石保順 吳一凡 練秋生
摘 要:
編碼衍射成像旨在利用衍射強(qiáng)度圖樣重建原始圖像,而現(xiàn)有基于人工設(shè)計(jì)先驗(yàn)的編碼衍射成像算法大都在低信噪比下成像質(zhì)量低。通過基于深度神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)的深度先驗(yàn)?zāi)軌蚪鉀Q上述問題,但有監(jiān)督學(xué)習(xí)需要大規(guī)模樣本對,不利于實(shí)際應(yīng)用。針對這一問題,本文提出一種基于無監(jiān)督學(xué)習(xí)的編碼衍射成像方法。該方法結(jié)合雙數(shù)據(jù)保真項(xiàng)、卷積稀疏編碼模型和深度圖像先驗(yàn)?zāi)P蜆?gòu)建了能夠融合互補(bǔ)先驗(yàn)的優(yōu)化模型,并利用交替優(yōu)化方法對其進(jìn)行有效求解。實(shí)驗(yàn)結(jié)果表明,提出的方法能夠在低信噪比下僅通過單幅編碼衍射強(qiáng)度圖樣重建出高質(zhì)量的圖像。
關(guān)鍵詞:
計(jì)算成像;衍射成像;無監(jiān)督學(xué)習(xí);深度圖像先驗(yàn);卷積稀疏編碼
中圖分類號: TP391.41, O436? 文獻(xiàn)標(biāo)識碼: A? DOI:10.3969/j.issn.1007-791X.2023.01.006
0 引言
編碼衍射成像作為相位恢復(fù)領(lǐng)域的一個研究熱點(diǎn),在生物學(xué)[1]、醫(yī)學(xué)[2]、光學(xué)[3-4]及天文學(xué)[5]領(lǐng)域有著廣泛應(yīng)用。編碼衍射成像系統(tǒng)在觀測物體后加入隨機(jī)相位板對物體結(jié)構(gòu)信息進(jìn)行調(diào)制,通過探測器記錄編碼衍射圖樣(Coded diffraction patterns, CDP)。由于采樣設(shè)備的限制,探測器只能記錄衍射圖樣的強(qiáng)度值,導(dǎo)致測量的數(shù)據(jù)丟失了相位信息。因此通過CDP重建原始圖像是一個高度不適定的非凸、非線性問題,如何對該問題進(jìn)行有效求解是一大挑戰(zhàn)。
解決該非凸優(yōu)化問題的一個思路是通過數(shù)學(xué)方法構(gòu)建較好的初始解并利用基于梯度的迭代算法進(jìn)行求解,主要代表算法有WF (Wirtinger flow)算法[6]、TWF (Truncated wirtinger flow)算法[7]、TAF (Truncated amplitude flow)算法[8]等。該類方法未利用先驗(yàn)信息,導(dǎo)致重建高質(zhì)量圖像需要多幅編碼衍射圖樣。為解決該問題,基于人工設(shè)計(jì)先驗(yàn)的算法利用圖像固有先驗(yàn)信息進(jìn)行編碼衍射成像,例如基于稀疏性的衍射成像算法[9-13]及基于非局部相似性的衍射成像算法[14-17]等。Tillmann等[9]利用圖像在自適應(yīng)字典下的稀疏性,提出了DOLPHIn算法用于編碼衍射成像。Chang等[10]通過構(gòu)造全變差正則項(xiàng)引入梯度稀疏性來提高相位恢復(fù)的重建質(zhì)量。Shi等[11]提出了一種正則化與交替投影框架相結(jié)合的編碼衍射成像算法,有效地將BM3D (Block matching and 3D collaborative filtering)框架下的稀疏性引入到圖像重建中。Katkovnik等基于物體相位和振幅在變換域的稀疏性,結(jié)合GS (Gerchberg-Saxton)[18]算法提出了能夠進(jìn)行像素分辨率衍射成像的SPAR (Sparse phase amplitude reconstruction)[14]算法及亞像素分辨率衍射成像的SR-SPAR (Super-resolution sparse phase amplitude retrieval)[15]算法。上述基于人工設(shè)計(jì)先驗(yàn)的編碼衍射成像算法利用的先驗(yàn)是通過解析方法刻畫的,難以充分描述所有圖像的統(tǒng)計(jì)分布信息。因此,利用人工設(shè)計(jì)先驗(yàn)的編碼衍射成像算法的重建質(zhì)量有很大提升空間。
近些年,深度學(xué)習(xí)技術(shù)引起了學(xué)者們的廣泛關(guān)注,該技術(shù)能夠利用大規(guī)模數(shù)據(jù)集和深度神經(jīng)網(wǎng)絡(luò)強(qiáng)大的表示能力學(xué)習(xí)對訓(xùn)練數(shù)據(jù)集最優(yōu)的網(wǎng)絡(luò)模型[19-20]。Hand等[21]提出了一種基于生成先驗(yàn)的相位恢復(fù)框架,證明了在相位恢復(fù)任務(wù)中利用生成模型的策略優(yōu)于基于稀疏性的相位恢復(fù)算法。針對相位恢復(fù)問題,Morales等[22]提出了一種端到端的方法,該方法通過聯(lián)合學(xué)習(xí)譜初始化和深度神經(jīng)網(wǎng)絡(luò)參數(shù)來解決相位恢復(fù)問題。上述基于有監(jiān)督學(xué)習(xí)方法訓(xùn)練的深度神經(jīng)網(wǎng)絡(luò)對特定信噪比是最優(yōu)的,但對其他信噪比是非最優(yōu)的。因而,在不同信噪比下,上述方法需要重新訓(xùn)練深度神經(jīng)網(wǎng)絡(luò),靈活性較差。為彌補(bǔ)該不足,即插即用方法將事先訓(xùn)練好的深度高斯去噪器引入到優(yōu)化過程中。其中Metzler等[23]將卷積神經(jīng)網(wǎng)絡(luò)去噪器引入到去噪正則化(Regularization by denoising, RED)[24]框架提出了prDeep算法,Shi等[25]利用圖像與去噪圖像的相似性提出了能夠融合深度去噪器先驗(yàn)的稀疏表示正則化模型,有效提高了編碼衍射成像的成像質(zhì)量及分辨率。
上述編碼衍射成像網(wǎng)絡(luò)或深度高斯去噪器都采用有監(jiān)督的學(xué)習(xí)方式進(jìn)行學(xué)習(xí),泛化能力較差,并且深度神經(jīng)網(wǎng)絡(luò)訓(xùn)練需要大規(guī)模數(shù)據(jù)集。為解決該問題,Ulyanov等[26]提出一種稱為深度圖像先驗(yàn)(Deep image prior, DIP)的無監(jiān)督深度學(xué)習(xí)框架。圖像的統(tǒng)計(jì)分布可以由深度神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)本身所表征,基于該觀測,DIP能夠僅通過退化圖像學(xué)習(xí)網(wǎng)絡(luò)參數(shù)并表征圖像。DIP的提出為深度學(xué)習(xí)提供了一種新的思路,即無監(jiān)督學(xué)習(xí),其參數(shù)并不需要通過大量的外部數(shù)據(jù)集進(jìn)行學(xué)習(xí)。然而,DIP框架存在過擬合的問題,解決該問題最直接的方式是早停技術(shù),但最優(yōu)的早停迭代次數(shù)難以精確獲得。為彌補(bǔ)該不足,Cheng等[27]提出了基于隨機(jī)梯度Langevin動力學(xué)進(jìn)行后驗(yàn)推斷的DIP,該方法不需要早停技術(shù)就能夠獲得滿意解。為進(jìn)一步提升DIP求解圖像反問題的性能,學(xué)者們將傳統(tǒng)先驗(yàn)與DIP相結(jié)合以提升重建質(zhì)量,例如Sun等[28]將即插即用先驗(yàn)與DIP相結(jié)合,解決了由過擬合導(dǎo)致的DIP算法重建質(zhì)量低的問題。Cascarano等[29]將梯度稀疏性與DIP相結(jié)合進(jìn)行了高質(zhì)量的圖像復(fù)原。Mataev等[30]通過引入RED模型,將非局部相似性與DIP相結(jié)合提出了DeepRED算法,該算法較原始RED和DIP算法的性能具有顯著提升。
上述無監(jiān)督學(xué)習(xí)方法針對線性、凸問題取得了較好的效果,但當(dāng)采用無監(jiān)督學(xué)習(xí)方法求解編碼衍射成像等非線性、非凸優(yōu)化問題時,仍存在以下挑戰(zhàn):1)現(xiàn)有基于深度圖像先驗(yàn)的方法并未充分利用圖像固有先驗(yàn)信息,如何挖掘并利用互補(bǔ)的圖像先驗(yàn)信息提升重建質(zhì)量是一大挑戰(zhàn);2) DeepRED算法能夠利用互補(bǔ)先驗(yàn)知識進(jìn)行圖像重建,但估計(jì)圖像過程中并未充分挖掘測量數(shù)據(jù)包含的待重建圖像信息。為解決這兩個問題,受卷積稀疏編碼(Convolutional sparse coding, CSC)[31-32]及雙數(shù)據(jù)保真項(xiàng)在單曝光壓縮成像[33]領(lǐng)域成功應(yīng)用的啟發(fā),本文通過雙數(shù)據(jù)保真項(xiàng)、卷積稀疏編碼模型和深度圖像先驗(yàn)?zāi)P吞岢鲆环N基于無監(jiān)督學(xué)習(xí)的編碼衍射成像算法。在本文提出的方法中,CSC模型通過線下訓(xùn)練好的卷積字典表示圖像,DIP通過退化圖像優(yōu)化深度神經(jīng)網(wǎng)絡(luò)的參數(shù),實(shí)現(xiàn)深度神經(jīng)網(wǎng)絡(luò)的“線上”訓(xùn)練。兩者有效結(jié)合,理論上能夠有效融合互補(bǔ)先驗(yàn)知識。雙數(shù)據(jù)保真項(xiàng)有利于進(jìn)一步挖掘編碼衍射強(qiáng)度圖樣中的信息,輔助DIP框架中參數(shù)的無監(jiān)督學(xué)習(xí)。因此,與現(xiàn)有編碼衍射成像算法相比,本文提出的算法有望在低信噪比情況下僅通過單幅編碼衍射圖樣重建更高質(zhì)量的圖像。
3 分析與討論
3.1 實(shí)驗(yàn)細(xì)節(jié)與參數(shù)設(shè)置
測試圖像如圖1所示。其中前6幅是prDeep算法軟件包(https://github.com/ricedsp/prDeep)中大小為512×512的標(biāo)準(zhǔn)化灰度圖像。后14幅圖像來源于細(xì)胞圖像庫(The Cell Image Library: http://www.cellimagelibrary.org/home),每幅原始細(xì)胞圖像被裁剪為512×512的標(biāo)準(zhǔn)化灰度圖像。
算法的參數(shù)通過經(jīng)驗(yàn)調(diào)整設(shè)置,具體地,實(shí)驗(yàn)過程中將外循環(huán)設(shè)置為i=40次,每次循環(huán)中DIP內(nèi)循環(huán)設(shè)置為2i次。使用ADAM作為優(yōu)化器,學(xué)習(xí)率初始值設(shè)置為0.001。投影次數(shù)J在信噪比(Signal-to-noise ratio, SNR)為5 dB,10 dB和15 dB三種情況下分別設(shè)為2,2和4。DCD-DIP算法其他參數(shù)設(shè)置如表1所示。本文通過小波域估計(jì)算法[42]對噪聲標(biāo)準(zhǔn)差進(jìn)行估計(jì),并采用與文獻(xiàn)[11]輸入噪聲標(biāo)準(zhǔn)差相同的的計(jì)算方式,即輸入標(biāo)準(zhǔn)差為估計(jì)標(biāo)準(zhǔn)差乘以常數(shù)Q。
所有算法都在同一個四元隨機(jī)掩模條件下進(jìn)行實(shí)驗(yàn)與比較。所有實(shí)驗(yàn)均采用隨機(jī)初始值,隨機(jī)種子固定。本文采用峰值信噪比(Peak signal to noise ratio, PSNR)和結(jié)構(gòu)相似性(Structure similarity index measure, SSIM)作為評價算法重建性能的客觀指標(biāo),PSNR值越高,說明圖像重建質(zhì)量越高;SSIM值越高,說明兩幅圖像的相似程度越高。所有實(shí)驗(yàn)均在Intel Core i9-10850k@3.60 GHz、內(nèi)存64 GB、NVIDIA GTX 3080Ti GPU的硬件平臺上進(jìn)行。
3.2 雙數(shù)據(jù)保真項(xiàng)對重建結(jié)果影響分析
為分析雙數(shù)據(jù)保真項(xiàng)的有效性,本節(jié)將僅利用DIP框架的成像算法、單數(shù)據(jù)保真項(xiàng)的成像算法及本文提出的采用雙數(shù)據(jù)保真項(xiàng)的DCD-DIP算法進(jìn)行編碼衍射成像實(shí)驗(yàn),并對實(shí)驗(yàn)結(jié)果進(jìn)行對比。對于單數(shù)據(jù)保真項(xiàng)的成像算法,去除式(8)中的第一項(xiàng),ρ設(shè)為1并將其他參數(shù)調(diào)制最優(yōu),其參數(shù)如表2所示。圖2中給出了在不同信噪比下三種算法重建測試圖像平均PSNR值和SSIM值的比較。從圖中可以看出,本文所提出的采用雙數(shù)據(jù)保真項(xiàng)的DCD-DIP算法在重建質(zhì)量及結(jié)構(gòu)相似性上明顯優(yōu)于另兩種算法,說明其能夠充分利用互補(bǔ)先驗(yàn)知識,挖掘測量數(shù)據(jù)包含的圖像信息進(jìn)行圖像重建。
3.3 低信噪比下多種編碼衍射成像算法的比較及分析
為驗(yàn)證本文算法在低信噪比下的有效性,本節(jié)在不同信噪比下與現(xiàn)有編碼衍射成像方法進(jìn)行對比。在信噪比分別為5 dB、10 dB和15 dB時,將本文提出DCD-DIP算法與DOLPHIn[9]、BM3D-prGAMP[17]、SPAR[14]和prDeep[23]算法重建圖像的性能進(jìn)行比較。表3給出了20幅測試圖像在不同信噪比下利用上述五種算法進(jìn)行圖像重建的平均PSNR值和SSIM值。由表3可知,本文所提出算法在不同信噪比下的PSNR值和SSIM值均明顯高于對比算法。DOLPHIn算法在CDP數(shù)量較少以及低信噪比的情況下重建性能較差。BM3D-prGAMP和SPAR算法利用圖像的非局部相似性及相似塊的稀疏性,能夠獲得比DOLPHIn算法高的重建質(zhì)量。prDeep算法通過有監(jiān)督訓(xùn)練的去噪器利用深度先驗(yàn)進(jìn)行重建,其重建質(zhì)量優(yōu)于基于BM3D的成像算法,但在信噪比較低情況下重建質(zhì)量仍然較差。而本文提出的算法通過融合互補(bǔ)的先驗(yàn)知識,采用雙數(shù)據(jù)保真項(xiàng)充分挖掘單幅編碼衍射強(qiáng)度圖樣中的信息,較上述四種對比算法的性能具有顯著的提升。
為進(jìn)一步說明算法的有效性,圖3~5展示了信噪比分別為5 dB,10 dB,15 dB時,不同算法重建Pollen、Alcea Rosea、Butterfly圖像的視覺效果及部分放大效果圖。由重建圖像及放大部分可以看出,DOLPHIn算法的重建圖像中存在著明顯的塊效應(yīng);BM3D-prGAMP算法的重建圖像缺失了大量細(xì)節(jié)且存在偽影;SPAR算法重建圖像細(xì)節(jié)較好,但偽影較為明顯;prDeep算法重建圖像較為平滑,導(dǎo)致大量的細(xì)節(jié)丟失。本文提出的DCD-DIP算法重建圖像視覺效果較好,保留了大量細(xì)節(jié)信息,有效消除了塊效應(yīng)和偽影。綜上所述,DCD-DIP算法在不同信噪比下均表現(xiàn)出了良好的性能。
3.4 收斂性分析
提出的DCD-DIP算法求解的是一個非凸優(yōu)化問題,雖然能夠取得較好的重建效果,但很難從理論上嚴(yán)格證明其收斂性。為說明該算法具有良好的收斂性能,圖6給出了在信噪比10 dB和15 dB情況下獲得的PSNR值和SSIM值隨迭代次數(shù)變化的曲線圖。從圖6中看出PSNR值和SSIM值隨著迭代次數(shù)的增加呈不斷上升直至平穩(wěn)的趨勢,SNR=10 dB時經(jīng)25次迭代PSNR值和SSIM值趨近于平穩(wěn),SNR=15 dB時迭代到20次時趨近于平穩(wěn)。由此可以說明DCD-DIP算法具有良好的收斂性能。此外,從圖中可以看出,SSIM值在第1次到第10次迭代過程中存在波動。由于求解優(yōu)化問題本身的非凸性,收斂曲線存在波動是合理的。
4 結(jié)論
本文將深度卷積稀疏編碼與深度圖像先驗(yàn)?zāi)P拖嘟Y(jié)合,利用雙數(shù)據(jù)保真項(xiàng)提出了一種面向編碼衍射成像的無監(jiān)督學(xué)習(xí)算法。該算法無需樣本標(biāo)簽及大規(guī)模數(shù)據(jù)集,僅通過單次觀測的編碼衍射強(qiáng)度圖樣就能夠同時實(shí)現(xiàn)低信噪比下的高質(zhì)量圖像重建及深度神經(jīng)網(wǎng)絡(luò)參數(shù)的優(yōu)化。提出的算法能夠融合互補(bǔ)先驗(yàn)知識,實(shí)驗(yàn)結(jié)果表明該算法重建質(zhì)量優(yōu)于現(xiàn)有編碼衍射成像算法。
參考文獻(xiàn)
[1]? WU J, LU Z, JIANG D, et al. Iterative tomography with digital adaptive optics permits hour-long intravital observation of 3D subcellular dynamics at millisecond scale[J]. Cell, 2021, 184(12): 3318-3332.
[2] AKCAKAYA M, TAROKH V. Sparse signal recovery from a mixture of linear and magnitude-only measurements[J]. IEEE Signal Processing Letters, 2015, 22(9): 1220-1223.
[3] MILLANE R P. Phase retrieval in crystallography and optics[J]. Journal of the Optical Society of America A, 1990, 7(3): 394-411.
[4] SHECHTMAN Y, ELDAR Y C, COHEN O, et al. Phase retrieval with application to optical imaging: a contemporary overview[J]. IEEE Signal Processing Magazine, 2015, 32(3): 87-109.
[5] JAGANATHAN K, OYMAK S, HASSIBI B. Sparse phase retrieval: uniqueness guarantees and recovery algorithms[J]. IEEE Transactions on Signal Processing, 2017, 65(9): 2402-2410.
[6] CANDS E J, LI X D, SOLTANOLKOTABI M. Phase retrieval via wirtinger flow: theory and algorithms[J]. IEEE Transactions on Information Theory, 2015, 61(4): 1985-2007.
[7] CHEN Y X, CANDS E J. Solving random quadratic systems of equations is nearly as easy as solving linear systems[J]. Communications on Pure and Applied Mathematics, 2017, 70(5): 822-883.
[8] WANG G, GIANNAKIS G B, ELDAR C Y. Solving systems of random quadratic equations via truncated amplitude flow[J]. IEEE Transactions on Information Theory, 2018, 64(2): 773-794.
[9] TILLMANN A M, ELDAR Y C, MAIRAL J. DOLPHIn—Dictionary learning for phase retrieval[J]. IEEE Transactions on Signal Processing, 2016, 64(24): 6485-6500.
[10] CHANG H B, LOU Y F, NG M K, et al. Phase retrieval from incomplete magnitude information via total variation regularization[J]. SIAM Journal on Scientific Computing, 2016, 38(6): A3672-A3695.
[11] SHI B S, LIAN Q S, HUANG X, et al. Constrained phase retrieval: when alternating projection meets regularization[J]. Journal of the Optical Society of America B, 2018, 35(6): 1271-1281.
[12] CHANG H B, MARCHESINI S. Denoising poisson phaseless measurements via orthogonal dictionary learning[J]. Optics Express, 2018, 26(16): 19773-19796.
[13] 練秋生, 李穎,陳書貞. 融合多種小波與全變差正則化的相位恢復(fù)算法[J]. 光學(xué)學(xué)報(bào), 2018, 38(2): 298-305.
LIAN Q S, LI Y, CHEN S Z. Phase retrieval algorithm fusing multiple wavelets and total variation regularization[J]. Acta Optica Sinica, 2018, 38(2): 298-305.
[14] KATKOVNIK V, EGIAZARIAN K. Sparse superresolution phase retrieval from phase-coded noisy intensity patterns[J]. Optical Engineering, 2017, 56(9): 094103.
[15] KATKOVNIK V, SHEVKUNOV I, PETROV N V, et al. Computational super-resolution phase retrieval from multiple phase-coded diffraction patterns: simulation study and experiments[J]. Optica, 2017, 4(7): 786-794.
[16] KOCSIS P, SHEVKUNOV I, KATKOVNIK V, et al. Single-shot pixel super-resolution phase imaging by wavefront separation approach[J]. Optics Express, 2021, 29(26): 43662-43678.
[17] METZLER C A, MALEKI A, BARANIUK R G. BM3D-PRGAMP: Compressive phase retrieval based on BM3D denoising[C]//Proceedings of the IEEE International Conference on Image Processing, Phoenix, AZ, USA, 2016: 2504-2508.
[18] GERCHBERG R W, SAXTON W. A practical algorithm for the determination of phase from image and diffraction plane pictures[J]. Optik, 1972, 35(2): 237-246.
[19] 姜萬錄, 李振寶,雷亞飛,等.基于深度學(xué)習(xí)的滾動軸承故障診斷與性能退化程度識別方法[J].燕山大學(xué)學(xué)報(bào),2020,44(6):526-536
JIANG W L,LI Z B,LEI Y F, et al.Deep learning based rolling bearing fault diagnosis and performance degradation degree recognition method[J].Journal of Yanshan University,2020,44(6):526-536.
[20] 李林祥,袁毅,溫淑煥.基于BAU-Net的高分辨率遙感圖像中建筑物的分割提取 [J].燕山大學(xué)學(xué)報(bào),2021,45(4):335-342.
LI L X,YUAN Y,WEN S H.BAU-Net for buildings segmentation in high resolution remote sensing images[J].Journal of Yanshan University,2021,45(4):335-342.
[21] HAND P, LEONG O, VORONINSKI V. Phase retrieval under a generative prior[C]//Proceedings of the International Conference on Neural Information Processing Systems, Montréal,Canada, 2018: 9154-9164.
[22] MORALES D, JEREZ A, ARGUELLO H. Learning spectral initialization for phase retrieval via deep neural networks[J]. Applied Optics, 2022, 61(9): F25-F33.
[23] METZLER C A, SCHNITER P, VEERARAGHAVAN A, et al. prDeep: Robust phase retrieval with a flexible deep network[C]//Proceedings of International Conference on Machine Learning, Stockholm, Sweden, 2018, 80: 3501-3510.
[24] ROMANO Y, ELAD M, MILANFAR P. The little engine that could: regularization by denoising (RED)[J]. SIAM Journal on Imaging Sciences, 2017, 10(4): 1804-1844.
[25] SHI B S, LIAN Q S, CHANG H B. Deep prior-based sparse representation model for diffraction imaging: a plug-and-play method[J]. Signal Processing, 2020,168:107350.
[26] LEMPITSKY V, VEDALDI A, ULYANOV D. Deep image prior[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018: 9446-9454.
[27] CHENG Z Z, GADELHA M, MAJI S, et al. A Bayesian perspective on the deep image prior[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 2019: 5438-5446.
[28] SUN Z D, LATORRE F, SANCHEZ T, et al. A plug-and-play deep image prior[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, Toronto, ON, Canada, 2021:8103-8107.
[29] CASCARANO P, SEBASTIANI A, COLOMBA C M, et al. Combining weighted total variation and deep image prior for natural and medical image restoration via ADMM[C]//Proceedings of theInternational Conference on Computational Science and Its Applications, Cagliari, Italy, 2021: 39-46.
[30] MATAEV G, ELAD M, MILANFAR P. DeepRED: Deep image prior powered by RED[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Korea, 2019.
[31] GARCIA-CARDONA C,WOHLBERG B. Convolutional dictionary learning: a comparative review and new algorithms[J]. IEEE Transactions on Computational Imaging, 2018, 4(3): 366-381.
[32] 熊嬌嬌,盧紅陽,張明輝,等. 基于梯度域的卷積稀疏編碼磁共振成像重建[J]. 自動化學(xué)報(bào), 2017, 43(10): 1841-1849.
XIONG J J, LU H Y, ZHANG M H, et al. Convolutional sparse coding in gradient domain for MRI reconstruction[J]. Acta Automatica Sinica, 2017, 43(10): 1841-1849.
[33] MENG Z Y, YU Z M, XU K, et al. Self-supervised neural networks for spectral snapshot compressive imaging[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, QC, Canada, 2021: 2622-2631.
[34] LIU J M, SUN Y, XU X J, et al. Image restoration using total variation regularized deep image prior[C]//Proceedings of International Conference on Acoustics, Speech and Signal Processing, Brighton, UK, 2019:7715-7719.
[35] KINGMA D, BA J. Adam: a method for stochastic optimization[C]//Proceedings of International Conference on Learning Representations, San Diego, USA,2015.
[36] TOFIGHI M, KOSE K, CETIN A E. Denoising using projections onto the epigraph set of convex cost functions[C]//Proceedings of the IEEE International Conference on Image Processing, Paris, France, 2014: 2709-2713.
[37] ZHANG J, ZHAO C, ZHAO D B. Image compressive sensing recovery using adaptively learned sparsifying basis via L0 minimization[J]. Signal Processing, 2014, 103: 114-126.
[38] BRISTOW H, ERIKSSON A, SIMON L. Fast convolutional sparse coding[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Portland, OR, USA, 2013: 391-398.
[39] WOHLBERG B. Efficient convolutional sparse coding[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, Florence, Italy, 2014: 7173-7177.
[40] FU X Y, ZHA J Z, WU F, et al. Jpeg artifacts reduction via deep convolutional sparse coding[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Korea, 2019: 2501-2510.
[41] ZHENG H Y, YONG H W, ZHANG L. Deep convolutional dictionary learning for image denoising[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, TN, USA, 2021: 630-641.
[42] DING F, CAO T J. Application of daubechies wavelet transform in the estimation of standard deviation of white noise[C]//Proceedings of the Second International Conference on Digital Manufacturing & Automation, Zhangjiajie, China, 2011: 212-215.
Research on coded diffraction imaging method based on unsupervised learning
SHI Baoshun1,2, WU Yifan1,2, LIAN Qiusheng1,2
(1. School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China;
2. Hebei Key Laboratory of Information Transmission and Signal Processing,Yanshan University,Qinhuangdao,Hebei 066004,China)
Abstract:
Coded diffraction imaging aims to reconstruct the original image by utilizing the intensity of diffraction patterns. However,in the case of low signal-to-noise ratios,most of the existing coded diffraction imaging algorithms based on hand-crafted priors usually suffer from low-quality reconstructions.
The aforementioned problem can be solved by using the deep priors based on deep neural network learning.However, supervised learning methods need massive sample pairs, which is impractical for applications. To address this issue, a coded diffraction imaging method based on the unsupervised learning is proposed in this paper. An optimization model which can fuse complementary priors is formulated by combining the double data fidelity terms, convolutional sparse coding, and deep image prior models.Meanwhile,the alternating optimization method is utilized to solve the optimization model effectively. Experimental results show that the proposed method can reconstruct high-quality images only from single intensity of coded diffraction pattern at low signal-to-noise ratios.
Keywords:
computational imaging;diffraction imaging;unsupervised learning;deep image prior;convolutional sparse coding