rmt-denoise

Image denoising via Random Matrix Theory — automatically pick the best generalized-covariance parameters (a, β) for every test image and reconstruct it.

Installation

pip install rmt-denoise[images]   # `images` extra pulls in Pillow for load_folder

Or from source:

git clone https://github.com/yu314-coder/rmt-denoise.git
cd rmt-denoise && pip install -e .[images]

Quick Start

GeneralizedCovDenoiser runs a single workflow: pick a folder of images, choose how many to use for training, choose which one to denoise, and the denoiser jointly optimises (a, β) via differential_evolution to maximise PSNR of that test image.

from rmt_denoise import GeneralizedCovDenoiser, add_gaussian_noise

gc = GeneralizedCovDenoiser()    # always centres + T(a, β) + color-resize
denoised, clean, noisy, names = gc.denoise_folder(
    folder      = "path/to/images",                 # folder of (H, W) images
    n_train     = 300,                              # 300 training images (0 = use all)
    test        = "s12_1.png",                      # one filename — or an integer index
    size        = (100, 100),                       # resize on load (optional)
    noise_fn    = lambda imgs: add_gaussian_noise(imgs, sigma=0.04),
)
print(gc.a, gc.beta, gc.rank, gc.psnr_test)
# 0.27  0.95  12  34.5

The chosen (a, β) are exposed as direct attributes (gc.a, gc.beta, gc.rank, gc.sigma2, gc.psnr_test); the full diagnostic dict is still in gc.info.

denoised is the full (n_train+1, H, W) reconstructed stack — the test image lives at index -1. clean and noisy are the (clean / noisy) versions of that test image, returned for convenience.

Lower-level API (already have an in-memory stack)

from rmt_denoise import GeneralizedCovDenoiser

# noisy_stack: (n, H, W)   clean_test: (H, W)
gc = GeneralizedCovDenoiser()
denoised = gc.denoise(noisy_stack, clean=clean_test, test_index=-1)

The test column at test_index drives the (a, β) search; the rest of the stack is reconstructed at the same chosen rank with the same post-processing.

Train + target convenience

# train: (n_train, H, W) noisy   test_noisy: (H, W)   test_clean: (H, W)
gc = GeneralizedCovDenoiser()
denoised_test = gc.denoise_test(train, test_noisy, test_clean)   # (H, W)
print(gc.a, gc.beta, gc.rank, gc.psnr_test)

GPU acceleration

device='auto' (default) picks Apple MPS, then CUDA, then CPU. CPU runs np.linalg.svd in float64; MPS / CUDA runs torch.linalg.svd in float32. The downstream optimisation runs on CPU regardless of device.

gc = GeneralizedCovDenoiser(device='mps')   # Apple Silicon
gc = GeneralizedCovDenoiser(device='cuda')  # NVIDIA
gc = GeneralizedCovDenoiser(device='cpu')   # force CPU (slower, float64)

After every .denoise() call the lib prints the chosen (â, β̂, r̂, σ̂², PSNR, device) to stdout. The same fields are available as attributes on the denoiser:

gc.a, gc.beta, gc.rank, gc.sigma2, gc.psnr_test

MPLawDenoiser

Standard Marčenko–Pastur baseline, no oracle search:

from rmt_denoise import MPLawDenoiser
mp = MPLawDenoiser()
denoised = mp.denoise(noisy_images)        # (n, H, W) -> (n, H, W)

How GeneralizedCovDenoiser works

End-to-end:

1. Centre the data: X̃ = X − X̄.
2. SVD once (uses dual when p > n).
3. scipy.optimize.differential_evolution over (log a, β) with bounds a ∈ [0.01, 1.0], β ∈ [0.01, 0.99]. Default settings: popsize=20, Sobol init, seed=42, polish=True.
4. Per-candidate evaluation: gen-cov acceptance test → rank r̂ → cached projection → centring re-add → clip → T(a, β) → color-resize → PSNR vs clean.
5. Reconstruct every column at the best rank, applying the same T and color-resize to all of them.

T(a, β) is the diagonal matrix whose first ⌊p·β⌋ entries are √a and the rest are 1 (p = H·W); it embeds the noise-model scaling H = β·δ_a + (1 − β)·δ_1 directly into the post-processed image.

color_resize rescales each image as y = (x − min(x)) / max(x_before_subtract) and clips to [0, 1].

Both steps are part of the algorithm and are always applied — there are no flags to disable them.

API Reference

GeneralizedCovDenoiser

GeneralizedCovDenoiser(
    a_bracket=(0.01, 1.0),
    beta_bracket=(0.01, 0.99),
    seed=42,
    de_kwargs=None,
    device='auto',     # 'auto' | 'cpu' | 'mps' | 'cuda'
)

MPLawDenoiser(sigma2=None)

Folder loaders

from rmt_denoise import load_folder, split_train_test

images, names = load_folder("path/", size=(H, W))            # (n, H, W) in [0, 1]
train_imgs, test_img, n_train = split_train_test(
    images, names, n_train=300, test="s12_1.png", seed=42,
)

Noise utilities

from rmt_denoise import (
    add_gaussian_noise, add_laplacian_noise,
    add_mixture_gaussian_noise, add_structured_noise,
)

Metrics

from rmt_denoise import compute_psnr, compute_ssim
psnr = compute_psnr(clean, denoised)   # dB
ssim = compute_ssim(clean, denoised)   # [-1, 1]

Mathematical Background

B_n = S_n T_n where T_n has spectral distribution converging to H = β·δ_a + (1 − β)·δ_1. The noise eigenvalue support is bounded by

g(t) = -1/t + y · ( β·a/(1 + a·t) + (1 − β)/(1 + t) )

with bulk edges

λ_lower = σ² · max_{t > 0} g(t),
λ_upper = σ² · min_{−1/a < t < 0} g(t).

When a = 1 or β = 0, this reduces to classical Marčenko–Pastur: [(1 − √y)², (1 + √y)²].

The oracle workflow above selects (a, β) so that the corresponding acceptance-test rank gives the maximum PSNR against a clean reference, then bakes the resulting T(a, β) directly into the post-processed image.

References

1. Yu, Yao-Hsing (2025). "Geometric Analysis of the Eigenvalue Range of the Generalized Covariance Matrix." 2025 S.T. Yau High School Science Award (Asia).
2. Gavish, M. & Donoho, D. L. (2017). "Optimal Shrinkage of Singular Values." IEEE Trans. Inf. Theory, 63(4), 2137–2152.
3. Marčenko, V. A. & Pastur, L. A. (1967). "Distribution of eigenvalues for some sets of random matrices." Math. USSR-Sbornik, 1(4), 457–483.
4. Veraart, J. et al. (2016). "Denoising of diffusion MRI using random matrix theory." NeuroImage, 142, 394–406.
5. Storn, R. & Price, K. (1997). "Differential Evolution — A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces." J. Global Optim., 11(4), 341–359.

License

MIT