MPE Mathematical Problems in Engineering 1563-5147 1024-123X Hindawi Publishing Corporation 10.1155/2015/427153 427153 Research Article A Seismic Blind Deconvolution Algorithm Based on Bayesian Compressive Sensing Li Yanqin 1,2 Zhang Guoshan 1 Liu Wanquan 1 School of Electrical Engineering & Automation Tianjin University Tianjin 300072 China tju.edu.cn 2 Department of Disaster Prevention Equipment Institute of Disaster Prevention Beijing 101601 China fzxy.edu.cn 2015 352015 2015 19 03 2015 20 04 2015 352015 2015 Copyright © 2015 Yanqin Li and Guoshan Zhang. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Compressive sensing in seismic signal processing is a construction of the unknown reflectivity sequence from the incoherent measurements of the seismic records. Blind seismic deconvolution is the recovery of reflectivity sequence from the seismic records, when the seismic wavelet is unknown. In this paper, a seismic blind deconvolution algorithm based on Bayesian compressive sensing is proposed. The proposed algorithm combines compressive sensing and blind seismic deconvolution to get the reflectivity sequence and the unknown seismic wavelet through the compressive sensing measurements of the seismic records. Hierarchical Bayesian model and optimization method are used to estimate the unknown reflectivity sequence, the seismic wavelet, and the unknown parameters (hyperparameters). The estimated result by the proposed algorithm shows the better agreement with the real value on both simulation and field-data experiments.

1. Introduction

Digital seismic signal receiving devices have been developed for many years. In traditional seismic signal receiving system, the accurate signal is obtained by Nyquist sampling theorem which claims that the seismic signal sampling rate should be at least twice the seismic signal’s highest frequency. Measurement reduction, such as compressive sensing (CS), can make the seismic record devices be simpler, smaller, and cheaper. CS is a signal processing method which can reconstruct the original signal from a small number of random incoherent liner measurements [1, 2]. CS uses the property of the signal’s sparseness or compressibility in some domain, which achieves the entire signal from relatively few measurements . CS has become very popular in recent years because it can reduce the number of data transmissions and give potential aid in many practical applications [4, 5].

The CS problem is to compress data from original data while the inverse CS problem is to reconstruct original data from the compressed data. There exist lots of effective methods to solve the CS problem. Some of these methods are based on energy limit [6, 7] and the others are based on the Bayesian framework [8, 9]. Bayesian compressive sensing gives a Bayesian framework for solving the inverse problem of compressive sensing (CS). The basic Bayesian CS algorithm exploits the relevance vector machine and then it marginalizes the noise variance with improved robustness [10, 11].

Deconvolution, as a seismic data processing technique, has been widely used to yield reflectivity sequence from the seismic records . The Bernoulli-Gaussian model introduced in , as the reflectivity sequence model, is popular. It is very intuitive and its statistical properties can easily apply in Bayesian methods  such as Gibbs sampling and MCMC (Markov Chain Monte Carlo) techniques. In statistical models, an expectation maximization (EM) algorithm is an iterative method for finding maximum likelihood (ML) or maximum a posteriori (MAP) estimates of parameters, where the model depends on unobserved latent variables . So, EM algorithm is usually used to estimate the unknown parameters (the wavelet, parameters of the BG model and additive noise variance) in blind seismic deconvolution. The EM algorithm iteration alternates between performing an expectation (E) step and a maximization (M) step. E step creates a function for the expectation of the log-likelihood while M step computes parameters maximizing the expected log-likelihood found on the E step. These parameter-estimates are then used to determine the distribution of the latent variables in the next E step.

The Bayesian method is model-based method which is applied to estimate unknown parameters using the prior information. So it can handle observation noise and missing data problems . Moreover, sparsity, as the important characteristic of a signal, has been exploited in many application areas, such as seismic deconvolution; see . It could be an alternative method for the deconvolution of sparse signals that model a sparse signal with a Gaussian random variable whose variance is an inverse-Gamma random variable . It exploits variational Bayesian techniques for a closed-form approximate posterior inference .

In this paper, the reflectivity sequence and the seismic wavelet are recovered through CS measurement of the seismic records. Bayesian framework is used in CS problem in seismic signal processing. Blind seismic deconvolution and optimization method are combined to estimate the unknown reflectivity sequence, the seismic wavelet, and the hyperparameters. The organization of this paper is as follows. In Section 2, the Bayesian model with CS measurement of the blind seismic restoration problem is given. In Section 3, the observation model and the prior model on the signals coefficients are presented. In Section 4, the Bayesian inference is presented. Optimization method is used to get the solution of the unknown reflectivity sequence and the parameters. In Section 5, comparison between ML and Bayesian deconvolution with CS measurement is given based on the both simulation and field data experiments. In the last section, we draw the conclusions.

2. Problem Formulation

Seismic record is the convolution with the seismic wavelet and the reflectivity sequence. A standard seismic record image model is given in a matrix-vector form by (1)y=Sr+n,where y(n)=[y(1),y(M)]T denotes a seismic observation vector, r(n)=[r(1),,r(N)]T denotes an unknown seismic reflectivity sequence vector, and n(n)=[n(1),,n(M)]T is an additive white Gaussian noise vector, respectively. S, which represents seismic wavelet, is a M×N block Toeplitz matrix obtained by a block-circulant matrix. The seismic wavelet and reflectivity vector are both unknown, so the deconvolution is a blind problem.

Considering the sparsity of the seismic reflectivity vector, r is compressible by a linear wavelet basis Ψ; that is, r=Ψw, where w is a N×1 sparse signal. Reference  proposed sparse schemes that capture the convolution structure of pulse streams by defining (2) which resorts to a CS matrix Φ to complete MN measurements.

Equation (1) can be rewritten in the following form: (2)y=Φw+n,where Φ=SΨ. Based on the sparsity of w, the l1 norm of w is used in the inverse problem. The estimation of the original signal can be obtained by solving the following optimization problem : (3)w^=argminwy-Φw22+τw1,where ·1 represents the l1 norm.

3. Bayesian Modeling

The prior distribution of unknown signal w is assumed to be pwα. The observation y is the random process with distribution pyw,β, where β-1 denotes the unknown noise variance. These seismic record and reflectivity depend on the unknown hyperparameters α and β. Following the Bayesian framework of the CS inverse problem, the following joint posterior distribution can be expressed by (4)pw,α,β,y=pyw,βpwαpαpβ.

In Bayesian estimation, all unknown parameters are assumed to be random variables with specific distribution. Our goal is to estimate the parameters S, w, α, and β from y. The parameter S can be estimated by the Stochastic expectation maximization algorithm of Rosec et al. . The following gives how to get the other parameters.

3.1. Seismic Observation Modeling

By assuming that the seismic data noise in (2) is Gaussian distribution with zero mean and variance equal to β-1, the conditional distribution of the seismic record is (5)pyw,β=NyΦw,β-1,

with a Gamma prior distribution on β as the following:(6)pβaβ,bβ=Γβaβ,bβ.

The Gamma distribution for the hyperpriors β is defined as (7)px=Gxax,bx=bxaxΓaxxax-1exp-bxx,where x>0 represents hyperparameter, ax is the scale parameter, and bx is the shape parameter. These parameters are also assumed to be random variables. We will show how they are calculated in the following subsection. The Gamma distribution has the following mean and variance:(8)Ex=axbxVarx=axbx2.

3.2. Signal Modeling

In , as the reflectivity sequence also has sparsity, w can use the following prior model:(9)pwα=i=1NNwi0,αi,where α=α1,α2,,αN. Then, hyperprior of the hyperparameter αi is as the following: (10)pαiλ=Γαi1,λ2=λ2exp-λαi2,αi0,λ0.

Based on (9) and (10), we have the following equation:(11)pwλ=pwαpαλdα=ipwiαipαiλdαi=λN/22Nexp-λ1/2iwi.

At last, hyperparameter λ is modeled as the following Gamma hyperprior:(12)pλν=Γλν2,ν2.

The modeling has three stage forms. The first stage is the prior distribution about w in (9). The second stage is the prior distribution about α in (10). Finally, (12) is used to calculate λ. So we can first sample a Γν/2,ν/2 from the prior distribution on w distribution to get λ and then sample a Γ1,λ/2 distribution N times to get λi, i=1,,N, and at last sample a N0,αi distribution to get wi.

Based on the above, the joint distribution can be given as the following:(13)pw,α,λ,β,y=pyw,βpβpwαpαλpλ,where pyw,β, pβ, pwα, pαλ, and pλ are defined in (5), (6), (9), (10), and (12), respectively.

The seismic record variables of the proposed model are yi. The parameter is wi while the hyperparameters are αi and βi. The parameter of the hyperparameters model is λ. The dependencies among the random variables that define the proposed Bayesian compressive sensing model are shown in Figure 1.

Graphical model describing the data-generation process for the blind deconvolution problem.

Given the initial estimates q1w, q1α, q1β, and q1λ, we now proceed to explicitly calculate the distributions qkw, qk+1α, qk+1β, and qk+1λ (k=1,2) until a convergence criterion is met in the above algorithm. Then, the means above can be used to recalculate the distributions of reflectivity in the above algorithm. The performance of the algorithm heavily depends on the level of the noise and the initial distributions used in the iterative algorithm.

4. Bayesian Inference

The following Bayesian inference posterior distribution is well known; that is,(14)pw,α,λ,βy=pw,α,λ,β,ypy.

However, the posterior pw,α,λ,βy cannot be obtained just because (15)py=pw,α,λ,β,ydwdαdλdβ

and py is not easy to be calculated directly. So, the proper method should be given to perform the estimation. Evidence procedure is exploited to complete Bayesian inference in this paper.

Bayesian inference is derived by the evidence procedure which is based on the following decomposition:(16)pw,α,λ,βy=pwy,α,β,λpα,β,λy.

For pwy,α,β,λpw,y,α,β,λ, the distribution of pwy,α,β,λ is assumed to be a multivariate Gaussian distribution Nwμ,Σ. The mean μ and the variance Σ are(17)μ=ΣβΦTyΣ=βΦTΦ+Λ,with (18)Λ=diag1αi.

Since the distribution pα,β,λy=py,α,β,λ/py is proportional to py,α,β,λ, the hyperparameters can be obtained by maximizing the joint distribution py,α,β,λ. So we get (19)py,α,β,λ=pyw,βpwαpαλpβdw=12πN/2β-1I+ΦΛ-1ΦT-1/2·exp-12yTβ-1I+ΦΛ-1ΦT-1ypαλpλ·pβ=12πN/2C-1/2exp-12yTC-1y=pαλpλpβ,where I is M×M identity matrix and C=βI+ΦΛ-1ΦT.

Taking the logarithm of (19) and then maximizing the result, we can get the following function:(20)L=12logC-12yTC-1y+Nlogλ2-λ2iαi+ν2logν2-logΓν2+ν2-1logλ-ν2λ+aβ-1logβ-bββ.

In order to get the solution of the maximization problem, we give the following equations below. Consider(21)C=Λ-1β-1IΛ+βΦTΦ=Λ-1β-1IΣ-1.

Taking the logarithm of (21),(22)logC=-logΛ-Nlogβ-logΣ,where (23)C-1=β-1I+ΦΛ-1ΦT-1=βI-βΦΛ+βΦTΦ-1ΦTβ=βI-βΦΣΦTβ.

The following function is obtained by (23) and (17): (24)yTC-1y=βyTy-βyTΦΣΦTβy=βyTy-Φμ=βy-Φμ2+βμTΦTy-Φμ=βy-Φμ2+μTΛμ.

Therefore, based on the above equations and differential of the equation L with respect to αi, we obtain(25)dLdαi=12-1αi+wi2αi2-λ,where wi2=μi2+Σii and Σii denotes the ith diagonal element of Σ. Making (25) be zero, we can obtain the solution of αi:(26)αi=-12λ+14λ2+wi2λ.

Similarly, the other hyperparameters are updated in the following manner, which is given by(27)λ=N-1+ν/2Σiαi/2+ν/2β=N/2+aβy-Φw2/2+bβ.

The hyperparameter ν satisfies the following equation:(28)logν2+1-ϕν2+logλ-λ=0,where φ·=lnΓx.

Generally speaking, the hyperparameters α,λ,β,ν are iterated by (26), (27), and (28). The parameter w is estimated by (17).

5. Experiments Results

In order to demonstrate the performance of the proposed Bayesian compressive sensing algorithm, we give experimental results on both simulated and field data experiments in this section.

5.1. Simulation Data

The proposed Bayesian compressive sensing algorithm can improve the estimations quality compared to ML method and standard Bayesian method in the situation where there are noise and incomplete data. In this section, we support these theoretical arguments with one dimensional synthetic signals simulation experiments. The proposed Bayesian compressive sensing algorithm is terminated when the criterion Lγk-Lγk+12<10-4 is satisfied. The estimation quality of the three algorithms is compared in the following three parts: the reflectivity sequence recovery, the parameters estimation, and performance indices nMSE. Let us define the signal-to-noise ratios (SNR) as(29)SNR=10logλαr-1Esαw-1,where Es is the seismic wavelet energy and N is the seismic wavelet length, Es=N-1Σi=1Nsi2.

A reflectivity sequence with 100 samples by BG model is selected randomly as shown in Figures 2(a) and 3(a). The Stochastic wavelet with 45 samples is shown in Figure 2(b) while the minimum phase wavelet with 35 samples is shown in Figure 3(b). Then, the reflectivity sequence in every figure can be convolved with the corresponding Ricker and minimum phase wavelet and added white Gaussian noise, with SNRs of 10 dB and 20 dB, respectively. The corresponding seismic records with the noise are shown in Figures 2(c) and 3(c), respectively. The reflectivity sequence estimated with 10 dB and 20 dB SNRs by the proposed Bayesian compressive sensing algorithm is shown in Figures 2(d) and 3(d), respectively. It can be seen that the Bayesian compressive sensing performs well in reflectivity deconvolution no matter the seismic wavelet is wavelet or minimum phase wavelet.

Synthetic reflectivity, wavelet, and data sets. (a) Synthetic 1D reflectivity sequence. (b) 1D seismic wavelet. (c) 1D seismic data (SNR = 10 dB). (d) Estimated reflection coefficients (Bayesian compressive sensing).

Synthetic reflectivity, wavelet, and data sets. (a) Synthetic 1D reflectivity sequence. (b) 1D seismic wavelet. (c) 1D seismic data (SNR = 20 dB). (d) Estimated reflection coefficients (Bayesian compressive sensing).

Figures 4, 5, and 6 show the comparison between the true reflectivity sequence and deconvolved reflectivity sequence by proposed Bayesian compressive sensing method, ML method, and standard Bayesian method with SNRs of 10 dB, 20 dB, and 40 dB, respectively. It can be seen in Figures 46 that the Bayesian compressive sensing gives a better reflectivity recovery than the other algorithms. The reflectivity sequence by the Bayesian compressive sensing is more accurate because it has less missed and false detection. When SNR increases, the reflectivity impulse miss number decreases, because both high-amplitude impulse and low-amplitude impulses are well detected and localized. It can be remarked that the tree algorithms show good robustness with regard to the SNR decrease.

(a) True reflectivity and estimated reflectivity by Bayesian compressive sensing (BCS) (SNR = 10 dB). (b) True reflectivity and estimated reflectivity by ML (SNR = 10 dB). (c) True reflectivity and estimated reflectivity by standard Bayesian (SNR = 10 dB).

(a) True reflectivity and estimated reflectivity by Bayesian compressive sensing (SNR = 20 dB). (b) True reflectivity and estimated reflectivity by ML (SNR = 20 dB). (c) True reflectivity and estimated reflectivity by standard Bayesian (SNR = 20 dB).

(a) True reflectivity and estimated reflectivity by Bayesian compressive sensing (SNR = 40 dB). (b) True reflectivity and estimated reflectivity by ML (SNR = 40 dB). (c) True reflectivity and estimated reflectivity by standard Bayesian (SNR = 40 dB).

Furthermore, to study improvements brought by the proposed Bayesian compressive sensing algorithm, we consider the following performance indices: (30)nMSE1=r-r^2r2,nMSE2=rs-r^s^2rs2,where nMSE1 represents the normalized error energy of the estimated wavelet, while nMSE2 represents the normalized error of noiseless data reconstruction. Comparison in nMSE1 and nMSE2 of the three deconvolution algorithms is shown in Figure 7. The results show the improvement of the performance indices with the Bayesian compressive sensing algorithm compared with the standard Bayesian algorithm and ML algorithm. Although all algorithms show larger errors while SNR decreased, the proposed Bayesian compressive sensing algorithm suffers from noise too, and it is still more reliable than the other algorithms. Moreover, the Bayesian compressive sensing behaves much better when there is almost no noise, because the Bayesian compressive sensing uses the signals prior knowledge and the linear wavelet transform.

nMS E 1 and nMSE2 performance indices for ML algorithm, standard Bayesian, and Bayesian compressive sensing.

5.2. Field Data Experiments

Figure 8(a) shows a real CDP stack seismic profile. The reflectivity estimates obtained by ML algorithm, variational Bayesian, and Bayesian compressive sensing are shown in Figures 8(b), 8(c), and 8(d), respectively. Because the true layer structure is unknown, the continuous nature of the channel estimates obtained by the proposed algorithm can be seen. New layers become much clearer in the deconvoluted seismic image by variational Bayesian algorithm. The translations between layers are much thinner and less blurred in the deconvoluted seismic image by Bayesian compressive sensing algorithm, especially near CDP 225 at 0.5 s.

(a) A real CDP profile. (b) Reflectivity estimates obtained by ML algorithm. (c) Reflectivity estimates obtained by variational Bayesian algorithm. (d) Reflectivity estimates obtained by Bayesian compressive sensing algorithm.

6. Conclusions

Convolution and additive Gaussian noises are the major problems in blind seismic signal processing. In order to solve two problems, a novel blind seismic deconvolution algorithm based on Bayesian compressive sensing has been proposed in this paper. Bayesian modeling and the optimization method are used to estimate the unknown reflectivity sequence, seismic wavelet, and hyperparameters simultaneously in the proposed algorithm. Furthermore, we compared the proposed Bayesian compressive sensing algorithm with ML algorithm and standard Bayesian algorithm. Seismic images are much clear and accurate by the proposed Bayesian compressive sensing algorithm on both simulation data and field data.

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Ma J. Compressed sensing for surface characterization and metrology IEEE Transactions on Instrumentation and Measurement 2010 59 6 1600 1615 10.1109/TIM.2009.2027744 2-s2.0-77952293052 Duarte M. F. Baraniuk R. G. Spectral compressive sensing Applied and Computational Harmonic Analysis 2013 35 1 111 129 10.1016/j.acha.2012.08.003 MR3053749 2-s2.0-84876575108 Erlich Y. Gordon A. Brand M. Hannon G. J. Mitra P. P. Compressed genotyping IEEE Transactions on Information Theory 2010 56 2 706 723 10.1109/tit.2009.2037043 MR2724402 2-s2.0-77649284211 Amir A. Zuk O. Bacterial community reconstruction using compressed sensing Journal of Computational Biology 2011 18 11 1723 1741 10.1089/cmb.2011.0189 MR2860073 2-s2.0-80955139597 Duarte M. F. Davenport M. A. Takbar D. Laska J. N. Sun T. Kelly K. F. Baraniuk R. G. Single-pixel imaging via compressive sampling: building simpler, smaller, and less-expensive digital cameras IEEE Signal Processing Magazine 2008 25 2 83 91 10.1109/msp.2007.914730 2-s2.0-41949104024 Herman M. A. Strohmer T. High-resolution radar via compressed sensing IEEE Transactions on Signal Processing 2009 57 6 2275 2284 10.1109/TSP.2009.2014277 MR2641823 2-s2.0-66849132093 McMackin L. Herman M. A. Chatterjee B. Weldon M. A high-resolution SWIR camera via compressed sensing Infrared Technology and Applications XXXVIII 2012 Baltimore, Md, USA 835303 835303-10 Proceedings of SPIE 10.1117/12.920050 Candès E. J. Romberg J. Tao T. Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information IEEE Transactions on Information Theory 2006 52 2 489 509 10.1109/tit.2005.862083 MR2236170 2-s2.0-31744440684 Donoho D. L. Compressed sensing IEEE Transactions on Information Theory 2006 52 4 1289 1306 10.1109/tit.2006.871582 MR2241189 2-s2.0-33645712892 Qisong W. Zhang Y. D. Amin M. G. Himed B. Multi-task bayesian compressive sensing exploiting intra-task dependency IEEE Signal Processing Letters 2015 22 4 430 434 10.1109/lsp.2014.2360688 Babacan S. D. Molina R. Katsaggelos A. K. Bayesian compressive sensing using Laplace priors IEEE Transactions on Image Processing 2010 19 1 53 63 10.1109/tip.2009.2032894 MR2729957 2-s2.0-72949095917 Minato S. Matsuoka T. Tsuji T. Singular-value decomposition analysis of source illumination in seismic interferometry by multidimensional deconvolution Geophysics 2013 78 3 Q25 Q34 Ju S. H. A deconvolution scheme for determination of seismic loads in finite-element analyses Bulletin of the Seismological Society of America 2013 103 1 258 267 10.1785/0120120034 2-s2.0-84873510576 Menanno G. M. Mazzotti A. Deconvolution of multicomponent seismic data by means of quaternions: theory and preliminary results Geophysical Prospecting 2012 60 2 217 238 10.1111/j.1365-2478.2011.00988.x 2-s2.0-84856216157 Wapenaar K. van der Neut J. Ruigrok E. Draganov D. Hunziker J. Slob E. Thorbecke J. Snieder R. Seismic interferometry by crosscorrelation and by multidimensional deconvolution: a systematic comparison Geophysical Journal International 2011 185 3 1335 1364 10.1111/j.1365-246x.2011.05007.x 2-s2.0-79956039235 Nsiri B. Chonavel T. Boucher J.-M. Nouzé H. Blind submarine seismic deconvolution for long source wavelets IEEE Journal of Oceanic Engineering 2007 32 3 729 743 10.1109/joe.2007.899408 2-s2.0-40449095741 Li Y. Zhang G. Blind seismic deconvolution using variational Bayesian method Journal of Applied Geophysics 2014 110 82 89 10.1016/j.jappgeo.2014.09.002 Yildirim S. Cemgil A. T. Aktar M. Özakin Y. Ertüzün A. A Bayesian deconvolution approach for receiver function analysis IEEE Transactions on Geoscience and Remote Sensing 2010 48 12 4151 4163 10.1109/tgrs.2010.2050327 2-s2.0-78649317243 Yuan S. Wang S. Spectral sparse Bayesian learning reflectivity inversion Geophysical Prospecting 2013 61 4 735 746 10.1111/1365-2478.12000 2-s2.0-84879461024 Candès E. J. Wakin M. B. An introduction to compressive sampling IEEE Signal Processing Magazine 2008 25 2 21 30 10.1109/msp.2007.914731 2-s2.0-41949092318 Robert C. P. The Bayesian Choice 2007 New York, NY, USA Springer MR2723361 Gholami A. Residual statics estimation by sparsity maximization Geophysics 2013 78 1 V11 V19 10.1190/GEO2012-0035.1 2-s2.0-84878793752 Wang D. L. Tong Z. F. Tang C. Zhu H. An iterative curvelet thresholding algorithm for seismic random noise attenuation Applied Geophysics 2010 7 4 315 324 10.1007/s11770-010-0259-8 2-s2.0-78651385214 Heimer A. Cohen I. Multichannel seismic deconvolution using Markov-Bernoulli random-field modeling IEEE Transactions on Geoscience and Remote Sensing 2009 47 7 2047 2058 10.1109/tgrs.2008.2012348 2-s2.0-67651174315 Cemgil A. T. Févotte C. Godsill S. J. Variational and stochastic inference for Bayesian source separation Digital Signal Processing 2007 17 5 891 913 10.1016/j.dsp.2007.03.008 2-s2.0-34547454757 Kim S. Yoo C. D. Underdetermined blind source separation based on subspace representation IEEE Transactions on Signal Processing 2009 57 7 2604 2614 10.1109/tsp.2009.2017570 MR2650176 2-s2.0-67650178769 Févotte C. Godsill S. J. A Bayesian approach for blind separation of sparse sources IEEE Transactions on Audio, Speech and Language Processing 2006 14 6 2174 2188 10.1109/TSA.2005.858523 2-s2.0-34347404627 Févotte C. Bayesian Audio Source Separation 2007 New York, NY, USA Springer Tzikas D. Likas A. Galatsanos N. Variational bayesian blind image deconvolution with student-t priors Proceedings of the 14th IEEE International Conference on Image Processing (ICIP '07) September 2007 San Antonio, Tex, USA IEEE I109 I112 10.1109/icip.2007.4378903 2-s2.0-48149086764 Beal M. Variational algorithms for approximate Bayesian inference [Ph.D. thesis] 2003 London, UK Gatsby Computational Neuroscience Unit, Univercity College Amizic B. Spinoulas L. Molina R. Katsaggelos A. K. Compressive blind image deconvolution IEEE Transactions on Image Processing 2013 22 10 3994 4006 10.1109/tip.2013.2266100 MR3105158 2-s2.0-84883748550 Rosec O. Boucher J.-M. Nsiri B. Chonavel T. Blind marine seismic deconvolution using statistical MCMC methods IEEE Journal of Oceanic Engineering 2003 28 3 502 512 10.1109/JOE.2003.816683 2-s2.0-0242667719 Kazemi N. Sacchi M. D. Sparse multichannel blind deconlution Geophysics 2014 79 5 143 152 10.1190/geo2013-0465.1