Nonconvex is attractive: L2/3 regularized thresholding algorithm using multiple sub-dictionaries

Yunyi Li, Yue Hao, Fei Dai, Yue Yin, Shangang Fan, Jie Yang, Guan Gui, Fumiyuki Adachi

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

    Abstract

    The L2/3-regularization is a typical nonconvex and nonsmooth optimization method, which can obtain more powerful performance than L1 regularization in some applications, such as computational imaging, sparse signal recovery and low-rank matrix completion, etc. This paper proposes an adaptive iteratively-weighted thresholding algorithm for L2/3-regularized problem based on the multiple analysis sub- dictionaries (MD) sparsifying transform strategy, the MD strategy can be employed to further exploit the prior knowledge of estimated signal for sparse recovery. What's more, we propose an adaptive updating scheme for regularization parameter to weight the contribution of each sub-dictionary. Experiments confirm that the proposed method could obtain higher image quality and achieve faster convergence than some corresponding prior work.

    Original languageEnglish
    Title of host publication2018 IEEE International Conference on Communications, ICC 2018 - Proceedings
    PublisherInstitute of Electrical and Electronics Engineers Inc.
    ISBN (Print)9781538631805
    DOIs
    Publication statusPublished - 2018 Jul 27
    Event2018 IEEE International Conference on Communications, ICC 2018 - Kansas City, United States
    Duration: 2018 May 202018 May 24

    Publication series

    NameIEEE International Conference on Communications
    Volume2018-May
    ISSN (Print)1550-3607

    Other

    Other2018 IEEE International Conference on Communications, ICC 2018
    Country/TerritoryUnited States
    CityKansas City
    Period18/5/2018/5/24

    Keywords

    • Image restoration
    • Iteratively thresholding
    • L regularization
    • Multiple subdictionaries sparsifying transform
    • Nonconvex

    ASJC Scopus subject areas

    • Computer Networks and Communications
    • Electrical and Electronic Engineering

    Fingerprint

    Dive into the research topics of 'Nonconvex is attractive: L2/3 regularized thresholding algorithm using multiple sub-dictionaries'. Together they form a unique fingerprint.

    Cite this