第 36 卷第 12 期 电 子 与 信 息 学 报 Vol.36No.12
2014 年 12 月 Journal of Electronics & Information Technology Dec. 2014
基于图稀疏正则化多测量向量模型的高光谱压缩感知重建
孙玉宝
①④
李 欢
②
吴 敏
③
吴泽彬
④
贺金平
②
刘青山
①*
①
(南京信息工程大学信息与控制学院 南京 210014)
②
(北京空间机电研究所 北京 100076)
③
(南京军区南京总医院医学工程科 南京 210002)
④
(南京理工大学计算机科学与工程学院 南京 210094)
摘 要:压缩感知重建是解决高光谱现有成像模式数据量大冗余度高问题的一个有效机制。针对高光谱图像的多通
道特性,该文建立了高光谱压缩感知的多测量向量模型,编码端使用随机卷积算子对各通道进行快速采样,生成测
量向量矩阵。解码端构建图稀疏正则化的联合重建模型,在稀疏变换域将高光谱图像分解为谱间的关联成分和差异
成分,通过图结构化稀疏度量表征关联成分的空谱相关性,并约束谱间差异成分的稀疏性。进一步提出模型求解的
交替方向乘子迭代算法,通过引入辅助变量与线性化技巧,使得每一子问题均存在解析解,降低了模型求解的复杂
度。对多个实测数据集进行了对比实验,实验结果验证了该文模型与算法的有效性。
关键词:高光谱图像;压缩感知;多测量向量;图稀疏;交替方向乘子法
中图分类号: TP751.1 文献标识码: A 文章编号:1009-5896(2014)12-2942-07
DOI: 10.3724/SP.J.1146.2014.00566
Compressed Sensing Reconstruction of Hyperspectral Image Using the
Graph Sparsity Regularized Multiple Measurement Vector Model
Sun Yu-bao
①④
Li Huan
②
Wu Min
③
Wu Ze-bin
④
He Jin-ping
②
Liu Qing-shan
①
①
(Department of Information and Control, Nanjing University of Information Science and Technology, Nanjing 210014, China)
②
(Beijing Institute of Space Mechanics and Electricity, Beijing 100076, China)
③
(Department of Medical Engineering, Nanjing General Hospital of Nanjing Area Command, Nanjing 210002, China)
④
(Department of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China)
Abstract: Compressed Sensing (CS) reconstruction of hyperspectral image is an effective mechanism to comedy the
traditional hyperspectral imaging pattern with the drawback of high redundancy and vast data volume. This paper
presents a new multiple measurement vector model for compressed sensing reconstruction of hyperspectral data in
consideration of its multiple channel character. In the encoding side, the random convolution operator is used to
rapidly obtain the measurement vector of each channel which is subsequently reorganized as a measurement vector
matrix. In the decoding side, a joint reconstruction model is proposed to reconstruct the hyperspectral data from
the multiple measurement vectors. The model decomposes the hyperspectral data into the inter-channel correlated
and differenced component in the sparsifying transform domain, where the correlated component with high spatial
and spectral correlation is constrained to be graph structured sparse and the differenced component is constrained
to be
1
l sparse. A numerical optimization algorithm is also proposed to solve the reconstruction model by the
alternating direction method of multiplier. Every sub-problem in the iteration formula admits analysis solution by
introducing the auxiliary variable and linearization operation. The complexity of the numerical optimization
algorithm is reduced. The experimental results demonstrate the effectiveness of the proposed algorithm.
Key words: Hyperspectral image; Compressed Sensing (CS); Multiple measurement vectors; Graph structured
sparsity; Alternated direction method of multiplier
1 引言
高光谱图像的压缩感知问题是当前的一个研究
热点,是解决高光谱现有成像模式数据量大冗余度
2014-04-30 收到,2014-07-25 改回
国家自然科学基金(61272223, 61300162, 81201161),江苏省自然科
学基金(BK2012045, BK20131003),中国博士后基金(20110491429),
江苏省博士后基金(1101083C), CAST 创新基金(201227)和江苏省
光谱成像与智能感知重点实验室基金资助课题
*通信作者:刘青山 qsliu@nuist.edu.cn
高问题的一个有效机制
[1 3]−
,不同于 2 维信号,高
光谱图像是一种 3 维体数据,包含空间与光谱两个
维度。由于光谱间的连续性以及地物分布的空间连
续性,呈现出空间和谱间的联合相关结构,应有效
利用这一联合结构先验进行压缩感知重建。文献[4]
将全变差模型独立应用于每一谱带数据,约束其空
间光滑性,文献[5]使用小波对每一谱带进行稀疏表
示,但均忽略了谱带间的相关性。文献[6]采用 3 维