勷勤数学•专家报告

题      目：Variational Bayesian Inference for Poissonian Image Denoising

报  告  人： 王超 副研究员  (邀请人：骆其伦)

                                              南方科技大学

时      间： 10月14日  10：30-11:30

地     点：数科院东楼401

报告人简介:

          王超，南方科技大学统计与数据科学系副研究员，博导，其研究方向主要为图像处理、科学计算与交叉学科的数据科学。以第一作者或通讯作者身份在Cell子刊、SIAM系列、IEEE汇刊等权威期刊及CCF-A会议发表论文30余篇。广东省青年人才，深圳市鹏城孔雀计划特聘岗位，获CVPR研讨会最佳论文奖，CSIAM年会学生论文奖，以及2次获得SIAM差旅奖。主持国家自然科学基金项目两项、省部级基金一项、深圳市科研项目一项。 担任期刊J. Math. Imaging Vision （中国数学会T2 期刊）的特刊客座主编。

摘      要：

          Poisson noise removal is a fundamental challenge in image processing due to its signal-dependent nature and inherent ill-posedness. While modern optimization models integrate multiple priors to improve restoration, the simultaneous selection of regularization parameters and optimization of variables introduce significant computational complexity. To address these issues, we propose a variational Bayesian framework that incorporates the Poisson data-fidelity term with total variation and nonlocal low-rank regularization without requiring manual parameter tuning. Specifically, we employ maximum a posteriori estimation for the expectation of the latent image to ensure consistency with the posterior mode, while restricting the quadratic approximation to its covariance computation.The regularization parameters are treated as variables in hyperprior and inferred through variational posterior updates. Our approach enables efficient uncertainty quantification while avoiding the inaccuracies of global quadratic approximations by decoupling the estimation of distributional parameters. Furthermore, we establish a high-probability error bound for the resulting estimator. Experimental results demonstrate that the proposed algorithm significantly improves denoising performance and can be naturally extended to Gaussian, Poisson--Gaussian, and Cauchy noise scenarios.

          欢迎老师、同学们参加、交流！