ndsafir 2026.1014a0

image denoising https://doi.org/10.1109/TMI.2009.2033991

ndsafir v2026.1014-alpha

Denoising of multi-dimensional microscopy images.

Command line

Compilation

Install dependencies on ubuntu or debian with

sudo apt update
sudo apt install build-essential cmake libtiff-dev

To compile the code with a traditional makefile:

cd src
make release

To compile the code with CMake:

mkdir build
cd build
cmake ../src
make

Usage

Type ndsafir -h for help.

ndsafir 2026.1004-alpha - https://doi.org/10.1109/TMI.2009.2033991
    -i               0                        input (TIFF file)
    -o               0                        output (TIFF file)
    -patch           3,3,0,0,0                patch radius x,y,z,c,t
    -3D              true                     use 3D
    -time            true                     use time
    -noise           0                        noise model: gain,offset or gain,offset,readout
    -max_iter        4                        maximum number of iterations
    -pval            0.01                     p-value (0,1)
    -nthreads        12                       number of available threads

The input and output are TIFF file and support the hyperstack from ImageJ.

The patch radius can be tuned for each axis (x,y,z,c and time).

3D and time option will activate the use of 3d and time.

The noise level (-noise) can be specified as

• the noise standard deviation for Gaussisan noise
• a triplet gain, offset, readout for Poisson-Gaussian noise
• a pair of gain and noise offset If not specified it will be estimated.

The pvalue pval determines the threshold for detecting similar block or patches in the images.

Python bindings

Installation

git clone this repo
pip install ./ndsafir

Usage in a script or notebook

from skimage import data
import ndsafir
img = data.cat()
denoised = ndsafir.denoise(img,"gaussian",noise_std=5,axes="YXC")

Usage as a command line

The module can be used as a command line tool too which allows to support more input format and use distributed computing using dask distributed.

For processing tif files:

python -m ndsafir --input input.tif --output output.tif

For processing ome-zarr with 16x256x256 chunks with 8 pixel overlap.

python -m ndsafir --input input.zarr --output output.zarr --chunk 1,1,16,256,256 --depth 0,0,8,8,8

For processing a czi file with 16,256x256 chunks with 8 pixel overlaps.

python -m ndsafir --input input.czi --output output.zarr --chunk 1,1,16,256,256 --depth 0,0,8,8,8

For processing an ome-zarr on a cluster with 12 workers (with 112 processors each):

python -m ndsafir --input input.czi --output output.zarr --chunk 1,1,16,256,256 --depth 0,0,8,8,8 --compute compute.yml

with compute.yml a file:

slurm:
   processes: 1
   cores: 112
   n_workers: 10
   memory: "64GB"
   queue: "cpu"
   walltime: '01:00:00'
   local_directory: $SLURM_SCRATCH_DIR

Benchmark

A benchmarking bash script is in the scripts folder.

cd scripts
./benchmark.sh

Results are saved in a benchmark.csv file and can be analyzed using the notebook 'Analyze Benchmark.ipynb'.

Average PSNR for several datasets:

dataset	5	10	15	20	25	30
BSD68	36.7516	32.6167	30.5264	29.1308	28.0673	27.1925
CBSD68	36.8811	32.9032	30.9022	29.563	28.5382	27.6981
Fluo4	43.8542	39.4669	36.8542	34.9359	33.4384	32.1515
Kodak24	37.5176	33.8983	31.9482	30.5841	29.528	28.6477
McMaster	38.1061	34.5695	32.5226	31.0209	29.8129	28.808
RNI15	35.3304	30.8142	28.9499	27.7385	26.8422	26.1281
RNI6	35.7485	31.5641	29.6211	28.342	27.3921	26.6005
Set12	37.2683	33.6293	31.6777	30.2513	29.0975	28.133

Reference

J. Boulanger, C. Kervrann, P. Bouthemy, P. Elbau, J. -B. Sibarita and J. Salamero, "Patch-Based Nonlocal Functional for Denoising Fluorescence Microscopy Image Sequences," in IEEE Transactions on Medical Imaging, vol. 29, no. 2, pp. 442-454, Feb. 2010, doi: 10.1109/TMI.2009.2033991.