您好,欢迎来到微智科技网。
搜索
您的当前位置:首页基于L1范数正则化的四元数信号重建

基于L1范数正则化的四元数信号重建

来源:微智科技网
Journal of Southeast University(English Edition) Vo1.29,No.1,PP.33—37 Mar.2013 ISSN 1O03—7985 —r ‘ ‘ - ● ^ ・ ■ - l L,-nOrm minimization tot quaternion signals Zhang Xu Wu Jiasong ’ Yang Guanyu ,。Lotif Senahdji , Shu Huazhong ’ ( Laboratory of Image Science and Technology,Southeast University,Nanjing 210096,China) ( LTSI,INSERM U 1099,Universit6 de Rennes 1,Rennes 35000,France) ( Centre de Recherche en Information Biom6dicale Sino—frnqaias,Nanjing 210096,China) Abstract:An algorithm for recovering the quatemion signals in both noiseless and noise contaminated scenarios by solving an squares problem: Ll—norm minimization problem is presented.The L1一norm imnimization problem over the quatemion number field is solved by converting it to all equivalent second—order cone programming problem over the real number field,which can be readily solved by convex optimization solvers like SeDuMi. Numerical experiments are provided to illustrate the effectiveness of the proposed algorithm. In a noiseless scenario,the experimental results show that under some practically acceptable conditions,exact signal recovery can be achieved.With additive noise contamination in measurements, the experimental results show that the proposed algorithm is robust to noise.The proposed algorithm can be applied in compressed--sensing--based signal recovery in the quaternion domain. Key words:quatemion;signal recovery;compressed sensing doi:10.3969/j.issn.1003—7985.2013.01.007 T hpeo rptarnotb lreomle  oifn  Lth1-en orremce nmitlyn imidezvaetlioopne dp lcaoymps anre sismed— sensing(CS)theory ,which is a new approach for da— ta acquisition and has wide applications in the field of sig— nal and image processing.CS has been conventionally used to real input data.Recently,special attention has al— so been paid to the complex input data such as blind source separationI and terahertz imaging【51. On the other hand,the theory and application of qua— ternion or hypercomplex algebra,invented by Hamil— ton【 have received much attention in recent yearst 一 . .Ell and Sangwine applied the quaternion Fourier trans— ofrm to color image processing.Jiang and Wei pro— posed all algorithm for solving the quaternion least Received 2012.10 29. Biographies:Zhang Xu(1984一),male,graduate;Shu Huazhong(cor— responding author),male,doctor,professor,shu.1ist@seu.edu.cn. FoundatiOil items:The National Basic Research Program of China f 973 program)(No.2011CB707904),the National Natural Science Founda・ tion of China(No.61073138,61271312,6120l344,81101104, 6091 1 1303701.the Research Fund for the Doctoral Program of Higher Education of Ministry of Education of China f No.20110092l10023. 20120o9212o0361.the Natural Science Foundation of Jiangsu Province (No.BK2012329,BK2O12743). Citation:Zhang Xu,Wu Jiasong,Yang Guanyu,et a1.Ll—norm mini— mization for quaternion signals[J].Journal of Southeast University (English Edition),2013,29(1):33—37.[doi:10.3969/j.issn.1003— 7985.2013.01.0071 min llXA—Y fl2 s.t.Bx=z (1) where A∈Q ,B∈Q“ ,X∈Q ,y∈Q” ,Q de— notes the quatemion number ifeld,and_I・lI2 denotes the L2-norm of quaternion vector.The algorithm reported in Ref.[9]was further extended by Jinag et a1. 10 3 to a two— dimensional scenario where the inputs are quaternion ma— tirces.Note that the algorithms reported in Refs.[9—10] solve the overdetermined system,that is,n≥m for meas・ urement matrix A. In this paper,we consider the recovery problem of quaternion signals for the case where n≤m,that is,we deal wiht the underdetermined linear system with meas— urement matrix A,which is quite different from that of (1).We will do that by solving an Ll—norm minimiza— tion problem.To the authors’knowledge,the L1一norrn minimization problem for the quaternion signals has not yet been investigated.Winter et a1.[ ]converted the L1一norm minimization of complex signals to the second— order cone programming(SOCP),which was then solved by SeDuMi software .In this paper,we extend hte algorithm to the quaternion signals with and without noise contamination,which are,respectively,defined by the following two optimization problems: min lIxlll s.t.Y=Ax (2) min lIx Jll s.t.IlAx—Yll:≤ (3) where I1.1J】denotes the L1.norm of the quatenrion vector and s is the noise penalty level for noise term e in noise measurements Y=Ax十e. 1 Deftnitions A quaternion q is a hypercomplex number which con— sists of one real part R(q)and three imaginary parts ,(q),J(q)and K(q)as follows: q=R(q)+,(q)i+J(q)J+K(q)k (4) where R(q),,(q),J(q),K(q)∈R;and i,J and k rae three imaginary units obeying the following rules: i =_/ =k =ijk=一1 (5) ij=一ji=k,jk=一 =i,ki=一f = (6) Zhang Xu,Wu Jiasong,Yang Guanyu,Lotif Senahdji,and Shu Huazhong The conjugate and modulus of a quaternion are,re spectively,defined as =[1,0,0,0,0,…,1,0,0,0,0] ∈R =(18) (19) [R(y),,(),),J(y), (),)] ∈R4 0 0 0 T q =R(q)一,(q)i—J(q)J~K(q)k (7) 0 q l= We also define (q)+ (q)+‘,2(q)+K2(q) (8) -R(a1) I(a1) J(a ) K(a.) I(a。) R(a ) K(a。) 一J(a ) J(a ) 一K(a。) R(a1) K(a ) J(a。) ● ● 一 ’ ● -A=[a ,…,a ] =(9) 一I(a.) ‘ 一I(a ) R(a.) ● ● [xl,…, ]T=R( )+,( )i+J(x)j+ ( )k (10) A= ∈R“ 0 0 0 0 Y=[Y。,…,Y ] =R(y)+I(y)i+J(Y)j+K(y)k (11) The Lp-norm of a quaternion vector x is given by , P=I,2 (12) In this section,we derive an algorithm based on SOCP f0r L.-norli1 minimization problems dealing with quaterni— on signals. 2.1 Noiseless case The minimization problem shown in(2)is equivalent to its epigraph form: min f∈R s.t.Y=Ax,fIxll≤f (13) By introducing auxiliary variables t,∈R ,where r= 1,2,…,m and R denotes the positive real number field, hte second constraint j1≤ cna be decomposed into a set of m constraints [R(x,),I(x,),J(x ), (xr)] l I≤f r=1,2,…,m (14) and(13)becomes minlrt∈R t s.t.Y=Ax [R(x ),I(x,),J(x,),K(x )] II ≤t, r=1,2,…,m (15) After some manipulation,(15)can be written as min 量∈R s.t. = [R(x ),I(x,),g(x,),K(x,)] ff ≤f, r:1,2,…,m (16) where 膏=[t1,R(x1),I(x1),J(x1),K(x。),…, t ,R(x ),I(x ),J(x ),K(x )] ∈R (17) R(a ) I(a ) J(a ) K(a ) -I(a ) R(a ) K(a ) 一J(a ) 一g(a ) 一K(a ) R(a ) I(a ) 一K(a ) J(a ) 一I(a ) R(a ) (20) 2.2 Noise case Similar to the previous case,the minimization problem shown in(3)can be reformulated as min ∈R s.t.f I一 l l≤8 [R(x,),I(x,),S(x,),K(x )] l I≤t, r=1,2,…,m (21) The first inequality constraint of(2 1)can be tten as 膏 A 叠一2 ATy+夕 ≤8 (22) or equivalently IIzl l≤2z。z (23) where z:缸,Zo=l,z。= 82 ATy一 (24) Here(z,zo,z1)belongs to a rotated second.order cone.Taking the linear relationship among叠,z,z0 and zl into account,we can obtain a linear constraint: AYe=b where = 】∈R4Ⅲ =[叠,z,z0,z1]∈R A~ 0=:『l1   ;TA  i: 一 .‘.. 1(J R(4n+2)x(4n +5m+2)一1_Negative identity matrix And the original problem can be turned into a second一0r— der cone problem as Ll—norm minimization for quatemion signals 35 min ∈R 1.O 0.9 O.8 s.t.b~=A~Ye II【n(x ),I(x,),J(x,),K(x )】 Il:≤t, IzII:≤2 z0z。 =r=1,2,…,m (25) O.7 O.6 0.5 O.4 0.3 [ ,0]∈R4 ~ =[ ,z,z0,z1]∈R4 O.2 0.1 (16)and(25)are the standard forms of the SOCP problem and can be solved by using several mature tool— boxes,such as SeDuMi[15]Then,we can easily obtain .0 20 4O 6o 80 1oo 120 14o 160 Spamity level of onginal signal 0 hte recovered quaternion signal xr from or . 3 Numerical Experiments In a noiseless scenario,just as in Ref.1 1 l,we present numerical experiments that indicate empiircal bounds on sparsity s(time domain suppo ̄of the input signa1)rela— tive to n(the number of measurements)for perfectly re. coveflng a quaternion signal x.The results can be seen as a set of practical guidelines for situations in which one ex. g】置 口 皇田田 ^日 0 0— HrI pects perfect recovery from random Gaussian quaternion measurements usi∞∞加ng SOCP ∞∞∞∞Numer加 ical experiments are carded out as follows: 1)An n by m(m=512 is the length of input signal。n is the number of measurements、random Gaussian meas. urement matirx A∈Q (n≤m)is produced wiht ran— dom entries sampled from an independent and identically distributed(i.i.d.)Gaussian process with a zero mean nad a variance equaling 1(in quatemion L2-norm sense). 2)Sparse quatemion input signal x∈Q~ is produced by selecting a suppo ̄set of size J J=s(sparsity) uniformly at random and sampling a vector x on T with i.i.d.Gaussian entries. 3)Quaternion output signal Y∈Q is obtianed by multiplying A with input x. 4)The vectors , , and matrix A are constructed as descnbed in(17)to(20). 5)SeDuMi toolbox【15j is called to solve SOCP problem (16)and the error is computed.which is the L,一nornl of hte difference between the recovered signal xr and the in— put signal x,i.e.,II 一x We perform experiments 100 times for each pair of s and n,then we save these errors and count the number of perfect recovered experiments.The criteflon for perfect recovery is chosen as IIxr—x ≤10~. Fig.1 shows the recovery success rate of 512一length signals with different measurement numbers and sparsity level configurations.The image intensity indicates a SUC— cess ratio with sparsity level s and measurements number n.Fig.2 depicts its cross section at five different values of n.It can be seen from this figure that for n≥32.the recovery rate is about 80%when s≤n/5 and practically 100%when s≤n/8.These results are very similar to those reported in Ref.[1]which deals with the real sig- nal with length m=512. Fig.1 Recovery experiment for m=512 ITl/n Fig.2 Cross section of Fig.1 at n=8,16,32,64 and 128 To further illustrate the recovery results,Fig.3 and Fig.4 depict the original quaternion signal x and its corre— sponding recovered signal xr.In Fig.3,m=512,n= 160.and s=60 and the quaternion signal is perfectly re— covered.In Fig.4,m=512, =160,and s=120 and in this case the recovery process failed. 喜一 竺0 tO0 200 300 400 500 =! 竺=!鬯喜一 0 100 200 300 400 500 !竺!=空±堂 (a) (b) 0 100 2oo 3o0 400 500 0 100 2oo 3oo 400 500 Index Index (e) (f) 耋一 E0 100 200 3oo 400 500 ===== ==:型 一 E0 100 200 30o 40o 500 !== 竺 Index Index (g) (h) Fig.3 Illustration of successful recovery of quaternion signal with sparsity level s=60.(a)Original signal,r part;(b)Recov— ered signal,r part;(c)Original signal,i part;(d)Recovered signal, i part;(e)Original signal,J part;(f)Recovered signal,J part; (g)Original signal,k part;(h)Recovered signal,k part In the noise case,the original signal x and measure— ment matirx A are generated as in the noiseless case n= 36 Zhang Xu,Wu Jiasong,Yang Guanyu,L0tfi Senahdji,and Sbu Huazhong 160.m=512.s=60.The measurements are corrupted by the white quaternion Gaussian noise vector comprised 一 of i.i.d.Gaussian variables with mean 0 and variance ,so the squared L,一nornl of noise vector l_e is a chi— j Index Index (a) (b) square random variable with mean 4no- and standard de— viation 2 耽r .Since the probability that l le exceeds its mean plus two or three standard deviations is smal1, 》一 一 Index Index here we choose the cons ̄aint parameter 8 to be s /4n+4 [『。2 耋一2 Index Index (a) (b) 。2 主一2 Index Index (g) (h) Fig.4 Illustration of failed recovery of quatemion signal with sparsity level s=120.(a)Original signal,r part;(b)Recovered signal, part;(c)Original signal,i part;(d)Recovered signal,i part;(e)Original signal, part;(f)Recovered signal, part;(g) Original signal,k part;(h)Recovered signal,k part We carried out ten recovery trials at different noise lev— els,and the average recovery errors with respect to corre— sponding noise strengths in those experiments are shown in Tab.1.The results in Tab.1 show that the proposed recovery algorithm is robust to noise contamination,and the recovery errors are about two times the noise strengths.An example of a recovery of sparse signals with noisy measurements is shown in Fig.5(o-=0.5). Tab.1 Recovery errors with respect to corresponding noise strengths 4 Conclusions and Future Work In this paper,we propose an algorithm for solving the L1一norl/1 minimization problem of quaternion signals, which is converted to SOCP and then solved by SeDuMi software.Numerical examples are provided to illustrate the feasibility of the algorithm.The results can be viewed as a set of practical guidelines for situations where one 0 一 Index Index (e) (f) ∞2 主一2 Index Index (g) (h) Fig.5 Example of sparse signal recovery with noisy measure— ments.(a)Original signal,rpart;(b)Recovered signal,rpart;(c) Original signal,i part;(d)Recovered signal,i part;(e)Original sig— nal, part;(f)Recovered signal, part;(g)Original signal,k part; (h)Recovered signal,k part expects perfect or stable recovery from random Gaussian quaternion measurement matirx information using SOCP. The main advantage of the proposed algorithm is that when converting the quaternion values optimization to that of rea1 values.many mature toolboxes can be applied. However,the converting process decouples the real and imaginary parts of the quaternion signals and no prior phase information is exploited.Further research includes: l、Considering the prior phase information in quaternion signals optimization ;2)Extending the quaternion sig。 nal vector L..norln minimization algorithm to that of qua- ternion matrix nuclear norifl minimization,that is,qua— temion matirx completion 16 J;3)Studying the problem of robust quaternion principal component analysis . References [1]Cand ̄s E,Romberg J,Tao T.Robust unce ̄ainty princi— ples:exact signal reconstruction from highly incomplete frequency nformation I 。\.IEEE Transactions on Infor- mation Theory。2006,52(2):489—509. 『2]Donoho D.Compressed sensing 1 J 1.IEEE mnsactions on Information Theory,2006,52(4):1289—1306. [3]Cand ̄s E,Romberg J,Tao T.Stable signal recovery rfom incomplete and inaccurate measurements[J].Corn— munications on P“ and Applied Mathematics.2006.59 (8):1207—1223. 『4]Winter S。Kellermann W,Sawada H,et a1.MAP—based underdetermined blind source separation of convolutive mixtures by hierarchical clustering and Ll—norin minimi— zation i 、.EURASIP Journat on Advances in Signal Processing.2007,2O07(1):81. [5]Yu S,Khwaja A S,Ma J.Compressed sensing of com— plex—valued data[J].Journal on Signal Processing, 2012.92(2):357—362. 『6]Hamilton W R.On quaternions[c]//Proceedings of Ll—ilorl/l minimization for quaternion signals 37 Royal Irish Academy.Dublin,Ireland,1844:1—16. component naalysis algorithm for quaternion signals[J]. [7]Ell T,Sangwine S.Hypercomplex Fourier transforms of 衄Transactions on Neural Networks,201l,22(12): color images[J].IEEE Transactions on Image Process— 1967—1978. ing,2007,16(1):22—35. [13]Le Bihan N,Buchholz S.Quaternionic independent com— [8]Via J,Ramirez D,Vielva L.Properness and widely line— ponent analysis using hypercomplex nonlinearities C\,, ar processing of quaternion rnadom vectors[J].IEEE 7th International Conference on Mathematics of Signal Transactions on Information Theory,2010,56(7):3502 Processing.Cirencester,UK,2006. —3515. [14] Bihan N,Sangwine S J.Quaternion principal component [9]Jinag T,Wei M.Equality consrtained least squares prob— naalysis of color images[C]//International Conference on lem over quatemion ifeld[J].Applied Mathematics Let- Iamge Processing.Barcelona,Spain,2003:808—812. ters,2003,16(6):883—888. [15]Sturrn J F.Using SeDuMi 1.02,a MATLAB toolbox for [1O]Jiang T,Zhao J,Wei M.A new technique of quaternion optimization over symmetric cones l J 1.Optimization equality constrained least squares problem[J].Journal Methods and Software,1999,11(1):625—653. of Computational and Applied Mathematics,2008,216 [16]Cand ̄s E,Recht B.Exact matrix completion via convex (2):509—513. optimization[J].Foundations of Computational Mathe— [11]Via J,Palomar D P,Vielva L,et a1.Quaternion ICA matics,2009,9(6):717—772. from second—order statistics[J].1EEE Transactions on [17]Cand ̄s E,Li X,Ma Y,et a1.Robust principal compo— Signal Processing,201I,59(4):1586—1600. nent analysis?[J].Journal ofACM,2011,58(3):1— [12]Javidi S,Took C C,Mandic D P.A fsat independent 37. 基于L 范数正则化的四元数信号重建 张旭 伍家松 ’。 杨冠羽 ' Lotif Senahdji ’ 舒华忠 , (。东南大学影像科学与技术实验室,南京210096) ( LTSI,INSERM U 1099,Universit6 de Rennes 1,Rennes 35000,Frnace) ( 中法生物医学信息研究中心,南京210096) 摘要:提出了一种通过求解L 范数最小化问题来重建四元数信号的算法,并且同时考虑了有噪声和没有噪 声2种应用场景.该算法首先将四元数域的L 范数最小化问题转化为实数域的二次锥规划问题,然后通过 工具包如SeDuMi来解决这个二次锥规划问题.为了验证所提出算法的正确性和有效性,进行了相关的数值 试验.试验结果表明:在没有噪声的情况下,在某些实际可接受的条件下原始信号的精确重建是可以实现 的;在有噪声的情况下,所提出的算法对于测量中的加性噪声具有鲁棒性.该算法可以被应用于四元数域基 于压缩感知理论的信号重建中. 关键词:四元数;信号重建;压缩感知 中图分类号:TP391 

因篇幅问题不能全部显示,请点此查看更多更全内容

Copyright © 2019- 7swz.com 版权所有 赣ICP备2024042798号-8

违法及侵权请联系:TEL:199 18 7713 E-MAIL:2724546146@qq.com

本站由北京市万商天勤律师事务所王兴未律师提供法律服务