肖宿
摘要:作為目前圖像處理領(lǐng)域的研究重點(diǎn),圖像復(fù)原可移除圖像中的模糊與噪聲,具有重要的理論價(jià)值和廣闊的應(yīng)用前景。為使圖像復(fù)原的研究被人們所了解,該文首先對(duì)圖像復(fù)原做了簡(jiǎn)單的描述,接著介紹了近年來(lái)出現(xiàn)的一些非盲圖像復(fù)原算法,包括基于總變分模型的算法、基于Bregman迭代的算法和基于稀疏表示的算法等,最后基于對(duì)現(xiàn)有算法的了解與分析,總結(jié)了圖像復(fù)原研究的難點(diǎn)與趨勢(shì)。
關(guān)鍵詞: 圖像復(fù)原; 總變分模型; Bregman迭代; 稀疏表示; 優(yōu)化問(wèn)題
中圖分類(lèi)號(hào):TP391 文獻(xiàn)標(biāo)識(shí)碼:A 文章編號(hào):1009-3044(2013)07-1642-03
由于噪聲、模糊等不利因素的影響,數(shù)字圖像的質(zhì)量通常難以令人滿意,無(wú)法對(duì)其進(jìn)一步進(jìn)行研究和利用。觀測(cè)到的退化與理想的原始圖像之間的關(guān)系可表示為:
式中,y表示觀測(cè)圖像;A表示模糊算子;x表示原始圖像;n表示加性噪聲。因此,圖像復(fù)原的目的是在模型(1)的框架下,估計(jì)最優(yōu)的原始圖像x,這是一種典型的線性逆問(wèn)題。圖像復(fù)原已成為圖像處理領(lǐng)域乃至計(jì)算機(jī)領(lǐng)域的研究熱點(diǎn),其研究可追溯到上個(gè)世紀(jì)六十年代,經(jīng)過(guò)逾半個(gè)世紀(jì)的發(fā)展,不斷有新的算法和技術(shù)涌現(xiàn)。目前,圖像復(fù)原算法主要分為非盲(non-blind)圖像復(fù)原算法和圖像盲復(fù)原算法兩大類(lèi)。當(dāng)模型(1)中的模糊算子A是已知的,圖像復(fù)原被稱(chēng)為非盲的圖像復(fù)原。非盲圖像復(fù)原算法主要有:基于總變分(TV,total variation)模型的算法、基于Bregman方法的算法和基于稀疏表示的算法等。
1 基于TV模型的圖像復(fù)原算法
TV模型亦稱(chēng)ROF模型,可表示為:
由L. I. Rudin、S. Osher和E. Fatemi提出[1],其最突出的優(yōu)點(diǎn)是抑噪的同時(shí)可保留圖像邊緣等重要信息,該模型已成為圖像復(fù)原、圖像去噪[2]和圖像修復(fù)[3]等領(lǐng)域使用最廣泛的先驗(yàn)?zāi)P椭?。由于TV模型是不可微分的,基于TV模型的算法需重點(diǎn)考慮圖像復(fù)原的數(shù)值解問(wèn)題。為求解經(jīng)典的TV/L2問(wèn)題,Y. L. Wang等人[4]建立了一種半二次化(quadratic minimization)模型,采用二次最小化方法計(jì)算該模型以提高圖像復(fù)原的效率。J. F. Wang等人[5]擴(kuò)展了Y. L. Wang等人的算法,將其用于彩色圖像復(fù)原問(wèn)題的處理。A. Beck等人[6]提出了一種基于Nesterov梯度法的TV圖像復(fù)原算法,該算法具有全局收斂速率?;谧兞糠蛛x和交替方向法(ADM,alternating direction method),M. K. Ng等人[7]將TV/L2模型分解為更簡(jiǎn)單的子問(wèn)題進(jìn)行求解,ADM方法的優(yōu)點(diǎn)是通過(guò)有限步迭代即可得到問(wèn)題的精確解。結(jié)合增廣拉格朗日法,S. H. Chan等人[8]提出了與M. K. Ng等人類(lèi)似的算法,該算法主要用于復(fù)原視頻圖像序列。盡管TV模型可保留圖像細(xì)節(jié),但復(fù)原圖像的平滑區(qū)域產(chǎn)生階梯效應(yīng)(staircase effect),因此一些研究提出在TV模型中加入高階項(xiàng)以克服其缺點(diǎn)。例如,F(xiàn). Li等人[9]將TV濾波器和四階PDE濾波器結(jié)合復(fù)原圖像,避免了階梯效應(yīng)的出現(xiàn)。原始-對(duì)偶法(primal-dual method)是一類(lèi)求解經(jīng)典TV圖像復(fù)原比較有效的方法。它最早由T. F. Chan等人提出[10],為了消除模型(2)目標(biāo)函數(shù)的奇異性和不可微性,T. F. Chan等人用新變量替換了其歐拉-拉格朗日方程中的?x/|?x|,得到如下的模型:
及其對(duì)偶模型,并通過(guò)半光滑牛頓法(semismooth Newton method)對(duì)模型(4)及其對(duì)偶模型進(jìn)行求解。M. Hintermuller等人的算法具有超線性的收斂速度,相比T. F. Chan等人的算法,該算法對(duì)對(duì)偶變量沒(méi)有過(guò)多的約束。E. Esser等人[12]修改了原始對(duì)偶混合梯度(PDHG,primal-dual hybrid gradient)算法,將TV最小化問(wèn)題轉(zhuǎn)化為等價(jià)的鞍點(diǎn)問(wèn)題(saddle point problem),并利用不精確Uzawa算法求解該鞍點(diǎn)問(wèn)題。由最終的實(shí)驗(yàn)結(jié)果,E. Esser等人的算法略優(yōu)于PDHG算法。A. Chambolle等人[13]將TV優(yōu)化問(wèn)題轉(zhuǎn)化為更通用的鞍點(diǎn)問(wèn)題,采用了與PDHG類(lèi)似的算法處理該鞍點(diǎn)問(wèn)題。針對(duì)不同的具體問(wèn)題,A. Chambolle等人的算法有著不同的收斂速率。
2 基于Bregman迭代的圖像復(fù)原算法
近年來(lái),在圖像復(fù)原領(lǐng)域,Bregman迭代因其簡(jiǎn)單、穩(wěn)定、速度快和效率高而受到越來(lái)越多的關(guān)注。Bregman迭代是一系列以Bregman距離為基礎(chǔ)的方法的總稱(chēng),包括了經(jīng)典的Bregman方法、線性(linerized)Bregman方法和分裂(split)Bregman方法等。Bregman迭代方法的基本思想是將優(yōu)化問(wèn)題分解為等價(jià)的非約束優(yōu)化子問(wèn)題,其中某些子問(wèn)題的目標(biāo)函數(shù)由Bregman距離定義。該方法最早S. Osher等人[14]引入圖像復(fù)原領(lǐng)域,用以改進(jìn)傳統(tǒng)方法對(duì)TV模型的處理。為提高經(jīng)典Bregman方法的性能,將該方法中的二次項(xiàng)0.5×||Ax-y||2 2替換為
3 基于稀疏表示的圖像復(fù)原算法
信號(hào)的稀疏表示來(lái)自對(duì)壓縮傳感問(wèn)題的研究,作為目前理論研究的一大熱點(diǎn),其在圖像處理領(lǐng)域已經(jīng)得到了較廣泛的應(yīng)用,基于稀疏表示的圖像復(fù)原算法亦越來(lái)越多。針對(duì)圖像復(fù)原問(wèn)題,K. Bredies等人[24]建立了以lp范式為約束項(xiàng)的優(yōu)化模型,并用迭代硬收縮(hard shrinkage)方法求解該優(yōu)化模型。A. Beck等人[25]為圖像復(fù)原問(wèn)題 minx{F(x) = f(x) + g(x)} 建立了二次逼近模型,并為計(jì)算該模型提供一種快速的迭代收縮-閾值算法,該算法比經(jīng)典的迭代收縮-閾值算法更快。F. X. Dupe等人[26]將圖像復(fù)原表示為凸泛函最小化問(wèn)題,該泛函由數(shù)據(jù)保真項(xiàng)(data-fidelity term),稀疏促進(jìn)項(xiàng)(sparsity-promoting term)和附加項(xiàng)組成。一種快速的向前-向后分裂方法被用于最小化泛函,為自適應(yīng)選擇參數(shù),文中還引入了廣義交叉檢驗(yàn)法(GCV,generalized cross validation)。基于邊界優(yōu)化的思想,婁帥等人[27]通過(guò)類(lèi)期望最大化獲得罰函數(shù)優(yōu)化問(wèn)題的解。該算法采用了輪廓波(contourlet)分解表示圖像,與基于小波變換的算法相比,運(yùn)算代價(jià)較低,能更好保護(hù)圖像的邊緣和細(xì)節(jié)信息。針對(duì)單一正則項(xiàng)存在的不足,N. Pustelnik等人[28]提出了包含多正則項(xiàng)的圖像復(fù)原問(wèn)題,一種快速的并行鄰近算法(parallel proximal algorithm)被用于處理該問(wèn)題。N. Pustelnik等人的算法可利用各種正則化技術(shù)的優(yōu)點(diǎn),并克服其相應(yīng)的缺點(diǎn)。小波基、小波框架、輪廓波等常作為字典用于分解圖像,然而它們無(wú)法自適應(yīng)地刻畫(huà)圖像的結(jié)構(gòu)特征,通過(guò)學(xué)習(xí)獲得的字典可具有更好的自適應(yīng)性。W. S. Dong等人[29]利用主成分分析(principal component analysis)技術(shù)學(xué)習(xí)圖像子字典(sub-dictionaries),并自適應(yīng)地選擇子字典組成字典分解圖像?;诜蔷植孔韵嗨萍s束(non-local self-similarity constraint)和自適應(yīng)的分段自回歸模型(piecewise autoregressive models),I. Daubechies等人[30]構(gòu)造了圖像復(fù)原的l1優(yōu)化問(wèn)題模型,并提出用迭代收縮算法處理該優(yōu)化問(wèn)題。
4 結(jié)論
圖像復(fù)原是一項(xiàng)具有廣闊發(fā)展空間和應(yīng)用前景的技術(shù)。隨著信號(hào)處理技術(shù)、控制技術(shù)、估計(jì)理論、數(shù)值分析方法的進(jìn)步,新的復(fù)原算法會(huì)不斷地涌現(xiàn)。對(duì)于未來(lái)可能出現(xiàn)的新算法,多通道圖像的復(fù)原、復(fù)原的實(shí)時(shí)性、退化參數(shù)的自適應(yīng)選擇、模糊的辨識(shí)、稀疏表示/壓縮傳感技術(shù)、Bregman迭代及其他優(yōu)化技術(shù)等將是其研究的重點(diǎn)內(nèi)容。
參考文獻(xiàn):
[1] Rudin L I, Osher S, Fatemi E. Nonlinear total variation based noise removal algorithms [J]. Physica D: Nonlinear Phenomena, 1992, 60(1-4): 259-268.
[2] 蔡澤民,賴(lài)劍煌.一種基于超完備字典學(xué)習(xí)的圖像去噪方法[J].電子學(xué)報(bào),2009,37(2):347-350.
[3] 張紅英,彭啟琮.數(shù)字圖像修復(fù)技術(shù)綜述[J].中國(guó)圖象圖形學(xué)報(bào),2007,12(1):1-10.
[4] WANG Y L, YANG J F, YIN W T, et al. A new alternating minimization algorithm for total variation image reconstruction [J]. SIAM Journal on Imaging Sciences,2008,1(3):248-272.
[5] YANG J F, ZHANG Y, YIN W T. An efficient TVL1 algorithm for deblurring multichannel images corrupted by impulsive noise [J]. SIAM on Journal Scientific Computing,2009,31(4): 2842-2865.
[6] BECK A, TEBOULLE M. Fast gradient-based algorithms for constrained total variation image denoising and deblurring problems [J].IEEE Transactions on Image Processing,2009,18(11): 2419-2434.
[7] NG M K, WEISS P, YUAN X M. Solving constrained total-variation image restoration and reconstruction problems via alternating direction methods [J]. SIAM Journal on Scientific Computing,2010,32(5):2710-2736.
[8] CHAN S H, KHOSHABEH R, GIBSON K B, et al. An augmented Lagrangian method for total variation video restoration[J].IEEE Transactions on Image Processing, 2011, 20(11): 3097-3111.
[9] LI F, SHEN C M, FAN J S, et al. Image restoration combining a total variational filter and a fourth-order filter[J].Journal of Visual Communication and Image Representation, 2007, 18(4): 322-330.
[10] CHAN T F, GOLUB G H, MULET P. A nonlinear primal-dual method for total variation-based image restoration [J].SIAM Journal on Scientific Computing,1999,20(6):1964-1977.
[11] HINTERMULLER M, STADLER G. An infeasible primal-dual algorithm for total bounded variation-based inf-convolution-type image restoration[J].SIAM Journal on Scientific Computing, 2006, 28(1): 1-23.
[12] ESSER E, ZHANG X Q, CHAN T F. A general framework for a class of first order primal-dual algorithms for convex optimization in imaging science [J].SIAM Journal on Imaging Sciences, 2010, 3(4): 1015-1046.
[13] CHAMBOLLE A, POCK T. A first-order primal-dual algorithm for convex problems with applications to imaging [J]. Journal of Mathematical Imaging and Vision, 2011,40(1):120-145.
[14] OSHER S, BURGER M, GOLDFARB D,et al. An iterative regularization method for total variation-based image restoration[J]. SIAM Journal on Multiscale Modeling & Simulation, 2005, 4(2): 460-489.
[15] YIN W T, OSHER S, GOLDFARB D, et al. Bregman iterative algorithms for l1-minimization with applications to compressed sensing, SIAM Journal on Imaging Sciences,2008,1(1):143-168.
[16] 李樹(shù)濤,魏丹.壓縮傳感綜述[J].自動(dòng)化學(xué)報(bào),2009,35(11):1369-1377.
[17] CAI J F, OSHER S, SHEN Z W. Linearized Bregman iterations for frame-based image deblurring [J]. SIAM Journal on Imaging Sciences,2009,2(1):226-252.
[18] GOLDSTEIN T, OSHER S. The split Bregman method for L1-regularized problems[J].SIAM Journal on Imaging Sciences,2009,2(2):323-343.
[19] SEZTER S. Operator splittings, Bregman methods and frame shrinkage in image processing [J]. International Journal of Computer Vision, 2011, 92(3): 265-280.
[20] CAI J F, OSHER S, SHEN Z W.Split Bregman methods and frame based image restoration [J]. SIAM Journal on Multiscale Modeling & Simulation, 2010, 8(2): 337-369.
[21] SETZER S, STEIDL G, TEUBER T. Deblurring Poissonian images by split Bregman techniques [J]. Journal of Visual Communication and Image Representation, 2010, 21(3): 193-199.
[22] LIU X W, HUANG L H. Split Bregman iteration algorithm for total bounded variation regularization based image deblurring [J]. Journal of Mathematical Analysis and Applications, 2010, 372(2): 486-495.
[23] ZHANG X Q, BURGER M, BRESSON X, et al. Bregmanized nonlocal regularization for deconvolution and sparse reconstruction [J]. SIAM Journal on Imaging Sciences, 2010, 3(3): 253-276.
[24] BREDIES K, LORENZ D A. Iterated hard shrinkage for minimization problems with sparsity constraints [J].SIAM Journal on Scientific Computing,2008,30(2):657-683.
[25] BECK A, TEBOULLE M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems[J].SIAM Journal on Imaging Sciences,2009,2(1):183-202.
[26] DUPE F X, FADILI J M, STARCK J L. A proximal iteration for deconvolving Poisson noisy images using sparse representations [J]. IEEE Transactions on Image Processing,2009,18(2): 310-321.
[27] 婁帥,丁振良,袁峰.基于Contourlet 變換的迭代圖像復(fù)原算法[J].光學(xué)學(xué)報(bào),2009,29(10): 2768-2773.
[28] PUSTELNIK N, CHAUX C, PESQUET J. Parallel proximal algorithm for image restoration using hybrid regularization [J]. IEEE Transactions on Image Processing,2011,20(9):2450-2462.
[29] DONG W S, ZHANG L, SHI G M, et al. Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization [J].IEEE Transactions on Image Processing, 2011, 20(7):1838-1857.
[30] DAUBECHIES I, DEFRIESE M, DE MOL C. An iterative thresholding algorithm for linear inverse problems with a sparsity constraint[J].Communications on Pure and Applied Mathematics, 2004, 57(11):1413-1457.