AstroFrame/docs/ngc7000-findings.md
2026-10-03 18:29:15 -07:00

5.4 KiB
Raw Permalink Blame History

NGC 7000 preview, morphology and performance findings

The supplied sample contains 147 Seestar S30 Pro lights: 2160×3840 GRBG Bayer, BITPIX 16 with BZERO 32768, 30 seconds, gain 200 and LP filter. FITS UTC timestamps span 2026-08-17 04:33–05:58. File names use the local previous date. Originals were read without modification.

Why the previews failed

The original renderer selected one superpixel out of each 3×3 region when reducing the image, then used a bright asinh stretch based on the unreduced image's percentiles. This discarded spatial information and displayed noise prominently. The reported KEEP exposure 20260816-213449.fit reproduced the failure: median display brightness 0.596, median adjacent-pixel difference 0.125 of the full display range.

Area averaging before a noise-aware, adjustable screen transfer lowers those measurements to 0.176 and 0.051. The nebula's broad structure and stars become discernible. Nebula view further averages display samples, while Frame view retains more stellar detail. This is an inspection display, not calibration or astrophotography post-processing. It cannot create faint signal absent from an individual 30-second exposure. The combined Bayer channel was retained; the red channel alone was substantially noisier in this dataset and did not clearly improve this display experiment.

Linear, compressed thumbnails have their own cache version. Older projects create updated previews on demand, without recalculating quality or losing manual decisions. First-time reads and display preparation run off the UI thread; rapid navigation cancels obsolete loads. Slider changes reuse prepared display statistics. All display operations leave detector pixels and original FITS unchanged.

Why the two examples are REVIEW

These were flagged for stellar morphology, not cloud/nebulosity:

Local filename suffix Automatic Ellipticity Nearby reference FWHM Reference FWHM Relative flux
20260816-213449 KEEP 0.15 0.16 3.31 px 3.40 px 0.986
20260816-222249 REVIEW 0.36 0.16 4.06 px 3.42 px 1.010
20260816-224840 REVIEW 0.37 0.16 4.13 px 3.40 px 1.008

Both REVIEW examples have zero cloud-mask coverage and normal stellar flux. Multiple bright stars in correctly aligned original source crops show horizontal elongation compared with rounder stars in the KEEP exposure. That supports a morphology warning, potentially tracking-related. The measured elongation is moderate rather than a catastrophic width/shape combination, so the conservative policy routes to REVIEW. A human may exclude these based on their tolerance for elongated stars. The inspection does not establish the cause with certainty or validate every emission-nebula frame.

The new Inspect stars window reads two originals and displays identical stars side by side against a nearby representative KEEP frame. It uses source Bayer-superpixel samples, enlarged without interpolation and individually normalized in brightness. Shape can be judged directly; absolute flux must be judged from the photometry. The primary flagged evidence is also visible above the main preview.

Results and performance

The complete new NGC 7000 analysis produces 141 KEEP, 6 REVIEW, 0 REJECT, identical to the owner's previous saved 147-frame analysis. No selection thresholds were changed for this target. The detector operates on linear source measurements; it never consumes the stretched preview. This independence does not imply all decisions are accurate; independent labels and controlled Siril stacks remain necessary.

Sequential process benchmarks on the same 48, evenly spaced NGC 7000 originals:

Measurement workers Seconds Peak resident memory
1 9.04 159.5 MiB
2 4.99 248.6 MiB
4 2.94 250.0 MiB

The star-count arrays are identical. This is a roughly 3.1× improvement for the stellar/background stage, not a whole-project speedup guarantee. Four workers are selected on this 16-logical-processor PC; machines with fewer processors use fewer. Raising a memory cap alone does not speed CPU work. Pixels remain streamed in bounded batches instead of filling RAM with the dataset.

The full fresh 147-frame run, including new previews, references and forced flux, took 14.76 seconds, peaking at 319.9 MiB resident, 294.5 MiB committed. Refreshing the combined 2,067-frame project reused old measurements, added NGC 7000 and generated the new display caches; it took 67.74 seconds, peaking at 661.2 MiB resident, 634.0 MiB committed. That cached refresh is not comparable to the earlier fresh full run.

Windows verifies a 6 GiB process cap and the runtime uses a 3 GiB managed-heap cap. Physical headroom checks still reserve a quarter of RAM up to 8 GiB; the 90 GB page file is not counted as free RAM. Larger caps provide room for parallelism and future projects without allocating memory unnecessarily.

Evidence: diagnostics/workers-*.json, diagnostics/ngc7000-before.png, diagnostics/ngc7000-after.png, screenshots/ngc7000-frame.png, screenshots/ngc7000-nebula.png, screenshots/ngc7000-stars.png, and the reports under projects/NGC 7000 and projects/Sample Survey. The 39 engine checks include display/source independence, noise suppression, retained conspicuous stars, linear cache round-trip and source-fingerprint invalidation. Sequential UI checks exercise cancellation, rapid navigation and display controls.