Dark-Structure Enhancement

Selectively increase the visibility/contrast of dark structures embedded in brighter surroundings -- galaxy dust lanes, dark filaments, faint outer structure -- via multiscale local-contrast redistribution. This is contrast redistribution over existing recorded structure, not recovery of missing signal, not background extraction, deconvolution, denoise, or HDR compression, and not the same operation as PixInsight LHE, CLAHE, or curves, even though all pursue related-sounding perceptual goals.

schema 1 · revision 1

Recognize it

A specific dark structure (dust lane, dark filament, faint outer structure) is visibly or quantitatively present in the stretched image but lacks local separation from its immediate surroundings.

Intended output

Dark structures gain local contrast/depth against their immediate surroundings, with midtone/highlight regions outside the intended mask preserved and no new ringing, noise/mottle amplification, unnatural black channels, or seam/background-extraction-residue exaggeration.

Required input state

A nonlinear, already-stretched image whose upstream background, color, denoise, and dynamic-range decisions are already stable, so this step is not compensating for an earlier defect. Current ontology places DSE in the nonlinear stage and recommends it after HDR compression.

What NOVA does

`seti_astro.dark_enhance(n_scales=6, boost_factor=5.0, mask_gamma=1.0)` loads through the normalizing shared loader and calls SASpro's `compute_wavescale_dse()`. `tool_params.compute_dark_enhance()` derives a heuristic `shadow_snr` from the whole-image SNR-like proxy scaled by the clamped ratio of the global stretched median to 0.20, computes boost as roughly `1.5 + shadow_snr/4` (increased ~25% for galaxies, reduced for clusters), and derives scales/gamma from other image-statistic proxies -- these are engineering heuristics keyed to a global stretched median and generic SNR estimate, not a measurement of any specific dark structure. In Experiment Mode, `_adapt_nonlinear_variants()` additionally computes and blends an external shadow luminance mask (lower bound ~median*0.5 clamped 0.03-0.10, upper bound ~median*2.0 clamped 0.20-0.45) after the SASpro transform runs, on top of SASpro's own internal darkness mask.

Use when

The parent image is already correctly background-modeled and stretched, a specific dark structure is visibly/quantitatively present but lacks local separation, denoise is stable enough that enhancement is not mostly revealing grain, and the structure's spatial scale fits the chosen wavelet depth. Galaxy dust lanes and selected nebular dark features are the strongest current-use candidates.

Skip when

The parent already resolves the dark structure naturally, the target is star-dominated with little meaningful dark extended structure, background noise or mottling is close to the scale/contrast of the desired feature, the apparent feature changes strongly across independent stacks or aligns with gradients, tile seams, drizzle artifacts, ringing, or denoise residuals, or additional local contrast would make the image more dramatic but less natural.

Limits

The current upstream SASpro implementation is the clearest algorithmic description available: `compute_wavescale_dse()` performs an a-trous multiscale decomposition, builds a darkness mask from the negative portions of wavelet planes on selected mid-scales, and increases the magnitude of those negative coefficients over two iterations with an exponentially decaying scale weight; on RGB data it works in Lab luminance only, then converts back. This is contrast redistribution -- it does not estimate a physical sky/background model and does not infer missing detail from a PSF. If weak structure is below the useful SNR of the parent data, making a pattern more visible can equally make correlated noise or processing residue more visible.
NOVA's `_load_fits()` globally min/max-normalizes the loaded array before the transform, so normalization is part of the effective treatment -- a nominally identical boost/scale/gamma applied to differently-ranged parent files is not necessarily equivalent, and a comparison against PixInsight LHE must account for this.
NOVA adds an external shadow luminance mask after the SASpro transform in Experiment Mode, even though SASpro's DSE already computes its own internal darkness mask from wavelet coefficients. The real effective treatment is therefore NOVA global normalization -> SASpro internal wavelet darkness selection -> DSE transform -> NOVA external luminance-mask blend, not simply "WaveScale DSE at boost X." That composite chain should be part of any cited experiment provenance.
Issue #722 traced the historical treatment collapse to adaptive scaling: with an M51 baseline of 2.1, old declarations 1.0, 1.3, and 1.8 all hit the 1.0 clamp. A direct retained-frame sweep then supported declarations 3.0, 3.5, 5.0, and 7.5, which adapt to distinct 1.26, 1.47, 2.10, and 3.15 treatments on that fixture. This proves treatment identity and spacing there, not a universal physical optimum; historical winner counts retain their original effective-treatment meaning and are not reinterpreted.
The `pi_lhe_dark` candidate now routes through engine `pixinsight` and function `lhe`, exposing only `lhe_amount`. The former declared kernel-radius and slope-limit keys were removed because the executed wrapper does not consume them. This makes LHE executable as a strategy alternative, but does not make its amount numerically commensurate with DSE boost or establish comparative superiority.
The calibrated ontology range is 3.0-10.0 with default 5.0. The adapter still permits values through 12.0, but that wider ceiling was not selected by the retained-M51 calibration and must not be presented as validated guidance.
n_scales is a pixel count, not an angular unit -- a native SeeStar stack, a drizzled stack, and a differently cropped/resampled mosaic receive materially different angular/spatial treatment at the same nominal scale value.

Scientific and aesthetic notes

A rigorous DSE comparison needs the same exact nonlinear parent FITS for every candidate, a true unchanged `none` arm, effective post-adaptation parameters recorded for each arm (not just candidate ID, given the _STEP_CFG mismatch), normalization held constant or explicitly factored in, the same outer luminance mask when comparing strength probes, fixed downstream processing, and target/background ROIs defined before seeing candidate labels when possible. Compare DSE to LHE only as a strategy comparison, never a single-knob one -- DSE's wavelet scales/boost/masks and LHE's kernel radius/slope/amount are not numerically commensurate, and the corrected route alone is not evidence of a fair comparison.
Entropy, high-frequency power, generic sharpness index, NOVA's whole-frame SNR-like score, a "detail level" score, perceptual preference, and historical win rate can all rise when noise, ringing, seams, or artificial texture are amplified. A run-local preferred DSE result is evidence of preference for that finishing treatment, not evidence that new astrophysical detail was recovered.
Clusters are already downweighted in NOVA's adaptive heuristic -- a sensible direction, but a heuristic policy rather than a validated universal rule.

Evidence and measurements

Measurements	Acceptance criteria
Local contrast across known dust-lane/adjacent-signal profiles, and percentile/profile separation between the dark feature and its immediate surroundings, in matched ROIs on the same parent. Background RMS/MAD in matched empty-sky ROIs, and a difference image showing where the transform actually acted. Edge/ring/overshoot measurements near sharp boundaries, and shadow clipping fraction / fraction pinned to zero. Preservation of bright/midtone regions outside the intended mask, and color differences/chroma drift on RGB parents. Effective post-adaptation boost/scales/gamma, distinct from candidate ID, preserving calibrated treatment identity.	The intended dark structure gains local contrast/depth against its immediate surroundings without noise/mottle amplification, dark ringing, unnatural black channels, or seam/gradient exaggeration, checked against a real unchanged control. Effective post-adaptation parameters are recorded, not just candidate ID; every retained DSE arm must remain a distinct effective treatment or be explicitly declared a control. No claim of a fair DSE-vs-PixInsight-LHE comparison is made merely because the corrected `pi_lhe_dark` route executes. Entropy, high-frequency power, generic sharpness, whole-frame SNR-like score, or historical win rate are never presented as standalone proof of recovered real structure.

Measurements

Acceptance criteria

Local contrast across known dust-lane/adjacent-signal profiles, and percentile/profile separation between the dark feature and its immediate surroundings, in matched ROIs on the same parent.
Background RMS/MAD in matched empty-sky ROIs, and a difference image showing where the transform actually acted.
Edge/ring/overshoot measurements near sharp boundaries, and shadow clipping fraction / fraction pinned to zero.
Preservation of bright/midtone regions outside the intended mask, and color differences/chroma drift on RGB parents.
Effective post-adaptation boost/scales/gamma, distinct from candidate ID, preserving calibrated treatment identity.

The intended dark structure gains local contrast/depth against its immediate surroundings without noise/mottle amplification, dark ringing, unnatural black channels, or seam/gradient exaggeration, checked against a real unchanged control.
Effective post-adaptation parameters are recorded, not just candidate ID; every retained DSE arm must remain a distinct effective treatment or be explicitly declared a control.
No claim of a fair DSE-vs-PixInsight-LHE comparison is made merely because the corrected `pi_lhe_dark` route executes.
Entropy, high-frequency power, generic sharpness, whole-frame SNR-like score, or historical win rate are never presented as standalone proof of recovered real structure.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	Increased local contrast/depth of the intended dark structure with midtone/highlight regions outside the intended mask, and background continuity, preserved.	nova-dse-source, nova-pi-lhe-dark-dispatch-gap
pixinsight	Functional alternative	An LHE output produced through the existing PixInsight dispatch branch.	nova-pi-lhe-dark-dispatch-gap
saspro	Same engine, adapted host	Multiscale negative-detail enhancement on the Lab luminance channel, converted back to RGB.	saspro-wavescale-dse-upstream, nova-dse-source
siril	Functional alternative	Not applicable to NOVA's current dark-structure-enhancement dispatch; documented for completeness only.	nova-dse-source

NOVA Python pipeline SASpro WaveScale DSE wrapper + external luminance-mask blend; shadow-SNR-heuristic adaptation

Record effective post-adaptation boost/scales/gamma and whether the external luminance mask actually applied -- candidate ID alone is not sufficient; preserve the post-adaptation treatment record.
Treat `pi_lhe_dark` as an executable strategy alternative, not as proof that DSE and LHE were fairly compared.
Do not silently treat the adapter's wider ceiling (up to 12.0) as validated guidance; issue #722 calibrated 3-10.

Controls and starting ranges

dse_whisper / dse_gentle / dse_mild: declared boost 3.0 / 3.5 / 5.0 -- calibrated parameter probes
dse_standard: declared boost 7.5 -- stronger calibrated probe
pi_lhe_dark: PixInsight LHE alternative; lhe_amount 0.4
none: true control

Mask behavior: Native SASpro internal darkness mask plus an external NOVA-computed luminance-mask blend layered on top -- both are part of the effective treatment.

Failure modes

Noise/mottle amplified as apparent structure.
Dark ringing around stars or high-contrast edges.
Duplicated enhancement from the internal SASpro darkness mask plus the external NOVA mask stacking.
Different appearance at native vs. drizzled sampling despite identical nominal `n_scales`.

Recovery

Revert to the parent/no-op; lower boost; reduce affected scales; change or soften the mask; fix the actual upstream defect rather than masking it with DSE.

PixInsight LocalHistogramEqualization LocalHistogramEqualization (LHE) as an executable strategy alternative

LHE can pursue a similar perceptual goal through a different local-histogram mechanism, but P24 (Local Contrast) owns LHE's own semantics -- do not duplicate that research here.
Do not reinterpret historical failed `pi_lhe_dark` attempts as DSE-vs-LHE evidence after the routing correction.

Controls and starting ranges

ontology value: lhe_amount 0.4 -- the only control consumed by the existing wrapper

Mask behavior: The existing PixInsight LHE path uses its background anchor; no unconsumed kernel/slope controls are declared.

Failure modes

Treating the corrected route as proof of a fair DSE-vs-LHE comparison.

Recovery

Use P24's own LHE tool guidance for a real PixInsight local-contrast comparison instead.

Seti Astro Suite Pro WaveScale DSE WaveScale DSE (compute_wavescale_dse); project pin 1.18.0

Current upstream source performs an a-trous multiscale decomposition with a darkness mask from negative wavelet-plane coefficients on mid-scales, two iterations with exponentially decaying scale weight, Lab-luminance-only processing on RGB -- treat as version-bound context, not confirmed proof of the pinned 1.18.0 install's exact numeric path.
Remember NOVA's shared FITS loader performs a global min/max normalization to [0,1] before this transform runs.

Controls and starting ranges

n_scales: 3-8, ontology default 6 -- pixel count, not an angular unit
boost_factor: calibrated ontology 3.0-10.0, default 5.0
mask_gamma: NOVA adaptive ~1.1 + SNR/80, clamped 1.0-1.6

Mask behavior: Internal darkness mask built from wavelet-plane negative coefficients; NOVA layers its own external luminance mask on top.

Failure modes

Presenting the adapter's wider computed ceiling (up to 12.0) as validated public guidance instead of the documented 1.0-5.0 range.

Recovery

Use the conservative documented range; verify pinned 1.18.0 source before presenting implementation details beyond the wrapper as production-runtime fact.

Siril (context only for this family) no dedicated multiscale dark-structure/DSE-equivalent tool independently confirmed in this repo

No packaged Siril tool matching WaveScale DSE's specific multiscale darkness-mask mechanism was confirmed in this repo's research -- Siril's general local-contrast/curve tools pursue a related perceptual goal by a different mechanism.

Controls and starting ranges

availability: not currently NOVA's dispatch path for this step

Mask behavior: Not applicable to NOVA's current dispatch for this step.

Failure modes

Assuming a general Siril local-contrast adjustment is evidence about NOVA's DSE candidates.

Recovery

N/A -- not NOVA's current dispatch path for this family.

Validation status

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

Sources

Phase-2 research packet P30 (Part A: Dark-Structure Enhancement), 2026-08-28 packet:P30:2026-08-28 — NOVA source-confirmed
NOVA dark_enhance() implementation and calibrated ontology candidates nas_server/seti_astro.py:dark_enhance; nas_server/tool_params.py:compute_dark_enhance; nas_server/processing_ontology.json dark_enhance; nas_server/experiments.py _STEP_CFG — NOVA source-confirmed
Issue #722 pi_lhe_dark dispatch correction and M51 DSE calibration nas_server/experiments.py:_run_variant (fn_name == "lhe" branch); nas_server/processing_ontology.json dark_enhance pi_lhe_dark — NOVA source-confirmed
Seti Astro Suite Pro current upstream WaveScale DSE source (version-bound, not the pinned install) setiastro/setiastrosuitepro src/setiastro/saspro/wavescalede.py, ref c2c941b88335c82cc42ced4a2016c43261610918 — 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.