multicolorfits.deblend_background#

multicolorfits.deblend_background(rgb, background='white', gamma=1.0, tol=0.0058823529411764705, alpha_smooth=0.0, alpha_lo=None, alpha_hi=None, alpha_gamma=1.0, fill='background', matte='none', soft_edge=0.0, crop='none', pad=0.05, size=None)[source]#

Recover an RGBA foreground from a composite flattened onto a known solid background color (un-premultiply) — the inverse of flatten_rgba() and the principled sibling of make_transparent_cutout().

Solves C = alpha * F + (1 - alpha) * B per pixel for the minimum alpha whose foreground F stays inside the RGB gamut (closed form; see _deblend_alpha()). Works for any background color: black reduces to max(C), white to max(1 - C), and dark features on a light background survive.

With the default settings the reconstruction is exact: flattening the returned RGBA back onto background (same gamma) reproduces the input. The matte-shaping knobs below share the vocabulary — and the implementation — of make_transparent_cutout(); using them trades exactness for presentation control.

Parameters:

  • rgb (array (ny, nx, 3)) – Flattened composite, float [0, 1] (e.g. a JPEG/PNG loaded and scaled).

  • background (color) – The known solid background it was flattened onto (any matplotlib color spec).

  • gamma (float) – Encoding gamma assumed for the blend. 1.0 (default) treats stored values as blend-space; use e.g. 2.2 if the composite was blended in linear light.

  • tol (float) – Snap alpha to exactly 0 where the pixel is within tol of the background in every channel (default 1.5/255). This absorbs 8-bit quantization / JPEG noise that would otherwise leave a faint near-transparent haze over the whole frame.

  • alpha_smooth (float) – Gaussian sigma (pixels) applied to the recovered alpha map before the foreground is re-solved (gamut-clipped) — same trick as in make_transparent_cutout(): mild smoothing trades exact reconstruction near edges for a less noisy matte. Requires scipy when > 0.

  • alpha_lo (float or None) – Optional percentile ramp re-applied to the recovered alpha, with the same meaning as in make_transparent_cutout(): alpha_hi is the full-opacity threshold, fading to transparent at alpha_lo. None (default) keeps the physical alpha.

  • alpha_hi (float or None) – Optional percentile ramp re-applied to the recovered alpha, with the same meaning as in make_transparent_cutout(): alpha_hi is the full-opacity threshold, fading to transparent at alpha_lo. None (default) keeps the physical alpha.

  • alpha_gamma (float) – Fade shape (> 0). With a ramp, shapes it exactly as in make_transparent_cutout(); without one, applied as a plain power on the physical alpha (< 1 bolder body, > 1 tighter matte).

  • fill ({'background', 'black', 'white'} or color) – Foreground color to store where alpha == 0 (invisible; affects only later edits). Default keeps the background color there.

  • matte ({'none', 'circle', 'ellipse'}) – Optional geometric window (same as make_transparent_cutout()).

  • soft_edge (float) – Gaussian feather (pixels) on the finished alpha (same as make_transparent_cutout()).

  • crop ({'none', 'auto'} or False) – 'auto' trims to the non-transparent bbox (+ pad). Default 'none' — deblending is a recovery operation, so the frame is preserved unless cropping is requested.

  • pad (float) – Fractional padding around the auto-crop bbox.

  • size (int or None) – Downsample so the longest side is at most this many pixels.

Returns:

array (ny, nx, 4) float in [0, 1]

Notes

The problem is underdetermined (4 unknowns, 3 equations); minimum alpha is the standard, maximally-transparent convention (as used by color-keying / un-premultiply tools). The recovered F touches the gamut boundary in at least one channel wherever 0 < alpha < 1 — that is inherent, not a bug.