超光谱图像去噪基准

根据是否联合利用超光谱图像的空间和光谱信息，高光谱图像去噪技术可以分为两类。第一类就是将传统 2-D 图像去噪的方法直接应用到超光谱图像的每个频带上去，称为逐带去噪。第二类就是联合利用空间和光谱信息来进行去噪，称为联合去噪，这又可以大致分为基于变换域的方法和基于空间域的方法。除此之外，由于深度理论的兴起，最近也出现了一些基于深度学习的超光谱图像去噪方法。

逐带去噪

• [BM3D] Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering, TIP2007, K. Dabov et al.
• [WNNM] Weighted nuclear norm minimization with application to image denoising, CVPR2014, S. Gu et al.
• [EPLL] From learning models of natural image patches to whole image restoration, ICCV2011, D. Zoran et al.

然而，这些逐带去噪方法通常导致更大的频谱失真，因为没有同时考虑不同频带之间的空间和频谱信息的相关性。

联合去噪

1) 基于变换域的方法

基于变化域的方法尝试通过不同的变换来将干净信号从噪声数据中分离出来，比如主成分分析、傅里叶变换、小波变换。

• Wavelet-based hyperspectral image estimation, IGARSS2003, I. Atkinson et al.
• Noise reduction of hyperspectral imagery using hybrid spatial-spectral derivative-domain wavelet shrinkage, TGRS2006, H. Othman et al.

这一类方法的主要缺点是它们对变换函数的选择很敏感，并且没有考虑超光谱图像几何特征的差异。

2) 基于空间域的方法

采用合理的假设或先验，如谱间全局相关性(Global Correlation along Spetrum) 、空间非局部自相似性(Non-local Self Similarity across space)、总变差(Total Variation)、非局部(Non-Local)、稀疏表示(Sparse Representation)、低秩模型(Low Rank models)等 ，基于空间域的方法可以将噪声超光谱图像映射到干净图像并且保持其空间和光谱特征。

• [GCS and NSS] Adaptive Spatial-Spectral Dictionary Learning for Hyperspectral Image Denoising, ICCV2015, Ying Fu et al.

• [GCS and NSS] Decomposable nonlocal tensor dictionary learning for multispectral image denoising, CVPR2014, P. Yi et al.

• [GCS and NSS] Multispectral images denoising by intrinsic tensor sparsity regularization, CVPR2016, Q. Xie et al.

• [GCS] Denoising of hyperspectral images using the parafac model and statistical performance analysis, TGRS2012, X. F. Liu, et al.

• [TV] Hyperspectral image denoising employing a spectral–spatial adaptive total variation model, TGRS2012, Q. Yuan et al.

• [TV] Hyperspectral image denoising with a combined spatial and spectral hyperspectral total variation model, CJRS2014, G. Chen et al.

• [NL] A nonlocal transform-domain filter for volumetric data denoising and reconstruction, TIP2012, M. Maggioni et al.

• [SR] Spectral–Spatial Adaptive Sparse Representation for Hyperspectral Image Denoising, TGRS2016, T. Lu et al.

• [SR] Noise removal from hyperspectral image with joint spectral-spatial distributed sparse representation, TGRS2016, J. Li et al.

• [LR] Denoising and dimensionality reduction using multilinear tools for hyperspectral images, GRSL2008, N. Renard et al.

• [LR] Hyperspectral image restoration using low-rank matrix recovery, TGRS2014, H. Zhang et al.

• [LR] Hyperspectral image denoising via sparse representation and low-rank constraint, TGRS2015, Y. Zhao et al.

• [LR] Hyperspectral image restoration via iteratively regularized weighted Schatten p-norm minimization, TGRS2016, Y. Xie et al.

• [LR] Hyperspectral image restoration using low-rank tensor recovery, J-STARS2017, H. Fan et al.

深度学习去噪

• Hyperspectral imagery denoising by deep learning with trainable nonlinearity function, GRSL2017, W. Xie et al.
• Hyperspectral Image Denoising Employing a Spatial-Spectral Deep Residual Convolutional Neural Network, TGRS2018, Q. Yuan et al. [code]

数据集

• CAVE dataset
• AVIRIS
• ROSIS
• HYDICE
• EO-1 Hyperion Data
• Harvard dataset
• iCVL dataset
• NUS datase
• NTIRE18 dataset

================================================================================

去噪效果评价指标

• Peak Signal to Noise Ratio (PSNR)
• Structural SIMilarity index (SSIM)
• Feature SIMilarity index (FSIM)
• Erreur Relative Globale Adimensionnelle de Synthèse (ERGAS)
• Spectral Angle Mapper (SAM)