图1 压缩波前探测的波前重构过程
Received:11 January 2023,
Revised:09 February 2023,
Published:05 June 2023
Scan QR Code
Cite this article
When the compressive sensoring is used in wavefront measurement, classic methods of slopes’ restoration has a relatively low precision, which make it difficult to measure the atmospheric turbulence wavefront. In the paper, a deep neural network is presented to improve the slopes’ restoration precision. The traditional compressive sensing technology does not take into account the relatively small slopes, which increases the wavefront measurement errors. To measure the complex wavefront induced by atmospheric turbulence with a high speed, the paper presents an improved deep neural network to restore the slopes from sparse ones with high precision, which improves the precision of wavefront reconstruction. When the compression ratio is ranged from 0.1 to 0.9, the wavefront error PV (Peak to valley) of the compressed wavefront detection algorithm based on depth neural network (DNNCWS) proposed in this paper is better than 0.014 μm, and the running time of the algorithm is 4.4 ms. In the case of low signal-to-noise ratio, the residual wavefront PV is better than 0.011 μm. In addition, the simulation results indicate that it has good anti-noise performance. The DNNCWS improves the detection accuracy of compressive sensing and overcomes the problem of low accuracy for complex aberration induced by atmospheric turbulence. It can also be used in other adaptive optical applications, such as laser communication and retinal imaging.
compressed wavefront sensing;
adaptive optics;
atmospheric turbulence
在使用地基光学望远镜进行天文观测时,人们利用自适应光学系统(Adaptive optics system, AOS)补偿大气湍流造成的像差[
近年来,人们将压缩感知技术引入波前测量,以少的斜率测量来提高波前测量速度。2014年,Polans提出了压缩波前探测,他利用Zernike多项式对斜率进行稀疏化并对由几个低阶Zernike模式组成的简单波前像差进行了仿真验证[
为了解决该问题,我们提出了深度神经网络来提高稀疏斜率的恢复精度。传统的算法可以从光点阵计算斜率并重构波前,但针对暗弱空间目标探测时,湍流扰动引起的光点亮度起伏会导致部分光点信噪比差,甚至不能用于波前重构,导致重构精度降低。近年来,人们利用深度学习技术直接从光点阵预测波前且具有较高的波前重构精度,但它们的算法严重依赖于硬件和数据训练,计算时间通常高达几十毫秒[
仿真过程如
图1 压缩波前探测的波前重构过程
Fig.1 Wavefront reconstruction process of compressive wavefront sensor
SHWFS通过测量光斑阵列并计算其沿x和y方向的偏移,根据斜率重构波前。在SHWFS中,波前通过微透镜阵列形成光点阵列。每个子孔径上的局部波前的平均倾斜和尖端与光点的偏移成比例。偏移可以通过比较从参考波前获得的光点的质心与畸变波前的质心之间的差异来计算。通常用传统的重心(CoG)算法计算:
(1) |
(2) |
其中I(i,j)是第i行和第j列中像素的光强。然后,根据光点的偏移x和y计算波前斜率。定义φ(x,y)为坐标(x,y)处的大气湍流波前相位,其斜率分别为x和y方向上的Sx和Sy。它们之间的关系如下:
(3) |
(4) |
其中f是透镜的焦距。计算斜率后,可以通过经典波前重构方法重构最终波前W,如
(5) |
其中:Zi是第i个Zernike模式,ai是第i个Zernike模式的系数。
定义Ψ∈RM×N作为M
(6) |
其中稀疏向量θx∈RN,θy∈RN分别是Sx和Sy的稀疏表示。离散余弦变换(DCT)矩阵被用作稀疏矩阵。
设计一个M×N探测矩阵来测量Sx和Sy并获得斜率测量值S'x、S'y。我们选择高斯矩阵作为探测矩阵:
(7) |
根据方程(5)和(6),S'x和S'y可以表示如下:
(8) |
其中A=ΦΨ。
矩阵Φ满足限制等距性质(RIP)标准[
(9) |
因为M比N小得多,所以S'x和S'y的值通过求解最小0范数来得到。而0范数问题是一个NP难问题,但Candes、Tao和Donoho等人已证明,当测量矩阵Φ满足约束等距性质(RIP)时,l0约束优化问题可转化为l1约束的凸优化问题:
(10) |
(11) |
恢复斜率
图2 深度神经网络结构
Fig.2 Deep neural network structure
本文在J波段进行了模拟,望远镜孔径D为3.6 m,r0范围为0.33~0.42 m。SHWFS中的微透镜数量为16×16,每个子区域的像素数为6×6。压缩比r表示所有斜率中有效斜率的百分比。
对于相同的压缩比r=0.7,分别使用DNNCWS和GCWS重构波前斜率及其波前。
图3 (a)原始波前的x方向斜率;(b)GCWS恢复的x方向斜率;(c)GCWS恢复的x方向斜率的误差;(d)DNNCWS恢复的x方向斜率;(e)DNNCWS恢复的x方向斜率的误差;(f)原始波前的y方向斜率;(g)GCWS恢复的y方向斜率;(h)由GCWS恢复的y方向斜率的误差;(i)由DNNCWS恢复的y方向斜率;(j)由DNNCWS恢复的y方向斜率误差。
Fig.3 (a) X-direction slopes of the original wavefront; (b) X-direction slopes recovered by GCWS; (c) Error of the x-direction slopes recovered by GCWS; (d) X-direction slopes recovered by DNNCWS; (e) X-direction slopes error recovered by DNNCWS; (f) Y-direction slopes of the original wavefront; (g) Y-direction slopes recovered by GCWS; (h) Error of the y-direction slopes recovered by GCWS; (i) Y-direction slopes recovered by DNNCWS; (j) Y-direction slopes error recovered by DNNCWS.
基于所获得的斜率进行波前重构,得到了如
图4 (a)原始波前;(b)DNNCWS算法恢复的斜率重构的波前;(c)DNNCWS算法的残差波前;(d)GCWS算法恢复的斜率重构的波前;(e)GCWS算法的残差波前。
Fig.4 (a) Original wavefront; (b) Reconstructed wavefront with DNNCWS; (c) Residual wavefront of DNNCWS; (d) Reconstructed wavefront with GCWS; (e) Residual wavefront of GCWS.
为了检验算法的稳定性,我们选择了30组数据进行了测试。
图5 30组斜率的残差波前的PV和RMS误差的比较。(a)误差的PV值;(b)误差的RMS值。
Fig.5 Comparison of PV and RMS errors of residual wavefront for 30 sets of slopes. (a) PV values of errors; (b) RMS values of errors.
图6 不同压缩比下的残差波前比较。(a)残差波前PV的比较;(b)残差波前RMS的比较。
Fig.6 Comparison of residual wavefront under different compression ratios. (a) Comparison of PV of residual wavefront; (b) Comparison of RMS of residual wavefront.
对于比较暗弱的11星等目标,在信噪比比较低(SNR=8.11 dB),压缩比r=0.7的情况下,
图7 (a)原始波前;(b)DNNCWS算法恢复的斜率重构的波前;(c)DNNCWS算法恢复的斜率重构的波前的残差波前;(d)GCWS算法恢复的斜率重构的波前;(e)GCWS算法恢复的斜率重构的波前的残差波前。
Fig.7 (a) Original wavefront; (b) Reconstructed wavefront with DNNCWS; (c) Residual wavefront of DNNCWS; (d) Reconstructed wavefront with GCWS; (e) Residual wavefront of GCWS.
图8 30组斜率的波前误差比较。(a)残差波前PV的比较;(b)残差波前RMS的比较。
Fig.8 Wavefront error comparison of 30 sets of slopes. (a) PV of residual wavefront; (b) RMS of residual wavefront.
图9 不同星等下的残差波前。(a)残差波前PV的比较;(b)残差波前RMS的比较。
Fig.9 Residual wavefront under different magnitude of star. (a) RMS of residual wavefront; (b) PV of residual wavefront.
本文提出DNNCWS的方法,采用深度神经网络从稀疏的斜率恢复斜率,提高了压缩波前的检测精度。与已有的压缩波前探测方法相比,该方法克服了测量复杂像差精度低的问题,可以用于大气湍流引起的复杂波前测量。此外,DNNCWS可以在不同压缩比和低信噪比的条件下以高精度和稳定性重构波前。在压缩比从0.1~0.9变化情况下,DNNCWS的PV范围为0.005~0.014 μm,残差波前的RMS均在0.001 μm左右,算法的运行时间为4.4 ms。在暗弱星等情况下,即低信噪比条件下,DNNCWS的残差波前PV在0.005~0.011 μm之间,RMS在0.000 9~0.0017 μm之间。本文所提出的方法在使用自适应光学的其他光学成像系统中也将有较好的应用。
ANGEL J R P. Ground-based imaging of extrasolar planets using adaptive optics [J]. Nature, 1994, 368(6468): 203-207. doi: 10.1038/368203a0 [Baidu Scholar]
HU L F, XUAN L, LIU Y J, et al. Phase-only liquid-crystal spatial light modulator for wavefront correction with high precision [J]. Optics Express, 2004, 12(26): 6403-6409. doi: 10.1364/opex.12.006403 [Baidu Scholar]
赵子云,顾虎,马文超,等.自适应光学系统误差分析与参数优化研究[J].液晶与显示,2021,36(5):663-672. doi: 10.37188/CJLCD.2020-0356 [Baidu Scholar]
ZHAO Z Y, GU H, MA W C, et al. Error budget and parameters optimization of adaptive optics system [J]. Chinese Journal of Liquid Crystals and Displays, 2021, 36(5): 663-672. (in Chinese). doi: 10.37188/CJLCD.2020-0356 [Baidu Scholar]
GU H, ZHAO Z Y, ZHANG Z G, et al. High-precision wavefront reconstruction from Shack-Hartmann wavefront sensor data by a deep convolutional neural network [J]. Measurement Science and Technology, 2021, 32(8): 085101. doi: 10.1088/1361-6501/abf708 [Baidu Scholar]
DUNCAN J L, ZHANG Y H, GANDHI J, et al. High-resolution imaging with adaptive optics in patients with inherited retinal degeneration [J]. Investigative Ophthalmology & Visual Science, 2007, 48(7): 3283-3291. doi: 10.1167/iovs.06-1422 [Baidu Scholar]
POYNEER L A, GAVEL D T, BRASE J M. Fast wave-front reconstruction in large adaptive optics systems with use of the Fourier transform [J]. Journal of the Optical Society of America, 2002, 19(10): 2100-2111. doi: 10.1364/josaa.19.002100 [Baidu Scholar]
陈浩,宣丽,胡立发,等.望远镜的紧凑型闭环液晶自适应光学系统设计[J].液晶与显示,2010,25(3):379-385. doi: 10.3969/j.issn.1007-2780.2010.03.017 [Baidu Scholar]
CHEN H, XUAN L, HU L F, et al. Design on compact type closed-loop liquid crystal adaptive optical system for telescope [J]. Chinese Journal of Liquid Crystals and Displays, 2010, 25(3): 379-385. (in Chinese). doi: 10.3969/j.issn.1007-2780.2010.03.017 [Baidu Scholar]
POLANS J, MCNABB R P, IZATT J A, et al. Compressed wavefront sensing [J]. Optics Letters, 2014, 39(5): 1189-1192. doi: 10.1364/ol.39.001189 [Baidu Scholar]
HOWLAND G A, LUM D J, HOWELL J C. Compressive wavefront sensing with weak value [J]. Optics Express, 2014, 22(16): 18870-18880. doi: 10.1364/oe.22.018870 [Baidu Scholar]
TIK E C M, WANG X, GUO N Q, et al. Investigation on compressed wavefront sensing in freeform surface measurements [C]. 2016 4th International Conference on Photonics, Optics and Laser Technology. Rome: IEEE, 2016: 1-6. doi: 10.5220/0005743901520157 [Baidu Scholar]
KE X Z, WU J L, HAO J X. Distorted wavefront reconstruction based on compressed sensing [J]. Applied Physics B, 2022, 128: 107. doi: 10.1007/s00340-022-07827-6 [Baidu Scholar]
NISHIZAKI Y, VALDIVIA M, HORISAKI R, et al. Deep learning wavefront sensing [J]. Optics Express, 2019, 27(1): 240-251. doi: 10.1364/oe.27.000240 [Baidu Scholar]
GUO H Y, XU Y J, LI Q, et al. Improved machine learning approach for wavefront sensing [J]. Sensors, 2019, 19(16): 3533. doi: 10.3390/s19163533 [Baidu Scholar]
VISHNIAKOU I, SEELIG J D. Wavefront correction for adaptive optics with reflected light and deep neural networks [J]. Optics Express, 2020, 28(10): 15459-15471. doi: 10.1364/oe.392794 [Baidu Scholar]
JIA P, MA M Y, CAI D M, et al. Compressive Shack-Hartmann wavefront sensor based on deep neural networks [J]. Monthly Notices of the Royal Astronomical Society, 2021, 503(3): 3194-3203. doi: 10.1093/mnras/staa4045 [Baidu Scholar]
SOUTHWELL W H. Wave-front estimation from wave-front slope measurements [J]. Journal of the Optical Society of America, 1980, 70(8): 998-1006. doi: 10.1364/josa.70.000998 [Baidu Scholar]
FRIED D L. Least-square fitting a wave-front distortion estimate to an array of phase-difference measurements [J]. Journal of the Optical Society of America, 1977, 67(3): 370-375. doi: 10.1364/josa.67.000370 [Baidu Scholar]
BARANIUK R, DAVENPORT M, DEVORE R, et al. A simple proof of the restricted isometry property for random matrices [J]. Constructive Approximation, 2008, 28(3): 253-263. doi: 10.1007/s00365-007-9003-x [Baidu Scholar]
CANDÈS E J. The restricted isometry property and its implications for compressed sensing [J]. Comptes Rendus Mathematique, 2008, 346(9/10): 589-592. doi: 10.1016/j.crma.2008.03.014 [Baidu Scholar]
DANDES E J, TAO T. Near-optimal signal recovery from random projections: Universal encoding strategies [J]. IEEE Transactions on Information Theory, 2006, 52(12): 5406-5425. doi: 10.1109/tit.2006.885507 [Baidu Scholar]
474
Views
162
Downloads
0
CSCD
Related Articles
Related Author
Related Institution