothing parameter (larger choice -> stronger smoothing; beta=inf -> piecewise polynomial model)
order: order of the applied Mumford-Shah or Potts model (polynomial trends up to the specified order will be preserved)
varargin: optional input parameters

References

M. Storath, L. Kiefer, A. Weinmann "Smoothing for signals with discontinuities using higher order Mumford-Shah models." Numerische Mathematik, 2019.