中山大学学报(自然科学版)
中山大學學報(自然科學版)
중산대학학보(자연과학판)
ACTA SCIENTIARUM NATURALIUM UNIVERSITATIS SUNYATSENI
2014年
2期
23-28
,共6页
迭代重加权%矩阵秩%压缩感知%Frobenius范数
迭代重加權%矩陣秩%壓縮感知%Frobenius範數
질대중가권%구진질%압축감지%Frobenius범수
iterative reweighted%matrix rank%compressed sensing%Frobenius norm
利用最稀疏表示重构原始信号是压缩感知理论的核心,而基于几何影射约束的最小l1范数凸优化算法是其实现的主要方法。目前,解决最小lp(p≤1)范数问题的关键是迭代重加权最小二乘算法(IRLS-p ,0<p≤1),但其收敛和实时性较差。为此,文中从最小化矩阵秩的角度出发对一类扩展迭代重加权最小二乘算法(EIRLS-p )进行性能实现分析,用以改进IRLS-p算法的连续迭代收敛性及其实时性能。验证结果表明,EIRLS-0和sEIRLS-0算法性能优于奇异值门限(SVT)算法。同时,在没有先验知识的情况下,sEIRLS-0算法性能也优于迭代硬阈值(IHT)算法。
利用最稀疏錶示重構原始信號是壓縮感知理論的覈心,而基于幾何影射約束的最小l1範數凸優化算法是其實現的主要方法。目前,解決最小lp(p≤1)範數問題的關鍵是迭代重加權最小二乘算法(IRLS-p ,0<p≤1),但其收斂和實時性較差。為此,文中從最小化矩陣秩的角度齣髮對一類擴展迭代重加權最小二乘算法(EIRLS-p )進行性能實現分析,用以改進IRLS-p算法的連續迭代收斂性及其實時性能。驗證結果錶明,EIRLS-0和sEIRLS-0算法性能優于奇異值門限(SVT)算法。同時,在沒有先驗知識的情況下,sEIRLS-0算法性能也優于迭代硬閾值(IHT)算法。
이용최희소표시중구원시신호시압축감지이론적핵심,이기우궤하영사약속적최소l1범수철우화산법시기실현적주요방법。목전,해결최소lp(p≤1)범수문제적관건시질대중가권최소이승산법(IRLS-p ,0<p≤1),단기수렴화실시성교차。위차,문중종최소화구진질적각도출발대일류확전질대중가권최소이승산법(EIRLS-p )진행성능실현분석,용이개진IRLS-p산법적련속질대수렴성급기실시성능。험증결과표명,EIRLS-0화sEIRLS-0산법성능우우기이치문한(SVT)산법。동시,재몰유선험지식적정황하,sEIRLS-0산법성능야우우질대경역치(IHT)산법。
The kernel technology of Compressed sensing theory is to find the sparsest representation to recover original signal data,in which the convex optimization algorithm of minimization the l1 norm is a important method.At present,a key algorithm solved minimization the lp(p≤1 )norm is iterative re-weighted least squares algorithm (IRLS-p ,0 <p≤1 )with affine constraints,but a crucial question of the IRLS-p Algorithm is to iterate convergence and real time performances.Therefore,the EIRLS-p and sEIRLS-p algorithms were proposed to extend IRLS-p as a family of algorithms for the matrix rank minimi-zation problem,and to improve IRLS-p implementations performances of successive iterates convergence and real time.Validating results show that both EIRLS-0 and sEIRLS-0 perform better than singular value thresholding (SVT)algorithm.At the same time,it was observed that sEIRLS-0 performs better than iter-ative hard thresholding algorithm (IHT)when there is no apriori information on the low rank solution.