Hyperspectral image denoising with bilinear low rank matrix factorization

Huixin Fan, Jie Li*, Qiangqiang Yuan, Xinxin Liu, Kwok Po NG

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

12 Citations (Scopus)

Abstract

Hyperspectral images (HSIs) have rich spectral information, but the various noises generated during the imaging process destroy the visual quality of images and lower the application precision. Therefore, it's crucial to denoise HSI for making better use of it. At present, low-rank-based methods have shown potential in mixture noises removal. While their limitations in rank function approximation, which affects the description of the low rank property in HSI, still need to be broken through. This paper puts forward a bilinear low rank matrix factorization (BLRMF) HSI denoising method, where the bi-nuclear quasi-norm is employed for constraining the low rank characteristic in HSI. The bi-nuclear quasi-norm is a closer approximation to the rank function and can be calculated by the nuclear norms of two smaller factor matrices, which respectively describe the spatial low rank and the spectral low rank. The Alternating Direction Method of Multipliers (ADMM) is employed for solving the optimization problem. A large number of experiments on HSI denoising are conducted to verify the superiority of the BLRMF over the mainstream denoising methods.

Original languageEnglish
Pages (from-to)132-152
Number of pages21
JournalSignal Processing
Volume163
DOIs
Publication statusPublished - Oct 2019

Scopus Subject Areas

  • Control and Systems Engineering
  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Electrical and Electronic Engineering

User-Defined Keywords

  • ADMM
  • Bi-nuclear quasi-norm
  • Bilinear low rank matrix factorization (BLRMF)
  • Hyperspectral images (HSIs) denoising
  • Low rank

Fingerprint

Dive into the research topics of 'Hyperspectral image denoising with bilinear low rank matrix factorization'. Together they form a unique fingerprint.

Cite this