BM3D ORNL

BM3D ORNL removes streak and ring artifacts from neutron-imaging sinograms. BM3D means block-matching and three-dimensional filtering.

The BM3D algorithm was originally proposed by K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian in the paper "Image Denoising by Sparse 3D Transform-Domain Collaborative Filtering" (2007).

The library provides a Python API backed by Rust. Key features include:

• Streak removal targets vertical detector artifacts.
• Multiscale processing handles streaks wider than one scale can capture.
• Fourier-SVD combines frequency detection with rank-one matrix decomposition.
• Stack processing batches three-dimensional sinogram data.
• Rust processing accelerates block matching and transforms.

Full documentation: https://bm3dornl.readthedocs.io

It covers the installation guide, a tutorial, the parameter reference, the GUI guide, and the API reference.

Use pip for published releases and Pixi for a clone.

Do not run pip install -e . in a clone. It is unsupported: Pixi pins every dependency in pixi.lock, and pip resolves its own set. pixi run build compiles the Rust extension and installs the package in editable mode.

The [gui] extra installs the separate bm3dornl-gui binary wheel from PyPI. It does not build the GUI in a clone. Run the clone's GUI with pixi run gui.

Using Pip (released packages)

# Core library only
pip install bm3dornl

# With GUI application
pip install bm3dornl[gui]

Supported Platforms

from bm3dornl import bm3d_ring_artifact_removal
import numpy as np

# Load sinogram data - 2D (H, W) or 3D stack (N, H, W)
sinogram = np.load("sinogram.npy")

# Standard BM3D denoising (generic white noise)
denoised = bm3d_ring_artifact_removal(sinogram, mode="generic", sigma_random=0.1)

# Streak artifact removal (recommended for ring artifacts)
denoised = bm3d_ring_artifact_removal(sinogram, mode="streak", sigma_random=0.1)

# With custom parameters (all parameters are flat, no dict wrapping)
denoised = bm3d_ring_artifact_removal(
    sinogram,
    mode="streak",
    sigma_random=0.1,
    patch_size=8,           # Patch size (7 or 8 recommended)
    step_size=4,            # Step size for patch extraction
    search_window=40,       # Max search distance
    max_matches=64,         # Similar patches per 3D group
    batch_size=32,          # Batch size for stack processing
)

# Multi-scale BM3D for wide streaks (v0.7.0+)
denoised = bm3d_ring_artifact_removal(
    sinogram,
    mode="streak",
    multiscale=True,        # Enable multi-scale pyramid processing
    num_scales=None,        # Auto-detect (or set explicitly)
    filter_strength=1.0,    # Filtering intensity multiplier
)

For an alternative streak-removal method:

from bm3dornl.fourier_svd import fourier_svd_removal

# Fourier-guided streak removal
denoised = fourier_svd_removal(
    sinogram,
    fft_alpha=1.0,          # Weight for detected vertical-frequency energy
    notch_width=2.0,        # Gaussian notch width in frequency bins
)

Processing time for one 720×725 sinogram, mean of 100 runs (measured 2026-08-31 and 2026-09-02):

The Rust backend runs in parallel across cores with Rayon. The benchmark study behind these numbers, including the TomoPy comparison, lives on the feature/joss-paper branch under study/. A smaller comparison notebook is notebooks/evaluation_performance.ipynb.

Key optimizations:

• Integral image pre-screening for fast block matching
• Early termination in distance calculations
• Pre-computed FFT plans
• Fast Walsh-Hadamard transform for 8×8 patches

Use Pixi for every command in a clone. See How to install for the install policy. Pixi supplies Python, Maturin, HDF5, and the other build dependencies. It does not supply Rust: install a stable Rust toolchain with rustup first, so that cargo is on your PATH.

Clone the repository, enter its directory, then run pixi run build, pixi run test, and pixi run gui, in that order.

# Prerequisites: Pixi and a stable Rust toolchain (rustup)
git clone https://github.com/ornlneutronimaging/bm3dornl.git
cd bm3dornl
pixi run build
pixi run test
pixi run gui

Each pixi run command installs the environment on demand. Run pixi install first only when you want to create the environment separately.

The first pixi run gui compiles the GUI application, which takes a few minutes; later runs start immediately.

tests/bm3dornl-data is a Git submodule with notebook reference data. It is not required to build the package or run tests. The clone command above is enough for pixi run build and pixi run test.

Fetch it only if you want to run the notebooks:

git submodule update --init tests/bm3dornl-data
git -C tests/bm3dornl-data lfs pull

The submodule is hosted on code.ornl.gov. It clones anonymously over HTTPS, without an ORNL account or SSH key. The tomostack_small.h5 file is about 1.4 GB. It uses Git Large File Storage (Git LFS). Without Git LFS, the file remains a small text pointer. The lfs pull step can take time.

BM3DORNL includes a standalone GUI application for interactive ring artifact removal.

Released binaries (Linux x86_64 and macOS Apple Silicon):

pip install bm3dornl[gui]

Or install the GUI separately:

pip install bm3dornl-gui

In a clone, run the GUI through Pixi. See How to install.

pixi run gui

The released binary installs a bm3dornl-gui command:

bm3dornl-gui

• Load HDF5 files with tree browser for dataset selection
• Interactive slice viewer with histogram
• Side-by-side comparison of original and processed images
• Adjust parameters before processing
• View the processed result after processing finishes
• ROI selection for histogram (Shift+drag to select region)
• Export processed data to TIFF or HDF5

• Contributing: see CONTRIBUTING.md for setup, coding standards, testing, and pull requests. Participation follows our Code of Conduct.
• Reporting issues: open an issue at https://github.com/ornlneutronimaging/bm3dornl/issues. Please include your platform, the bm3dornl version, and a minimal reproduction.
• Getting support: for usage questions, start a GitHub Discussion or email the maintainer at zhangc@ornl.gov.