变形L_(1)正则化的高光谱图像稀疏解混  

Transformed L_(1)regularization for hyperspectral sparse unmixing

作　　者：李璠[1,2] 吴朝明 张绍泉 胡蕾 邓承志[1] LI Fan;WU Zhao-ming;ZHANG Shao-quan;HU Lei;DENG Cheng-zhi(Jiangxi Province Key Laboratory of Water Information Cooperative Sensing and Intelligent Processing,Nanchang Institute of Technology,Nanchang 330099,China;Computer Information Engineering College,Jiangxi Normal University,Nanchang 330022,China)

机构地区：[1]南昌工程学院江西省水信息协同感知与智能处理重点实验室,江西南昌330099 [2]江西师范大学计算机信息工程学院,江西南昌330022

出　　处：《激光与红外》2021年第4期515-522,共8页Laser & Infrared

基　　金：江西省教育厅科技项目(No.GJJ190956,No.GJJ180962,No.GJJ170992);国家自然科学基金资助项目(No.61865012,No.61662033);江西省自然科学基金项目(No.20192BAB217003);江西省重点研发计划项目(No.20202BBGL73081,No.20181ACG70022)资助。

摘　　要：如何准确地刻画易于求解的稀疏正则化函数是高光谱图像稀疏解混的难点。变形L_(1)正则化函数是一类由绝对值函数组成的双线性变换的单参数族,类似于L_(p)p∈0,1范数,通过调整参数a∈0,可以准确表征L_(0)和L_(1)之间的任意范数,并具有无偏、稀疏和Lipschitz连续性。论文首先研究变形L_(1)正则化函数,然后提出变形L_(1)正则化的高光谱稀疏解混变分模型,最后提出变形L_(1)正则化高光谱稀疏解混模型的凸函数差分求解算法。通过模拟和真实的高光谱数据实验,与经典的SUnSAL算法相比,表明提出的算法能够更准确地刻画丰度系数的稀疏性,并获得更高的解混精度。How to accurately characterize the sparse regularization function which is easy to be solved is difficult for sparse hyperspectral unmixing.Transformed L_(1)TL_(1)regularization function is a one parameter family of bilinear transformations composed with the absolute value function,similar to L_(p) norm with p∈0,1,it represents any norms from L_(0) to L_(1)exactly through a nonnegative parameter a∈0,,and satisfies unbiasedness,sparsity and Lipschitz continuity properties.In this paper,we study TL_(1)regularization function and propose sparse hyperspectral unmixing model with TL_(1)regularization.Meanwhile,a difference of convex algorithm for TL_(1)in computing TL_(1)regularized sparse unmixing problems is presented.Experimental results on both simulated and real hyperspectral data demonstrate that the TL_(1)regularization function describes the sparsity of endmembers more accurately and the proposed algorithm is much more accurate than SUnSAL algorithm.

关 键 词：高光谱图像 稀疏解混 变形L_(1)正则化 凸函数差分算法 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]