信号恢复
计算机科学
信号处理
人工智能
信号重构
稳健性(进化)
模式识别(心理学)
算法
稀疏矩阵
数据恢复
信号(编程语言)
稀疏逼近
压缩传感
噪音(视频)
探测理论
多维信号处理
信噪比(成像)
数据挖掘
语音识别
噪声测量
作者
Keyuan Hu,Zhi-Yong Wang,Hing Cheung So
标识
DOI:10.1109/tsp.2026.3684761
摘要
In the context of sparse signal recovery, a particularly challenging scenario arises when both dense noise and outliers are presented. The choice of regularizers plays a pivotal role in effectively addressing this problem, influencing both the accuracy and robustness of the recovery process. In this paper, we introduce a novel class of implicit sparsity-inducing regularizers (ISIRs) generated within a unified framework to achieve a balance among robustness, solution uniqueness, and unbiasedness. Structurally, ISIR exhibits weak convexity while satisfying three fundamental properties of an ideal regularizer: (i) sparsity, (ii) approximate unbiasedness, and (iii) continuity. Furthermore, through the double null space property (DNSP), ISIR ensures solution uniqueness under mild conditions. Building on this foundation, we establish the theoretical guarantees: in the absence and presence of dense noise, the global solutions achieve exact recovery and ensure stable recovery with bounded error, respectively. To facilitate computational efficiency, we propose a solver based on alternating direction method of multipliers, where most subproblems admit closed-form solutions via the proximity operator. Comprehensive analyses of convergence and computational complexity are also presented. Finally, numerical results validate the superior performance of the developed algorithm compared to existing schemes.
科研通智能强力驱动
Strongly Powered by AbleSci AI