fullseye

anscombe_transform — PHOTON transform op

使い方

Anscombe variance-stabilising transform: Poisson counts -> ~unit-variance.

Photon-limited data has signal-dependent noise, which every classical denoiser (Gaussian, bilateral, NLM, wavelet, BM3D) assumes away. The Anscombe transform A(x) = 2*sqrt(x + 3/8) makes the variance approximately 1 independently of the signal, so the standard route is transform -> denoise with a unit-sigma Gaussian denoiser -> :func:anscombe_inverse.

The generalised form (Starck/Murtagh/Bijaoui) also absorbs the sensor’s analogue chain, and takes exactly the parameters :func:backends_aug.aug_read_noise injects::

A(x) = (2/g) * sqrt(g*(x - offset) + (3/8)*g^2 + sigma_r^2)

with gain g in ADU per photon, read_sigma the Gaussian read noise in ADU and offset the black level in ADU. The defaults g=1, sigma_r=0, offset=0 reduce it to the classical form exactly.

Measured stabilisation — var(A(X)) for X ~ Poisson(lambda), computed exactly by summing the Poisson pmf (no sampling, so anyone can reproduce these; tests/test_photoncount.py pins them and the sampled versions):

======== ======== lambda var(A) ======== ======== 1 0.717443 2 0.924297 4 0.998754 10 1.000910 100 1.000006 ======== ========

So “variance 1” is true from about 4 photons/pixel upward and false below it — at 1 photon/pixel the variance is 0.717, a 28% shortfall, which is the honest statement of the transform’s low-count limit. Below a few photons an exact Poisson method (or the exact unbiased inverse, see :func:anscombe_inverse) is required.

It does not help a linear smoother, and the tests say so. Measured on a two-level scene (4 and 64 photons/pixel, seed 5): a plain Gaussian filter applied to the raw counts reaches RMSE 2.387, and the same filter through the Anscombe route reaches 2.459 — i.e. slightly worse. That is expected: averaging is already the right thing to do to Poisson counts, so stabilising the variance first buys nothing. The transform pays off for denoisers whose parameter is an absolute noise scale — thresholds, sigma filters, wavelet shrinkage, NLM, BM3D — because that parameter becomes one constant instead of a per-pixel function. Measured with a 5x5 sigma filter at a 3-sigma threshold on that same scene: 1.191 through the transform against 2.307 in the raw domain using the same 3-sigma rule with a globally estimated sigma. (An oracle raw threshold, swept against ground truth one does not have in practice, reaches 1.080 — so the honest headline is “one principled constant instead of a tuned guess”, not “always better”.)

Returns a float64 array of the same shape as image.

Raises ValueError: non-finite image, non-positive gain, negative read_sigma, and — unless clip=True — any pixel whose argument under the square root is negative (which can happen for real read-noise data dipping below the black level). clip=True floors the argument at 0 and is the documented, opt-in behaviour; the default refuses rather than quietly manufacturing a value.

詳しい使い方ガイド

参考(サンプルデータ・文献)

実行できる例(この op を実際に呼ぶ検証済みサンプル)

型が繋がる次の op(image2d を入力に取れる)

photon_sample · photon_statistics · photon_uncertainty · anscombe_inverse

同カテゴリ(transform)

anscombe_inverse


Provenance: photoncount.py — PHOTON 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.