SARA New YouTube video loaded: ezRA Analysis 5 Interference Filter
20 views
Skip to first unread message
Dr. Rich Russel
Mar 9, 2023, 9:50:00 PM3/9/23
to sara...@googlegroups.com
Marcus D. Leech
Mar 10, 2023, 12:04:35 AM3/10/23
to sara...@googlegroups.com
It struck me that 1d and 2d median filtering could do a lot of what
was discussed in that video, without manual
intervention.
Impulse removal via median filtering has been a technique for decades. 1d is typically used for things like audio files, or if
you pretend your FFT bin collection is a "time series" and apply a median filter with an appropriate window length.
2d median filtering has been used for "de-noising" images, where the median evaluation considers each "pixel" in a 2d context,
rather than a 1d context. If one considers a set of spectral estimates over some convenient time-interval as a 2d "image",
then one can apply 2d median filtering to that image to produce "better" data. Consider, for example a process that
delivers let's say 30 spectral estimates per second, and we consider 1 second worth of these estimates as our 2d
"image". A 2d median filter can remove short-lived artifacts in both dimensions. This is what some of the WVU RAIL
applications do, and I've been using it (or a version of it) in some of my applications also.
For "sharp" impulse-like frequency domain artifacts, 1d median filtering is quite effective at dramatically lowering amplitude.
If one is using spectral data to do *total power* plots (which are useful in their own right quite apart from spectral information),
then very narrow spikes in the frequency domain that aren't overwhelmingly loud tend not to have much of an effect
on the total power calculation. Consider a 1-bin wide rfi spike in a 2048-bin spectral plot that is only a couple or three dB
out of the noise. It barely modulates the total-power at all (and, actually, if that spike is always present, it doesn't "modulate"
the total-power at all).
Anyway. It just struck me that some of the techniques discussed needn't be manual...
intervention.
Impulse removal via median filtering has been a technique for decades. 1d is typically used for things like audio files, or if
you pretend your FFT bin collection is a "time series" and apply a median filter with an appropriate window length.
2d median filtering has been used for "de-noising" images, where the median evaluation considers each "pixel" in a 2d context,
rather than a 1d context. If one considers a set of spectral estimates over some convenient time-interval as a 2d "image",
then one can apply 2d median filtering to that image to produce "better" data. Consider, for example a process that
delivers let's say 30 spectral estimates per second, and we consider 1 second worth of these estimates as our 2d
"image". A 2d median filter can remove short-lived artifacts in both dimensions. This is what some of the WVU RAIL
applications do, and I've been using it (or a version of it) in some of my applications also.
For "sharp" impulse-like frequency domain artifacts, 1d median filtering is quite effective at dramatically lowering amplitude.
If one is using spectral data to do *total power* plots (which are useful in their own right quite apart from spectral information),
then very narrow spikes in the frequency domain that aren't overwhelmingly loud tend not to have much of an effect
on the total power calculation. Consider a 1-bin wide rfi spike in a 2048-bin spectral plot that is only a couple or three dB
out of the noise. It barely modulates the total-power at all (and, actually, if that spike is always present, it doesn't "modulate"
the total-power at all).
Anyway. It just struck me that some of the techniques discussed needn't be manual...
Reply all
Reply to author
Forward
0 new messages