Block
Combination
- What it does
- It makes one colour image from the stacks of several filters of a monochrome camera, or from the channels of a dual-band image.
- What it shows
- One colour image, which goes on through the chain as any stack does, with the scale of each channel given in the log.
- What can be set
- For S-H-O and H-O, the hue of each filter; the defaults suit most images, and Channel weights sets the share of each filter.
A Combination block makes one colour image from the images of several workflows: one workflow per filter of a monochrome camera, or the channels of one colour image shot through a dual-band filter. What it does is written in a recipe, a short file made of a few steps from a fixed list: the sky of every channel brought to zero, a linear match of the channels to a reference, an optional continuum subtraction. A recipe then makes red, green and blue from the channels, either with a hue per filter or with a matrix. Four recipes give the blocks S-H-O, H-O, RGB and LRGB.
- Stacking
- Denoise
- Combination
- Color
- Stars/background
1What goes in, what comes out
A combination is the first block of its own workflow. Each of its inputs, called a port (S II, Hα and O III for S-H-O; R, G and B for RGB), is wired to one workflow, one per filter. The block reads that workflow’s Output: its linear image, with the sky taken off. When the colour is made on separate images, a port may instead read the starless image or the stars alone of the workflow’s Stars/background block. Any workflow may feed any port. A port with no workflow is empty and the block waits for it.
Each port reads one channel of its image: the luminance of an image, its red, green or blue, or the mean of green and blue. A monochrome image is read as its grey. The mean of green and blue is the O III of a dual-band filter on a colour sensor, whose green and blue pixels both see the 501 nm line.
The block writes a stack, which its own Output block renders as it renders any stack. It handles no star: a Stars/background block placed after it does, and so does the Combine layers block in Post. The valid area of the result is where every input has data. Its plate solution is that of the first input that has one.
2One grid for all the channels
The block aligns nothing. Its inputs must be on one pixel grid, that of the first port. The channels of one colour image are. Images of other workflows are, once an Align on block, placed after their Stacking, has brought them onto the first one’s grid.
Align on has two methods. Stars, the default, matches triangles of stars (Valdes et al.,[1] Groth,[2] with the implementation of Astroalign[3]), then fits the cubic polynomial of Registration on every star matched. Plate reprojects the reference grid through the plate solutions of both images, when they carry one. If the inputs are not on one grid, the block refuses them and names the ports concerned: “Hα is not on the grid of S II: place Align on after its Stacking”.
3The steps of a recipe
A recipe lists steps in order, from three operations. The first is always the background neutralisation.
Background neutralisation
An empty sky must map to (0, 0, 0) whatever the hues. Each channel’s sky level is therefore measured on the same pixels in every channel and subtracted in the linear data. Rector et al.[4] set the background of each filter’s image to zero before the layers are coloured and added. The sky pixels are the faintest fifth of the blocks of 16 × 16 pixels, judged in all channels at once, each in units of its own spread. The level is the median of these pixels after a 3σ clipping.
Linear match
Each channel is fitted by least squares to a reference channel, reference = a × channel + b, on the pixels where the reference lies between 1e-7 and 0.875 of the sensor’s full scale. This is the linear match of Siril and of its RGB compositing tool,[5] with the bounds of that tool. The offset b moves the sky, which is taken back to zero after the fit. The reference is the H-alpha for narrow-band recipes, the strongest signal, so that a faint O III weighs as much as a strong H-alpha. It is the red for RGB and the luminance for LRGB, the top layer of the compositing tool.[5] A fit with a negative slope is refused: the channels do not show the same field.
Continuum subtraction
A narrow-band filter records its line and the continuum under it. A broadband filter around the line records the continuum and a small share of the line. The stellar method[6][7][8] takes the stars of the field, which have no line emission: the ratio of their fluxes in the two images is the scale of the continuum. The line image minus the broadband image times that scale gives the line. FLOAT measures the stars by aperture photometry, a circle whose radius is 1.5 times the stars’ full width at half maximum, the local sky taken in an annulus from 1.25 to 2 radii, and keeps the median of their ratios. A star with a neighbour in its annulus, a saturated star and a star under 10 times its flux error are left out. At least 5 stars are needed. The line is then divided by the scale, so that a star is equal in both images. A recipe may also give the scale as a number.
After the steps, the weights of the channels (1 by default) scale each channel.
4A hue per filter, lightness and colour
The narrow-band recipes give each filter a hue, in degrees on the colour wheel (0 red, 120 green, 240 blue). Each filter’s greyscale image is tinted with its hue and the layers are added, as Rector et al. describe.[4] The colour balance is solved so that a single filter’s signal has exactly its own hue and a neutral input stays neutral. Three hues solve it exactly. Two hues, for H-O, are solved by least squares, which is exact when the hues face each other. A recipe may give a matrix instead: three rows (red, green, blue), one number per channel.
The default hues of S-H-O are taken from five public S-H-O composites of the Hubble teams (ESA/Hubble heic1501a, opo9544a, heic1007a, heic1520a, heic0707a), read in the perceptual space Oklch on their nebula pixels: S II 15.5°, H-alpha 31.4° and O III 202.2° on the colour wheel of the linear data, which land on the warm areas’ lower quartile, the warm areas’ median and the blue areas’ median in Post’s default display. The defaults of H-O are the usual mapping of a dual-band filter: H-alpha red (0°), O III cyan (180°).
A narrow-band image is composed as lightness and colour, the way an LRGB is.[9]
- The lightness is the mean of the channels, each weighted by 1 over the square of its sky noise (the least noisy image weighs most), or an equal mean. Equal channels give that same value: a neutral source keeps its level.
- The colour is each pixel’s colour ratios to its own luminance, smoothed. The eye resolves colour more coarsely than lightness,[10] so the lightness keeps every detail. In Auto, the smoothing radius is where the noisiest channel’s noise, averaged by the Gaussian, comes down to the lightness’s own; below half a pixel it is 0. The radius is shown in the results.
- The noise floor. Where the colour’s luminance is under 3 times its noise, measured on the sky pixels, the pixel is grey. From 1 to 2 floors the colour comes in smoothly. The noise of a weak filter, raised by the balance, no longer paints the sky.
- The strength. The colour’s departure from grey is multiplied by 1.9, the luminance kept, so that the nebula’s median chroma in Post’s default display matches the 0.054 measured on the same five composites. Where the brightest channel would be shown above 0.9 of the display, the factor is lowered pixel by pixel down to 1: gamut compression at constant lightness,[11] so that a bright core is not flattened.
5RGB and LRGB
The broadband recipes give the colour as it was recorded. Their colour balance is that of a Color block placed after the combination, as in the templates Mono RGB and Mono LRGB: a photometric white balance of the stars, one factor per colour.[12] The linear match only puts the channels on one scale.
- RGB is the red, green and blue channels, matched to the red, and nothing else.
- LRGB is the colour of RGB and the lightness of L at every pixel. R, G and B are linearly matched to L. Each pixel’s three channels are scaled so that its colour ratios keep their own luminance times L, which keeps the pixel’s chromaticity on a linear image. Where the colour’s luminance is under its noise floor, the pixel is grey at L’s level: the depth of L is kept on the faint areas and on the sky, where the colour has no signal. The colour is not smoothed and its strength is not raised.
Both hand on to the Color block the saturation map of the stacks of R, G and B, so that the Color block leaves out the saturated stars by that map and not by an estimate from the brightest pixels. L is not in it: it scales the three channels alike, so a star saturated in L keeps the colour ratios of its R, G and B.
6The shipped recipes
A recipe is a file of the application’s folder. It names its inputs, its parameters with their defaults, its steps and its output; a file with an error is left out of the list with a message naming the file, the line and the setting. When a combination is placed in a project, its recipe is copied into the project, so that the project computes the same whatever happens to the file. A project whose copy is older than the file is told so, and changes only when the file is applied.
| Recipe | Inputs | Steps | Output |
|---|---|---|---|
| S-H-O | S II, Hα, O III | Sky; linear match to Hα | One hue per filter; lightness from the channels, weighted by noise; colour smoothed; strength 1.9 |
| H-O | Hα, O III (two filters, or a dual-band image’s channels) | Sky; linear match to Hα | Hues 0° and 180°; same composition as S-H-O |
| RGB | R, G, B | Sky; linear match to R | The channels as they are; saturation maps handed on |
| LRGB | L, R, G, B | Sky; linear match to L | Colour of R, G, B, lightness of L; colour not smoothed; saturation maps handed on |
| HOO (colour camera dual-band) | Hα read from the red, O III from the mean of green and blue | Sky; linear match to Hα | Hα in red, O III in green and blue, as they are; no lightness composition |
| HaRGB | R, G, B, Hα | Sky; G and B matched to R; continuum of Hα taken off by R, scale measured on the stars | The line added to the red, a fifth of it to the blue; saturation maps of R, G, B handed on |
For HaRGB, the line image is on the red’s scale, a star having the same value in both, so it adds to the red. The amount added is the weight of the Hα channel, 1 by default. The HOO recipe reads the H-alpha from the red pixels of a dual-band image and the O III from the green and blue ones, as Siril’s extraction does.[13]
7Settings
| Setting | Value | Range | What it does |
|---|---|---|---|
| S II hue | 15.5° | 0 – 360 | The colour of the S II filter on the colour wheel (S-H-O only). |
| Hα hue | 31.4° (S-H-O), 0° (H-O) | 0 – 360 | The colour of the H-alpha filter. |
| O III hue | 202.2° (S-H-O), 180° (H-O) | 0 – 360 | The colour of the O III filter. |
| Lightness | noise | noise, equal | How the filters make the lightness: weighted by their noise, or equally. |
| Colour smoothing | Auto | Auto, Radius | Auto takes the smoothing radius from the filters’ noise; Radius takes the one set below. |
| Smoothing radius | 0 px | 0 – 50 | The colour’s smoothing radius when Colour smoothing is Radius; 0: none. |
| Channel weights | 1 | per channel | Each channel’s weight, after the balance. A channel too faint for the balance is lowered by hand. |
A recipe other than the four shipped types has its own parameters, each with its default and its range, written in its file; the block shows them under Recipe settings and refuses a value outside the range.
8Limits, and what was tried
- No alignment inside the block. Inputs that are not on one grid are refused, naming them. Align on does the work, before the combination.
- The continuum scale needs stars. With fewer than 5 isolated, unsaturated stars measured in both images, the step stops and asks for the scale to be written in the recipe.
- A background taken from the input’s own sky level was tried and left. It differs from one filter to the other by what the nebula adds, and the balance multiplied the difference into a cast of the whole sky (section 3).
- The photometric calibration of stars on a combination. An estimate of the saturated stars from the brightest pixels refused most of a field’s stars on an LRGB (160 of 195); the saturation map of the stacks of R, G and B, handed on by the combination, replaces it (129 stars measured).
- No chrominance smoothing in LRGB. Each filter’s workflow has its own Denoise; the combination does not smooth the colour and does not raise its strength, which varies by a factor of five from one subject to another among broadband composites of professional observatories (nine ESO and ESA/Hubble images).
References
- Valdes, F. et al. 1995, FOCAS Automatic Catalog Matching Algorithms, PASP 107, 1119. doi:10.1086/133667
- Groth, E. J. 1986, A pattern-matching algorithm for two-dimensional coordinate lists, AJ 91, 1244. doi:10.1086/114099
- Beroiz, M., Cabral, J. B. & Sanchez, B. 2020, Astroalign: A Python module for astronomical image registration, Astronomy and Computing 32, 100384. arXiv:1909.02946
- Rector, T. A., Levay, Z. G., Frattare, L. M., English, J. & Pu’uohau-Pummill, K. 2007, Image-Processing Techniques for the Creation of Presentation-Quality Astronomical Images, AJ 133, 598. doi:10.1086/510117
- Siril documentation, RGB compositing and Commands (linear_match). siril.readthedocs.io; Siril FAQ, siril.org/faq
- Waller, W. H. 1990, Emission-line and continuum fluxes from narrow- and broad-band imagery, PASP 102, 1217. doi:10.1086/132752
- Koopmann, R. A., Kenney, J. D. P. & Young, J. 2001, An Atlas of H-alpha and R Images and Radial Profiles of 63 Bright Virgo Cluster Spiral Galaxies, ApJS 135, 125. arXiv:astro-ph/0106335
- Spector, O., Finkelman, I. & Brosch, N. 2012, Continuum removal in H-alpha extragalactic measurements, MNRAS 419, 2156. arXiv:1110.5636
- Gendler, R. 2013, Lessons from the Masters, Springer. doi:10.1007/978-1-4614-7834-8
- Mullen, K. T. 1985, The contrast sensitivity of human colour vision to red-green and blue-yellow chromatic gratings, J. Physiol. 359, 381. doi:10.1113/jphysiol.1985.sp015591
- Morovic, J. & Luo, M. R. 2001, The fundamentals of gamut mapping: a survey, J. Imaging Sci. Technol. 45(3), 283. doi:10.2352/j.imagingsci.technol.2001.45.3.art00013
- Siril tutorial, RGB composition. siril.org/tutorials/rgb_composition
- Siril documentation, Extraction. siril.readthedocs.io