You are on page 1of 16

Measurement Science and Technology

PAPER You may also like


- T2* phase imaging and processing for
Comparative study of phase unwrapping brain functional magnetic susceptibility ()
mapping
algorithms based on solving the Poisson equation Zikuan Chen and Vince Calhoun

- Weighted regularized preconditioned


conjugate gradient (PCG) phase
To cite this article: Zixin Zhao et al 2020 Meas. Sci. Technol. 31 065004 unwrapping method
Yong Bian and Bryan Mercer

- Quantitative measurement of high intensity


focused ultrasound pressure field by
optical phase contrast method applying
View the article online for updates and enhancements. non-continuous phase unwrapping
algorithm
Mohd Syahid, Seiji Oyama, Jun Yasuda et
al.

This content was downloaded from IP address 134.147.53.54 on 09/11/2022 at 09:55


Measurement Science and Technology

Meas. Sci. Technol. 31 (2020) 065004 (15pp) https://doi.org/10.1088/1361-6501/ab63e8

Comparative study of phase unwrapping


algorithms based on solving the Poisson
equation
Zixin Zhao1 , Hangying Zhang1, Changqing Ma2, Chen Fan1
and Hong Zhao1
1
  State Key Laboratory for Manufacturing Systems Engineering, School of Mechanical Engineering,
Xi’an Jiaotong University, Xi’an 710049, Shaanxi, People’s Republic of China
2
  Department of Mechanical and Electronic Engineering, Shandong University of Science
and Technology, Tai’an 271019, Shandong, People’s Republic of China

E-mail: zixinzhao@xjtu.edu.cn and zhaohong@xjtu.edu.cn

Received 4 September 2019, revised 14 December 2019


Accepted for publication 19 December 2019
Published 12 March 2020

Abstract
Phase unwrapping is a crucial process to obtain the absolute phase profile in many optical
phase measurement techniques such as interferometry, holography, profilometry, etc. In
this paper, we have studied several phase unwrapping algorithms based on solving the
discrete Poisson equation. The differences among those algorithms lie in two aspects:
one is the calculation of the input for the Poisson equation using the wrapped phase
data and the other is the way to compute the output (unwrapped phase data) using the
corresponding input. Firstly, the method to compute the input for the Poisson equation was
investigated using the finite difference and fast Fourier transform (FFT) methods. Then
different methods, based on FFT or discrete cosine transform, were employed to calculate
the unwrapped phase, and their performances were compared in terms of accuracy
and efficiency. To enhance the precision of those algorithms, an iteration strategy was
introduced and its performance was investigated under different noise conditions. Finally,
several pieces of real phase data was tested by using the direct and iterative methods.
The detailed software package can be found online (www.mathworks.com/matlabcentral/
fileexchange/71810-phase-unwrapping-algorithms-by-solving-the-poisson-equation).

Keywords: phase unwrapping, poisson equation, least square, transport of intensity equation

(Some figures may appear in colour only in the online journal)

1. Introduction due to its high accuracy. Nevertheless, the phase is wrapped


between  −π and π for the most phase demodulation methods.
Plenty of information that we are interested in is coded in the As a result, phase unwrapping is an important process to get
phase of light. To obtain the information we want, the phase the final absolute phase profile. To date, many unwrapping
needs to be decoded and different methods are proposed [1–10]. methods have been proposed to solve the problem. They can
Generally, these methods can be clarified into two groups: the be clarified into two groups: time domain and spatial domain.
interferometric methods [1, 6] and non-interferometric ones In the former group, several wrapped phase maps are needed
[2–5, 7–10]. Specifically, the non-interferometric methods can and the phase unwrapping is accomplished using the same
be divided into three types: the iterative ones [2], the deter- spatial located pixels obtained at different times. Temporal
ministic ones [3, 4, 7, 9, 10] and the combination of the itera- phase unwrapping searches the phase jumps through the time
tive and deterministic methods [5, 8]. Among these methods, axis [11, 12]. If the phase difference between two consecutive
the interferometric methods are the most representative ones points is greater than a predetermined threshold, the phase is

1361-6501/20 /065004+15$33.00 1 © 2020 IOP Publishing Ltd  Printed in the UK


Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Table 1.  Different combinations for methods to calculate the input and the output of the Poisson equation.

Calculate input Least square Transport of intensity equation


Calculate output Equation (3) Equation (6) Equation (7)
Equation (8) LS-FD-FFT [43] TIE-FD-FFT TIE-FFT-FFT [49–56]
Equation (9) LS-FD-DCT [44–47] TIE-FD-DCT [57] TIE-FFT-DCT

Figure 1.  Unwrapping result and the corresponding residual errors for the six direct methods under different noise levels. (a1)–(a6) denote
the LS-FD-FFT, (b1)–(b6) denote the LS-FD-DCT, (c1)–(c6) denote the TIE-FD-FFT, (d1)–(d6) denote the TIE-FD-DCT, (e1)–(e6) denote
the TIE-FFT-FFT, (f1)–(f6) denote the TIE-FFT-DCT (the image size is 512  ×  512 and the noise std means noise standard deviation).

Figure 2.  The residual RMS errors and the consumed time for the six direct methods under different noise levels. (a) The residual RMS
errors, (b) the consumed time.

added or subtracted by 2π at the latter point. It provides a solu- interval in which the phase of each pixel changes from  −π
tion for absolute phase determination and phase unwrapping to π [17]. While in the latter, only one wrapped phase map is
on a discontinuous phase map [13–16], but it should ensure needed and the phase unwrapping is completed using adjacent
that more than two phase maps are obtained during each pixels in different spatial locations (first or second derivative).

2
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 3.  The input for (a) the LS-FD based methods, (b) the TIE-FD based methods, (c) the TIE-FFT methods, and (d) the difference
between (a) and (b), (e) the difference between (a) and (c).

Figure 4.  Flowchart of the iteration process using different direct Poisson solvers.

As a result, the spatial domain methods are more widely used equation (TIE)-based algorithms [49–57], and the polynomial
in the actual applications and they can be divided into two fitting-based algorithms [58–63].
categories: path dependent and path independent. The path- Here, we briefly introduce the development of the LS- and
dependent methods are based on a path integration strategy TIE-based methods. In early times, scholars have sought to
to obtain the unwrapped phase map where a lot of strategies find the absolute phase using the phase differences [41, 42],
have been proposed, such as the mask cut algorithms [18–25], which inspires other researchers to solve the phase unwrap-
the quality-guided path following algorithms [26–37], and the ping problem using the wrapped phase differences. By
Flynn’s algorithms [38–40]. The path-independent methods constructing a discrete Poisson equation, researchers have
seek to find an unwrapped phase whose spatial differences adopted the FFT [43] or DCT [44, 45] to solve the equation,
match the differences in the wrapped phase data as closely which were termed the direct methods. The direct methods
as possible, in which the phase unwrapping is completed by consume very little time, but the performance is not satisfac-
minimization of a global function. Plenty of methods have tory especially for the heavy noise cases. As a result, some
been proposed to accomplish this task, such as the least square iterative strategies [46–48] have been proposed to improve
(LS)-based algorithms [41–48], the transport of intensity the performance of the direct ones at the cost of consuming
3
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 5.  Unwrapping result and the corresponding residual errors for the six iterative methods under different noise levels. (a1)–(a6)
denote the LS-FD-FFT, (b1)–(b6) denote the LS-FD-DCT, (c1)–(c6) denote the TIE-FD-FFT, (d1)–(d6) denote the TIE-FD-DCT, (e1)–(e6)
denote the TIE-FFT-FFT, and (f1)–(f6) denote the TIE-FFT-DCT.

more time. The development of the TIE-based methods is 2.  Direct methods
fascinating since it was not named the TIE in early times in
[49–51] until [52] presented a detailed mathematical deriva- Generally, phase unwrapping can be described as the fol-
tion of the TIE-based phase unwrapping algorithm. We found lowing equation:
that those methods [49–51, 53] are mathematically the same φ(x, y) = ϕ(x, y) + 2k(x, y)π,
(1)
with the TIE-based methods (see detailed analysis in sec-
tion  2). That is why we clarify the methods in [49–51, 53] where φ(x, y) is the unwrapped phase, ϕ(x, y) is the wrapped
into the TIE-based methods. Like the LS-based methods, the phase, and k(x, y) is the integer to be solved. Based on the
TIE-based methods can also be clarified into two groups: the assumption that the unwrapped phase and the wrapped
direct [49–56] and iterative ones [57]. Generally, the iterative phase have the same second derivative in the x and y  direc-
method performs better than the direct ones at the expense of tions, the Poisson equation as shown in equation (2) has been
consuming more time. It is worth noting that [56] proposed a constructed:
TIE-based unwrapping method using two wavelengths, which ∂ 2 φ(x, y) ∂ 2 φ(x, y)
can be viewed as a combination of the spatial and temporal (2) + = ρ(x, y),
∂x2 ∂y2
domain methods. Note that the LS- and TIE-based methods
have in common that they all complete the phase unwrapping where ρ(x, y) is the input for the Poisson equation. Currently,
by establishing a discrete Poisson equation [64]. The differ- two ways are adopted to estimate the input for the Poisson
ence between them lies in that the method of calculating the equation using the wrapped phase data. One is the finite dif-
input for the discrete Poisson equation is different, which is ference (FD) method as shown by equation (3):
discussed comprehensively in our paper. ρ(x, y) = ∇2 ϕ(x, y)
The paper is organized as follows. The second section intro- = W(ϕ(x + 1, y) − ϕ(x, y))
duces six direct methods by combining different methods to −W(ϕ(x, y) − ϕ(x − 1, y))
(3)
calculate the input and the output of the Poisson equation. +W(ϕ(x, y + 1) − ϕ(x, y))
The third section explains an iterative strategy to enhance the
−W(ϕ(x, y) − ϕ(x, y − 1)),
noise performance of six direct methods and the performance
of those iterative methods is investigated under different noise where ∇2 is the Laplacian operator and W is the wrapping
levels. The fourth section  presents some real unwrapping operator satisfying that W(x) = arctan(sin(x)/cos(x)). It
results using the direct and iterative methods respectively. is widely used in the traditional least square (LS) method
Finally, the fifth section concludes the paper. [43–47]. The other is based on the analytical solution of the

4
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 6.  The iteration process of (a) LS-FD-DCT and (b) TIE-FD-DCT with a noise standard deviation equal to 1.

transport of intensity equation  (TIE) and the input for the


Poisson equation can be calculated by equation (4) [52]:
ρ(x, y) = Im(e−iϕ(x,y) ∇2 (eiϕ(x,y) )),
(4)
where Im indicates the imaginary part. Note that if we
expanded equation (4) according to the Euler’s formula, equa-
tion (4) is derived as the following equation [53]:
ρ(x, y) = cos(ϕ(x, y)) · ∇2 (sin(ϕ(x, y)))

− sin(ϕ(x, y)) · ∇2 (cos(ϕ(x, y))). (5)
Since equations (4) and (5) are equivalent, we have classified
these Fourier-based methods [49–51, 53] as the TIE-based
methods. However, using equation  (5) as the input to solve
the Poisson equation, more Fourier transforms (six in [51]
and eight in [50]) are needed. As a result, equation (4) is used
in our paper. To solve equation (4), we can use two different
strategies. One is the same as we adopted in equation (3). We
only need to replace ϕ(x, y) by eiϕ(x,y) and ρ(x, y) can be esti- Figure 7.  The iteration process of TIE-FFT-DCT with a noise
standard deviation equal to 1.
mated by equation (6):
ρ(x, y) = Im{e−iϕ(x,y) [W(eiϕ(x+1,y) − eiϕ(x,y) ) − W(eiϕ(x,y) − eiϕ(x−1,y) ) where DCT and DCT−1 are the discrete cosine transform and
+W(eiϕ(x,y+1) − eiϕ(x,y) ) − W(eiϕ(x,y) − eiϕ(x,y−1) )]}. (6) inverse discrete cosine transform, respectively. M, N are the

dimensional size in x and y  directions. Note that the input
The other strategy is to use the fast Fourier transform (FFT) ρ(x, y) can be computed according to equations  (3), (6) or
to estimate ρ(x, y): (7). Since equation (3) is based on the traditional least square
ρ(x, y) = Im{e−iϕ(x,y) FFT−1 [( fx2 + fy2 )FFT(eiϕ(x,y) )]},
(7) method and equations  (6) and (7) are based on the trans-
port of intensity equation. We can see that the method using
where FFT and FFT−1 represent the Fourier transform and its equations  (3) and (8) is based on the least square using the
inverse transform respectively. finite difference and fast Fourier transform (LS-FD-FFT), the
After the input for the discrete Poisson equation  was method using equations (6) and (8) is based on the transport of
determined, the way to solve the equation is quite important. intensity equation using the finite difference and fast Fourier
Typically, the direct method based on a fast Fourier transform transform (TIE-FD-FFT), and the method using equations (7)
(FFT) or discrete cosine transform (DCT) is widely used, and and (8) is based on the transport of intensity equation using
the unwrapped phase can be obtained as follows: the fast Fourier transforms (TIE-FFT-FFT). By using equa-
ñ ô tion  (9) to compute the unwrapped phase (output), another
−1 FFT(ρ(x, y))
(8) φ(x, y) = FFT , three methods are shown in table  1 with their abbreviated
fx2 + fy2
form. Note that the LS-FD-FFT, LS-FD-DCT, TIE-FFT-FFT
ñ ô were developed by other researchers, especially the latter
DCT (ρ(x, y))
−1 two methods that have been widely used (see the detailed
φ(x, y) = DCT
(9)       ,
2 cos π Mx + cos π Ny − 2 references in table  1). The TIE-FD-DCT was developed by

5
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 8.  Unwrapping result and the corresponding residual errors for the six iterative methods under complex noise levels (additive and
multiplicative/speckle noise). (a1)–(a6) denote the LS-FD-FFT, (b1)–(b6) denote the LS-FD-DCT, (c1)–(c6) denote the TIE-FD-FFT,
(d1)–(d6) denote the TIE-FD-DCT, (e1)–(e6) denote the TIE-FFT-FFT, and (f1)–(f6) denote the TIE-FFT-DCT.

Figure 9.  The iteration process of the TIE-FD-DCT method under complex noise conditions (speckle size equals 1 pixel and the additive
noise deviation is 1). (a) Unwrapped phase maps and (b) residual error maps (wrapped) with the corresponding iteration number N.

us in [57] and the remaining two methods are developed by the same and worst performance over the six methods. As
us for the first time in the manuscript. What is more, all the the noise increases, the traditional LS-based methods per-
FFT-based Poisson solvers have applied the Volkov’s mirror- form better than the TIE-based methods, which shows that
padding scheme [65] to the wrapped phase. The employment the LS-based methods have some anti-noise performance.
of this kind of padding strategy will generate a good approx­ Nevertheless, this anti-noise performance is quite limited,
imation to the Neumann boundary conditions [66–68] and it since all the six methods fail when the noise standard devia-
also fulfils the cyclic conditions of the FFT-based solvers [52]. tion equals 1.
To investigate the performance of the above six methods Specifically, we have computed the root mean square
under different noise levels (Gaussian white noise), we con- (RMS) of the residual error (the difference between the
ducted a simulation using a phase map generated by the unwrapped phase and the ideal one) for the six methods
MATLAB built-in function peaks. The original phase equals with respect to different noise levels and the result is shown
8·peaks (512). We first rewrapped the phase map and then in figure  2(a). It can be seen that the traditional LS-based
different methods were used to unwrap the wrapped phase methods perform best, followed by the TIE-FFT methods, and
map. The unwrapped result is shown in figure  1. It can be the TIE-FD-based methods perform the worst. Figure  2(b)
seen from figure  1 that the LS-FD-DCT performs the best shows the consumed time for six methods under different
for the noise-free condition. The LS-FD-FFT, TIE-FFT-FFT, noise levels. As can be seen, the FD-DCT-based methods use
and TIE-FFT-DCT also have good performance in this condi- less time than the FD-FFT-based methods, which is consistent
tion, while the TIE-FD-FFT and TIE-FD-DCT have almost with the result in [69].

6
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

It can be seen from figure  2(a) that the TIE-FD-based


methods produce significant error when there is no noise at
the wrapped phase map. To find the reason for that, the input
ρ(x, y) for the LS-FD, TIE-FD, and TIE-FFT methods were
calculated and are shown in figures 3(a)–(c) respectively.
Since the LS-FD-based methods have the best accuracy
for the noise free condition, we use the input for the LS-FD
methods as the benchmark, and the difference between
­figures 3(a) and (b) is shown in figure 3(d). Figure 3(e) shows
the difference between figures 3(a) and (c). It can be seen that
calculating the input for the Poisson equation  using equa-
tion (6) has significant errors, which is the main error source of
the TIE-FD-based methods in the noise free condition. We can
also find that calculating the input using equation (7) also has
some errors, but they mainly exist in the border of the map. As
a result, the TIE-FFT-based methods work better than the TIE-
FD-based methods but worse than the LS-FD-based methods.

3.  Iterative methods

It can be seen from the above simulation that the performance Figure 10.  The wrapped phase map obtained with a Twyman–
Green interferometer (a) by measuring an optical planar surface (b)
of the direct methods is not so satisfactory, especially for the
and spherical surface (c).
noisy phase data. To improve the performance of the direct
method, we have introduced an iteration strategy based on also find that the TIE-FFT-DCT performs the worst in the
an unwrapped phase correction process (integer correction). noisy conditions. To find the reason for that, we analyzed
In our previously proposed iterative method [57], the TIE- the detailed iteration process of the TIE-FD-DCT method
FD-DCT was used to calculate the initial unwrapped phase (best performance), the LS-FD-DCT method (poor perfor-
and the stopping condition is that the integer k(x, y) is no mance) and the TIE-FFT-DCT (worst performance) with a
longer changed for all the pixels. Note that the other five direct noise standard deviation equal to 1. Figure  6(a) shows the
methods can also be incorporated into this iteration strategy. iteration process of the LS-FD-DCT. Note that the RMS
The detailed iteration process is shown in figure 4. By using value that represents the difference between the adjacent
different Poisson solvers (different direct methods), the corre­ unwrapped phase maps was used to determine whether the
sponding iterative algorithms can be constructed. Specifically, iteration needs to be terminated or not. It can be seen that
the iterative procedure uses the differences between the first the iteration was terminated after five iterations, while the
unwrapped phase obtained by equation  (8) or (9) and the unwrapped phase map remains almost unchanged during the
second unwrapped phase obtained by equation (1). Thus, the iteration process, which indicates that the proposed iteration
errors in integer k(x, y) can be corrected as much as possible strategy cannot significantly improve the performance of the
with the iteration process. What is more, we have updated the LS-based method under a heavy noise condition. However,
iterative termination criteria since the old condition in [57] is the performance of the TIE-FD-DCT is greatly enhanced by
a quite critical one that makes the iteration process very time- employing the same iteration strategy. As can be seen from
consuming. The new stop criteria are that the integer k(x, y) figure 6(b), the unwrapped phase becomes better and better
is no longer changed for all the pixels or the RMS difference until the iteration is stopped.
between the adjacent residual errors is less than a specified Figure 7 shows the iteration process of the TIE-FFT-DCT
threshold, e.g. 10−5. method. We found an interesting phenomenon that the
unwrapped phase map became better in the first several steps
and then rapidly got worse and worse. Note that the unwrapped
3.1.  Unwrapping performance of the iterative methods under
different Gaussian noise
phase maps are obtained with the same termination threshold
of 10−5. To obtain satisfactory unwrapping results by using
We firstly investigated the performance of the six iterative the iterative method based on TIE-FFT-DCT, it is recom-
methods under different Gaussian noise levels. The stopping mended to use a threshold of 10−1.
threshold was set to 10−5 for all the methods. The result is
shown in figure  5. It can be seen that the performance of
3.2.  Unwrapping performance of the iterative methods under
the six methods is almost the same and perfect in the noise-
different combined noise
free condition. As the noise increases, the TIE-FD-FFT, TIE-
FD-DCT, and TIE-FFT-FFT perform better than the LS-based When measuring the rough surface/object with interferom-
methods. Note that the LS-FD-FFT and LS-FD-DCT per- etry, the speckle or coherent noise widely exists, such as
form almost the same in the three noise conditions. We can in ESPI [70–72], OCT [53, 73–75], etc. The speckle is the

7
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 11.  Unwrapping result of the direct and iterative methods for the wrapped phase data in figure 10. (a1)–(a4) denote the LS-FD-FFT,
(b1)–(b4) denote the LS-FD-DCT, (c1)–(c4) denote the TIE-FD-FFT, (d1)–(d4) denote the TIE-FD-DCT, (e1)–(e4) denote the TIE-FFT-
FFT, and (f1)–(f4) denote the TIE-FFT-DCT.

Figure 12.  The obtained wrapped phase map. (a) Low speckle noise and (b) high speckle noise.

result of the interference among multiple beams with random noise standard deviation equals zero (only the speckle noise
phase differences [76], which heavily degrades the fringe exists), the TIE-FD-FFT, TIE-FD-DCT, and TIE-FFT-FFT
or image contrast. To test the performance of the six itera- methods can successfully unwrap the wrapped phase map.
tive methods under such a complex noise condition. We first The TIE-FFT-DCT method has worse performance than other
generated a wrapped phase map containing a speckle noise TIE-based methods since it has significant errors in the area
with an average size of 1 pixel according to the [77]. Then with a relatively high fringe density as shown in figure 8(f2).
different additive Gaussian noise was added to the wrapped Note that the LS-based methods have the same and worst per-
phase map as shown in the first column of figure 8. Six itera- formance. When the additive noise deviation increases to 0.5,
tive methods were adopted to unwrap the noisy phase maps as can be seen in figures 8(a4)–(f4), only the TIE-FD-based
respectively. Note that the iteration termination threshold was methods can unwrap successfully, the performance of other
set to 10−5 for all the methods except for the TIE-FFT-DCT- four methods become worse. It should be noted that all the six
based method. As we stated before, the TIE-FFT-DCT will methods fail to unwrap in the condition that the additive noise
perform very badly for such a threshold and, as a result, it was deviation is 1. The reason why iterations will not work well
set to 10−1 for the TIE-FFT-DCT method. The unwrapping with high noise level may be that the heavy noise affects the
result is shown in figure 8. It can be seen that the TIE-based calculation of the input of the Poisson equation. Note that the
methods perform better than the LS-based methods under phase obtained by the TIE-FD-based method is closer to the
this complex noise condition. Specifically, when the additive true phase. Specifically, as the iteration number increases, the

8
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 13.  Unwrapping result of the direct and iterative methods for the wrapped phase data in figure 12. (a1)–(a4) denote the LS-FD-FFT,
(b1)–(b4) denote the LS-FD-DCT, (c1)–(c4) denote the TIE-FD-FFT, (d1)–(d4) denote the TIE-FD-DCT, (e1)–(e4) denote the TIE-FFT-
FFT, and (f1)–(f4) denote the TIE-FFT-DCT.

unwrapped phase maps and residual error maps obtained by


the iterative TIE-FD-DCT method are shown in figures 9(a)
and (b) respectively.
It can be seen that, as the iteration number increases to
6 and 7, the residual error maps change a little, causing the
unwrapped phase map almost unchanged. This indicates that
the improvement of the iteration process is very limited when
the heavy noise exists. To successfully unwrap the wrapped
phase with such a heavy noise level, a pre-denoising opera-
tion might be needed. Overall, the TIE-FD-based methods
have the best performance under such a complex noise
condition.

4.  Case studies Figure 14.  The simulated SAR phase data.

4.1.  Application to the interferometry for measuring optical


surfaces with the simulation results shown in figures  1(a1)–(f1) and
5(a1)–(f1). Since the wrapped phase contains little noise, the
To evaluate the performance of the direct and iterative methods direct LS-FD-DCT may be a better choice as it consumes the
on the real data, several experiments were conducted. The first least amount of time. As a matter of fact, this method is widely
experiment was to measure an optical planar and spherical used in quite low noise conditions.
surface using the Twyman–Green interferometer as shown
in figure  10(a). Using the generalized principal component
4.2.  Application to the electronic speckle interferometry
analysis method [78] to demodulate the phase shifted fringe
(ESPI)
patterns, the wrapped phase map with a size of 512  ×  512,
as shown in figures 10(b) and (c), can be obtained. It can be Other experiments were conducted to measure the deforma-
seen that the image contains very little noise. We then applied tion of a roughly planar surface using the phase shifting elec-
the direct and iterative methods to the data respectively, and tronic speckle interferometry. The obtained wrapped phase
the unwrapped phase maps were acquired and are shown in maps with a size of 512  ×  512 are shown in figure 12. It can
figure 11. be seen that the wrapped phase maps are corrupted by speckle
As can be seen from figure  11 that all the methods can noise and the noise in figure  12(b) is stronger than that in
successfully finish the unwrapping task, which is consistent figure 12(a).

9
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 15.  Unwrapping result of the direct and iterative methods for the wrapped phase data in figure 14. (a1) and (a2) denote the LS-FD-
FFT, (b1) and (b2) denote the LS-FD-DCT, (c1) and (c2) denote the TIE-FD-FFT, (d1) and (d2) denote the TIE-FD-DCT, (e1) and (e2)
denote the TIE-FFT-FFT, and (f1) and (f2) denote the TIE-FFT-DCT.

Using the direct and iterative methods to unwrap the


phase maps, the corresponding results are shown in figure 13.
Specifically, figures  13(a1)–(f1) and (a3)–(f3) show the
unwrapped phase maps obtained by the six direct methods
and figures  13(a2)–(f2) and (a4)–(f4) depict the unwrapped
phase maps by using the iterative methods. It can be seen
that the direct methods have a satisfactory performance when
the noise is low, but they fail when the noise is quite heavy.
Besides, the iteration strategy can improve the performance of
the corresponding direct methods. Nevertheless, the iterative
methods perform almost the same in a low noise condition
while they perform differently under a high noise condition.
Concretely, the TIE-FD-based iterative methods perform the
best, which is consistent with the result in simulation. Figure 16.  1D transfer function of the Poisson equation.

10
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Figure 17.  Cosine map of the corresponding unwrapped phase maps shown in figure 15.

Table 2.  The RMS errors for the direct and iterative methods (Rad).

Methods LS-FD-FFT LS-FD-DCT TIE-FD-FFT TIE-FD-DCT TIE-FFT-FFT TIE-FFT-DCT


Data Direct Iterative Direct Iterative Direct Iterative Direct Iterative Direct Iterative Direct Iterative
−3
Figure 10(b) 10 0 10−10 0 1.191 0 1.190 0 0.034 0 0.037 0
Figure 10(c) 0.050 0.012 0.044 0.012 0.427 0 0.426 0 0.394 0 0.397 0.012
Figure 12(a) 0.532 0.174 0.511 0.173 1.286 0.066 1.273 0 1.432 0.010 1.435 0.069
Figure 12(b) 4.768 4.663 4.764 4.659 5.582 0.368 5.573 0 6.745 1.702 6.742 3.904
Figure 14 2.636 2.746 2.628 2.871 9.808 2.666 9.807 0 4.538 2.602 4.538 2.197
Figure 14 1.745 1.048 1.742 0 8.525 3.926 8.523 2.871 3.548 3.831 3.547 2.064

4.3.  Application to the synthetic aperture radar (SAR) where F represents the Fourier transform. If we consider the
interferometry 1D case, the transfer function H( f ) = 1/f 2  , which is similar
to a low pass filter as shown in figure 16. As a result, all the
The data we used is a simulated interferometric SAR (InSAR)
direct methods have the smooth effect.
example supplied in [79] (see details in chapter 3 of the book).
To clearly see this smooth effect, we have presented the
It was generated based on a real digital elevation model of
cosine map of the corresponding unwrapped phase data in
mountainous terrain around Longs Peak, Colorado, using a
figure 17. As depicted by the red arrows in figure 17 the fringe
high-fidelity InSAR simulator that models the SAR point
density in the result obtained by direct methods becomes
spread function, the InSAR geometry, the speckle noise, and
sparse in the area with sudden phase changes. Note that the
the layover and shadow phenomena [80]. The wrapped phase
smooth effect is more significant for TIE-FD-based methods.
with a size of 152  ×  458 is shown in figure 14. It can be seen
Fortunately, the smooth effect can be suppressed by employing
that the wrapped phase contains some high slope features,
our proposed iteration strategy.
which results in high-density phase fringes.
To quantitatively evaluate the performance of those methods,
The unwrapping result obtained using the direct and itera-
we choose the result obtained by the iterative TIE-FD-DCT
tive methods are shown in figures 15(a1)–(f1) and (a2)–(f2)
method as the benchmark. As a result, the RMS error and con-
respectively.
sumed time for those methods are presented in tables 2 and 3
It can be seen that the high slope features in the phase data
respectively. As can be seen in table 2, the iterative methods
obtained by direct methods are lost. As a result, the direct
can get a result with better accuracy. Note that the RMS values
Poisson solvers can be regarded as a low pass filter. As a
in the last two rows are obtained by using a different bench-
matter of fact, if we make the Fourier transform of the Poisson
mark (the zero value indicates the corresponding benchmark).
equation as shown in equation (2), we can get
The error of some iterative methods is slightly larger than the
direct ones, which is shown in red. The reason may be that the
1 two benchmarks are not ideal enough, which indicates that our
F(φ(x, y)) = 2
(10) F(ρ(x, y)),
fx + fy2 iteration strategy for the data with high slopes needs further

11
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

Table 3.  The consumed time for the direct and iterative methods (seconds).

Methods LS-FD-FFT LS-FD-DCT TIE-FD-FFT TIE-FD-DCT TIE-FFT-FFT TIE-FFT-DCT


Data Direct Iterative Direct Iterative Direct Iterative Direct Iterative Direct Iterative Direct Iterative
Figure 10(b) 0.122 0.226 0.056 0.119 0.152 0.470 0.067 0.213 0.133 0.285 0.140 0.303
Figure 10(c) 0.111 0.233 0.057 0.121 0.149 0.615 0.068 0.210 0.133 0.582 0.144 0.485
Figure 12(a) 0.122 1.480 0.058 0.237 0.155 1.387 0.070 0.490 0.137 1.070 0.144 0.585
Figure 12(b) 0.121 2.130 0.058 0.302 0.149 10.857 0.067 3.163 0.140 10.721 0.144 0.440
Figure 14 0.045 0.283 0.025 0.071 0.049 4.583 0.018 0.235 0.059 5.807 0.045 0.125

improvement. The consumed time for the unwrapping process noise condition (as shown in figures 8(c6)–(d6)), a weighting
was calculated using the code implemented with MATLAB on [81–83] or pre-filtering [37, 84–89] strategy can be adopted to
a computer with an Intel Core i5 processor with a clock rate further improve the performance of those methods. Besides,
of 1.8 GHz, and a memory size of 8 GB. It can be seen from with a booming development of deep learning [90–97], using
table 3 that the time needed for the direct methods depends on a proper network to enhance or accomplish the phase unwrap-
the image size and the type of transform, since DCT is faster ping is also a promising research direction.
than FFT, while the factors affecting the time consumed by
the iterative methods lie in two aspects: the time needed in Acknowledgments
one loop (direct method) and the number of loops. Note that
the number of loops depends on the noise level of the data and This work is financially supported by the National Natural
the pre-determined threshold. That is why the iterative TIE- Science Foundation of China (Grant No. 51705404), National
FFT-DCT method with a threshold of 10−1 consumed much Key Research and Development Programs of China (Grant No.
less time than other iterative TIE-based methods (threshold is 2018YFC1603500), China Postdoctoral Science Foundation
10−5) for the data in figure 12(b). (Grant No. 2016M602806), the Fundamental Research Funds
for the Central Universities (Grant No. xjj2017093), the China
Scholarship Council Foundation (Grant No. 201806285004),
5. Conclusions and the Funding for the Growth Plan of Young Teachers in
Shandong.
In this paper, we have studied the phase unwrapping problem,
which is modeled as a discrete Poisson equation. Different
methods to compute the input and output of the equation have ORCID iDs
been presented. As a result, six direct methods (Poisson
solvers) are developed. Simulation and experimental results Zixin Zhao https://orcid.org/0000-0002-6597-8873
show that all six direct methods can unwrap successfully in Hong Zhao https://orcid.org/0000-0003-0927-8634
very low noise conditions. Specifically, the LS-FD-DCT
method performs the best in terms of accuracy and efficiency.
Besides, the traditional LS-FD-based methods show some References
anti-noise properties though it is very limited. What is more,
the direct methods have a smooth effect on the phase data with [1] de Groot P J 2019 A review of selected topics in
high slope features, which also restricts its use for the real interferometric optical metrology Rep. Prog. Phys.
82 056101
applications, especially for the InSAR data. Since the direct
[2] Situ G-H and Wang H-C 2017 Phase problems in optical
methods are very sensitive to noise and high slope features, imaging Frontiers Inf. Technol. Electron. Eng. 18 1277–87
we have developed an iteration strategy to improve the per- [3] Chakraborty T and Petruccelli J C 2017 Source diversity
formance of the direct methods on the phase data under high for transport of intensity phase imaging Opt. Express
noise levels as well as on the phase data with high slopes. 25 9122–37
[4] Li J, Chen Q, Sun J, Zhang J, Pan X and Zuo C 2018 Optimal
Simulation and experimental results show that the itera-
illumination pattern for transport-of-intensity quantitative
tion strategy significantly improves the performance of the phase microscopy Opt. Express 26 27599–614
six direct methods. However, the improvement of the itera- [5] Martinez-Carranza J and Kozacki T 2018 Quantitative phase
tion on the six direct methods is different especially for the imaging with increased spatial coherence based on Fourier
high noise cases. The TIE-FD-based iteration methods have filtering Opt. Lett. 43 5435–8
[6] Zhang J, Chen Q, Li J, Sun J and Zuo C 2018 Lensfree
the best performance as they can get a better result than the
dynamic super-resolved phase imaging based on active
other four methods for the phase data with relatively high micro-scanning Opt. Lett. 43 3714–7
noise. Note that our proposed iteration strategy may also [7] Hu J, Wei Q, Kong Y, Jiang Z, Xue L, Liu F, Kim D Y, Liu C
be applied to other unwrapping algorithms to improve their and Wang S 2019 Higher order transport of intensity
performance. Overall, the iterative methods, especially the equation methods: comparisons and their hybrid application
for noise adaptive phase imaging IEEE Photonics 11 1–14
TIE-FD-DCT-based method, are recommended to be used
[8] Li J, Xu J, Zhong L, Zhang Q, Wang H, Tian J and Lu X 2019
in the actual applications. However, the performance of the An orthogonal direction iterative algorithm of the transport-
iterative methods was not so satisfactory under a very heavy of-intensity equation Opt. Lasers Eng. 120 6–12

12
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

[9] Lu L, Sun J, Zhang J, Fan Y, Chen Q and Zuo C 2019 [31] Zhong H, Tang J, Zhang S and Chen M 2011 An improved
Quantitative phase imaging camera with a weak diffuser quality-guided phase-unwrapping algorithm based on
Frontiers Phys. 7 77 priority queue IEEE Geosci. Remote Sens. Lett. 8 364–8
[10] Wen Y and Asundi A 2019 3D profile measurement for [32] Zhao M, Huang L, Zhang Q, Su X, Asundi A and Kemao Q
stepped microstructures using region-based transport of 2011 Quality-guided phase unwrapping technique:
intensity equation Meas. Sci. Technol. 30 2 comparison of quality maps and guiding strategies Appl.
[11] Servin M, Padilla M and Garnica G 2018 Super-sensitive Opt. 50 6214–24
two-wavelength fringe projection profilometry with [33] Liu G, Wang R, Deng Y, Chen R, Shao Y and Yuan Z 2013 A
2-sensitivities temporal unwrapping Opt. Lasers Eng. new quality map for 2D phase unwrapping based on gray
106 68–74 level co-occurrence matrix IEEE Geosci. Remote Sens. Lett.
[12] Wu S, Zhu L, Pan S and Yang L 2016 Spatiotemporal three- 11 444–8
dimensional phase unwrapping in digital speckle pattern [34] Zhao M, Wang H and Kemao Q 2015 Snake-assisted quality-
interferometry Opt. Lett. 41 1050–3 guided phase unwrapping for discontinuous phase fields
[13] Huntley J M and Saldner H 1993 Temporal phase-unwrapping Appl. Opt. 54 7462–70
algorithm for automated interferogram analysis Appl. Opt. [35] Arevalillo-Herráez M, Villatoro F R and Gdeisat M A 2016
32 3047–52 A robust and simple measure for quality-guided 2D phase
[14] Zhao H, Chen W and Tan Y 1994 Phase-unwrapping unwrapping algorithms IEEE Trans. Image Process.
algorithm for the measurement of three-dimensional object 25 2601–9
shapes Appl. Opt. 33 4497–500 [36] Lopez-Garcia L, Garcia-Arellano A and Cruz-Santos W 2018
[15] Saldner H O and Huntley J M 1997 Temporal phase Fast quality-guided phase unwrapping algorithm through
unwrapping: application to surface profiling of a pruning strategy: applications in dynamic interferometry
discontinuous objects Appl. Opt. 36 2770–5 Appl. Opt. 57 3126–33
[16] Zuo C, Huang L, Zhang M, Chen Q and Asundi A 2016 [37] Xie X, Zeng Q, Liao K and Liu Q 2019 Efficient phase
Temporal phase unwrapping algorithms for fringe unwrapping algorithm based on cubature information
projection profilometry: a comparative review Opt. Lasers particle filter applied to unwrap noisy continuous phase
Eng. 85 84–103 maps Opt. Express 27 9906–24
[17] Servin M, Padilla J M, Gonzalez A and Garnica G 2015 [38] Flynn T J 1997 Two-dimensional phase unwrapping with
Temporal phase-unwrapping of static surfaces with minimum weighted discontinuity J. Opt. Soc. Am. A
2-sensitivity fringe-patterns Opt Express 23 15806–15 14 2692–701
[18] Goldstein R M, Zebker H A and Werner C L 1988 Satellite [39] Xu J, An D, Huang X and Yi P 2016 An efficient minimum-
radar interferometry: two-dimensional phase unwrapping discontinuity phase-unwrapping method IEEE Geosci.
Radio Sci. 23 713–20 Remote Sens. Lett. 13 666–70
[19] Gutmann B and Weber H 2000 Phase unwrapping with the [40] Gdeisat M 2018 Performance evaluation and acceleration of
branch-cut method: role of phase-field direction Appl. Opt. Flynn phase unwrapping algorithm using wraps reduction
39 4802–16 algorithms Opt. Lasers Eng. 110 172–8
[20] Chavez S, Xiang Q-S and An L 2002 Understanding phase [41] Takajo H and Takahashi T 1988 Noniterative method for
maps in MRI: a new cutline phase unwrapping method obtaining the exact solution for the normal equation in
IEEE Trans. Med. Imaging 21 966–77 least-squares phase estimation from the phase difference
[21] Gao D and Yin F 2011 Mask cut optimization in two- J. Opt. Soc. Am. A 5 1818–27
dimensional phase unwrapping IEEE Geosci. Remote Sens. [42] Takajo H and Takahashi T 1988 Least-squares phase
Lett. 9 338–42 estimation from the phase difference J. Opt. Soc. Am. A
[22] Zheng D and Da F 2011 A novel algorithm for branch cut 5 416–25
phase unwrapping Opt. Lasers Eng. 49 609–17 [43] Pritt M D and Shipman J S 1994 Least-squares two-
[23] Zhang Y-T, Huang M-J, Liang H-R and Lao F-Y 2012 Branch dimensional phase unwrapping using FFT’s IEEE Trans.
cutting algorithm for unwrapping photoelastic phase map Geosc. Remote Sens. 32 706–8
with isotropic point Optics Lasers Eng. 50 619–31 [44] Ghiglia D C and Romero L A 1994 Robust two-dimensional
[24] de Souza J C, Oliveira M E and dos Santos P A 2015 Branch- weighted and unweighted phase unwrapping that uses
cut algorithm for optical phase unwrapping Opt. Lett. fast transforms and iterative methods J. Opt. Soc. Am. A
40 3456–9 11 107–17
[25] Zhou L, Chai D, Xia Y, Ma P and Lin H 2018 Interferometric [45] Kerr D, Kaufmann G and Galizzi G 1996 Unwrapping of
synthetic aperture radar phase unwrapping based on sparse interferometric phase-fringe maps by the discrete cosine
Markov random fields by graph cuts J. Appl. Remote Sens. transform Appl. Opt. 35 810–6
12 015006 [46] Xia H-T, Guo R-X, Fan Z-B, Cheng H-M and Yang B-C 2012
[26] Lim H, Xu W and Huang X 1995 Two new practical methods Non-invasive mechanical measurement for transparent
for phase unwrapping Int. Geoscience and Remote Sensing objects by digital holographic interferometry based on
Symp., IGARSS’95. Quantitative Remote Sensing for iterative least-squares phase unwrapping Exp. Mech.
Science and Applications (IEEE) 1 196–8 52 439–45
[27] Herráez M A, Burton D R, Lalor M J and Gdeisat M A 2002 [47] Xia H, Montresor S, Guo R, Li J, Yan F, Cheng H and Picart P
Fast two-dimensional phase-unwrapping algorithm based 2016 Phase calibration unwrapping algorithm for phase
on sorting by reliability following a noncontinuous path data corrupted by strong decorrelation speckle noise Opt.
Appl. Opt. 41 7437–44 Express 24 28713–30
[28] Su X and Chen W 2004 Reliability-guided phase unwrapping [48] Guo Y, Chen X and Zhang T 2014 Robust phase unwrapping
algorithm: a review Opt. Lasers Eng. 42 245–61 algorithm based on least squares Opt. Lasers Eng.
[29] Kemao Q, Gao W and Wang H 2008 Windowed Fourier- 63 25–9
filtered and quality-guided phase-unwrapping algorithm [49] Volkov V V and Zhu Y 2003 Deterministic phase unwrapping
Appl. Opt. 47 5420–8 in the presence of noise Opt. Lett. 28 2156–8
[30] Kemao Q, Gao W and Wang H 2010 Windowed Fourier [50] Schofield M A and Zhu Y 2003 Fast phase unwrapping
filtered and quality guided phase unwrapping algorithm: on algorithm for interferometric applications Opt. Lett.
locally high-order polynomial phase Appl. Opt. 49 1075–9 28 1194–6

13
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

[51] Jeught S, Sijbers J and Dirckx J 2015 Fast Fourier-based phase interferometry for real-time in-plane rotation analysis Opt.
unwrapping on the graphics processing unit in real-time Express 26 8744–55
imaging applications J. Imaging 1 31–44 [71] Li B, Tang C, Zheng T and Lei Z 2019 Fully automated
[52] Martinez-Carranza J, Falaggis K and Kozacki T 2017 Fast extraction of the fringe skeletons in dynamic electronic
and accurate phase-unwrapping algorithm based on the speckle pattern interferometry using a U-Net convolutional
transport of intensity equation Appl. Opt. 56 7079–88 neural network Opt. Eng. 58 023105
[53] Pijewska E, Gorczynska I and Szkulmowski M 2019 [72] Pokharna H and Schajer G S 2019 Quasi single-frame
Computationally effective 2D and 3D fast phase measurements with phase-stepped ESPI Opt. Lasers Eng.
unwrapping algorithms and their applications to Doppler 121 181–8
optical coherence tomography Biomed. Opt. Express 10 [73] Golde J, Kirsten L, Schnabel C, Walther J and Koch E 2018
1365–82 Handbook of Advanced Non-Destructive Evaluation
[54] Pandey N, Ghosh A and Khare K 2016 Two-dimensional (Cham: Springer) pp 1–44
phase unwrapping using the transport of intensity equation [74] Leitgeb R A 2019 En face optical coherence tomography: a
Appl. Opt. 55 2418–25 technology review Biomed. Opt. Express 10 2177–201
[55] Pandey N, Singh M, Ghosh A and Khare K 2018 Optical [75] Zhang P, Manna S K, Miller E B, Jian Y, Meleppat R K,
surface measurement using accurate carrier estimation in Sarunic M V, Pugh E N and Zawadzki R J 2019 Aperture
Fourier transform fringe analysis and phase unwrapping phase modulation with adaptive optics: a novel approach
based upon transport of intensity equation J. Opt. for speckle reduction and structure extraction in optical
47 389–95 coherence tomography Biomed. Opt. Express 10 552–70
[56] Cheng H, Wang J, Gao Y, Zhang Q and Wei S 2019 Phase [76] Goodman J W 2007 Speckle Phenomena in Optics: Theory
unwrapping based on transport-of-intensity equation with and Applications (Englewood, NJ: Roberts and Company
two wavelengths Opt. Eng. 58 054103 Publishers) (https://doi.org/10.1117/3.2548484)
[57] Zhao Z, Zhang H, Xiao Z, Du H, Zhuang Y, Fan C and Zhao H [77] Equis S and Jacquot P 2006 Simulation of speckle complex
2018 Robust 2D phase unwrapping algorithm based on amplitude: advocating the linear model Proc. SPIE 6341
the transport of intensity equation Meas. Sci. Technol. 634138
30 015201 [78] Vargas J, Sorzano C, Estrada J and Carazo J 2013
[58] Liang Z-P 1996 A model-based method for phase unwrapping Generalization of the principal component analysis
IEEE Trans. Med. Imaging 15 893–7 algorithm for interferometry Opt. Commun. 286 130–4
[59] Langley J and Zhao Q 2009 Unwrapping magnetic resonance [79] Ghiglia D C and Pritt M D 1998 Two-Dimensional Phase
phase maps with Chebyshev polynomials Magn. Reson. Unwrapping: Theory, Algorithms, and Software vol 4 (New
Imaging 27 1293–301 York: Wiley)
[60] Langley J and Zhao Q 2009 A model-based 3D phase [80] Bioucas-Dias J, Katkovnik V, Astola J and Egiazarian K 2008
unwrapping algorithm using Gegenbauer polynomials Phys. Absolute phase estimation: adaptive local denoising and
Med. Biol. 54 5237 global unwrapping Appl. Opt. 47 5358–69
[61] Zhao Z, Zhao H, Zhang L, Gao F, Qin Y and Du H 2014 [81] Wang X, Fang S and Zhu X 2017 Weighted least-squares
2D phase unwrapping algorithm for interferometric phase unwrapping algorithm based on a non-interfering
applications based on derivative Zernike polynomial fitting image of an object Appl. Opt. 56 4543–50
technique Meas. Sci. Technol. 26 017001 [82] Yan L, Zhang H, Zhang R, Xie X and Chen B 2019 A robust
[62] Kulkarni R and Rastogi P 2018 Phase unwrapping algorithm phase unwrapping algorithm based on reliability mask and
using polynomial phase approximation and linear Kalman weighted minimum least-squares method Opt. Lasers Eng.
filter Appl. Opt. 57 702–8 112 39–45
[63] Tellez-Quinones A, Legarda-Saenz R, Salazar-Garibay A, [83] Li B, Tang C, Zhou Q and Lei Z 2019 Weighted least-squares
Valdiviezo N J and Leon-Rodriguez M 2019 Direct phase phase-unwrapping algorithm based on the orientation
unwrapping method based on a local third-order polynomial coherence for discontinuous optical phase patterns Appl.
fit Appl. Opt. 58 436–45 Opt. 58 219–26
[64] Zuo C 2017 Connections between transport of intensity [84] Xie X 2016 Iterated unscented Kalman filter for phase
equation and two-dimensional phase unwrapping unwrapping of interferometric fringes Opt. Express
(arXiv:1704.03950) 24 18872–97
[65] Volkov V, Zhu Y and De Graef M 2002 A new symmetrized [85] Waghmare R G, Sukumar P R, Subrahmanyam G R,
solution for phase retrieval using the transport of intensity Singh R K and Mishra D 2016 Particle-filter-based phase
equation Micron 33 411–6 estimation in digital holographic interferometry J. Opt. Soc.
[66] Roddier F 1990 Wavefront sensing and the irradiance transport Am. A 33 326–32
equation Appl. Opt. 29 1402–3 [86] Xia H, Montresor S, Picart P, Guo R and Li J 2018
[67] Martinez-Carranza J, Falaggis K, Kozacki T and Comparative analysis for combination of unwrapping and
Kujawinska M 2013 Effect of imposed boundary conditions de-noising of phase data with high speckle decorrelation
on the accuracy of transport of intensity equation based noise Opt. Lasers Eng. 107 71–7
solvers Modeling Aspects in Optical Metrology IV [87] Guo R, Zhang W, Liu R, Duan C and Wang F 2018 Phase
8789 87890N unwrapping in dual-wavelength digital holographic
[68] Zuo C, Chen Q, Li H, Qu W and Asundi A 2014 Boundary- microscopy with total variation regularization Opt. Lett.
artifact-free phase retrieval with the transport of intensity 43 3449–52
equation II: applications to microlens characterization Opt. [88] Krishnan J P, Figueiredo M A and Bioucas-Dias J M 2019
Express 22 18310–24 SURE-fuse WFF: a multi-resolution windowed Fourier
[69] Shi W, Zhu Y and Yao Y 2010 Discussion about the DCT/ analysis for interferometric phase denoising IEEE Access
FFT phase-unwrapping algorithm for interferometric 7 120708–23
applications Optik 121 1443–9 [89] Gao Y, Zhang S, Li T, Guo L, Chen Q and Li S 2019 A novel
[70] Wang S, Lu M, Bilgeri L M, Jakobi M, Bloise F S and two-step noise reduction approach for interferometric phase
Koch A W 2018 Temporal electronic speckle pattern images Opt. Lasers Eng. 121 1–10

14
Meas. Sci. Technol. 31 (2020) 065004 Z Zhao et al

[90] Dardikman G, Turko N A and Shaked N T 2018 Deep learning [94] Yin W, Chen Q, Feng S, Tao T, Huang L, Trusiak M, Asundi A
approaches for unwrapping phase images with steep spatial and Zuo C 2019 Temporal phase unwrapping using deep
gradients: a simulation 2018 IEEE Int. Conf. on the Science learning Sci. Rep. 9 20175
of Electrical Engineering in Israel (ICSEE) (IEEE) pp 1–4 [95] Yin W, Zuo C, Feng S, Tao T and Chen Q 2019 Bi-frequency
[91] Dardikman G and Shaked N T 2018 Phase unwrapping using temporal phase unwrapping using deep learning Proc. SPIE
residual neural networks Computational Optical Sensing 10991 109910D95
and Imaging (Optical Society of America) p CW3B.5 [96] Zhang J, Tian X, Shao J, Luo H and Liang R 2019 Phase
[92] Spoorthi G, Gorthi S and Gorthi R K S S 2018 Phasenet: a unwrapping in optical metrology via denoised and
deep convolutional neural network for two-dimensional convolutional segmentation networks Opt. Express 27
phase unwrapping IEEE Signal Process. Lett. 26 54–8 14903–12
[93] Wang K, Li Y, Kemao Q, Di J and Zhao J 2019 One-step [97] Zhang T, Jiang S, Zhao Z, Dixit K, Zhou X, Hou J, Zhang Y
robust deep learning phase unwrapping Opt. Express and Yan C 2019 Rapid and robust two-dimensional phase
27 15100–15 unwrapping via deep learning Opt. Express 27 23173–85

15

You might also like