BM3D ORNL

A high-performance BM3D denoising library for neutron imaging, optimized for streak/ring artifact removal from sinograms.

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).

BM3D ORNL provides a Python API with a Rust backend for efficient, parallel processing of tomography data. Key features:

• Streak/Ring Artifact Removal: Specialized mode for removing vertical streak artifacts common in neutron and X-ray imaging
• Multi-Scale Processing: True multi-scale BM3D for handling wide streaks that single-scale cannot capture (based on Mäkinen et al. 2021)
• Fourier-SVD Method: Alternative fast algorithm (~2.6x faster than BM3D) combining FFT-based energy detection with rank-1 SVD
• Stack Processing: Efficient batched processing of 3D sinogram stacks
• High Performance: Rust backend with optimized block matching (integral images, early termination) and transforms (Hadamard, FFT)

How to install

Using Pip

# Core library only
pip install bm3dornl

# With GUI application
pip install bm3dornl[gui]

Supported Platforms

Platform	Architecture	Library	GUI
Linux	x86_64	✅	✅
macOS	ARM64 (Apple Silicon)	✅	✅

Using Pixi (Development)

pixi install
pixi run build

Usage

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
)

Fourier-SVD Method (v0.7.0+)

For faster processing with excellent results on many datasets:

from bm3dornl.fourier_svd import fourier_svd_removal

# Fast streak removal (~2.6x faster than BM3D)
denoised = fourier_svd_removal(
    sinogram,
    fft_alpha=1.0,          # FFT-guided trust factor (0.0 disables FFT guidance)
    notch_width=2.0,        # Gaussian notch filter width
)

Performance

The Rust backend provides high performance for tomography stacks:

Metric	Value
Speed	~0.63s per frame (512×512) on Apple Silicon
Memory	>50% reduction via chunked processing
Parallelism	Zero-overhead parallel processing via Rayon

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

Development

We use pixi for development environment management.

1. Clone repo.
2. Run pixi run build to compile the Rust backend and install in editable mode.
3. Run pixi run test to run tests.
4. Run pixi run bench to run performance benchmarks.

git clone https://github.com/ornlneutronimaging/bm3dornl.git
cd bm3dornl
pixi run build
pixi run test

GUI Application

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

Installation

pip install bm3dornl[gui]

Or install the GUI separately:

pip install bm3dornl-gui

Launching

bm3dornl-gui

Features

• Load HDF5 files with tree browser for dataset selection
• Interactive slice viewer with histogram
• Side-by-side comparison of original and processed images
• Real-time parameter adjustment
• ROI selection for histogram (Shift+drag to select region)
• Export processed data to TIFF or HDF5

Keyboard Shortcuts

Shortcut	Action
Shift+Drag	Select ROI for histogram
Scroll	Zoom in/out on image
Drag	Pan image

Parameter Reference

Parameter	Default	Description
mode	"streak"	"generic" for white noise, "streak" for ring artifacts
sigma_random	0.1	Noise standard deviation
patch_size	8	Patch size (7 or 8 recommended)
step_size	3	Step size for patch extraction
search_window	39	Maximum search distance for similar patches
max_matches	16	Maximum similar patches per 3D group
batch_size	32	Batch size for stack processing
streak_sigma_smooth	1.0	Smoothing for streak mode (streak mode only)
multiscale	False	Enable multi-scale processing for wide streaks
num_scales	None	Number of scales (auto-detected if None)
filter_strength	1.0	Filtering strength multiplier for multi-scale
debin_iterations	30	Debinning iterations for multi-scale

Fourier-SVD Parameters

Parameter	Default	Description
fft_alpha	1.0	FFT-guided trust factor (0.0 disables FFT guidance)
notch_width	2.0	Gaussian notch filter width in frequency domain