From 1c8ca4c1fb5d6d806b5ea8f3db76139e5ea2c37c Mon Sep 17 00:00:00 2001 From: Dimitri Yatsenko Date: Fri, 14 Nov 2025 00:50:02 -0600 Subject: [PATCH 1/2] minor doc fixes --- docs/user-guide/overview.md | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/docs/user-guide/overview.md b/docs/user-guide/overview.md index 028cd50..7c273af 100644 --- a/docs/user-guide/overview.md +++ b/docs/user-guide/overview.md @@ -2,7 +2,8 @@ ## The Anscombe Transform -The Anscombe Transform is a variance-stabilizing transformation specifically designed for data with Poisson noise. In photon-limited imaging, the noise variance equals the signal mean (characteristic of Poisson statistics), which makes compression difficult because different intensity levels have different noise characteristics. +The Anscombe Transform is a variance-stabilizing transformation specifically designed for data with Poisson noise. In photon-limited imaging, the noise variance grows linearly with the signal mean (characteristic of Poisson statistics), +which makes compression difficult because different intensity levels have different noise characteristics. ### The Problem @@ -93,9 +94,9 @@ The codec is designed for speed: ### Accuracy The codec is designed to be **nearly lossless** for photon-limited data: -- Typical error: < 1 photon per pixel -- Error scales with `conversion_gain` and quantization (`beta` parameter) -- For well-chosen parameters, reconstruction error is below the noise floor +- Mean absolute error: < 0.25 noise-sigma per pixel (for beta = 0.5) +- Error scales with quantization (`beta` parameter) +- For default parameters (beta=0.5), the noise variance is increased by a ~1 % with respect to the original noise variance and no bias is introduced. ## When to Use This Codec From aaf88af8bf4f81575a8d971929a4a59328b96a79 Mon Sep 17 00:00:00 2001 From: Dimitri Yatsenko Date: Fri, 14 Nov 2025 01:07:13 -0600 Subject: [PATCH 2/2] minor doc fixes and typos --- docs/user-guide/overview.md | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/docs/user-guide/overview.md b/docs/user-guide/overview.md index 7c273af..d0ec49b 100644 --- a/docs/user-guide/overview.md +++ b/docs/user-guide/overview.md @@ -94,9 +94,9 @@ The codec is designed for speed: ### Accuracy The codec is designed to be **nearly lossless** for photon-limited data: -- Mean absolute error: < 0.25 noise-sigma per pixel (for beta = 0.5) +- Max absolute error: ~ 0.25 noise-sigma per pixel (for `beta=0.5`) - Error scales with quantization (`beta` parameter) -- For default parameters (beta=0.5), the noise variance is increased by a ~1 % with respect to the original noise variance and no bias is introduced. +- For default parameters (`beta=0.5`), the noise variance is increased by a ~1 % with respect to the original noise variance and no bias is introduced. ## When to Use This Codec @@ -110,10 +110,9 @@ The codec is designed to be **nearly lossless** for photon-limited data: ### Not Recommended ❌ -- Data without Poisson noise (e.g., pre-processed images) -- Data where camera parameters are unknown and can't be estimated -- Data with very high dynamic range (> 16-bit) -- Data that has already been normalized or transformed +- Data with non-Poisson or non-stationalry noise (e.g., pre-processed images) +- Data where detector parameters are unknown and can't be estimated +- Data that has been transformed with a non-linear function (e.g. gamma correction) ## Next Steps