typed op• データ種: signal → feature
• 呼び出し: fullseye.apply(img, "tb_equivalent_level", a=0.5, b=0.5) (2-D は 1 画像 + 2 スカラつまみ a,b∈[0,1] のモデル)
*図なし: この op は signal を入力に取る。画像から始まる Studio のプログラムでは型が届かないので、下の「実行できる例」で使い方を見ること。*
記録のエネルギー等価レベル（`ref` を基準とするdB値）。
> 以下の詳細説明は原文のままです —— 要約と見出しは訳出済み。
`L_eq = 10 log10(mean(x_w**2) / ref**2) where x_w` is the signal after
the chosen weighting. Returns a plain float.
The reference is yours to supply. The default `ref=1.0` means dB
relative to one unit of the signal's own units; it is not dB SPL, because
this library never sees your calibration. Pass `ref=20e-6` for pascals.
Measured: a 1 kHz sine of amplitude 1.0 at 16 kHz over exactly 1000 periods
gives `L_eq = -3.010300` dB with Z weighting, against the closed form
`10*log10(1/2) = -3.010300` (difference 2.2e-15 dB), and the same value
under A weighting (difference 8.9e-16 dB), because A is 0 dB at 1 kHz.
Doubling the amplitude adds 6.020600 dB. Silence returns -200.0.
Silence returns `floor_db (default -200) rather than -inf; an -inf`
in a list of levels destroys every average taken over it afterwards.
**A weighted level is only as good as the weighting, and the weighting has a
leakage limit this operator inherits in full.** A pure tone that is not a
whole number of periods in the record comes back too loud — measured
+7.7986 dB at 31.5 Hz (a nominal one-third-octave centre) over 0.5 s at
48 kHz, and up to +17.2116 dB at 20.5 Hz — with no exception, no NaN and
no warning. `weighting="Z" is exempt (it does no filtering) and "C"` is
nearly so (+0.0493 dB on the same tone); it is `"A"`, whose curve spans
about 40 dB across the audio band, that is exposed.
`window="hann"` is the opt-in remedy: the record is multiplied by a Hann
window and the mean square divided by the window's own mean square, which
suppresses the leakage almost entirely — the 31.5 Hz error goes from
+7.7986 to +0.0534 dB and the 20.5 Hz error from +17.2116 to
+0.1841 — while costing about 0.15 dB on records that were exact
before. It is not the default, and must not be used when a transient's
level is the point: a window makes the answer depend on *where in the
record the sound happened*. Measured on a 50 ms burst inside a 0.5 s record
(Z weighting, so only the window acts; unwindowed all three are -13.0103 dB
as they must be): at the start -36.0587, at the centre -8.8218, at
the end -36.0587 — a 27 dB spread produced by nothing but position.
Use `"hann"` for a stationary tonal record, which is exactly the case the
leakage ruins, and `"none"` (the default, an honest energy average) for
everything else. The full measurement is in :func:apply_weighting.
Raises `ValueError: everything :func:_as_signal` refuses, an unknown
`weighting or window, ref <= 0` (a decibel needs a positive reference; a zero
reference makes every level `+inf` and a negative one makes the ratio
negative), `rate <= 0`.
Typed bridge of the acoustics op `equivalent_level into the 2-D evolution registry: the same implementation, called under the op(v, a, b) convention. a drives ref (default 1); b` is unused.
• 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。
• (まだありません)
feature を入力に取れる)typed)tb_points_to_voxel · tb_estimate_point_normals · tb_iss_keypoints · tb_angle_3points · tb_project_points · tb_render_point_depth · tb_statistical_outlier_removal · tb_radius_outlier_removal
*Provenance: ops.py — 2D 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.