imaging opimage2d → pairsimport fullseye as fs; fs.ledger.psf_to_mtf(psf, pixel_pitch_um=1.0) (実装を直接呼ぶなら import optics; optics.psf_to_mtf(psf, pixel_pitch_um=1.0)、台帳から引くなら opsoptics.get("psf_to_mtf"))Radially-averaged MTF of a measured point-spread function.
OTF = FFT{PSF}, MTF = |OTF| / |OTF(0)|, then averaged over annuli of
constant spatial frequency out to the Nyquist limit 1/(2*pitch). This
is the measurement side of resolution: image a point source (or a slit, or
differentiate a knife edge), hand the spot here, and compare the curve with
the diffraction limit from :func:mtf_diffraction.
Returns an (n, 2) float64 pairs array: column 0 the spatial
frequency in cycles per millimetre, column 1 the MTF in [0, 1]. One row
per non-empty radial bin (a very anisotropic array can leave a bin empty;
those rows are dropped rather than filled with a NaN).
Ground truth it reproduces (measured): a delta PSF gives MTF == 1 at every
frequency exactly (max deviation 0.0); a Gaussian PSF of sigma pixels
gives the closed form exp(-2*pi^2*sigma^2*f^2) — the maximum absolute
deviation over the whole curve is 4.1e-4 at sigma = 2 px on 128x128,
8.3e-4 at sigma = 1.5 px on 64x64 and 2.4e-4 at sigma = 3 px on 256x256
(the residual is the radial average over a square grid, not an error in the
transform). Doubling pixel_pitch_um halves every reported frequency and
leaves the MTF column bit-identical.
The PSF is not re-normalised or re-centred: a PSF whose energy is not centred carries a linear phase, which the modulus discards, so the MTF is unaffected — but the phase transfer function, which is where a decentred/asymmetric PSF shows up, is deliberately not summarised here.
Raises ValueError: psf is not 2-D / smaller than 2x2 / over the
size cap / complex / masked / non-finite; a PSF that sums to zero or less
(the DC normalisation would be 0/0 — an all-zero “PSF” is not a PSF);
non-positive or non-finite pixel_pitch_um.
optics の全 op は入力を検証してから計算する(黙って通さない):
_mm / _um / _deg / _mrad。mm と µm の取り違えは crash ではなく「もっともらしく間違った答え」なので、名前で防ぐ。大きさから単位を推測する処理は一切しない。ValueError — float('50') は成功してしまうため、未パースの設定値が長さとして通り抜ける(実測: thin_lens('50', '200') がもっともらしい 66.667 mm を返していた)。bool も True == 1 の暗黙昇格として拒否。ValueError(実数枠のみ。虚部の無言切り捨て・マスク剥がしを拒否)。NaN/Inf は全入力で ValueError。depth_of_field の過焦点距離以遠の far_mm = inf(それが過焦点距離の定義)と gaussian_beam のウエストでの wavefront_radius_mm = inf(平面波面の曲率半径)。どちらも有限の相棒(far_is_infinite / curvature_per_mm)を併せて返す。それ以外の無言 NaN/Inf は内部で検出して ValueError —「float64 が溢れた」と「答えが無限大」は別の主張なので、後者の顔で前者を返さない。optics.MAX_GRID(4096)、供給された場/PSF/開口は optics.MAX_FIELD_ELEMENTS(2^24)、ABCD 素子列は optics.MAX_SYSTEM_ELEMENTS(1024)、Zernike は MAX_ZERNIKE_TERMS(512)/ MAX_ZERNIKE_ORDER(40)/ MAX_ZERNIKE_BASIS(2^25)。小さな引数から巨大な内部確保が起きる経路(実測: n_max=40 × 4096² で 108 GB)を fail-closed で塞ぐ。| 物理的に不可能な状態も拒否: 偏光度 > 1 の Stokes ベクトル、負の透過率、負の強度、n- | m | が奇数などの不正な Zernike 添字。 |
py -3.11 examples/optics_imaging.pypy -3.11 examples/poc_veiling_glare.pypairs を入力に取れる)—
imaging)mtf_diffraction · wavefront_stats
Provenance: optics.py — OPTICS 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.