destripe 0.1.2

PDHG-based universal stripe-noise removal for images, backed by PyTorch

destripe

Universal stripe-noise removal for NumPy images. Decomposes an image into a TV-regularized clean component and directional ℓ²-penalized stripe components via a PDHG (primal-dual hybrid gradient) solver, backed by PyTorch.

Features

• Removes vertical and diagonal stripe patterns (5 directions).
• Preserves input dtype and value range (uint8, float32, float64).
• Supports grayscale, single-channel, and RGB; color is preserved by estimating stripes on the luminance channel.
• Tiled processing with cosine blending for large images that do not fit in memory.
• Auto-selects CUDA when available; falls back to CPU.

Install

pip install destripe

Quick Start

from destripe import destripe

image = ...  # numpy.ndarray, shape [H, W] or [H, W, 3]

clean = destripe(
    image,
    mu1=0.33,
    mu2=0.003,
    iterations=500,
    tiles=1,        # >1 for n x n tiled processing
    device="cpu",   # "cpu", "cuda", or None to auto-select
)

Integer inputs are clipped to their dtype range on output.

Parameters

• mu1 (default 0.33): TV weight. Higher smooths more and removes stronger stripes; loses fine detail.
• mu2 (default 0.003): ℓ² stripe penalty. Higher extracts stripes more eagerly; can leak real structure.
• iterations (default 500): max PDHG iterations. Early-stops on relative change below tol (default 1e-5), checked every 20 iterations.
• tiles (default 1): n × n tiling with cosine-blended overlap. Use when the image does not fit in memory or stripes are locally non-stationary. overlap (default 64) sets blending width.
• device (default None): "cpu", "cuda", a torch.device, or None to auto-select CUDA when available.
• proj (default True): project the clean component onto [0, 1].

Suggested mu Pairs

• Conservative / subtle stripes: [0.1, 0.001], [0.1, 0.0017]
• Light, thin stripes: [0.17, 0.003], [0.23, 0.003]
• Typical to strong stripes: [0.33, 0.003], [0.4, 0.007]
• Severe corruption / short stripes: [0.5, 0.017]

Starting points, not universal optima.

Reference

• https://github.com/NiklasRottmayer/General-Stripe-Removal