surface op• データ種: zscan → depth
• 呼び出し: import interferometry; interferometry.csi_height_map(stack, z_step_um=0.05, z_start_um=0.0, wavelength_um=0.6, mode='gaussian', remove_bias=True, min_visibility=0.3, max_edge_envelope=0.05, carrier_tolerance=2.0, on_invalid='raise', fill_value=nan) (または opsinterferometry.get("csi_height_map"))
`(Z, H, W)` のコヒーレンス走査スタックからの高さマップ —— CSI の逆問題。
> 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。
The per-pixel :func:csi_peak_position, vectorised. **The scan axis is
first**; see :func:csi_stack_simulate for why that is checked rather than
inferred.
A pixel is *invalid* when any of three things is true, and all three are the
same trap in different shapes — each would otherwise yield a finite, plausible,
wrong height:
1. its envelope prominence is below *min_visibility* — no fringes ever
formed there (a hole, a dark or specular-dropout pixel, a facet tilted
out of the aperture);
2. its envelope peaks on the first or last plane — the surface is at or past
the end of the scan;
3. its envelope is still above *max_edge_envelope* of its peak at the ends
of the scan — most of the coherence peak is outside the scan even though
the argmax is on an interior plane. This is the one that has no natural
alarm: measured, a surface at 0.500 um with an edge level of 0.639 reads
0.119 um and nothing else about the result looks wrong. See
:func:_edge_level.
What happens then is a decision, not a default:
• `on_invalid="raise"` (the default) — raise, naming how many pixels
failed which of the three checks. Fail-closed: a height map with silently
wrong pixels in it is worse than no height map.
• `on_invalid="fill"` — write *fill_value* (default NaN) at those pixels
and return. Opt in to this when you intend to mask afterwards; a NaN
height poisons every downstream reduction, which is precisely why it is
not the default.
There is deliberately no third option that quietly reports the boundary plane.
Returns a float64 `(H, W)` height map, in the units of *z_start_um* /
*z_step_um*.
Ground truth (measured, 32x32 pixels, 241 planes x 0.05 um, 0.60 um
wavelength, 2.83 um envelope FWHM). On a tilted plane spanning 5.0-7.0 um,
i.e. comfortably inside a 0-12 um scan, the RMS height error is 1.42e-02 um
for `"peak", 3.81e-05 um for "centroid"`, 4.02e-06 um for
`"parabolic" and 2.08e-06 um for "gaussian"`. Widen the same plane to
2.0-10.0 um — pixels now within 2 um of the ends of the scan — and the
errors become 1.91e-02 / 5.98e-02 / 7.06e-03 / 7.06e-03 um: the local fits
lose three decades and the centroid loses four, entirely to envelope
truncation at the scan ends. Accuracy here is a property of the *scan layout*,
not of the estimator, and this is the number to look at when a real
measurement disappoints.
On that same surface, phase-shifting interferometry via :mod:fringe is exact
below a lambda/4 = 0.15 um step and wrong by exact multiples of
lambda/2 = 0.30 um above it.
Raises `ValueError`: a non-3-D stack, fewer than 3 planes, an empty
spatial extent, a stack over :data:MAX_STACK_ELEMENTS (checked before the
float64 promotion), a non-finite / complex / masked stack, an unknown *mode*
or *on_invalid*, a *min_visibility* / *max_edge_envelope* outside `[0, 1]`,
a *z_step_um* at or past the `wavelength_um/4` Nyquist ceiling, a
non-numeric *fill_value*, and — under the default `on_invalid="raise"` —
any invalid pixel.
• 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。
• coherence_scanning — py -3.11 examples/coherence_scanning.py
• poc_focus_stacking — py -3.11 examples/poc_focus_stacking.py
• poc_interferometry_step — py -3.11 examples/poc_interferometry_step.py
depth を入力に取れる)surface)*Provenance: interferometry.py — INTERFEROMETRY 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.