peak_subbin — ONED signal op

資料種類:signal × indicesmeasurement

呼叫: import fullseye as fs; fs.ledger.peak_subbin(x, idx=None, mode='parabola')(要直接呼叫實作,import dsp; dsp.peak_subbin(x, idx=None, mode='parabola');從台帳取用則 ops1d.get("peak_subbin"))

用法

> 該運算子的說明尚無譯文,以下照原文給出。

Peak position between samples — the vertex of a fit through 3 points.

:func:find_peaks returns integer indices, so the position of a spectral

line or a correlation peak is quantised to the bin grid no matter how

finely the peak itself is resolved. This refines each index to a float by

fitting the sample and its two neighbours.

*mode* picks what is fitted:

`"parabola"`

`a x^2 + b x + c` through the three raw values. The classic estimator;

exact for a genuinely parabolic top and cheap.

`"gauss"`

the same parabola through `log` of the three values, which is exact for

a Gaussian peak. Requires all three values positive — a magnitude

spectrum qualifies, a signed correlation does not. Non-positive

neighbourhoods raise `ValueError` rather than silently returning the

integer index, because "the refinement quietly did nothing" is the

failure this operator exists to remove.

Measured on a Gaussian of width 1.7 bins centred at 40.37: the integer

argmax gives 40 (error 0.37 bins), `parabola` gives 40.3553 (error

0.0147) and `gauss` gives 40.3700 (error 0, to machine precision).

**The parabola is biased on a Gaussian, and the bias grows as the peak gets

narrower** — the same measurement at widths 6.0 / 3.0 / 1.7 / 1.0 bins errs

by 0.0012 / 0.0047 / 0.0147 / 0.0434. That is the whole reason `gauss`

exists: for a spectral line the log-parabola is not an approximation, it is

the exact model.

The refinement is still an assumption about shape, not a measurement of it.

On a triangular peak, whose top is not smooth, `parabola` errs by 0.0857

bins where the integer index errs by 0.30 — better, but three times worse

than on the Gaussian it was designed for.

The shift is not clamped. A vertex more than half a sample away from the

index means the index was not a local maximum in the first place, and that is

information: clamping it to +/-0.5 would return a plausible number for a bin

that has no peak in it. Run :func:find_peaks first, or clamp deliberately

at the call site when evaluating bins that are not maxima (a harmonic comb,

for instance).

A peak sitting on the first or last sample has no neighbour on one side and

keeps its integer position (there is nothing to interpolate against); the

returned array says so by being exactly equal to the input index there.

Parameters

----------

x : array_like

The 1-D signal the peaks were found in.

idx : int, sequence of int, or None

Peak indices. `None means "the argmax", so peak_subbin(mag)` is the

one-liner for a single line.

mode : {"parabola", "gauss"}

Returns

-------

float or ndarray

Refined position(s) in samples. Scalar in, scalar out.

See also

--------

find_peaks : which indices to refine.

參考(範例資料・文獻)

• 範例資料目錄(下載 URL / 授權) —— 2-D 用 skimage.data(BSD/公有領域)加合成圖,3-D 給出真實資料源(Stanford/PDS 等)的下載 URL。

• 運算子來歷與參考文獻 —— 該運算子族所依據的研究/方法出處。

• 演算法的正典(作者・年份)與用途見上面的族使用指南

可執行的範例(實際呼叫該運算子並已驗證的樣例)

poc_multibeam_bathymetrypy -3.11 examples/poc_multibeam_bathymetry.py

poc_print_registrationpy -3.11 examples/poc_print_registration.py

poc_web_roll_periodicitypy -3.11 examples/poc_web_roll_periodicity.py

型別可銜接的下一個運算子(可接受 measurement 作為輸入)

同類別(signal)

lowpass · highpass · bandpass · envelope · rms · local_std · quantize · companding_mu_law


*Provenance: dsp.py — ONED 運算子登記表。本條目由 tools/opdocs.py md 自動產生(請勿手動編輯)。*

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