color op• Datenarten: qimage → qimage
• Aufruf: import quatimage; quatimage.quat_color_filter(qimage, direction_rgb, mode) -> 'np.ndarray' (oder opsquat.get("quat_color_filter"))
Behält oder entfernt exakt eine Farbrichtung. → (H, W, 4).
> Die ausführliche Beschreibung unten ist der Originaltext — Zusammenfassung und Überschriften sind übersetzt.
`mode="remove" returns v - (v.g) g for the unit RGB direction g`:
the component along `g` is exactly zero everywhere afterwards, to
machine precision (measured max residual 5.8e-16 and 6.5e-16 on two random
colour images — seed-dependent only at the 1e-16 level),
and `remove + keep` reproduces the input to 0.0 exactly.
`mode="keep" returns the complementary (v.g) g`. The scalar part is
passed through untouched in both.
There is no default mode. The two are opposites, both return a valid
picture, and neither raises — so choosing for the caller would be a coin flip
that never announces itself.
Not a new algorithm, and this docstring will not pretend otherwise
-----------------------------------------------------------------
The `remove` branch is the specular-invariant projection of Mallick et
al. (2005), which this repository already implements as
`specularity.specular_free_transform for the rgbimage` sort. Rather
than write the same three lines twice, this operator *delegates* to it — so
agreement between the two sorts is by construction rather than by luck, and a
future fix in one is a fix in both. What is added here is the `keep`
branch (which has no counterpart there) and the `qimage` sort, so the
projection composes with :func:quat_color_rotate and :func:qft2.
What this can do that a channelwise pipeline cannot
---------------------------------------------------
A per-channel filter applies a diagonal matrix, and `I - g g^T` is diagonal
only when `g is a coordinate axis. For g = (1,1,1)/sqrt(3)` — remove
the grey axis, i.e. keep only chromatic content — the *best possible*
diagonal approximation is off by `||P - diag(P)||_2 = 0.666667` in operator
norm. Concretely, a pure red pixel `(1, 0, 0)` must become
`(0.666667, -0.333333, -0.333333)`; the best diagonal filter can only reach
`(0.666667, 0, 0), an error of 0.471405` — it cannot put anything into
the green and blue channels, because it multiplies each channel by a number
and both start at zero. The impossibility is structural, not a tuning gap.
(A full 3x3 colour matrix, of course, does it exactly; see
:func:quat_color_rotate for that half of the accounting.)
Raises `ValueError: *qimage* is not a valid (H, W, 4)` field;
*direction_rgb* is not a finite non-zero 3-vector; *mode* is not
`'remove' / 'keep'`.
• Leitfaden zur Familie quaternion_monogenic
• Katalog der Beispieldaten (Download-URLs / Lizenzen) — 2-D nutzt skimage.data (BSD/Public Domain) plus synthetische Bilder, 3-D nennt Download-URLs echter Datenquellen (Stanford, PDS, …).
• Herkunft und Literatur der Operatoren — die Quellen der Forschung/Verfahren, auf denen diese Operatorfamilie beruht.
• Der kanonische Algorithmus (Autor, Jahr) und seine Anwendungen stehen im Familienleitfaden oben.
• quaternion_monogenic — py -3.11 examples/quaternion_monogenic.py
qimage als Eingabe)quaternion_to_rgb · quat_norm · quat_conjugate_image · quat_normalize_image · quat_image_multiply · monogenic_amplitude · monogenic_phase · monogenic_orientation
color)*Provenance: quatimage.py — QUAT Operator-Registry. Diese Notiz wird von tools/opdocs.py md erzeugt (nicht von Hand bearbeiten).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.