光子学报 ›› 2019, Vol. 48 ›› Issue (10): 1010002-1010002.doi: 10.3788/gzxb20194810.1010002

• 图像处理 • 上一篇    下一篇

L1-2空谱全变差正则化下的高光谱图像去噪

曾海金, 蒋家伟, 赵佳佳, 王艺卓, 谢晓振   

  1. 西北农林科技大学 理学院, 陕西 杨凌 712100
  • 收稿日期:2019-07-15 出版日期:2019-10-25 发布日期:2019-08-13
  • 通讯作者: 谢晓振(1984-),男,副教授,博士,主要研究方向为数字光信息处理及遥感图像处理.Email:xiexzh@nwafu.edu.cn E-mail:xiexzh@nwafu.edu.cn
  • 作者简介:曾海金(1996-),男,硕士研究生,主要研究方向为高光谱图像处理.Email:zeng_navy@163.com
  • 基金资助:

    国家自然科金项目(No.61401368),中央高校基本科研业务费专项资金(No.2452019073)

L1-2 Spectral-spatial Total Variation Regularized Hyperspectral Image Denoising

ZENG Hai-jin, JIANG Jia-wei, ZHAO Jia-jia, WANG Yi-zhuo, XIE Xiao-zhen   

  1. College of Science, Northwest A & F University, Yangling, Shaanxi 712100, China
  • Received:2019-07-15 Online:2019-10-25 Published:2019-08-13
  • Contact: 2019-08-13 E-mail:xiexzh@nwafu.edu.cn
  • Supported by:

    The National Natural Science Foundation of China (No.61401368), the Fundamental Research Funds for the Central Universities(No.2452019073)

摘要:

针对高光谱图像的复原问题,提出了一种基于局部核范数最小化和全局L1-2空谱全变差正则化的高光谱复原模型.首先,将高光谱图像划分成局部交叠的三维图块,在提高局部低秩性的同时减少核范数最小化带来的信息损失;然后,建立稀疏表达能力更强的L1-2空谱全变差正则项,不仅能表示空间稀疏先验,而且还能发掘光谱稀疏先验;最后联合两者的优势,在局部上利用核范数最小化惩罚光谱低秩性,在全局上利用L1-2空谱全变差约束高光谱的空间和光谱稀疏性,建立新的高光谱图像复原模型.该模型不仅能够有效去除高斯噪声、脉冲噪声、死线噪声及其混合噪声,而且减少了对噪声独立同分布假设的依赖,能部分抑制与结构相关的噪声.通过对模拟的和真实的高光谱图像进行大量的实验仿真,并与经典的基于低秩和全变差的复原方法相比,本文模型复原结果的平均峰值信噪比提高1.36 dB,平均结构性相似指标提高0.004,而Q-测度降低1.35,平均光谱角降低0.64,复原精度大幅度提高.

关键词: 局部低秩, 交替方向乘子法, L1-2空谱全变差, 高光谱图像, 凸函数差算法

Abstract:

A model for Hyperspectral Image (HSI) restoration was proposed, which combines the L1 norm minimization of local patches and L1-2 Spatial-Spectral Total Variation (L1-2 SSTV) of global image. Firstly, the HSI was divided into local overlapping 3D patches, to reduce information loss caused by minimization of nuclear norm while improving local low-rank property. Secondly, the L1-2 SSTV regularization term with stronger sparse expression ability was proposed, to explore both spectral and spatial sparse prior simultaneously. Finally, the advantages of both worlds are combined, and a new HSI restoration model is proposed. The local norm is used to penalize spectrum low-rank term, and the L1-2 SSTV is used globally to constrain the spatial and spectral sparsity term of HSI, The model can not only effectively remove Gaussian noise, impulse noise, deadline and its mixture noise, but also reduce the dependence on the noise independent and identical distribution hypothesis, and can partially suppress the structure-related noise. Through a large number of experiments of simulated and real HSIs, and compared with the classical low-rank and total-variation-based restoration methods, experimental results were conducted to illustrate the advantage of the proposed method in HSI restoring, from visual/quantitative evaluations.

Key words: Difference of convex algorithm, Local low-rank, Alternating direction method of multipliers, L1-2 spatial-spectral total variation, Hyperspectral image

中图分类号: