返回列表
结构张量总变分正则化加权核范数最小化高光谱图像混合去噪
算法
2026/09/21
摘要：针对高光谱图像（（ HSI）混合去噪，已提出了多种基于逐波段TV正则化的低秩模型，这些模型可以同时利用光谱 和空间信息。然而，这些方法可能会由于噪声而导致较大的波动，也会产生油画效果。此外，它们仅以单独的方式利用 空间信息，这可能会对去除具有明显结构的噪声（例如截止噪声）的性能产生负面影响。针对上述问题，提出了一种新 的结构张量全变分（（ STV）正则化的加权核范数最小化（（ STWNNM）模型。为了获得所需的性能，包括三个问题。首先， 采用加权核范数最小化（（ WNNM）通过以不同的权重收缩不同的特征值来利用光谱信息。其次，使用结构张量同时利用 所有波段内的全局空间结构信息。第三，加入卷积核以从邻域像素中获取更多的局部结构信息。然后，提出了两种不同 的优化策略来解决导出的优化问题。通过与其他最先进的基于 TV 正则化的低秩方法进行比较，模拟和真实数据实验都 证明了所提出的 STWNNM 对 HSI 混合去噪的更高性能。
关键词：高光谱图像去噪；低秩；结构张量总变分 (STV)；核范数最小化 (NNM)；交替方向乘子法 (ADMM)
期刊：Signal Processing
DOI：http://dx.DOI.org/10.1016/j.sigpro.2016.07.031
设备：GaiaField
Structure Tensor Total Variation-Regularized Weighted Nuclear Norm Minimization For Hyperspectral Image Mixed Denoising
Abstract：Several band-by-band TV-regularized low rank based models have been proposed for Hyperspectral image (HSI) mixed denoising, which can exploit the spectral and spatial information simultaneously. However, these methods may lead to large fluctuations due to the noise and also create oil painting effects. Moreover, they only exploit the spatial information in a separated manner, which may negatively affect the performance of removing the noise with obvious structure, e.g., the deadline noise. To cope with the above problems, a novel Structure tensor Total Variation (STV)-regularized Weighted Nuclear Norm Minimization (STWNNM) model is proposed. To obtain the desired performance, three issues are included. First, the Weighted Nuclear Norm Minimization (WNNM) is adopted to utilize the spectral information by shrinking different eigenvalues with different weights. Second, the structure tensor is used to exploit the global spatial structure information within all bands simultaneously. Third, a convolution kernel is incorporated to obtain more local structure information from neighborhood pixels. Then, two different optimization strategies are proposed to solve the derived optimization problem. Both simulated and real data experiments illustrate the higher performance of the proposed STWNNM for HSI mixed denoising, by comparing with other state-of-the-art TV- regularized low rank based methods.
Keywords：Hyperspectral Image Denoising; Low Rank; Structure Tensor Total Variation (STV); Nuclear Norm Minimization (NNM); Alternating Direction Method of Multipliers (ADMM)
