TotalVariationImageFiltering.jl

TotalVariationImageFiltering.jl is a Julia package for total-variation (TV) denoising and reconstruction on N-dimensional arrays.

Original implementation: GPUFilter.jl

Documentation

Full manual: https://urlicht.github.io/TotalVariationImageFiltering.jl/

Recommended entry points:

• Quick Start
• Problem & API
• ROF Solver
• PDHG Solver
• Lambda Selection
• Batch & CUDA
• API Reference

Features

• ROF denoising (L2 + TV) with a Chambolle-style dual projected-gradient method
• PDHG / Chambolle-Pock for L2 + TV and Poisson KL + TV
• PDHG primal constraints: non-negativity and box constraints
• Isotropic and anisotropic TV
• Overlapping group-sparse anisotropic TV for 2D images and 3D volumes
• Neumann and periodic boundaries
• Single-image and batched solves
• Automatic lambda selection for ROF (discrepancy principle and MC-SURE)
• Optional CUDA acceleration via package extension

Installation

From the registry:

] add TotalVariationImageFiltering

From this repository:

julia --project=.

From another Julia environment (local path):

import Pkg
Pkg.develop(path="/absolute/path/to/TotalVariationImageFiltering.jl")

From a hosted repository:

import Pkg
Pkg.add(url="https://github.com/urlicht/TotalVariationImageFiltering.jl")

Minimal Example

using TotalVariationImageFiltering

f = rand(Float32, 128, 128)
problem = TotalVariationImageFiltering.TVProblem(
    f;
    lambda = 0.1f0,
    data_fidelity = TotalVariationImageFiltering.L2Fidelity(),
    tv_mode = TotalVariationImageFiltering.IsotropicTV(),
)

u, stats = TotalVariationImageFiltering.solve(problem, TotalVariationImageFiltering.ROFConfig())

Group-sparse TV uses the low-memory ADMM/MM solver:

volume = rand(Float32, 96, 96, 64)
mode = TotalVariationImageFiltering.GroupSparseTV(3)
problem = TotalVariationImageFiltering.TVProblem(
    volume;
    lambda = 0.08f0,
    tv_mode = mode,
)
u, stats = TotalVariationImageFiltering.solve(
    problem,
    TotalVariationImageFiltering.GSTVConfig(),
)

Benchmarking

Benchmark instructions and scripts are documented in benchmark/README.md.

Testing

import Pkg
Pkg.test()

CUDA tests run only when CUDA is installed and functional.

License

MIT. See LICENSE.