报告题目: TRGN: Guaranteed Tensor Recovery by Decoupling the Gradient—— A Coupled Low-Rank and Sparse Framework

报告人: 孙佳宁 东北师范大学

报告时间：2025年11月22日 10：00-11:00

报告地点：#腾讯会议：769-436-585

校内联系人：伍铁如 [email protected]

报告摘要：Tensor data, such as hyperspectral images and videos, are frequently corrupted by missing entries or noise, making their recovery a fundamental challenge. In this presentation, we introduce a novel paradigm: the Tensor Robust Gradient Norm (TRGN) framework. Departing from holistic constraints, TRGN is the first method to explicitly decouple the tensor gradient into distinct low-rank and sparse components. We will demonstrate how this precise modeling enables guaranteed exact recovery of both components under mild incoherence conditions, a theoretical assurance previous methods lack. We will then explain how TRGN is integrated into tensor completion and robust PCA models, and solved via an efficient ADMM-based algorithm with provable convergence. Comprehensive experiments on color images, hyperspectral data, and videos confirm that TRGN consistently and significantly outperforms SOTA methods—achieving PSNR gains of 0.87-1.30 dB at an extreme 0.5% sampling rate and superior denoising under high-intensity noise—while remaining computationally efficient. This work bridges low-rank and sparse priors in the gradient domain, offering a theoretically sound, versatile, and powerful solution for the most challenging tensor recovery tasks.

报告人简介：孙佳宁，目前任教于东北师范大学数学与统计学院计算数学专业，早年毕业于吉林大学炸金花游戏 ，研究方向主要包括数据建模、优化计算方法、小波分析理论等，相关成果曾发表于Science子刊Advances、IEEE Transactions on Pattern Analysis and Machine Intelligence、IEEE Journal of Seleted Topics in Signal Processing、Information Sciences和Signal Processing等，还曾在CCF A类会议IEEE CVPR、ICCV发表文章，并在算法领域获得发明专利一项。