In this paper, we consider computational super-resolution inverse diffraction phase retrieval. The optical setup is lensless, with a spatial light modulator for aperture phase coding. The paper is focused on experimental tests of the super-resolution sparse phase amplitude retrieval algorithm. We start from simulations and proceed to physical experiments. Both simulation tests and experiments demonstrate good-quality imaging for super-resolution with a factor of 4 and a serious advantage over diffraction-limited resolution as defined by Abbe’s criterion.
© 2017 Optical Society of America
A. Problem Setup
The phase retrieval problem concerns the reconstruction of 2D and 3D complex-domain objects provided that only the intensity of light radiation is measured. Such problems appear in science and engineering with a wide area of applications from astronomy and optics to medicine and biology (e.g., [1–3]).
The standard formalization of the phase retrieval problem is of the form
In this paper, we consider the setup with aperture modulation such that
The phases in can be generated as deterministic or random. It results in observations known as coded diffraction patterns (e.g., ):
Various methods have been developed to make these sufficiently different in order to achieve observation diversity to be able to find from observations . A defocussing of the registered images is one of the instruments to achieve sufficient phase diversity [5–9]. In a recent development, a spatial light modulator (SLM) is exploited for the defocussing (e.g., [10–12]).
For noisy observations, Eq. (3) is changed for
In this paper, we assume that the observations have a Poissonian distribution typical for optical photon counting. This process is well modeled by independent Poissonian random variables of the form1), at pixel .
The parameter in Eq. (7) is a scaling factor of the Poisson distribution. Defining the observation signal-to-noise ratio (SNR) as the ratio between the square of the mean and the variance of , we have . It follows that the relative noisiness of the observations becomes stronger as () and approaches zero when (). The latter case corresponds to the noiseless scenario, i.e., with probability 1. The scale parameter is of importance for modeling as it allows controlling a level of noise in observations.
Reconstruction of the complex-valued object (phase and amplitude) from noiseless or noisy observations is the phase retrieval problem. Here phase emphasizes that in the object phase is a variable of first priority, whereas amplitude may be an auxiliary variable often useful only in order to improve the phase imaging.
B. Phase Retrieval Algorithms
There is a flow of publications on phase retrieval. Various versions of the Gerchberg–Saxton technique are quite universal for different setups (e.g., [7–9,17,18]). These algorithms, based on alternating projections between the object and observation planes, are flexible in order to incorporate any information available for variables in these planes. Recent developments in this area as well as a review are provided in .
Contrary to this type of intuitive heuristic algorithms, the variational formulations have a stronger mathematical background and are based on numerical algorithms for solving some optimization problems (e.g., [20–22]).
A new variational algorithm for phase retrieval from noisy data based on transform domain sparsity for object phase and amplitude is developed in . The sparsity concept as a tool for phase retrieval is a topic of publication , where the original sparsifying learning transform is developed.
The spatial resolution of phase retrieval is limited by two factors: low-pass filtering by the propagation operator and the size of the sensor’s pixels. High-frequency spatial information is lost in intensity observations, which can be treated as corresponding to a sub-sampled true object . Super-resolution imaging allows compensating for this sub-sampling effect.
The computational approach to super-resolution is different assuming that the optical setup is fixed and a super-resolution image should be extracted from the data available. Note that computational imaging is a hot topic for the real domain, with a flow of publications, whereas for phase/amplitude imaging, in the complex domain, the situation is quite different, with a few publications, which can be referred to, in particular, for phase retrieval.
Compressed sensing (CS) is one of the comparatively new computational techniques for restoration of sub-sampled data. Application of this sort of techniques in optics can be found in [30,31]. CS can be treated as super-resolution; however, there is a deep difference between these two problems. CS is focused on the sub-sampling design with the smallest number of observations as a target, whereas in super-resolution, the observations are given on a usually regular and fixed grid.
C. Contribution and Structure of This Paper
The transform domain phase/amplitude sparsity introduced in  was generalized for the super-resolution phase retrieval in . This paper is focused on experimental testing of the general sparsity-based approach developed in  and , where only simulation results have been demonstrated. We modify the algorithms from  to the lensless optical setup used in the experiments with AS wavefront propagation [Eqs. (5) and (6)]. Contrary to it, a lenslet and the FT propagation model are used in . We provide simulation and experimental tests demonstrating that the developed algorithm enables super-resolution phase retrieval with an up-sampling factor of up to 4 and with a significant advantage over the diffraction-limited system with resolution defined by Abbe’s criterion.
2. SUPER-RESOLUTION PHASE RETRIEVAL
A. Complex Domain Sparsity
The complex domain sparsity for the object as used in this paper assumes that there are small fragments (patches) of the object , which have similar features. These similar patches can be found in different parts of the image. It follows that these patches taken together admit sparse representations: they can be well approximated by linear combinations of a small number of basic functions.
The complex domain images, such as the object , are defined by two variables, phase and amplitude . Respectively, the sparse representation can be imposed on the following pairs of real-valued variables:
The success of the sparsity approach depends on how rich and redundant are the dictionaries (sets of basic functions) used for image analysis and synthesis. In this paper, we use the block-matching and 3D filtering (BM3D) algorithm for the sparsity implementation .
Let us mention the basic steps of this advanced technique known as nonlocal self-similarity sparsity. At the first stage, the image is partitioned into small overlapping square patches. For each patch, a group of similar patches is collected, which are stacked together to form a 3D array (group). This stage is called grouping. The entire 3D group array is projected onto a 3D predefined transform basis. The spectral coefficients obtained as a result of this transform are hard-thresholded (small coefficients are zeroed) and the inverse 3D transform gives the filtered patches, which are returned to the original position of these patches in the image. This stage is called collaborative filtering. This process is repeated for all pixels of the entire image and the resultant overlapped filtered patches are aggregated in the final image estimate. This last stage is called aggregation. Details of this algorithm are provided in .
The links of this algorithm with the general sparsity concept are revealed in . It is shown that the grouping operations result in data adaptive analysis and synthesis of image transforms (frames), which when combined with the hard-thresholding of these transforms define the BM3D algorithm. It is emphasized that sparsity is achieved mainly due to the grouping, which allows to jointly analyze similar patches and in this way to guarantee sparsity (self-similarity of patches) at least for each of the 3D groups.
Note that the standard BM3D algorithm as presented in  is composed of two successive steps, thresholding and Wiener filtering. In this paper, we use a simplified version of BM3D including grouping, transforms, and thresholding without Wiener filtering. In this form, this algorithm is obtained from the variational setting of the wavefront retrieval problem.
It is shown in  that the nonlocal block-wise sparsity imposed on phase and amplitude results in separate BM3D filtering of phase and amplitude:8) can be interferometric or absolute depending on the sparsity formulation.
B. Super-resolution SPAR Algorithm
Conventionally, the pixels are square and for the SLM and sensor, respectively. A continuous object is discretized by computational pixels . This discretization is necessary for both digital data processing and modeling of wavefront propagation and image formation. Contrary to the pixels of the SLM and sensor, which are defined by the corresponding optical-electronic devices, the object pixels are computational , which maybe be taken to be of arbitrary small sizes.
Assuming for a moment , the reconstruction of from the observations is the conventional phase retrieval with the resolution for the object dictated by the pixel size of the sensor. Let us term this case pixel (or pixel-wise) resolution imaging.
If , we arrive at a much more interesting case of sub-pixel resolution (or super-resolution). It is convenient to assume that , where is an integer pixel up-sampling factor, and , where . Then and are numbers of smaller computational object pixels covering a single pixel of the sensor and SLM, respectively. These SLM pixels give the same modulation phase shift to the object pixels in these groups.
In order to study the effects of SLM phase modulation, we also introduce an SLM superpixel as a larger size pixel , which is composed of adjacent SLM pixels. In the phase modulation, the corresponding SLM superpixels provide the same phase shifts to the object pixels covered by the superpixel.
In our experiments, we use for the SLM’s superpixels the random phases generated as follows:
Then the corresponding phase modulation masks in Eq. (2) are defined as
The computational wavefront reconstruction goes from the continuous domain wavefront propagation [Eqs. (5) and (6)] to the corresponding discrete model based on pixelation of the object distribution ; thus, we arrive at the discrete modeling of the system at hand linking the discrete values of the sensor output (observations) with the discrete values of the object with the integral FT replaced by the fast FT. Note that all these calculations are produced with the computation sampling period .
A discrete modeling of the continuous propagation in Eqs. (5) and (6) imposes restrictions on the computational sampling period with connection to the distance , the wavelength , and the computational aperture size. With reference to [37–39], for the considered scenario, these restrictions can be given as the inequality
In order to simplify the algorithm presentation, we preserve the notation for this discrete model initially introduced for continuous variables and integrals. The abbreviation SR-SPAR (super-resolution sparse phase amplitude retrieval) is used for this algorithm.
Let us emphasize that SR-SPAR has originated from the variational formulation introduced for the optimal reconstruction of from the Poissonian observations . The corresponding minus log-likelihood for the Poissonian observations according to Eq. (7) is as follows:
The derivation of the algorithm is based on the technique developed in  for pixel-wise resolution phase retrieval and FT propagation. We show the block scheme of this algorithm in Fig. 1 referring to  and  for details.
The inputs in this algorithm are the observations up-sampled by factor . We use the zero-order up-sampling giving as piece-wise invariant, with the invariant values for computational pixels corresponding to each of the larger size pixels of the sensor. At the initialization step of the algorithm, the initial estimate of the object wavefront is created. At Step 1, the object wavefront estimate is multiplied by the phase mask and propagates using the operator to the sensor plane; the result is denoted with . These wavefronts are calculated for the computational diffraction area denoted with .
At Step 2, this wavefront is updated to the variable by filtering the amplitude of according to the given observations . The following formula as derived in  defines the rule how the amplitude is calculated:
At Step 3, the estimates back-propagate to the object plane and update the object wavefront estimate . Here means a complex conjugate of and means inverse of the model Eq. (5).
The unwrapping of phase with reconstruction of the absolute phase in Step 4 is necessary only if the object phase range exceeds . Sparsification (filtering on the base of the sparse approximations) is produced in Step 5. At Step 6, the object reconstruction is updated.
C. Parameters of SR-SPAR
The performance of the SR-SPAR algorithm essentially depends on its parameters. Optimization can be produced for each amplitude/phase distribution and noise level. However, in our study, the parameters are fixed for all the experiments. The image patches in BM3D are square . The group size is limited by 39 patches. The step size between neighboring patches is equal to 2. Discrete cosine (for patches) and Haar (for group length) transforms are used for 3D group data processing in BM3D.
The parameters defining the iterations of the algorithm are as follows: , , . The number of iterations is fixed to 30.
For our experiments, we use MATLAB R2016b on a computer with 32 GB of RAM and CPU with a 3.40 GHz Intel(R) Core(TM) i7-3770 processor.
The computation complexity of the algorithm is characterized by the time required for processing. For 30 iterations, , and images, zero-padded to , this time is equal 2150 s.
We make the MATLAB demo codes (see Code 1, Ref. ) of the developed SR-SPAR algorithm publicly available at http://www.cs.tut.fi/sgn/imaging/sparse. These can be used to reproduce the experiments presented in this paper as well as for further tests.
3. SIMULATION AND EXPERIMENTS
A. Experimental Setup
The experimental setup, presented in Fig. 2, contains a laser LASOS GLK 3250 TS with a wavelength of , a digital 8-bit CMOS camera EVS with a pixel size of , and a phase-only reflective SLM Holoeye LETO with a pixel size of and a fill factor of 93%. The polarization of the light beam emitted from the laser source is changed from vertical to horizontal by the retardation plate and the polarizer . Further, the beam expands in the collimators of the two lenses and and passes through the beamsplitter (BS) to the SLM, where the light wavefront obtains the desired phase retardation, and propagates to the CMOS camera through the same BS. The optical distance between the SLM and CMOS is 21.55 mm. According to Eq. (2), on the SLM we implement the sum of the object phase and the phase modulation masks . The SR-SPAR algorithm has been tested on two phase objects with the invariant amplitude and the phase images “cameraman” and a logo “TUT” (see Figs. 3 and 4, middle images in the top rows). All the results presented further in Figs. 3–10 are shown in images of identical structure—top line: reconstructed and original phases in the left and middle images, respectively, and reconstructed amplitude in the right image; bottom line: phase (left) and amplitude (right) longitudinal cross sections.
In our experiments, both the phase object and phase modulation masks are implemented on the SLM. Thus, the object’s and SLM’s pixels are equal, i.e., . The object size in pixels is (in SLM’s pixels). In the modeling and experiments, we use the same values of the parameters: propagation distance , , .
The objects and sensor are zero-padded to pixels, i.e., . It is the largest dimension allowed by the RAM of our computer. The experiments have been produced with super-resolution factors of . In what follows, we show the results only for (pixel resolution) and (sub-pixel resolution). The computational pixels for pixel resolution and sub-pixel resolution take values of and , respectively.
The propagation distance is minimal for our optical setup restricted by the used reflective SLM. For , the distance in Eq. (12) is . This upper bound is smaller than the used , but because the difference is not significant, the theoretical requirement [Eq. (12)] is approximately fulfilled.
We evaluate the diffraction-limited resolution using Abbe’s criterion:
In the inverse imaging, the phase retrieval resolution is defined by the computational pixel provided good-quality imaging is achieved.
Let us compare these experimentally observed resolutions with the theoretical upper bounds for super-resolution calculated according to Abbe’s criterion.
A square part of the sensor is used for the pixel resolution imaging; then according to Eq. (15), while , and it follows that .
A smaller central part of the sensor is used for the sub-pixel resolution imaging with . According to Eq. (15) then, while and once more .
Thus, super-resolution imaging, in terms of Abbe’s criterion, is demonstrated by SR-SPAR for both pixel and sub-pixel resolution imaging of course provided that the reconstructed images are of good quality.
The ratio compares the diffraction-limited and computational resolution. In the case , which is always fulfilled in our tests, the computational resolution is essentially higher than the diffraction-limited resolution as defined by Abbe’s criterion.
B. Simulation Experiments
In this section, we show the results for nearly noiseless observations with the Poissonian scale parameter . The way that the SPAR algorithm suppresses the Poissonian noise can be seen in , where it is done for pixel resolution and the FT propagation model. Similar strong filtering can be demonstrated for the super-resolution by SR-SPAR. In what follows, the accuracy of reconstruction is evaluated by the root-mean-square error (RMSE) criterion.
1. Pixel Resolution
In Figs. 3 and 4, we present the results for pixel resolution, i.e., the computational and sensor pixels have equal sizes: , , . The latter means that for the modulation phase masks, we use SLM superpixels with . It follows from the figures that SR-SPAR demonstrates perfect results in phase reconstruction and negligible errors in amplitude reconstructions. The amplitude reconstructions have clearly visible squares forming grids. These squares are a mapping of the SLM’s superpixels. Each of these squares has a size of in SLM pixels. Thus, the total object size is in SLM pixels and in computational pixels .
We can also note that the amplitude errors are definitely correlated with the phase variations: features of “cameraman” and “TUT” can be seen in the amplitude reconstructions. In the intensity measurements, the amplitude and phase of the object are tangled due to propagation. The algorithm tries to untangle them. The observed cross-talk in the phase/amplitude reconstructions means that this work is not done perfectly.
In order to optimize the accuracy of the reconstructions, we have studied the dependence of RMSE with respect to the following four parameters: number of observations, , number of iterations, , standard deviation of the random phase in the phase modulation masks, and size of the SLM’s superpixels, .
In what follows, we illustrate these results calculating the RMSE for pixel-wise phase reconstruction, , in experiments with the “cameraman” phase test image.
More or less similar results have been observed for the “TUT” phase test image with various super-resolution factors .
The dependence on the number of observations is presented in Fig. 11. It is a decreasing function taking a steady-state value starting from . We take this value of for our simulation. However, for processing of experimental data, we take a larger value of .
The standard deviation (STD) of the random zero-mean Gaussian phase masks is a crucial parameter for reconstruction because it influences both the accuracy and convergence of the algorithm. Figure 12 illustrates RMSE as a function of STD. The curve takes a minimal value at , acceptable for our experiments.
Another significant parameter is the SLM superpixel size . Figure 13 clearly indicates that in simulation, the best value is . However, it was observed during processing of experimental data that larger values of provide better results. It happens due to the cross-talk between adjacent SLM pixels . We accepted as an optimal value and used it for processing of simulation and experimental data.
Figure 14 illustrates the performance of the algorithm with respect to the number of iterations, . The RMSE curve decreases monotonically with steady-state values achieved after about 15 iterations. Nevertheless, for experimental data, we take .
2. Sub-pixel Resolution
In this subsection, we show the results for the super-resolution performance of SR-SPAR with a super-resolution factor of , i.e., , , , and . This computational pixel is smaller than the wavelength . In this way, we arrive at the sub-wavelength resolution. Figures 5 and 6 demonstrate nearly perfect reconstructions, which are similar to the reconstructions obtained for pixel resolution.
C. Physical Experiments
The parameters used in the physical experiments exactly correspond to those for the simulation tests. The phase objects “cameraman” and “TUT” are modeled by the SLM with phase ranges of and , respectively. The computational pixels have sizes of and , respectively, for pixel- and sub-pixel resolution.
In the pixel resolution case, we use all pixels of the sensor; therefore, Abbe’s resolution is . In the sub-pixel resolution case, the resolution of means that we need to operate with computational pixels, which is not possible on our computer. As a result, we use only a central part of the sensor ; therefore, Abbe’s resolution for this case is .
1. Pixel Resolution
The pixel resolution images are shown in Figs. 7 and 8. Despite a large distance between the object and sensor planes and strong diffraction effects in the registered diffraction patterns, SR-SPAR gives reliable results with a clearly visible “cameraman” image in Fig. 7. The phase images “cameraman” and “TUT” are clearly visible in the amplitude reconstruction. It happens due to the effect discussed above. Let us repeat these arguments here. In the intensity measurements, the amplitude and phase of the object are tangled due to propagation. The algorithm should untangle them. The observed cross-talk in the phase/amplitude reconstructions means that this work is not done perfectly as well as the propagation model used is not accurate. In the experimental data, the additional effects follow from the cross-talk between the programmed phase and amplitude in the SLM.
It happens, in particular, due to a non-ideality of the SLM with some leakage from phase to amplitude modulation. The corresponding phase cross section demonstrates that the reconstruction follows the respective changes in the original image. The reconstructions for the “TUT” phase object are also quite acceptable as can be seen from the phase cross section showing the locations of the peaks quite accurately. We can also see the features of the phase object in the amplitude reconstruction. Again, at least partially, it can be attributed to the non-ideality of the SLM. The blurred shape of the reconstructions as compared with the shape of the reconstructions obtained in the simulation tests is due to an approximate modeling of wavefront propagation imbedded in SR-SPAR as defined by Eqs. (5) and (6).
2. Sub-pixel Resolution
The results for sub-pixel super-resolution are presented in Figs. 9 and 10. Again we can see quite good reconstructions for both the “cameraman” and “TUT” phase objects. The comments on amplitude reconstructions are identical to those given above for the pixel resolution case. It is interesting that the reconstructions for sub-pixel resolution look very similar to the reconstructions obtained for pixel resolution. The only difference visible clearly in the cross sections is that the sub-pixel resolution images are smoother than the pixel resolution ones. The blurred character of these reconstructions can be attributed to an approximate modeling of wavefront propagation by Eqs. (5) and (6), as well as to the fact that a much smaller number of sensor pixels are used in these experiments. For the pixel-wise imaging, we use sensor pixels and only sensor pixels for sub-pixel imaging. This reduction in the used sensor size happens due to an increased number of computational pixels required for sub-pixel resolution with . In this case, is the largest data array accepted by the RAM of our computer.
Note also that for sub-pixel resolution , and provided fixed image support in computational pixels, , according to Eq. (12). It is much less than the used ; thus, the condition Eq. (12) is severely violated. It means a discrepancy between reality and modeling of wavefront propagation is what results in the additional blurring in Figs. 9 and 10.
4. CONCLUSION AND DISCUSSION
Computational sub-pixel super-resolution phase retrieval is considered for phase-coded intensity observations. An iterative algorithm is proposed for a lensless setup with the scalar Rayleigh–Sommerfeld model for wavefront propagation. One of the essential instruments of the algorithm is a sparsity hypothesis applied to both phase and amplitude. The efficiency of the algorithm is confirmed by simulation tests and experiments. It is shown that high-level super-resolution can be achieved with a super-resolution factor of up to 4, i.e., the pixel size of the reconstructed object is four times smaller than the pixel size of the sensor. In comparison with the wavelength, the achieved super-resolution is up to two-thirds the wavelength. The demonstrated computational resolution is essentially higher than the diffraction-limited resolution as defined by Abbe’s criterion.
Further work is planned. First, we are going to use a transmissive (not a reflective) SLM, which allows making the propagation distance much smaller and in this way to fulfill the condition Eq. (12). Second, we plan to use in our experiments a physical phase test object, which can allow observing details with resolution smaller than the SLM’s pixel size. Third, in the algorithm development, an adaptive correction of the transfer function for wavefront propagation is to be designed.
Finally, we wish to demonstrate that more impressive super-resolution can be achieved for a smaller freespace propagation distance. It is one of the arguments to use a transmissive SLM instead of a reflective one in our future work.
We show the simulation results obtained for and . The USAF test is used in the experiments for phase modeling. In contrast to the experiments considered above, we assume that the object pixels are smaller than the SLM’s pixels, . Remember that and . For the experiments, we use only a central part of the USAF test, , in order to reduce the size of the reconstructed images and to deal with only the smallest elements of the USAF test chart. The results are shown in Fig. 15 for sub-pixel resolution with . As in our previous figures, we show the phase reconstructions and the true phase in the first row of images and the cross sections in the second row. It is clear that the resolution of the reconstruction for a smaller distance of is much sharper and much more accurate as compared with the reconstruction for , which is actually quite blurred. A comparison of the cross sections allows making further quantitative conclusions on the achieved resolution.
The left cross-section plot is done for the horizontal cross section in areas 2–4 of the USAF image. These vertical lines have a width equal to two object pixels. It can be seen from the plots that SR-SPAR successfully resolved these lines for both the distances and . While the accuracy for is essentially better, we can claim that in both the cases a resolution of 0.7 μm (distance between two vertical lines) is shown.
The right cross section plot is done for the vertical cross section in areas 3–5 of the USAF image, where the distance between the horizontal lines is equal to one pixel, i.e., 0.35 μm. We can see from this plot that SR-SPAR successfully resolved these lines for and failed to do so for . Thus, better resolution of 0.35 μm is achieved for as compared with a lower resolution of 0.7 μm achieved for .
Suomen Akatemia (287150); Ministry of Education and Science of the Russian Federation (Minobrnauka) (3.1893.2017/4.6); Horizon 2020 Framework Programme (H2020) (687328 - HOLO).
The authors would like to thank anonymous reviewers for helpful and stimulating comments.
1. R. K. Tyson, Principles of Adaptive Optics, 4th ed. (CRC Press, 2014).
2. L. Wang and H. Wu, Biomedical Optics: Principles and Imaging (Wiley, 2007).
3. B. Kress and P. Meyrueis, Applied Digital Optics: from Micro-Optics to Nanooptics (Wiley, 2009).
4. E. Candès, X. Li, and M. Soltanolkotabi, “Phase retrieval from coded diffraction patterns,” Appl. Comput. Harmon. Anal. 39, 277–299 (2015).
5. D. L. Misell, “An examination of an iterative method for the solution of the phase problem in optics and electron optics: I. Test calculations,” J. Phys. D 6, 2200–2216 (1973). [CrossRef]
6. W. O. Saxton, “Correction of artefacts in linear and nonlinear high resolution electron micrographs,” J. Microsc. Spectrosc. Electron. 5, 665–674 (1980).
7. G. Pedrini, W. Osten, and Y. Zhang, “Wave-front reconstruction from a sequence of interferograms recorded at different planes,” Opt. Lett. 30, 833–835 (2005). [CrossRef]
8. P. Almoro, G. Pedrini, and W. Osten, “Complete wavefront reconstruction using sequential intensity measurements of a volume speckle field,” Appl. Opt. 45, 8596–8605 (2006). [CrossRef]
9. P. Almoro, A. M. S. Maallo, and S. Hanson, “Fast-convergent algorithm for speckle-based phase retrieval and a design for dynamic wavefront sensing,” Appl. Opt. 48, 1485–1493 (2009). [CrossRef]
10. C. Kohler, F. Zhang, and W. Osten, “Characterization of a spatial light modulator and its application in phase retrieval,” Appl. Opt. 48, 4003–4008 (2009). [CrossRef]
11. L. Camacho, V. Micy, Z. Zalevsky, and J. Garcha, “Quantitative phase microscopy using defocussing by means of a spatial light modulator,” Opt. Express 18, 6755–6766 (2010). [CrossRef]
12. M. Agour, P. Almoro, and C. Falldorf, “Investigation of smooth wave fronts using SLM-based phase retrieval and a phase diffuser,” J. Eur. Opt. Soc. Rapid. Publ. 7, 12046 (2012). [CrossRef]
13. P. Almoro, G. Pedrini, P. N. Gundu, W. Osten, and S. G. Hanson, “Enhanced wavefront reconstruction by random phase modulation with a phase diffuser,” Opt. Lasers Eng. 49, 252–257 (2011). [CrossRef]
14. E. J. Candès, X. Lib, and M. Soltanolkotabi, “Phase retrieval from coded diffraction patterns,” Appl. Comput. Harmon. Anal. 39, 277–299 (2015). [CrossRef]
15. P. Almoro, D. P. Quang Duc, I. Serrano-Garcia David, H. Satoshi, H. Yoshio, T. Mitsuo, and Y. Toyohiko, “Enhanced intensity variation for multiple-plane phase retrieval using spatial light modulator as convenient tunable diffuser,” Opt. Lett. 41, 2161–2164 (2016). [CrossRef]
16. J. W. Goodman, Introduction to Fourier Optics, 3rd ed. (Roberts & Company, 2005).
17. R. W. Gerchberg and W. O. Saxton, “A practical algorithm for the determination of phase from image and diffraction plane pictures,” Optik 35, 237–246 (1972).
18. J. R. Fienup, “Phase retrieval algorithms: a comparison,” Appl. Opt. 21, 2758–2769 (1982). [CrossRef]
19. C. Guo, S. Liu, and J. T. Sheridan, “Iterative phase retrieval algorithms. I: optimization,” Appl. Opt 54, 4698–4708 (2015). [CrossRef]
20. Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev, “Phase retrieval with application to optical imaging: a contemporary overview,” IEEE Signal Process. Mag. 32(3), 87–109 (2015). [CrossRef]
21. E. J. Candès, X. Li, and M. Soltanolkotabi, “Phase retrieval via Wirtinger flow: theory and algorithms,” IEEE Trans. Inf. Theory , 61, 1985–2007 (2015). [CrossRef]
22. Y. Chen and E. J. Candès, “Solving random quadratic systems of equations is nearly as easy as solving linear systems,” 2015, http://statweb.stanford.edu/~candes/papers/TruncatedWF.pdf.
23. V. Katkovnik, “Phase retrieval from noisy data based on sparse approximation of object phase and amplitude,” 2015, http://www.cs.tut.fi/~lasip/DDT/pdfs/Single_column.pdf.
24. Q. Lian, B. Shi, and S. Chen, “Transfer orthogonal sparsifying transform learning for phase retrieval,” Digit Signal Process. 62, 11–25 (2017). [CrossRef]
25. A. Fannjiang, “Absolute uniqueness of phase retrieval with random illumination,” Inverse Probl. 28, 075008 (2012). [CrossRef]
26. W. Bishara, U. Sikora, O. Mudanyali, T. W. Su, O. Yaglidere, S. Luckhart, and A. Ozcan, “Holographic pixel super-resolution in portable lensless on-chip microscopy using a fiber-optic array,” Lab Chip 11, 1276–1279 (2011). [CrossRef]
27. K. Guo, S. Dong, P. Nanda, and G. Zheng, “Optimization of sampling pattern and the design of Fourier ptychographic illuminator,” Opt. Express 23, 6171–6180 (2015). [CrossRef]
28. A. Greenbaum, W. Luo, B. Khademhosseinieh, T. W. Su, A. F. Coskun, and A. Ozcan, “Increased space-bandwidth product in pixel super-resolved lensfree on-chip microscopy,” Sci. Rep. 3, 1717 (2013). [CrossRef]
29. P. Almoro, G. Pedrini, and W. Osten, “Aperture synthesis in phase retrieval using a volume-speckle field,” Opt. Lett. 32, 733–735 (2007). [CrossRef]
30. S. Gazit, A. Szameit, Y. C. Eldar, and M. Segev, “Super-resolution and reconstruction of sparse sub-wavelength images,” Opt. Express 17, 23920–23946 (2009). [CrossRef]
31. J. Song, C. L. Swisher, H. Im, S. Jeong, D. Pathania, Y. Iwamoto, M. Pivovarov, R. Weissleder, and H. Lee, “Sparsity-based pixel super resolution for lens-free digital in-line holography,” Sci. Rep. 6, 24681 (2016). [CrossRef]
32. V. Katkovnik and K. Egiazarian, “Sparse super-resolution phase retrieval from phase-coded noisy intensity patterns,” 2017, http://www.cs.tut.fi/sgn/imaging/sparse/.
33. V. Katkovnik and J. Astola, “High-accuracy wavefield reconstruction: decoupled inverse imaging with sparse modeling of phase and amplitude,” J. Opt. Soc. Am. A 29, 44–54 (2012). [CrossRef]
34. V. Katkovnik and J. Astola, “Sparse ptychographical coherent diffractive imaging from noisy measurements,” J. Opt. Soc. Am. A 30, 367–379 (2013). [CrossRef]
35. K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by sparse 3D transform-domain collaborative filtering,” IEEE Trans. Image Process. 16, 2080–2095 (2007). [CrossRef]
36. A. Danielyan, V. Katkovnik, and K. Egiazarian, “BM3D frames and variational image deblurring,” IEEE Trans. Image Process. 21, 1715–1728 (2012). [CrossRef]
37. N. Verrier and M. Atlan, “Off-axis digital hologram reconstruction: some practical considerations,” Appl. Opt. 50, H136–H146 (2011). [CrossRef]
38. F. Zhang, G. Pedrini, and W. Osten, “Reconstruction algorithm for high-numerical-aperture holograms with diffraction-limited resolution,” Opt. Lett. 31, 1633–1635 (2006). [CrossRef]
39. A. F. Doval and C. Trillo, “Dimensionless formulation of the convolution and angular spectrum reconstruction methods in digital holography,” Proc. SPIE 7387, 73870U (2010). [CrossRef]
40. J. M. Bioucas-Dias and G. Valadão, “Phase unwrapping via graph cuts,” IEEE Trans. Image Process. 16, 698–709 (2007). [CrossRef]
41. Matlab Demo code, http://dx.doi.org/10.6084/m9.figshare.4833932.
42. P. Gemayel, B. Colicchio, A. Dieterlen, and P. Ambs, “Cross-talk compensation of a spatial light modulator for iterative phase retrieval applications,” Appl. Opt. 55, 802–810 (2016). [CrossRef]