OpenBXT — Open-source AI Deconvolution for Astronomical Images

OpenBXT is an open-source PyTorch reimplementation of the ideas behind BlurXTerminator: an AI-based deconvolution / aberration-correction tool specifically designed for linear astronomical images.

It is not a generative super-resolution model. The goal is to recover detail that is actually present in the data (within the band-limit of the optics) without inventing structure.

Features

• U-Net with PSF-conditioned residual prediction
• Joint stellar / non-stellar handling via a learned star map
• Support for spatially-varying PSF and optical aberrations:
  • Defocus, coma, astigmatism, trefoil, spherical (Zernike-based)
  • Atmospheric seeing (Moffat/Gaussian)
  • Guiding errors (motion blur)
  • Lateral / longitudinal chromatic aberration
• Synthetic training pipeline — no real ground truth needed
• Tiled inference for arbitrarily large frames
• FITS / TIFF / PNG I/O

Installation

pip install -r requirements.txt

Quick Start

Train

python train.py \
    --data_dir /path/to/sharp/linear/images \
    --out_dir runs/openbxt_v1 \
    --epochs 200 \
    --batch_size 8

The training set should contain sharp linear (un-stretched) astronomical images. The pipeline synthesizes blurs and aberrations on-the-fly.

Infer

python infer.py \
    --input my_image.fits \
    --output my_image_sharp.fits \
    --weights runs/openbxt_v1/best.pt \
    --sharpen_nonstellar 0.9 \
    --sharpen_stellar 0.5

Project Layout

openbxt/
  model.py       — UNet, PSF encoder, dual-branch heads
  psf.py         — Moffat / Gaussian / Zernike PSFs
  aberrations.py — spatially-varying aberration synthesis
  dataset.py     — synthetic blur/aberration dataset
  star_detect.py — DAOFIND-style star detection
  losses.py      — flux-conserving + edge-preserving losses
  inference.py   — tiled inference engine
  utils.py       — FITS / TIFF / PNG IO, normalization
train.py
infer.py

Notes on faithfulness

Deconvolution is mathematically ill-posed; many sharp images can produce the same blurry input. OpenBXT mitigates the "hallucination" risk in three ways:

1. The training distribution only contains realistic astronomical blurs applied to real sharp astrophotos — the network never learns generic image priors.
2. Loss functions enforce flux conservation (total energy preserved inside aperture-like regions).
3. Output is a bounded residual added to the input, so under-confident predictions degrade gracefully toward the input rather than fabricating new structure.

License

MIT.