Background Neutralization

Correct a measured residual sky color cast -- a channel imbalance in a chosen background reference -- after the main color-calibration and stretch path, without modeling a spatial gradient and without acting as a substitute for physical/catalog color calibration.

schema 1 · revision 1

Recognize it

Dark-corner sky patches read a noticeably non-neutral color (green, blue, or magenta tint) even after SPCC/SSSC and stretch have run.
The rest of the image otherwise looks correctly calibrated -- this is a residual background offset, not a broken calibration path.

Intended output

A reduction in the measured dark-corner sky channel imbalance, with target color, sky level/headroom, stars, and extended structure preserved closely enough that the result remains a valid continuation of the same processing path.

Required input state

Nonlinear/stretched RGB, after the main background-extraction/gradient step, after the color-calibration path and its success/failure state are known, preferably after a linked stretch when preservation of calibrated color is intended, and with a meaningful region of relatively clean sky/background available to measure. Not applicable to monochrome data -- the gate explicitly skips it.

What NOVA does

NOVA gates on measured dark-corner sky ratios (image_analyzer._color()'s sky_g_over_r/sky_b_over_r), skipping when the maximum deviation from 1.0 is <= 0.10 to preserve an already-correct SPCC balance, and running when it exceeds 0.10. When it runs, tool_params.compute_background_neutralize() selects the established scale correction. seti_astro.background_neutralize() then calls SASpro's auto_rect_50x50() to find a background patch and background_neutralize_rgb() to correct it, applies a NOVA-specific dark-corner completeness pass (driving all four frame corners toward the darkest channel's median, motivated by a case where the single auto patch read neutral while the corners still carried a cast), and finally blends the whole result back toward the original pixels in signal-bearing regions using a reused luminance mask -- failing closed (skipping the step) if that masking step itself fails. The objective post-check requires the measured dark-corner cast to drop by more than 0.02 with the SNR-proxy retained above 0.85, and can request a stronger retry when residual imbalance stays above 0.15.

Use when

A nonlinear RGB image has a measured residual background color cast after the upstream calibration/stretch state is understood.
Multiple plausible sky regions agree on the direction of the cast.
The image has a trustworthy clean-background reference, or the auto reference has been inspected/validated.
A fallback color path left a background pedestal/cast and a later cleanup is explicitly intended, not misrepresented as catalog calibration.

Skip when

Monochrome data.
The representative background is already neutral within the measured 0.10 engineering tolerance -- SPCC-calibrated color should be preserved, not disturbed, when there is no measured problem.
No clean background reference exists and the darkest areas are likely real extended signal (frame-filling nebulae, IFN, galaxy halos).
The apparent issue is a spatial gradient or vignetting rather than a global channel bias -- that is a background-extraction problem, not this step's job.
The upstream color or stretch failure has not been diagnosed -- this step should not be used to mask a problem that belongs earlier in the pipeline.

Limits

This is a channel-balancing transform based on a reference region, not gradient extraction and not physical color calibration. A neutral corner after this step is not proof that stars or nebula have physically correct color, and it does not repair a spatial gradient -- a flatter-looking corner should never be credited as gradient removal without an independent, matched P11 comparison.
The run/skip gate uses dark-corner sky ratios specifically because whole-frame color can be dominated by target signal. Ordinary processing keeps the established scale correction; Experiment Mode compares that real scale/pivot branch with SASpro's explicit offset branch and a no-treatment control.
As of 2026-09-04, NOVA's wrapper reuses Sky-Selective Green Rebalance's own luminance mask to blend the SASpro-corrected result back toward the original pixels wherever the mask says signal, not background -- so the correction is no longer applied uniformly to the whole frame. A real M42 run recorded in the source: the unmasked correction neutralized the sky (corner G/R 0.673 -> 0.939) but dragged the bright nebula core -- whose true color is not neutral, dominant OIII lands in G -- from G/R 0.914 to a visibly green 1.133; the masked correction restored the core to G/R 0.918 (matching its pre-correction value) while leaving the corner correction (0.941) essentially unchanged. If mask construction itself fails, NOVA skips neutralization entirely rather than falling back to the unmasked correction (fail-closed).
The auto-selected background reference is a heuristic, not a semantic "clean sky" detector: it searches for the darkest available patch. Dark does not guarantee uncontaminated -- extended faint nebulosity, IFN, dust, mosaic seams, vignette remnants, and narrowband structure can all occupy the darkest region.
Historical run evidence has sometimes credited this step with dramatically improving "gradient" or spatial flatness. That inference is rejected here: those runs also contained major upstream palette/stretch/background differences, and a channel-balancing transform is not a gradient model -- see the linked background extraction article for gradient-specific evidence.

Scientific and aesthetic notes

A rigorous comparison must hold constant the exact input FITS, the color-calibration success/failure path, the linked/unlinked stretch state, the narrowband/palette state, crop and mosaic coverage state, and ideally use a fixed validated background reference rectangle if testing transform semantics rather than reference selection.
Before treating a background region as valid, check that it is consistently dark across multiple scales, contains no detected stars or compact objects, does not reveal diffuse nebulosity/IFN/dust under a deep stretch, is not near a mosaic seam or vignette edge, and agrees with several other candidate sky regions.
Offset and scale are the two real correction semantics exposed by NOVA; pedestal removal remains a separate, unexposed axis.
A single historical run reporting a large numeric improvement from one pivot mode is confounded artifact evidence at best -- not statistical confidence, and not proof that mode is universally preferable.

Evidence and measurements

Measurements	Acceptance criteria
Defined background ROI channel ratios (G/R, B/R) before and after, in the exact region used to evaluate the step -- not whole-frame medians. Spatial consistency across several independently selected sky ROIs, not only one corner aggregate. Change map / residual map by channel, showing where the step altered the frame. Sky-level/headroom/clipping change per channel. Matched unsaturated-star color ratios, to check whether a background-only correction distorts calibrated stellar colors. Target-region color ratios, checked separately from background, especially in emission/narrowband fields. A no-op paired comparison on the exact same parent input.	Measured dark-corner sky cast drops without stronger corner neutrality being achieved by crushing one channel. Matched unsaturated star colors and target-region color ratios show no unexplained systematic shift. No new opposite-sign cast (e.g., a corrected green cast reappearing as magenta) elsewhere in the frame. The effective mode, selected reference rectangle, and whether the signal-preserving mask activated are recorded -- candidate ID alone is not sufficient evidence. A flatter-looking corner is never accepted as gradient-correction evidence without an independent, matched background-extraction comparison.

Measurements

Acceptance criteria

Defined background ROI channel ratios (G/R, B/R) before and after, in the exact region used to evaluate the step -- not whole-frame medians.
Spatial consistency across several independently selected sky ROIs, not only one corner aggregate.
Change map / residual map by channel, showing where the step altered the frame.
Sky-level/headroom/clipping change per channel.
Matched unsaturated-star color ratios, to check whether a background-only correction distorts calibrated stellar colors.
Target-region color ratios, checked separately from background, especially in emission/narrowband fields.
A no-op paired comparison on the exact same parent input.

Measured dark-corner sky cast drops without stronger corner neutrality being achieved by crushing one channel.
Matched unsaturated star colors and target-region color ratios show no unexplained systematic shift.
No new opposite-sign cast (e.g., a corrected green cast reappearing as magenta) elsewhere in the frame.
The effective mode, selected reference rectangle, and whether the signal-preserving mask activated are recorded -- candidate ID alone is not sufficient evidence.
A flatter-looking corner is never accepted as gradient-correction evidence without an independent, matched background-extraction comparison.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	A measurable reduction in dark-corner sky channel imbalance with target color, sky headroom, and structure preserved.	nova-background-neutralize-source, nova-m42-background-neutralize-mask-2026-09-04
saspro	Same engine, adapted host	A background-referenced channel correction toward neutral in the selected rectangle.	saspro-backgroundneutral-source-p22
siril	Functional alternative	Channel-equalized background toward neutral gray in the selected region.	siril-manual-color-calibration-p22
pixinsight	Functional alternative	Channel-referenced background correction toward neutral in the selected reference.	pixinsight-resources-p22

NOVA Python pipeline SASpro-backed wrapper with NOVA completeness pass and signal-preserving mask (2026-09-04)

Confirm the measured dark-corner cast that triggered this step, not just whether the frame looks non-neutral overall.
Treat offset and scale as distinct correction semantics, and surface any candidate artifacts that collapse in a run.
Persist the selected reference rectangle, effective mode, pre/post patch medians, pre/post dark-corner ratios, and whether the signal-preserving mask activated -- candidate ID alone is not sufficient provenance for this step.

Controls and starting ranges

bg_neutral_offset / bg_neutral_scale: the two real Experiment Mode correction semantics
none: true control -- essential for proving correction was needed
0.10 run/skip gate: source-confirmed engineering threshold on dark-corner max(|G/R-1|,|B/R-1|); not a scientifically validated optimum

Mask behavior: Signal-preserving luminance mask built in (2026-09-04); fails closed (skips the step) if mask construction itself fails.

Failure modes

Crediting this step with gradient correction because corners look flatter.
Treating a large historical improvement from one pivot mode as validated superiority rather than confounded artifact evidence.
Assuming the darkest available patch is guaranteed clean sky.

Recovery

Revert to the parent input; validate or choose a different background reference; skip if no clean sky exists.

Seti Astro Suite Pro background neutralization backgroundneutral.py; production-installed 1.18.0 core semantics not independently re-verified in this repo

Select a background rectangle that represents true, uncontaminated sky -- the upstream source's own auto-finder searches for darkness, not cleanliness.
Be aware that the inspected upstream core's default `remove_pedestal=True` subtracts each channel's whole-image minimum before the reference patch is even measured, which can alter channel relationships ahead of the correction itself -- this is unresolved/version-bound for the installed NOVA deployment.

Controls and starting ranges

mode: offset (explicit additive branch) or scale (NOVA's stable name for the pivot-around-1 branch)

Mask behavior: No signal-preserving mask at the inspected upstream core; NOVA's wrapper adds one on top.

Failure modes

Treating arbitrary non-offset strings as additional correction modes.

Recovery

Fall back to a manually validated reference rectangle, or skip.

Siril manual background-reference color calibration manual color calibration; exact version not independently confirmed in this repo

Select a background region that represents true background and avoid crowded or contrasty/stellar/target-dominated regions, per Siril's own documentation.
Treat this as the same process goal -- background-reference channel equalization -- with a different selection/transform implementation than NOVA/SASpro, not a pixel-identical algorithm.

Controls and starting ranges

background selection: manual region selection representing true sky

Mask behavior: Not applicable -- manual region selection is itself the targeting mechanism.

Failure modes

Selecting a background region that is dark but not actually clean sky.

Recovery

Choose a different background region; compare against a no-op.

PixInsight BackgroundNeutralization BackgroundNeutralization process; not currently dispatched by NOVA for this step

Select a background reference region using the same true-sky discipline as any other background-neutralization tool -- darkest is not automatically cleanest.
Treat this as the same process goal with a different host implementation; this packet did not establish a production NOVA-to-PixInsight dispatch path for this step, so no direct execution equivalence is claimed.

Controls and starting ranges

background reference: manually or ROI-selected region representing true sky

Mask behavior: Not independently confirmed in this repo.

Failure modes

Selecting a background reference that is dark but not actually clean sky.

Recovery

Choose a different reference region; compare against a no-op.

Validation status

Not yet Jeff-validated. Use this as sourced guidance, not a certification.

Sources

Phase-2 research packet P22, 2026-08-28 packet:P22:2026-08-28 — NOVA source-confirmed
NOVA background neutralization wrapper, gate, and signal-preserving mask nas_server/seti_astro.py:background_neutralize; nas_server/tool_params.py:compute_background_neutralize; nas_server/auto_process.py:_physics_should_run/background_neutralize; nas_server/processing_ontology.json background_neutralize — NOVA source-confirmed
Real M42 run: whole-frame correction dragged the nebula core green before the signal-preserving mask; masked correction restored it nas_server/seti_astro.py:background_neutralize (2026-09-04 code comment, real run values) — NOVA execution record
Seti Astro Suite Pro backgroundneutral.py (upstream source, commit 6df4a28) — Vendor-documented
Siril 1.4.4 Manual Color Calibration documentation — Vendor-documented
PixInsight official resources — Vendor-documented

Related evidence

The M66 recipe and Jeff-validation ledger will link here when their static exports are available. Until then, this article remains explicitly unvalidated.