Processing method
- Read every frame. No temporal sampling is used.
- Register stars. Star points, phase correlation and affine RANSAC estimate translation and small rotation. Ordinary MP4 has no WCS metadata, so point registration is used.
- Preserve the field. Transformed corners define an expanded canvas. Missing pixels are white; overlapping frame levels are matched by a robust mean.
- Track motion. A temporal median stellar background is subtracted, candidates are scored, and a linear path is fitted.
- First calibration: stationary sensor pattern. Persistent bad pixels are repaired in original sensor coordinates. A registered sky model is projected back into every source frame so stationary additive horizontal, vertical and non-separable sensor structure can be estimated and subtracted. A final temporal-median residual pass removes remaining fixed bands before a broad multiplicative flat is applied.
- Second calibration: dynamic cloud pseudoflat. Every star-aligned frame is compared with the robust temporal sky model. A sigma-24 weighted illumination field estimates broad cloud and transparency variation. The unrest index combines robust spatial variation with change from the previous frame and controls correction strength.
- Aggregate all frames. Arithmetic mean and per-pixel median are exported as PNG and RGB FITS for every sequence.
Video-derived sensor maps and pseudoflats do not replace measured dark and flat calibration frames. Cloud correction can normalise broad background and transparency changes but cannot reconstruct astronomical signal physically blocked by cloud.
AI was used as a development tool, not as an image-processing step. All processing described above uses deterministic local algorithms.