next up previous
Next: Implementation via STFTs for Up: Time-Frequency Reassignment (Charpentier 1986, Previous: Quick Intro

Framework: the Wigner-Ville Distribution

The Wigner-Ville distribution: A High Resolution Time-Frequency Distribution:
$\displaystyle WV(x;t,\omega) =
\int x(t+\tau/2)\cdot x^*(t - \tau/2)e^{-j\omega\tau} d\tau$     (8)

Notice this looks DFT-like.

Despite High resolution, ``cross-terms may lead to an erroneous visual interpretation of the signal's time-frequency structure.'' (See overhead figures.)

So apply lowpass filter to get a Cohen-class Time-Frequency representation:
$\displaystyle TFR(x;t,\omega) =\int\int \phi_{TF}(u,\Omega)
WV(x;t-u,w-\Omega) du \ensuremath{\frac{d\Omega}{2\pi}}$     (9)

where $\phi_{TF}(u,\Omega)$ acts as a ``two-dimensional lowpass filter," or time-frequency window.

Important concept: This is smoothing over the WV distribution, which itself is DFT-like. So one must remember that we are smoothing over a region in the time-frequency grid.

Now, to do reassignment: find smoothed center of gravity in time and in frequency:
$\displaystyle \hat{t}(x;t,\omega)$ $\textstyle =$ $\displaystyle t - \ensuremath{\frac{\int \int u \cdot \phi_{TF}(u,\Omega) WV(x;...
...{TF}(u,\Omega) WV(x;t-u,\omega - \Omega) du \ensuremath{\frac{d\Omega}{2\pi}}}}$ (10)
$\displaystyle \hat{\omega}(x;t,\omega)$ $\textstyle =$ $\displaystyle \omega - \ensuremath{\frac{\int \int \Omega \cdot \phi_{TF}(u,\Om...
...{TF}(u,\Omega) WV(x;t-u,\omega - \Omega) du \ensuremath{\frac{d\Omega}{2\pi}}}}$ (11)


In our current application, this smoothing will simply be the WV distribution of the familiar time domain windowing function $h(t)$, or:

\begin{displaymath}
\phi_{TF}(u,\Omega) = WV(h;u;\Omega)
\end{displaymath}



NEAT TRICK: when we do this, it can be shown (see Auger and Flandarin) that we can make equations 10 and 11 into equations in terms of just three modified STFTs!

Hand-wavy explanation of how to massage those equations into STFTs: recall Fourier theory:

multiply by the time variable $\rightleftharpoons$ differentiate in frequency
multiply by frequency variable $\rightleftharpoons$ differentiate in time

Since we may arrange the equations around the window $h(t)$ and its transform, we see that these operations need apply only to the window. This leads to a convenient implementation...
next up previous
Next: Implementation via STFTs for Up: Time-Frequency Reassignment (Charpentier 1986, Previous: Quick Intro
Aaron S. Master 2003-02-04