Hi all,
Over the past months I've been developing a custom reduction pipeline focused on background gradient removal. The motivation was simple: I wanted to stretch faint structures (galaxy halos, tidal features, IFN) more aggressively without amplifying background gradients or other artifacts.
Existing tools didn't quite get me there. I find DBE tedious and error-prone in practice — manual sample placement is a lot of work and easy to get wrong, especially around large galaxies or extended nebulosity. GraXpert and multiscale gradient tools are faster but often don't give me the result I want, and it's hard to understand why the algorithm decided what it decided. And none of them address the per-sub flat-fielding stage where many gradient problems originate.
So I built something based on the ABYSS approach from professional deep imaging surveys (Borlaff et al. 2019, A&A 621, A133): multi-pass background fitting with source-masked sky flats per sub, conservative gradient removal designed to preserve extended low-surface-brightness signal. Implementation with GPU acceleration.
The practical result: backgrounds are flat enough that aggressive stretching no longer reveals processing artifacts. Faint structures survive the background fit instead of being subtracted with it.
Comparison images — Pipeline vs PixInsight WBPP on identical subs:
📷 Coma_Abyss.jpg
📷 Coma_Wbpp.jpg
The Coma Cluster comparison shows the most striking visual difference — hundreds of cluster member galaxies that are barely visible in the WBPP output emerge cleanly in the pipeline output, along with faint diffuse signal in the cluster core consistent with intracluster light reported in deep imaging surveys.
📷 M65_Abyss.jpg
📷 M65_Wbpp.jpg
The Leo Triplet output shows similar improvements with documented structures — the NGC 3628 stellar halo and southern tidal plume are clearly visible in the pipeline output but not in WBPP at identical stretch.
Both comparisons use the same calibrated subs after identical quality filtering. Stretching applied identically. No BXT, NXT, or any post-processing on either side — what you see is the stack output directly.
Quantitative summary:
5σ detection depth: +0.7 to +1.2 mag deeper than WBPP (Gaia DR3 photometric calibration)
Background spread: 8-14× flatter
For Leo Triplet: 95% of the pipeline's additional detections have a direct PanSTARRS DR1 counterpart, confirming they're real sources rather than artifacts
Pipeline tested with pure-noise input (no spurious structures created) and mock-source injection (correct faint-source recovery)
Setup: SkyWatcher Esprit 120ED, QHY268M Mono, Bortle 5 skies. Coma: 432 ×120s subs. Leo Triplet: 376×120s L after quality filtering.
For now I'd genuinely appreciate critical feedback. Especially:
Anyone seeing issues in the comparison images I'm missing?
Suggestions for additional validation targets ?
Experiences with similar approaches?
Happy to discuss methodology in the thread.
Clear skies, Dirk
