silx.gui.colors: Add "percentile" mode for autoscaling

[clemisch]

It autoscales from 1st to 99th (nan)percentiles of the data. In my experience this usually gives a sensible value window for realistic data.

[woutdenolf]

Adding PERCENTILE_1_99 to the tests would be a bonus. Look for STDDEV3 in the tests.

[vallsv]

Hi and thx for the feature.

What do you mean by realistic data? Is that a specific use case or domain?
Could it be possible to have a screenshot with and without?

Also a percentile is like a mean, it's a sort on the values of millions of pixels, right?

The following this processes twice the same sort

numpy.nanpercentile(i, 1)
numpy.nanpercentile(i, 99)

Could you please check if numpy can do something faster the following way?

numpy.nanpercentile(i, [1, 99])

[clemisch]

What do you mean by realistic data? Is that a specific use case or domain?

There is no specific usecase or domain. With "realistic" I mean that it works well on noisy data, data which has artefacts outside of the region of interest, or a simulation which has extreme values only in a very small ratio of pixels. Min/max scaling would scale to the extreme values and not provide good contrast on the "main" content.

In other words, you get good linear contrast for the central 98% of the data.

The extisting mean+-std method only works for values around one mean value. This doesn't apply to many imaging examples where you have a value range inside some min/max but said "extreme" values on top.

---

Could it be possible to have a screenshot with and without?

Simulation of Fresnel propagation without noise (min/max and percentile scaling):

Noisy tomo reco:

Admittedly not the most impressive examples. I'm on a conference without good access to real data. In any case, 1-99 percentile scaling is the go-to autowindow of everyone in my x-ray imaging / tomography bubble.

---

Could you please check if numpy can do something faster the following way?

Yeah, numpy is indeed able to compute both percentiles simultaneously. This doesn't seem to be a recent addition so I'm not sure why I didn't use it so far. I changed the code to use it.

[t20100]

Hi @clemisch,

I've added some basic tests and reworked the code a bit, it's in this PR on your branch: clemisch#1

[clemisch]

Oh dang, I forgot to add the tests! Thank you for adding them, and props for committing into my branch despite my neglect 😲