quat_color_filter — QUAT color op

データ種: qimageqimage

呼び出し: import quatimage; quatimage.quat_color_filter(qimage, direction_rgb, mode) -> 'np.ndarray' (または opsquat.get("quat_color_filter"))

使い方

Keep or remove one colour direction, exactly. → (H, W, 4).

`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'`.

詳しい使い方ガイド

quaternion_monogenic ファミリ ガイド

参考(サンプルデータ・文献)

• サンプルデータ カタログ(DL URL / ライセンス) — 2-D は skimage.data(BSD/public)+ 合成、3-D は実データ源(Stanford/PDS 等)の DL URL。

• 演算子の来歴・参考文献 — この op 族の元になった研究/手法の出典。

• アルゴリズムの正典(著者・年)と用途は上記ファミリ使い方ガイドに記載。

実行できる例(この op を実際に呼ぶ検証済みサンプル)

quaternion_monogenicpy -3.11 examples/quaternion_monogenic.py

型が繋がる次の op(qimage を入力に取れる)

quaternion_to_rgb · quat_norm · quat_conjugate_image · quat_normalize_image · quat_image_multiply · monogenic_amplitude · monogenic_phase · monogenic_orientation

同カテゴリ(color)

quat_color_rotate


*Provenance: quatimage.py — QUAT operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.