Scholarlink

<p>Richardson-Lucy deconvolution is widely used to restore imaged objects blurred by a point spread function and corrupted by noise and is known to readily overfit noise, leading to high-frequency artifacts. Practical use therefore relies on hand-tuned stopping criteria or ad hoc regularization with limited physical justification. To resolve this problem, we present DeBayes: a rigorous Bayesian deconvolution framework that builds upon a physically accurate image formation model. DeBayes performs deconvolution in the spatial domain, jointly models accurate noise sources, and infers full posterior distributions over the underlying object. It avoids assumptions of sparsity or continuity, yields strictly positive reconstructions, and converges stably without user-tuned regularization parameters or iteration cutoffs. Our method is unsupervised and designed for fast, parallelizable computation, providing a principled alternative to Richardson-Lucy for robust, physics-informed image reconstruction. We demonstrate DeBayes’ stable convergence and minimal noise amplification on simulated and experimental images of mitochondria networks in HeLa cells.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A physics-informed alternative to Richardson-Lucy deconvolution across SNR regimes without iteration cutoffs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月30日

查看全文
  • Zachary H. Hendrix,
  • Peter T. Brown,
  • Rory Kruithoff,
  • Tim Flanagan,
  • Douglas P. Shepherd,
  • Ayush Saurabh,
  • Steve Pressé

摘要

Richardson-Lucy deconvolution is widely used to restore imaged objects blurred by a point spread function and corrupted by noise and is known to readily overfit noise, leading to high-frequency artifacts. Practical use therefore relies on hand-tuned stopping criteria or ad hoc regularization with limited physical justification. To resolve this problem, we present DeBayes: a rigorous Bayesian deconvolution framework that builds upon a physically accurate image formation model. DeBayes performs deconvolution in the spatial domain, jointly models accurate noise sources, and infers full posterior distributions over the underlying object. It avoids assumptions of sparsity or continuity, yields strictly positive reconstructions, and converges stably without user-tuned regularization parameters or iteration cutoffs. Our method is unsupervised and designed for fast, parallelizable computation, providing a principled alternative to Richardson-Lucy for robust, physics-informed image reconstruction. We demonstrate DeBayes’ stable convergence and minimal noise amplification on simulated and experimental images of mitochondria networks in HeLa cells.