基于群稀疏正则化的高光谱图像去噪
Hyperspectral Image Denoising Based on Group Sparse Regularization
姜斌 1叶军1
作者信息
- 1. 南京邮电大学 理学院,江苏 南京 210023
- 折叠
摘要
高光谱图像(HSI)具有良好的光谱识别能力,但在采集过程中易受到混合噪声的污染,严重影响了后续任务的精度,因此HSI去噪是重要的预处理过程.针对现有去噪方法对空间-光谱先验信息利用不足、条纹噪声建模不合理的问题,提出一种新的基于群稀疏正则化的高光谱图像去噪算法.该算法将干净HSI的空间-光谱低秩特性和各波段上条纹噪声的低秩结构融入一个新框架,实现了干净HSI与高强度结构化条纹噪声的分离;同时为了有效保持图像的边缘信息,在去噪模型中引入新的群稀疏正则化,即基于L2,1 范数的增强型三维全变分正则化(enhanced 3D total variation,E3DTV),充分挖掘HSI差分图像的稀疏先验信息,进一步提升了图像的分段平滑性.采用交替方向乘子法对变量优化求解,在仿真和真实数据集上进行数值实验表明,所提模型具有更好的去噪和去条纹性能,在视觉效果和定量评价结果上都明显优于其他对比算法.
Abstract
Hyperspectral image(HSI)has good spectral recognition ability,but it is easily polluted by mixed noise during the acquisition process,which seriously affects the accuracy of subsequent tasks,so HSI denoising is an important preprocessing process.Aiming at the problems of insufficient utilization of spatial-spectral prior information and unreasonable modeling of stripe noise in existing denoising methods,a new hyperspectral image denoising algorithm based on group sparse regularization is proposed.The algorithm integrates the spatial-spectral low-rank characteristics of clean HSI and the low-rank structure of stripe noise in each band into a new framework,and realizes the separation of clean HSI and high-intensity structured stripe noise;at the same time,in order to effectively maintain the edge information of the image,a new group sparse regularization is introduced into the denoising model,that is,enhanced 3D total variation(E3DTV)based on the L2,1 norm,which can fully explore the sparse prior information of HSI difference images.Alternate direction mul-tiplier method is used to optimize the solution of variables.Numerical experiments on simulation and real data sets show that the proposed model has better performance in denoising and destriping,and its visual effect and quantitative evaluation results are significantly better than that of other comparison algorithms.
关键词
高光谱图像/图像去噪/群稀疏正则化/低秩约束/条纹噪声Key words
hyperspectral images/images denoising/group sparse regularization/low-rank constraint/stripe noise引用本文复制引用
基金项目
国家自然科学基金(61971234)
南京邮电大学校内基金(NY220209)
出版年
2023
国家科技期刊平台
NSTL
维普
万方数据