bm3d_core

Pure Rust implementation of the BM3D (Block-Matching and 3D filtering) denoising algorithm, optimized for streak/ring artifact removal in tomographic imaging.

Features

• Generic Float Support: Works with both f32 and f64 precision via the Bm3dFloat trait
• Streak Artifact Removal: Specialized mode for vertical streak artifacts common in neutron and X-ray imaging
• Multi-Scale Processing: Pyramid-based processing for wide streaks that single-scale cannot capture
• Fourier-SVD Method: Fast alternative algorithm combining FFT-based detection with rank-1 SVD
• High Performance: Parallelized with Rayon, optimized block matching with integral images

Installation

Add this to your Cargo.toml:

[dependencies]
bm3d_core = "0.7"

Quick Start

use bm3d_core::{bm3d_ring_artifact_removal, Bm3dConfig, RingRemovalMode};
use ndarray::Array2;

// Create a noisy 2D image (H x W)
let image: Array2<f32> = /* your image data */;

// Configure BM3D for streak removal
let config = Bm3dConfig {
    mode: RingRemovalMode::Streak,
    sigma_random: 0.1,
    patch_size: 8,
    step_size: 3,
    search_window: 39,
    max_matches: 16,
    ..Default::default()
};

// Run denoising
let denoised = bm3d_ring_artifact_removal(&image, &config);

Main API

High-Level Functions

• bm3d_ring_artifact_removal - Main entry point for ring/streak artifact removal
• multiscale_bm3d_streak_removal - Multi-scale processing for wide streaks

Configuration

• Bm3dConfig - Configuration struct for BM3D parameters
• RingRemovalMode - Generic (white noise) or Streak (directional artifacts)
• MultiscaleConfig - Configuration for multi-scale processing

Low-Level Components

• run_bm3d_kernel - Core BM3D kernel for a single image
• run_bm3d_step - Single BM3D step (hard/wiener threshold)
• estimate_noise_sigma - Automatic noise level estimation

Performance

Optimizations include:

• Integral image pre-screening for fast block matching
• Early termination in distance calculations
• Pre-computed FFT plans (Bm3dPlans)
• Fast Walsh-Hadamard transform for 8×8 patches
• Zero-overhead parallelism via Rayon

References

• Dabov, K., Foi, A., Katkovnik, V., & Egiazarian, K. (2007). Image denoising by sparse 3D transform-domain collaborative filtering. IEEE TIP.
• Mäkinen, Y., et al. (2021). Collaborative Filtering of Correlated Noise: Exact Transform-Domain Variance for Improved Shrinkage and Patch Matching.

License

MIT License - see LICENSE for details.

Related

This crate is part of the bm3dornl project, which also provides:

• Python bindings via PyO3
• GUI application for interactive denoising