OriginStack

Phase 4, step by step

After the frames are stacked, OriginStack finishes the picture in a chain of steps. This page shows each one on a real stack, with a slider so you can drag across the change it makes.

How to read the sliders The whole chain Hot pixels Background Deconvolution Green noise Star reduction Local contrast What did not run

How to read the sliders

The pictures are the sample Whirlpool session (41 frames of 25 s from a Celestron Origin), stacked with the default settings, so you can repeat everything here yourself. Phase 4 works on the linear stack, which looks almost black until it is stretched, so every picture below is shown with the same fixed display curve: the sky level of each picture is set to black, and the brightness above it is scaled the same way. That way the slider shows what the step changed, not a difference in how it was displayed.

Drag, or use the arrow keys

Drag across a picture to move the divider: Before is on the left of it and After on the right. With a keyboard, tab to a slider and use the left and right arrow keys.

Times

The times are from one run on an 8-core laptop. All of Phase 4 took about 40 seconds of the roughly 70 seconds the whole run took.

Try it yourself

Run python originstack.py --from-stack stack.fits -o again.fits on a linear stack to redo only Phase 4, and add --skip-step NAME to leave a step out (the names are in the headings below). In the desktop app, the Post-processing steps panel switches each step off and on after a run and redoes only what follows it.

The whole chain

First the two ends: the stack as Phase 3 leaves it, and the finished picture after every step below. Between them the sky turns from a green-blue haze to black, the galaxy gains its spiral structure and the noise drops.

The finished picture after Phase 4: a black sky, many stars and the galaxy with its spiral arms. The linear stack before Phase 4: a flat blue-green sky with the galaxies faint and small. BeforeAfter
The stack before Phase 4 (left) and after it (right). The haze and colour cast come out; the stars and the galaxy stay.

Hot pixels (hot_pixel, 0.8 s)

A pixel that stands far above its neighbours in one colour channel is pulled back toward them. Most hot pixels are already removed when the frames are calibrated against the dark; this step catches the ones that survive.

On this session it changed 16 pixels, each by at most about a quarter of its value, and all of them on or next to bright stars. There is nothing to see, so there is no slider here. It matters for sessions that have no darks.

Background (background, 7.4 s)

The sky in a stack is never perfectly flat or neutral: sky glow brightens one corner, and the red, green and blue channels sit at different levels. The background step measures the sky in many small patches between the stars (47 px here), fits a smooth surface through them and subtracts it, then sets all three channels to the same low level so the sky is neutral grey.

After background extraction: a dark, neutral sky with the galaxies and stars standing out. Before background extraction: a blue-green sky, brighter toward the lower right. BeforeAfter
The blue-green cast and the brighter corner go; the galaxy and stars stay. The pictures are drawn with the same curve, so the difference is real: before, the red channel sat about a quarter below green and blue.

Switch it off with --no-background-extraction; the surface is fitted from sky patches between the stars, so the stars are not subtracted with it.

Sharpening and noise: neural-network deconvolution (nn, 26 s)

A stack is blurred a little by the atmosphere and the optics, and every pixel carries noise. This step runs a small neural network that was trained on real sessions to turn a noisy, soft image into a cleaner, sharper one in a single pass. On galaxies --auto keeps 35% of the deconvolver’s output and mixes in 65% of the plain denoiser’s, so faint outer arms keep their contrast. It is the longest step and the one with the biggest effect.

After deconvolution: the same galaxy, smoother, with the arms and the companion defined. Before deconvolution: a grainy galaxy with soft spiral arms. BeforeAfter
The galaxy pair before and after, at full resolution. The grain in the sky is the noise of a 17-minute stack; the arms and the bridge toward the companion become much clearer.

Choose it with --deconvolve nn. Other denoisers and classical deconvolution (--deconvolve rl and others) are available; --auto picked this one for a galaxy on Origin data. It also works on other cameras: it rescales the image to the star width it was trained on first.

Green noise (scnr, 0.2 s)

A colour sensor has twice as many green pixels as red or blue, and the sky ends up with green-tinted specks. SCNR (subtractive chromatic noise reduction) lowers green where it exceeds red and blue.

After SCNR: the same patch of sky with the green removed, leaving red and blue grain. Before SCNR: sky noise with green and blue speckle. BeforeAfter
A patch of sky from the same picture, enlarged. The stars keep their colour; the green speckle in the sky is gone.

Star reduction (star_reduce, 0.5 s)

Bright stars spread their light into soft halos that bury the faint detail around them. Star reduction shrinks each star’s profile toward a tighter one (here the typical star width goes from 4.3 px to about 3.5) without touching what is not star. Turn it off with --no-star-reduce.

The widest wings of the very brightest stars are a different problem: they stay above the sky noise for hundreds of pixels. The opt-in --halo-removal with --config halo_removal_method=psf models them from a radial profile measured on your stack, scaled by each star’s Gaia magnitude, and subtracts them beyond the core without removing more than the light that is there. See advanced features.

After star reduction: the same star, tighter, with a smaller halo. Before star reduction: a bright star with a wide, soft halo. BeforeAfter
The brightest star in the field, enlarged three times. The core stays; the halo around it pulls in.

Local contrast and the final flatten (local_contrast 1.8 s, sky_neutralize)

The last two steps shape the picture. Local contrast raises the contrast of structure at three scales (about 2, 12 and 40 px, at strength 0.8); --no-local-contrast skips it. The final flatten then measures what is left of the sky, evens it out and sets it to a neutral grey, so the picture’s black point sits in the same place in every channel.

After local contrast and the final flatten: the spiral arms stand out more. Before local contrast: the galaxy pair, smooth and soft. BeforeAfter
The same crop as above, after the previous steps (left) and after these two (right). The change is smaller than the earlier ones, but it is what gives the arms their definition.

What did not run

Phase 4 chooses its steps from the data, so another target runs a different chain. On this session these did not run:

Colour noise reduction (chroma_nr)

Smooths colour noise where the denoiser has not. The neural network already covers it here.

Sky residual

Takes out a faint leftover sky pattern. It is skipped for extended targets such as a galaxy, because it could mistake the target’s own outer light for sky.

Other denoisers (wavelet, bilateral, acdnr, aniso)

The directional wavelet denoiser is the default when the neural network is not used; the others are opt-in with --denoiser.

Gray-locus colour calibration (photo_cal)

A fallback colour calibration. It is skipped because the colours were already set from Gaia stars on the linear stack, before Phase 4.

Star removal

Off unless you ask for it with --remove-stars.

After Phase 4 the picture is still linear. The stretch that makes it look like a photograph is applied when the preview is made; see the stretch gallery, and Colour and finishing for how each step is done.