See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/299618290

Phase shift estimation in interferograms with unknown phase step

Article in Optics Communications · April 2016

DOI: 10.1016/j.optcom.2016.03.063

CITATIONS READS
11 247

3 authors:

Oscar S Dalmau-Cedeño Mariano Rivera

Centro de Investigación en Matemáticas (CIMAT) Centro de Investigación en Matemáticas (CIMAT)
68 PUBLICATIONS 341 CITATIONS 136 PUBLICATIONS 1,735 CITATIONS

SEE PROFILE SEE PROFILE

Christhian Adonai González

Universidad Iberoamericana León
16 PUBLICATIONS 162 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Diabetic Retinopathy Detection using Image Processing View project

Computational Methods for Optical Metrology View project

All content following this page was uploaded by Oscar S Dalmau-Cedeño on 07 April 2016.

The user has requested enhancement of the downloaded file.

 Phase shift estimation in interferograms with unknown
phase step
Oscar Dalmaua,∗, Mariano Riveraa , Adonai Gonzalezb
a
Centre for Mathematical Research, CIMAT A.C. Jalisco S/N Valenciana, CP 36240,
Guanajuato,Gto, Mexico.
b
Universidad Iberoamericana Leon, Dpto de Ciencias e Ingenieria, Blvd. Jorge Vertiz
Campero 1640, Col. Cañada de Alfaro. C.P. 37238, Leon, Gto, Mexico.

Abstract
We first present two closed formulas for computing the phase shift in in-
terferograms with unknown phase step. These formulas obtain theoretically
the exact phase step in fringe pattern without noise and only require the
information in two pixels of the image. The previous formulas allows us
to define a functional that yields an estimate of the phase step in interfer-
ograms corrupted by noise. In the experiment we use the standard Least
Square formulation which also yields a closed formula, although the general
formulation admits a robust potential. We provide two possible implemen-
tations of our approach, one in which the sites can be randomly selected and
the other in which we can scan the whole image. The experiments show that
the proposed algorithm presents the best results compared with state of the
art algorithms.
Keywords: Phase-shifting, two-step demodulation algorithm, phase
retrieval, fringe pattern analysis

1. Introduction
In the last years there has been an increasing interest in the two step
demodulation problem with an arbitrary unknown phase step; see, for exam-
ple, Refs. [1, 2, 3, 4, 5] and references therein. On the other hand, Gao et al.

∗
Corresponding author
Email addresses: [email protected] (Oscar Dalmau ), [email protected] (Mariano
Rivera), [email protected] (Adonai Gonzalez)

Preprint submitted to Optics Communications February 29, 2016

in [6] propose an algorithm which computes the unknown phase shift from
three phase shifting interferograms. H.V. Brug in [7] presents an algorithm
that estimates the phase shift which is based on the correlation between two
interferograms, see also Ref. [8]. On the other hand, Vargas et al. propose
a Gram-Schmidt orthonormalization algorithm for solving the two-step de-
modulation problem [3]. Although, this is a very fast algorithm, it could be
very sensitive to noise. In order to solve the main limitation of the provi-
ous algoritmm Trusiak and Patorski [5] propose a preproccesing technique,
based on the Hilbert-Huang transform, for filtering the fringe patterns. Even
though, the last two previous algorithms do not estimate the phase shift, re-
cently Niu et. al in [9] present an algorithm, also based on the Gram-Schmidt
orthonormalization, that computes the phase step. Using the same idea of
orthonormalization of two vectors, Chunshu et al. [2] build a parallelogram
in which the diagonals are perpendicular and are defined as the sum and the
difference of the interferogram vectors. Finally, the phase step can be ap-
proximated by using the ratio between the length of the two diagonal vectors.
In [10], Deng et al. propose another alternative to approximate the phase
step based on the Euclidean distance between the interferogram vectors. The
previous works obtain, in general, an approximation of the wrapped phase
under the corresponding assumption of each algorithm. Some of them also
provide an explicit estimation of the phase step. Even in the case of fringe
patterns without noise, all these alternative gives an approximation of the
wrapped phase and/or the phase step. On the other hand, Deng et al. [11]
present an algorithm that computes the phase shift, which is based on the
ratio of the extreme values of interferograms. The main drawback of the
previous approach is its sensitivity to noise, due to the location of extreme
values in noisy fringe patterns do not necessarily allow us to compute an
accurate phase shift.
The main contribution of this work is to present closed formulas for com-
puting the phase shift in interferograms with unknown phase shift. We also
provide proof that these formulas obtains theoretically exact phase shift in
fringe pattens without noise. To the best of our knowledge, this is the first
work that presents closed formulas for computing the phase step between
interferograms. Finally, we present an algorithm based on the previous for-
mulas that obtains exact phase shift in fringe pattens without noise and
also presents excellent phase step estimation in interferograms with noise.
We demonstrate experimentally that our proposal presents the best results
compared with state of the art algorithms.

2
 The rest of the paper is organized as follows: Section 2 presents closed
formulas for computing the phase step. Then, based on the previous re-
sults we present an algorithm for estimating the phase step; Section 3 shows
experimental results; and Section 4 presents the conclusions.

2. Proposed Algorithm
First, we introduce two closed formulas for computing the exact phase
step in interferograms without noise. Based on the previous formulas, we
propose a functional that allows us to estimate the phase step in interfero-
grams corrupted by additive noise.

2.1. Closed formulas for computing the exact phase step

The observation models for the two-step demodulation problem are given
by the following equations
Ii (r) = a(r) + b(r) cos(φ(r) + δi ) + ηi , i ∈ {1, 2}, (1)
where r represents a pixel in the image with r = (x, y); x = 0, 1, . . . , Nx −
1; y = 0, 1, . . . , Ny − 1 and Nx , Ny are the number of columns and rows
respectively. a(r) is the background intensity variation, b(r) modulation
strength, φ(r) is the phase, δ = δ1 − δ0 is the phase shift and ηi ∼ N (0, σi ) is
Gaussian noise with standard deviation σi . From now on, we define, without
loss of generality, δ0 = 0 and therefore δ = δ1 . Taking into account the
sign ambiguity of the cosine function we also consider δ ∈ (0, π) which is
sufficient in practical situations. These models can be simplified by using a
normalization process [12, 13, 14, 15]. The normalization consists basically
in filtering out the background intensity variation a(r) and the modulation
factor b(r). Normalization is a very important process in the estimation of the
phase step, due to the spatially varying background and modulation factor
may be error sources. After normalizing the interferograms one obtains the
following models:
I(r) = cos(φ(r)), (2)
J(r) = cos(φ(r) + δ). (3)
On the other hand, it is well known that the wrapped phase of φ(r), i.e.
W[φ(r)] can be obtained by using the formula [8, 11]
 
αI(r) − J(r)
W[φ(r)] = arctan , (4)
βI(r)

3
where

α = cos(δ), (5)
2 1/2
β = sin(δ) = (1 − α ) . (6)

Therefore, the challenge is to compute or estimate α, which is equivalent to

estimate the phase shift, due to δ = arccos(α).
The following propositions provide two closed formulas for computing
α. The main assumption is that the fringe patterns have been previously
normalized [12, 13, 14].

Proposition 1. Let r, s be two different sites in the image that satisfy the
following condition
I(r)J(r) 6= I(s)J(s), (7)
then

α = mode{α1 (r), α2 (r), α1 (s), α2 (s)}, (8)

where
def 1/2
αi (r) = I(r)J(r) + (−1)i [(1 − I(r)2 )(1 − J(r)2 )] (9)

for i ∈ {1, 2} and I(r), J(r) are defined in Eqs. (2)-(3) respectively.

Proof. The proof of the previous proposition is straightforward. However,

we provide some details. First, let us define K(r) = sin(φ(r)). Using also
the definitions (5)-(6), the Eqs. (2)-(3) can be transformed into the following
set of quadratic system of equations:

αI(r) − βK(r) = J(r),
S(r) = (10)
α2 + β 2 = 1,

for all r ∈ L. The number of system of equations is equal to number of pixels

of the image |L|.
Transforming the system of equations S(r) and taking into account that
I(r)2 + K(r)2 = 1, one obtains that α satisfies the following quadratic equa-
tion

α2 − 2I(r)J(r)α + I(r)2 + J(r)2 − 1 = 0, ∀r. (11)

4
Solving the previous equation for ‘α’ one obtains:
1/2
α1,2 (r) = I(r)J(r) ± [(1 − I(r)2 )(1 − J(r)2 )] . (12)

As cos(δ) is solution of the system (10) for each r then at least one of the
following equalities hold: α1 (r) = cos(δ) or α2 (r) = cos(δ) for all r ∈ L.
That is, the value cos(δ) appears at least |L| times in the dataset {I(r)J(r)±
1/2
[(1 − I(r)2 )(1 − J(r)2 )] | r ∈ L}. Note that, if I(r) = 1 or J(r) = 1 then
α1 (r) = α2 (r) = cos(δ) and in this case cos(δ) appears more than |L| times.
Therefore, we can compute α = cos(δ) with the information in two sites
r, s ∈ L that satisfy (7), in order to have two different quadratic equation
(11), by selecting the mode of the values α1,2 (r), α1,2 (s).
Note that due to the symmetry of αi (r) with respect to I(r), J(r) we do
not require to know the precedence of the images to compute the phase shift
δ. This is only needed for computing the wrapped phase of φ(r), (4).
If the fringe patterns satisfy the models in Eqs. (2)-(3) the previous
proposition gives an exact formula for computing α. Additionally, this for-
mula does not require any other assumption, as for example algorithms in
Refs [3, 4, 5, 2] which should fulfill other assumptions that in practice are
very difficult to verify, see Refs. [3, 4, 5, 2] for details. However, if we corrupt
the images with Gaussian noise, then the fringe patterns do not satisfy the
models (2)-(3). Therefore, in order to use the formula (8) the fringe pat-
terns need to be previously normalized, guaranteeing that |I(r)| ≤ 1 and
|J(r)| ≤ 1. In the last case we should use more that two pixel for estimat-
ing α. In fact we can use up to |L| pixels and one can estimate α as the
1/2
mode of the dataset {I(r)J(r) ± [(1 − I(r)2 )(1 − J(r)2 )] | r ∈ K} where
K ⊂ {1, 2, · · · , |L|} with |K| > 1. Of course, the computational time in-
creases while increasing |K|. Another problem of the previous approach is
that the estimation of α deteriorates rapidly with the increasing of noise
levels.
Based on the previous ideas, we propose the next proposition which pro-
vides a more practical formula for computing α with the advantage that this
new formula can also be extended to the case of fringe patterns with noise.

Proposition 2. Let r, s be two different sites in the image that satisfy con-
dition (7). If the images I, J are normalized fringe patterns; i.e., satisfy

5
Eqs. (2)-(3), and δ ∈ (0, π) then

I(r)2 − I(s)2 + J(r)2 − J(s)2

α= . (13)
2(I(r)J(r) − I(s)J(s))

Proof. Let us consider two sites r, s that satisfy the condition (7). Then,
we have two systems S(r), S(s) and both systems share the same α, β and
equation α2 +β 2 = 1. Solving the both system of equations S(r) and S(s) for
α, one obtains two quadratic equations (11) for r and s respectively. Solving
the new system of quadratic equations for α we obtain the formula (13).
It is important to remark that for computing α with the closed formula
(13), one only requires to find, among |L|(|L| − 1)/2 possible combinations,
two pixels r, s in the images for which I(r)J(r) 6= I(s)J(s). These two pixels
can be found by selecting randomly two sites in the image that satisfy the
previous condition. In general, the probability to find two pixels that satisfy
condition (7) is very high, see Fig. 1 for illustration purposes.

1.00000
φ1
0.99995 φ2
φ4
0.99990 φ3

0.99985
probability

0.99980

0.99975

0.99970

0.99965

0.0 0.5 1.0 1.5 2.0 2.5 3.0

δ

Figure 1: Probability of pair of pixels that satisfy condition (7) for different phase shift
using interferograms in Fig. 2.

6
 Another alternative is to consider that r, s are neighboring pixels. In this
case we can define new images A, B, C such that

A(r) = I(r)2 , B(r) = J(r)2 , C(r) = I(r)J(r). (14)

Therefore, for computing α according to (13), one can use any discrete ap-
proximation of the directional derivative of A, B and C, or simply a partial
derivative. For example, we can use the partial derivative in the x-direction
and by selecting a pixel r for which ∂x C(r) 6= 0, one can also compute α with
the following formula
∂x A(r) + ∂x B(r)
α= . (15)
2∂x C(r)
2.2. Phase step estimation
The formula (13), or the alternative (15), allows us to compute the exact
phase step α in interferograms obtained from the models (2)-(3). However,
this formula could be sensitive in fringe patterns with noise. In order to
overcome this limitation, we propose the following optimization problem
X
α∗ = arg min ρ(εr,s (α)), (16)
α
r,s∈L

where
def
εr,s (α) = a(r, s)α − b(r, s), (17)
def
a(r, s) = 2(I(r)J(r) − I(s)J(s)), (18)
def
b(r, s) = I(r)2 − I(s)2 + J(r)2 − J(s)2 . (19)

that allows us to obtain an estimate α∗ of the phase step. Note that the most
important element is this formulation is the error or residual term defined in
Eqs. (17)-(19). This error term is based on the formula (13). On the other
hand, the optimization problem (16) also requires to define the function
ρ(·). According to our experiments ρ(x) = x2 (the standard Least Square
formulation) produces good results, although we can use robust alternatives
[16] with the price of increasing the computational cost. In this case α∗ can
be computed using the following closed formula
P
r,s∈L a(r, s)b(r, s)
α∗ = P 2
. (20)
r,s∈L a(r, s)

7
In the previous equation, we do not require to apply the sum over all possible
pairs r, s ∈ L. In practice, we can select a predefined number of pairs k >
1, see Algorithm 1, in the experiments we use k = |L|. Note also that
2
P
we only need to guarantee that r,s∈L a(r, s) 6= 0, it means that a(r, s)
could be equal to zero for a particular pair r, s because in this case the
contribution in both the numerator and denominator is zero and therefore,
this does not affect the estimation of α∗ . The formula (20) can be even used
in interferograms without noise, i.e., that satisfy the models (2)-(3), because
in this case εr,s (α) = 0, ∀r, s; according to Proposition 2, and therefore α∗ =
α. That is, the formula (20) yields the exact phase step in interferograms
without noise.
Algorithm 1 Random Two Step (RTS) algorithm
Procedure RTS(I, J, k) . k > 1 is the number of pairs r, s
n ← |L| . Number of pixels in the image
i, ν, δ ← 0
while i < k or δ = 0 do
r ← random(1, n)
s ← random(1, n)
ν ← ν + a(r, s)b(r, s) . Use Eqs. (18)-(19)
δ ← δ + a(r, s)2 . Use (18)
i←i+1
end while
return νδ . An estimator of α∗

Another alternative is by using (15), i.e., considering pairs of neighboring

pixels r, s. In this case, the error is defined as follows
def
ε̂r (α) = â(r)α − b̂(r), (21)
where
def
â(r) = 2∂x C(r), (22)
def
b̂(r) = ∂x A(r) + ∂x B(r). (23)
If we use the standard Least Square formulation, i.e. ρ(x) = x2 , then
P
∗ r∈L â(r)b̂(r)
α = P 2
. (24)
r∈L â(r)

8
We can apply the previous formula by selecting random pixels in the image,
similar to Algorithm 1, or alternatively we can just scan the whole image.

3. Numerical experiment and simulation

First we present numerical comparisons using state of the art algorithms
and our approach. Then we present an example based on simulated fringe
patterns with spatially background intensity variation and modulation factor
(fringe patterns without normalization). Finally, we present an experiment
with real data.
3.1. Numerical comparisson

(a) φ1 (b) φ2 (c) φ3 (d) φ4

Figure 2: Test fringe patterns used in the experiments. These are obtained from (25) with
η = 0 with test phases φi (·, ·) defined in Eqs. (27)-(30).

In order to evaluate the performance of the proposed algorithm, we sim-

ulate fringe patterns using the following models
I(x, y) = cos(φi (x, y)) + η, (25)
J(x, y) = cos(φi (x, y) + δ) + η, (26)
where δ is the phase shift and η ∼ N (0, σ) is Gaussian noise with standard
deviation σ. The fringe patterns are obtained using the following test phases:

φ1 (x, y) = π(x2c + yc2 ), (27)

φ2 (x, y) = π(x2c + yc3 ), (28)
1
φ3 (x, y) = φ4 (x, y) + 5x + y, (29)
8
X9
φ4 (x, y) = ai g(xc , yc ; µi , γi ), (30)
i=1

9
 0.14 0.14
RTS RTS
0.12 0.12
DDV DDV
0.10 0.10
EMN EMN
0.08 0.08
Error

Error
TGS TGS
0.06 0.06
EVI EVI
0.04 COR 0.04 COR
0.02 0.02
0.00 0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
δ δ

(a) φ1 (b) φ2

0.14 0.14
RTS RTS
0.12 0.12
DDV DDV
0.10 0.10
EMN EMN
0.08 0.08
Error

Error

TGS TGS
0.06 0.06
EVI EVI
0.04 COR 0.04 COR
0.02 0.02
0.00 0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
δ δ

(c) φ3 (d) φ4

Figure 3: Errors between the estimated α̂ and the cosine of the true phase shift δ for
a discretization of δ ∈ (0, π), using the models (27)-(30) without noise, see also Fig. 2.
The results for the EVI algorithm is presented for δ ∈ (0, π/2) and the RTS algorithm for
δ ∈ (π/2, π), both algorithms obtain close-to-zero errors.

10
 0.14 0.14
RTS RTS
0.12 0.12
DDV DDV
0.10 EMN 0.10 EMN
Error

Error
0.08 TGS 0.08 TGS
0.06 EVI 0.06 EVI
0.04 COR 0.04 COR
0.02 0.02
0.00 0.00
0.00 0.02 0.04 0.06 0.08 0.10 0.00 0.02 0.04 0.06 0.08 0.10
Noise Level Noise Level

(a) φ1 (b) φ2

0.14 0.14
RTS RTS
0.12 0.12
DDV DDV
0.10 EMN 0.10 EMN
Error

Error

0.08 TGS 0.08 TGS

0.06 EVI 0.06 EVI
0.04 COR 0.04 COR
0.02 0.02
0.00 0.00
0.00 0.02 0.04 0.06 0.08 0.10 0.00 0.02 0.04 0.06 0.08 0.10
Noise Level Noise Level

(c) φ3 (d) φ4

Figure 4: Average of the errors between the estimated α̂ and the cosine of the true phase
shift δ. The average of the errors are computed over 100 runs of the algorithms using the
models (25)-(26) with different level of noise, see Fig. 2.

11
with [xc , yc ] = 6[ Nxx − 12 , Nyy − 21 ], and g(·; ·) is a Gaussian function

(x − µx )2 + (y − µy )2
 
g(x, y; µ, γ) = exp − , (31)
2γ 2

where µ = [µx , µy ]T .

In particular, for the phase in (30) we use the following parameters:

a = [ -20, 12, -24, 16, -20, 20, -4, 32, -8], µ = [ (0,-1), (-0.5,1), (2,-
2), (1.3,0.5), (-1.8, 0.2), (-2,1.5), (1.8,1.3), (-1.2,-1.4), (1.6,0)] and γ = [1,
0.6,0.6,0.6,0.8,0.6,1,0.6,0.6]. Fig. 2 depicts the test fringe patterns in (25)
without noise, i.e., with η = 0.
In the experiments we compare our proposal, Algorithm 1, with different
algorithms for phase shifts extraction with two interferograms: the algorithm
based on the orthogonality of diamond diagonal vectors (DDV) [2], the al-
gorithm based on Euclidean matrix norm (EMN) [10], the algorithm based
on the Gram-Schmidt orthonormalization of two vectors (TGS) [9], the algo-
rithm based on the extreme value of interference (EVI) [11] and the algorithm
based on correlation between fringe patterns (COR) [7, 8]. The results ob-
tained by using the formula (24) are practically equal to the one obtained by
(20), for this reason we do not present such results. In order to assess the
quality of the algorithms, we report the error obtained for each algorithms.
The error is computed as the absolute value of the difference between the
estimated α̂ and the cosine of the true phase shift δ, i.e., error = |α̂ − cos(δ)|.
In the first experiment we use interferograms obtained from models (25)-
(26) with σ = 0, using the test phases φi (·, ·) defined in (27)-(30). In this
case we change the phase shift parameter with the following discretization,
δn = Nn π; n = 1, 2, · · · , N − 1; and N = 100. Fig. 3 shows the results for
each method. Note that only two algorithms obtain close-to-zero errors, these
are the EVI and RTS algorithms. For the previous reason, Fig. 3 presents
the results of the EVI algorithm for δ ∈ (0, π/2) and the results of RTS
algorithm for δ ∈ (π/2, π). The maximum errors obtained by the previous
algorithms for each interferogram are [0.2108, 0.0000, 0.0480, 0.0161] ∗ 10−3
and [0.7911, 0.3941, 0.0200, 0.0441] ∗ 10−12 respectively. Then our proposal,
the RTS algorithm, obtains a better precision than the EVI algorithm. Al-
though, we do not report in Fig. 3 the errors obtained by (8), in this case
the results are also close-to-zero. It is clear that the EVI algorithm, together
with our proposal, presents excellent results. However, the main limitation of

12
the EVI algorithm is the high sensitivity to noise that yields a miscalculation
of the phase step. This is illustrated in the next experiment.
In the second experiment we evaluate the noise-sensitivity of the algo-
rithms. Now, and using the same models (25)-(26) with test phases φi (·, ·)
defined in (27)-(30), we change the noise standard deviation with the follow-
n
ing discretization σn = 200 , where n = 0, 1, · · · , 20. Then, for each n, we
compute the average of the errors over 100 runs of the algorithms. Fig. 4
depicts the results obtained for each algorithm. Similar to the previous ex-
periment, our proposal presents the best results. Note also that the EVI
algorithm is very sensitive to noise, this is justified by the fact that the lo-
cation of maximums and minimums of the interferograms change due to the
noise.

Figure 5: First and second columns correspond to fringe patterns without normalization.
The remainder columns depict the corresponding normalized fringe patterns using the
Gabor Filter Bank approach and the quality map.

3.2. Fringe patterns without normalization

Here we present a numerical experiment using simulated fringe patterns
with background and modulation terms, see first and second columns in
Fig. 5. In order to apply our algorithm, we first need to apply a normalization
algorithm. In this case we use a Gabor Filter Bank approach proposed in
[15]. The result is depicted in third and fourth columns of Fig. 5. The last
column represents a quality map where pixels in white are sites with high
confidence of being well normalized, and pixels in black have low confidence.
Our algorithm only takes into account pixels with high confidence of being
well normalized. The true phase shift is δ = 2.8274 and the estimation
obtained by our algorithm is δ̂ = 2.8266. As we can see the result is very good,
although this result depends, of course, on the quality of the normalization
process.

13
Figure 6: The experimental setup consists in a typical fringe projection configuration.
First and second columns correspond to fringe-pattern projection with π2 phase shift. The
third column corresponds to windows extracted from the previous interferograms. Last
column corresponds to the normalized fringe patterns and its quality map obtained by
using a Gabor Filter Bank.

3.3. Experiment with real data

The experimental setup consists in a typical fringe projection configu-
ration. As in the case of the triangular configuration, the camera and the
specimen under test are arranged face to face. The fringe patterns projected
by the DLP (digital light processing) over a 3D specimen were generated by
using a routine coded in the computer. In particular, the computer video
output is projected onto the specimen surface by means of the DLP by an
angle of 14 degrees with respect to the surface normal of the specimen. The
applied phase shift to the fringe patterns is π2 = 1.57. The distance between
the camera and the specimen is 1.5 m. The specimen under test correspond
to a central dash panel, that cover an area of 60 × 25 cm. The surface con-
tains tiny details (buttons) along with big brightness areas, see Fig. 6. For
applying our algorithm, we first we select a window from from each image,
see third column in Fig. 6. Then we apply a normalization procedure based
on Gabor Filter Bank, see fourth column in Fig. 6. Finally, we apply the
proposed algorithm for estimating the phase shift and algorithm obtained
δ̂ = 1.53, which is very close to the applied phase shift.

14
4. Conclusions
First, we propose two closed formulas for computing the phase shift be-
tween two interferograms without noise. These formulas only require the
information in two pixels of the interferograms and theoretically obtain the
exact phase step. The main limitation of the previous formulas is the sen-
sitivity in interferograms with noise. Therefore, and based on the previous
formulas, we propose an algorithm for interferograms corrupted by noise.
This algorithm yields, in general, an approximation of the phase step and
experimentally obtains the best results compared with state of the art algo-
rithms, even in interferograms corrupted by noise. In the experiment we use
the standard Least Square formulation, what also yields a closed formula,
although the general formulation admits robust potentials. In particular, we
provide two possible implementations of our approach, one in which the sites
can be randomly selected and the other in which we can scan the image.

Acknowledgments
This work was supported in part by CONACYT (Mexico), grant 258033.

References
[1] T. Kreis, W. Jueptner, Fourier transform evaluation of interference pat-
terns: demodulation and sign ambiguity, in: Proceedings of the SPIE
1553, Laser Interferometry IV: Computer-Aided Interferometry, Vol.
263, 1992, pp. 263–273.

[2] C. Luo, L. Zhong, P. Sun, H. Wang, J. Tian, X. Lu, Two-step demodula-

tion algorithm based on the orthogonality of diamond diagonal vectors,
Appl. Phys. B 270 (2) (2007) 161–168.

[3] J. Vargas, J. A. Quiroga, T. Belenguer, M. Servı́n, J. C. Estrada, Two-

step self-tuning phase-shifting interferometry, Opt. Express 19 (2) (2011)
638–648. doi:10.1364/OE.19.000638.

[4] J. Vargas, J. A. Quiroga, C. O. S. Sorzano, J. C. Estrada, J. M. Carazo,

Two-step demodulation based on the gram schmidt orthonormalization
method, Opt. Lett. 37 (3) (2012) 443–445. doi:10.1364/OL.37.000443.

15
 [5] M. Trusiak, K. Patorski, Two-shot fringe pattern phase-amplitude
demodulation using gram-schmidt orthonormalization with hilbert-
huang pre-filtering, Opt. Express 23 (4) (2015) 4672–4690.
doi:10.1364/OE.23.004672.
[6] P. Gao, B. Yao, N. Lindlein, K. Mantel, I. Harder, E. Geist, Phase-
shift extraction for generalized phase-shifting interferometry, Opt. Lett.
34 (22) (2009) 3553–3555.
[7] H. V. Brug, Phase-step calibration for phase-stepped interferometry.,
Appl. Opt. 38 (1999) 3549–3555.
[8] L. Muravsky, O. Ostash, A. Kmet, T. Voronyak, I. Andreiko, Two-
frame phase-shifting interferometry for retrieval of smooth surface and
its displacements, Optics and Lasers in Engineering 49 (2011) 305–312.
[9] W. Niu, L. Zhong, P. S. ans Xiaoxu Lu, Phase shifts extraction algorithm
based on gram-schmidt orthonormalization of two vectors, Opt. Quant
Electron 47 (2015) 2803–2810.
[10] J. Deng, H. Wang, D. Zhang, L. Zhong, J. Fan, X. Lu, Phase shift
extraction algorithm based on euclidean matrix norm, Opt. Lett. 38 (9)
(2013) 4669–4671.
[11] J. Deng, H. Wang, F. Zhang, D. Zhang, L. Zhong, X. Lu, Two-step phase
demodulation algorithm based on the extreme value of interference, Opt.
Lett. 37 (22) (2012) 4669–4671.
[12] J. A. Quiroga, M. Servin, Isotropic n-dimensional fringe pattern nor-
malization, Optics Communications 224 (4-6) (2003) 221 – 227.
[13] J. A. Guerrero, J. L. Marroquin, M. Rivera, J. A. Quiroga, Adaptive
monogenic filtering and normalization of espi fringe patterns, Opt. Lett.
30 (22) (2005) 3018–3020.
[14] N. A. Ochoa, A. Silva-Moreno, Normalization and noise-reduction algo-
rithm for fringe patterns, Optics Communications 270 (2) (2007) 161 –
168.
[15] M. Rivera, O. Dalmau, A. Gonzalez, F. Hernandez-Lopez, Two-step
fringe pattern analysis with a gabor filter bank, Submitted to Optics
and Lasers in Engineering.

16
 [16] M. Black, A. Rangarajan, On the unification of line processes, outlier
rejection, and robust statistics with applications in early vision, Int’l J.
Computer Vision 19 (1) (1996) 57–92.

17

View publication stats