Received July 31, 2019, accepted August 17, 2019, date of publication August 20, 2019, date of current

version September 4, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2936528

An Improved Fine-Resolution Method for

Frequency Estimation of Real-Valued
Single-Tone Using Three DFT Samples
SHUAI YAO , QISONG WU, AND SHILIANG FANG
Key Laboratory of Underwater Acoustic Signal Processing of Ministry of Education, Southeast University, Nanjing 210096, China
Corresponding author: Shuai Yao (yaoshuaiseu@seu.edu.cn)
This work was supported in part by the National Natural Science Foundation of China under Grant 11704069, Grant 61701109, and Grant
11674057, in part by the Fundamental Research Funds for the Central Universities under Grant 2242019K30021, and in part by the
Science and Technology on Sonar Laboratory under Grant 6142109180202.

ABSTRACT With respect to the frequency estimation of a real-valued single-tone, a direct frequency
estimation method is proposed to acquire fine-resolution parameter by only utilizing three discrete Fourier
transform (DFT) samples around the peak. Inspired by the Quinn, Rife and Candan methods, a novel Candan
combined Quinn Rife estimator (CCQRE) is proposed, and the theoretical analysis of the corresponding MSE
is also given. It has been verified that the proposed method is more robust to the interbin frequency and has
better performance than these existed state-of-the-art methods. Both simulations and real data experiments
demonstrate the effectiveness of the proposed method. In addition, the proposed methods are suitable for
real-time applications due to the simplicity, efficiency and low computational burden.

INDEX TERMS Interpolated DFT, frequency estimation, single-tone, mean square error, interbin frequency.

I. INTRODUCTION fine-resolution frequency parameter. In general, the largest

Frequency estimation of a real-valued single-tone represents
a fundamental problem of signal processing, and can be
used in many applications including spectrum estimation,
communications, speech, radar and sonar signal processing,
etc. [1], [2]. The parameter estimation of a single-tone prob-
lem was initially formulated by Slepian [3], and his works
focused on the continuous-time observation model [4]. The
maximum likelihood (ML) estimators were examined and
developed by Rife and Boorstyn in the discrete-time case [5].
It has been demonstrated that the peak location of the peri-
odogram of a single-tone has the maximum likelihood (ML)
in an additive white Gaussian noise condition. Due to the
intrinsic discrete property of the DFT based the periodogram,
its maximum is often deviated from the true location in the
noisy condition.
The inherent defects of the picket fence and the leakage
in DFT will lead to the deviations of parameter estimation.
Two-step procedures, including coarse estimation and refined
estimation, are often taken into consideration to acquire
The associate editor coordinating the review of this article and approving
it for publication was Qingchao Jiang.
magnitude in the periodogram is selected in the coarse fre-
quency interval case in the first step, and then the refined
search procedure is subsequently followed with the finer
interval in the region of interest in the second step. A num-
ber of methods had been provided for the refined estima-
tion, and they are generally regarded as Interpolated Fast
Fourier Transform (IpFFT). It has been proved that the MSE
is generally plagued by the threshold due to a non-linear
operation in the IpFFT, and the MSE often degrades rapidly
with the decreases of SNR [5]. In general, when the SNR
is high, the MSE is only caused by the deviation of true
location, which is less than 1/2 bin, due to the discretiza-
tion in the DFT [7], [8]. In this paper, we will focus on
the fine-resolution frequency estimation problem in the
high SNR case.
In recent decades, various methods have been proposed
for fine-search implementation by interpolating two or more
DFT complex samples or modules. These methods can be
divided into two categories: direct approaches and iterative
approaches [9]. Direct methods apply a simple formula on a
few DFT outputs obtained in the first stage and generate the
final frequency estimate with an almost negligible additional

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 117063
 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone
computational cost. The direct method was originally studied
by Rife and Vincent [10]. The peak value and the adjacent
submaximal component in the main lobe are used to estimate
the interbin frequency. When the interbin frequency is far
away from 0, the MSE of the Rife estimator approaches
the Cramer–Rao Lower Bound (CRLB) as SNR increases.
However, there is a high probability that the incorrect sign
problem will occur when the interbin frequency is close to 0
[11]. Instead of exploiting the amplitudes of the maximum
component and the adjacent submaximal components in the
main lobe, Quinn in [12] and [13] proposes a number of
classical estimators that interpolates the true signal frequency
employing the phase information of three DFT coefficients.
The performance of Quinn estimator does not degrade when
the interbin frequency is close to 0. However, these algo-
rithms also have a frequency-dependent performance that is
worst for a signal frequency coinciding with a bin center
[14]. In [15], a new three-point interpolated Fourier algo-
rithm relying on complex DFT samples estimator has been
proposed by Jacobsen. The Jacobsen estimator is based on
empirical observations and presented without a proof. In [8],
specifically designed for signal records with a small number
of samples, Candan presents a derivation for the Jacobsen for-
mula and present a bias correction. In [7], the bias and MSE
analysis of the frequency estimator are given and an improved
version of the estimator, with the removal of estimator bias,
were suggested in [8]. It approaches the Jacobsen estimator
when the number of analyzed samples is quite high [6]. Also,
in [16], Orguner and Candan use more DFT samples to reduce
the gap between the performance of the estimator proposed
in [7] and the related unbiased CRLB.
Different from the direct methods, the iterative methods
carry out iterations treating the refined estimate of an earlier
iteration as the coarse estimate of the next until conver-
gence [16]. The effect of wideband noise on the estimates
returned by two or three interpolation points algorithms can
be minimized when using the iterative methods. In [14],
two iterative Fourier interpolation algorithms was proposed
by Aboutanios and Mulgrew (A&M). One of the A&M
algorithms uses two complex DFT coefficients located in the
middle of neighbor FFT coefficients, and another works on
the moduli of the DFT coefficients. In [6], the A&M iterative
interpolation algorithms are generalized by introducing an
additional parameter to allow for selection of the Fourier
interpolation coefficients relative to the true frequency. The
two generalized algorithms are both capable of employing
more information from the FFT results and consequently
improve the performance of the original algorithms in either
accuracy or efficiency, depending on the choice of the
additional parameter. These iterative algorithms can achieve
identical asymptotic performance to approach the CRLB.
However, the required number of iterations depends on the
resolution as well as the operating SNR and can be quite large,
so these iterative methods are time consuming with a high
complexity [17]. The frequency of a real-valued single-tone
117064
can also be estimated by extracting the instantaneous fre-
quency on the basis of the signal maximum distributed energy
calculated by time-frequency distribution (TFD) [18].
In this paper, a direct frequency estimation method is
proposed to acquire fine-resolution frequency parameter by
leveraging three DFT samples around the peak. According
to analyzing the characteristics of Rife, Quinn and Candan
estimators, we propose a novel Candan combined Quinn-Rife
estimator (CCQRE) by jointly inheriting the advantages
of these three methods. An improved Quinn-Rife estima-
tor (CQRE) is firstly given and the corresponding MSE is
theoretically analyzed. It is shown that the CQRE has the
lower MSE than these in original Quinn and Rife estimators.
It is also found that the CQRE method has smaller MSE
with a small interbin frequency, whereas higher MSE with
a large one, compared to the Candan estimator. To inherit
the advantages of CQRE and Candan estimator, a novel and
improved CCQRE is proposed to acquire fine-resolution fre-
quency parameter, and the theoretical analysis of the MSE
in the CCQRE is also given. It has been demonstrated that
the CCQRE is more robust against interbin frequency, and
has better estimation performance, compared to the original
Quinn, Rife and Candan estimators. The proposed CCQRE
only utilizes three DFT samples for the frequency estimation,
and thus has comparable computation to three estimators.
The rest of the paper is organized as follows. The real-
valued single-tone model and the problem statement are
introduced in Section II. The proposed method is described
and the theoretic analysis of MSEs is given in Section III.
In Section IV, simulations and real data experiments are car-
ried out to analyze the performance of the proposed method.
Conclusions are drawn in Section V.
II. PROBLEM DESCRIPTION
The discrete-time observed signal model for real-valued
single-tone is given as
x (n) = s (n) + w (n) , n = 0, 1, · · · , N − 1 (1)
 
n
s (n) = A0 cos 2π f0 + φ0 , n = 0, 1, · · · , N − 1
fs
(2)
where x(n) is the single-tone, A0 > 0, 0 < f0 < fs /2, and
φ0 ∈ [0, 2π ) are an unknown constants, which represent
the amplitude, frequency, and initial phase of the signal,
respectively. The noise term w(n) is assumed to be a zero-
mean, additive white Gaussian process with variance σ 2 , fs
is the sampling frequency, and N is the number of samples.
The signal-to-noise ratio (SNR) definition used
 in this paper
is the input SNR which is defined as SNR=A20 2σ 2 .
It had been proved that a poor frequency estimate would
lead to the poor performance of amplitude and phase esti-
mates [19], and thus we will focus on the problem of the
fine-resolution frequency estimate. The problem of frequency
VOLUME 7, 2019
S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

estimation f0 can be written as Our goal is also utilizing the three DFT samples X (kp ),

fs X (kp−1 ) and X (kp+1 ) to produce a precisely estimate of δ.
f0 = kpN + δN (3) Once δ is estimated, the final frequency can be estimated by
N
taking estimated δ̂ into Eq. (4).
where kpN indicates the index of the bin that is closest to To acquire the statistical performance of the estimators,
the true frequency, and δN is a fractional correction term in we firstly analyze the statistical distribution of W (k). W (k)
the interval [−0.5, 0.5]. The subscript N indicates the depen- can be expressed as follow,
dence of the various parameters on N [17]. In the rest of the
N −1
paper, unless the dependence on N needs to be emphasized, X  n 
we drop the subscript for the sake of simplicity of notation. W (k) = w (n) exp −j2π k , k = 0, 1, 2, ..., N −1
N
It’s assumed that the coarse search returns the index of the n=0

true maximum bin. Thus our goal is to estimate the fractional = WR (k) − jWI (k) , k = 0, 1, 2, ..., N − 1 (10)
correction term δ. The estimated frequency is then given by where WR (k) and WI (k) are respectively the real and imagi-
nary part of W (k), denoted as,
 f
s
fˆ0 = kp + δ̂ (4) N
X −1  n 
N WR (k) = w (n) cos 2π k (11)
The signal x(n) can be transformed into X (k) by DFT, and N
n=0
it can be expressed as N
X −1  n 
N −1 WI (k) = w (n) sin 2π k (12)
X  n  N
X (k) = x (n) exp −j2π k n=0
N According to Eq. (11), it’s clear that WR (k) can be regarded
n=0
= S (k) + W (k) , k = 0, 1, · · · , N 2 − 1 (5) as the linear combinations of N independent Gauss white

noise sequences. Therefore, WR (k) is also Gaussian dis-
where S(k) and W (k) are the DFTs of the signal s(n) and tributed. The mean value and variance of WR (k) are given by,
noise w(n) respectively. It’s well known that the N -point DFT
N −1
samples of the real sequence is symmetrical on N /2, so only X  n 
E [WR (k)] = E [w (n)] cos 2π k = 0 (13)
N /2 DFT samples from 0 to N /2-1 is considered in this paper. N
n=0
S(k) is expressed as
N −1  n  N
A0 sin π k − f0 1f
X
var [WR (k)] = var [w (n)] cos2 2π k ≈ σ 2
 

S (k) =
  ej[φ0 −π(k−f0 /1f )(1−1/N )] N 2
2 sin π k − f0 1f N
 
n=0

(6) (14)

where 1f = fs N .
 where E [•] denotes the expectation operator. Therefore,
From Eq. (5) and (6), the peak bin and its immediate left WR (k) follows Gaussian distribution with zero mean and
the variance of N σ 2 2, i.e. WR (k) ∼ N 0, N σ 2 2 .


and right neighbor bins can be expressed as follows:



In the following, the correlation of WR (k) and WR (k + l) is
X kp − 1 = S kp − 1 + W kp − 1 considered as,
NA sin (π δ) j[φ (kp )−π ]
E [WR (k) WR (k + l)]


= e + W kp − 1
2π (−1 − δ) (N −1
 n  NX −1 h m
)
NA sin (π δ) jφ (kp )
X i
= e + W kp − 1


(7) =E w(n) cos 2π k w(m) cos 2π (k +l)
2π (1 + δ) N N



n=0 m=0
X kp = S kp + W kp
( "N −1  #
X 2π nk
NA sin (π δ) jφ (kp )
= E Re w(n) exp −j
= e + W kp (8) N
2π δ
n=0
"N −1


# )
X kp + 1 = S kp + 1 + W kp + 1

X 2π m
× Re w(m) exp j (k + l)
NA sin (π δ) j[φ (kp )+π ]
N
= e + W kp + 1 m=0
2π (1 − δ) (N −1 N −1
2π (m − n) k
 
NA sin (π δ) jφ (kp )

= Re
XX
E [w(m)w(n)] exp j
=− e + W kp + 1 (9)
2π (1 − δ) n=0 m=0
N
where φ kp = φ0 + π δ 1 − 1 N . For the direct


  )
2π ml
approaches, the most popular ones, such as Rife and Vin- × exp j (15)
N
cent [10], Quinn [12], [13] and Candan [7], [8] methods,
are to directly interpolate from the three DFT bins where where Re[x] denotes taking the real part of x. Due to the
the peak occurs and its immediate left and right neighbors. Gaussian property of w(n), E [w(m)w(n)] = σ 2 1 (m − n).

VOLUME 7, 2019 117065

 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

1 (m − n) denotes the unit impulse function. Therefore,

Eq. (15) can be rewritten as

E [WR (k) WR (k + l)]

(N −1 N −1
2π (m − n) k
XX  
= Re E [w(m)w(n)] exp j
N
n=0 m=0
 
2π ml
× exp j
N
(N −1 N −1
2π (m − n) k
XX  
= Re σ 1 (m − n) exp j
2
N
n=0 m=0
 
2π ml
× exp j FIGURE 1. The MSE of the rife estimator versus |δ| under different SNRs.
N
(N −1  )
X 2πml
= Re σ exp j
2
= σ 2 1 (l) In [20], it is proven that the Rife estimator is unbiased and
N
m=0
the MSE of the Rife estimator is given as follows,
(16)
  (1 − |δ|)2 (1 − |δ|)2 + |δ|2 
Eq. (16) implies that, for each different discrete frequency MSE δ̂R = + (2δ)2 PE (δ)
N sin c2 (|δ|) SNR
k, WR (k) are independent with each other, and the similar (18)
conclusions can be made for WI (k), and between WI (k)
and WR (k). where PE (δ) denotes the probability of the incorrect interpo-
lation direction. For Rife estimator, it is given as [20],
III. STATISTICAL PERFORMANCE ANALYSIS " #
1 |δ| |sin (π δ)| p
OF PROPOSED METHOD PE (δ) = erfc
R

N (SNR) (19)
In this section, by introducing the statistical performance 2 π 1 − |δ|2
of the state-of-the-art direct estimation methods including
where erfc(x) denotes the complementary error function and
Rife, Quinn and Candan estimator, an improved Quinn-Rife
the superscript R representatives abbreviation of Rife. The
estimator (CQRE) is firstly given and the corresponding MSE
MSE of the Rife estimator versus the fractional correction
is theoretically analyzed. Inspired by the Candan estimators,
term |δ| under different SNRs are presented in Fig.1. In Fig.1,
a novel Candan combined Quinn-Rife estimator (CCQRE)
the CRLB is also considered which sets the lowest vari-
is subsequently proposed by jointly exploiting the merits of
ances of frequency measurement for the optimal method. The
these two methods.
CRLB for the frequency estimation is [5]
A. IMPROVED QUINN-RIFE ESTIMATOR 12fs2
var (f0 ) ≥ (20)
(2π )2 SNRN N 2 − 1


Rife and Quinn estimator are the most commonly used ones
to estimate the frequency of single-tone. Therefore, Rife and
Quinn estimator are firstly considered. By analyzing their From Fig. 1, it has been found that there exists a threshold
statistical performances, the CQRE is given by combing Rife of δ, when |δ| is smaller than the threshold, the MSE of the
and Quinn estimator. The MSE for the CQRE method is also Rife estimator increases rapidly with |δ| increasing; while |δ|
derived. is smaller than the threshold, the MSE of the Rife estimator
decreases with the δ increasing. When |δ| approaches 0.5,
the MSE of the Rife estimator is very close to the CRLB.
1) RIFE ESTIMATOR
This results from the fact that when δ is small, the adjacent
The Rife estimator was proposed in [10], the modulus of
submaximal component in the main lobe is small, the ampli-
the peak value and the adjacent submaximal component in
tude of the first side lobe on the other side of the peak
the main lobe are used to estimate the fractional correction
value may exceeds the adjacent submaximal component in
term δ as follows:
the main lobe. In other words, the probability of the incorrect
δ̂R  interpolation sign of Rife estimator is large for small |δ|.


− X kp − 1


2) QUINN ESTIMATOR




X kp − 1 ≥ X kp + 1
X kp − 1 +
X kp

= It’s known that only the modulus of the peak value and the
X kp + 1 adjacent submaximal component in the main lobe are used to
X kp − 1 < X kp + 1







X kp + 1 + X kp estimate δ by Rife estimator. The natural question to ask is

(17) whether the phases can be used to estimate δ. The answer to

117066 VOLUME 7, 2019

S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

W3 = W kp + 1 = R3 + jI3 , R1 = Re W kp ,



this question is that the phases are used by Quinn estimator
[12]. For the procedure of Quinn estimator: the moduli of the I1 = Im W kp , R2 = Re W kp − 1 ,

 

complex Fourier coefficients are used to select kp , and the
I2 = Re W kp − 1 ,


real parts of two successive ratios of these coefficients are
R3 = Re W kp + 1 , I3 = Im W kp + 1 ,

 

then used to construct the estimator. In [12], the real parts
U1 = R1 cos φ kp + I1 sin φ kp ,



of two successive ratios
 of these coefficients are denoted as
α1 = Re X (kp − 1) X (kp ) , α2 = Re X (kp + 1) X (kp ) .
    
U2 = R2 cos φ kp + I2 sin φ kp .



The estimated fractional correction term δ of Quinn estimator
is given as: Im[x] denotes taking the imaginary part of x.
( Considering the fact that the SNR of FFT coefficients in
δα1 δα1 > 0 and δα2 > 0 the main lobe is generally high, under high SNR, it can be
δ̂Q = (21)
δα2 else found that
.
|W1 |2 |B1 |2  1, R1 R2 B1 B2  1, I1 I2 B1 B2  1 (25)
 
where δα1 = α1 (1 − α1 ), δα2 = −α2 (1 − α2 ).
 

It’s clear that when |δ| is small and there exists noise, for Substitution Eq. (25) into (24), α1 is approximately given as
|S(kp − 1)| is close to |S(kp + 1)|, the probability of the  
incorrect interpolation direction of Rife estimator is high; B2 1 + U1 B1 − U2 B2
α1 ≈ −  (26)
while the phase difference between S(kp − 1) and S(kp +1) B1 1 + 2U1 B1
is π , the probability of the incorrect interpolation direction of
Expanding α1 by the Taylor series yields:
Quinn estimator is small [21].   
For comparative analysis, the probability of the incorrect B2 U1 U2 2U1
Q α1 ≈ − 1+ − 1−
interpolation sign of Quinn estimator PE (δ) will be derived. B1 B1 B2 B1
 
From Eq. (21), we can derive B2 U1 U2
≈− 1− − (27)
Q
PE (δ) = 1 − P δα1 > 0, δα2 > 0|δ > 0

B1 B1 B2
Similarly, we can derive
−P δα1 < 0, δα2 < 0|δ < 0


S (k0 + 1) + W3 −B3 ejφ0 + W3
   
= 1−P α1 (1−α1 ) > 0, −α2 (1 − α2 ) > 0|δ > 0
 
α2 = Re = Re
S (k ) + W1 B1 ejφ0 + W1
 0 
−P α1 (1 − α1 ) < 0, −α2 (1−α2 ) < 0|δ < 0 B3 U1 U3
 

≈− 1− − (28)
B1 B1 B3
(22)
where U3 = R3 cos φ kp + I3 sin φ kp . For 0 < δ < 0.5,



h i h i
X (kp −1) X (kp +1)
For α1 = Re < 1 and α2 = Re
X (kp ) X (kp ) < 1, we can acquire
Eq. (22) can be rewritten as B2 NA sin (π δ) NA sin (π δ)

δ
Q =− =− <0
PE (δ) = 1 − P (α1 > 0, α2 < 0|δ > 0) B1 2π (1 + δ) 2π δ 1+δ
NA sin (π δ) NA sin (π δ) δ

B3
− P (α1 < 0, α2 > 0|δ < 0) (23) = = >0 (29)
B1 2π (1 − δ) 2π δ 1−δ
Now, we analyze α1 and α2 , Substitution Eq. (29) into (27) and (28), for 0 < δ < 0.5,
−B2 ejφ (kp ) + W2 we can get that α1 > 0, α2 < 0 is equivalent to
"
# " #
S kp − 1 + W2
α1 = Re = Re
B1 ejφ (kp ) + W1 U1 U2 U1 U3


S kp + W1 1− − > 0, 1 − − >0 (30)
( ) B1 B2 B1 B3
B2 cos φ kp − R2 + j B2 sin φ kp − I2



= −Re Similarly, for −0.5 < δ < 0, we can get that α1 < 0, α2 > 0
B1 cos φ kp + R1 + j B1 sin φ kp + I1



is also equivalent to,
B1 B2 + B2 U1 − B1 U2 − R1 R2 − I1 I2 U1 U2 U1 U3
=− 1− − > 0, 1− − >0 (31)
B21
+ |W1 + 2B1 U1 |2 B1 B2 B1 B3
    Therefore,
B2 1+U1 B1 −U2 B2 − R1 R2 B1 B2 − I1 I2 B1 B2
=− Q
B1 1 + |W1 |2 |B1 |2 + 2U1 B1
  PE (δ)
= 1 − P (α1 > 0, α2 < 0|δ > 0)
(24)
−P (α1 < 0, α2 > 0|δ < 0)
where
 
U1 U2 U1 U3
= 1−P 1− − > 0, 1 − − > 0|δ > 0
NA sin (π δ) NA sin (π δ) NA sin (π δ) B1 B2 B1 B3
B1 = , B2 = − , B3 =
2π δ 2π (1 + δ) 2π (1 − δ)
 
U1 U2 U1 U3
−P 1 − − > 0, 1 − − > 0|δ < 0
W1 = W kp = R1 + jI1 , W2 = W kp − 1 = R2 + jI2 ,



B1 B2 B1 B3
VOLUME 7, 2019 117067
 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

 
U1 U2 U1 U3
= 1−P 1− − > 0, 1 − − >0
B1 B2 B1 B3
= 1 − P (1 − ξ1 > 0, 1 − ξ2 > 0)
(32)
U3
where ξ1 = U B1 + B2 , ξ2 = B1 + B3 . According to the anal-
1 U2 U1

ysis in Eq. (13) and Eq. (14), R1 , I1 , R2 , I2 , R3 and I3 , follow

Gaussian distribution with zero mean and the variance of
N σ 2 2, and they are independent


with each other. There-
fore, we have U1 ∼
N 0, N σ 2 2 , U2 ∼ N 0, N σ 2 2 ,


U3 ∼ N 0, N σ 2 2 . U1 , U2 and U3 are also independent

with each other. Therefore, we can obtain
s 1
 ξ1 ∼ N (0, 1) Q
FIGURE 2. PE (δ) and PER (δ) versus δ under different SNRs.
σ 1
+ 1 N
B21 B22 2
(33)
s 1
 ξ2 ∼ N (0, 1) TABLE 1. The procedure of CQRE.

σ 1
+ 1 N
B21 B23 2

sin(πδ) sin(πδ)
Substitution Eq. (33) and B1 = NA 2πδ , B2 = − NA
2π(1+δ) ,
NA sin(πδ)
B3 = 2π(1−δ) into (32), we have
Q
PE (δ) = 1 − P (1 − ξ1 > 0, 1 − ξ2 > 0)
= 1 − P (ξ1 < 1) P (ξ2 < 1)
 
 
1 1
ξ1 <
 
= 1−P 
 s  s 
 
σ 1 1 N
σ 1
+ 12 N

+
B21 B22 2 B21 B2 2
 
 
1 1
ξ2 <
 
•P 
 s  s 


σ 1 1 N
σ 1 1 N
 
+ +
B21 B23 2 B21 B23 2
" √ !#
Based on the above analysis, an improved method by
1 |sin (π δ)| NSNR
= 1 − 1 − erfc √ p combing the Quinn and Rife estimator (CQRE) is consid-
2 2π 2δ 2 + 2 |δ| + 1
√ ered. We denote the estimated fractional correction term
of CQRE as δ̂CQRE . The procedure of CQRE is as shown
" !#
1 |sin (π δ)| NSNR
• 1 − erfc √ p (34) in Table 1.
2 2π 2δ 2 − 2 |δ| + 1
From Table 1, it is found that both the phases and modulus
Q
Both PRE (δ) and PE (δ) versus δ under different SNRs are are used to determine the interpolation direction while only
Q
presented in Fig.2. From Fig. 2, it’s easy to see that PE (δ) is the modulus is used to construct the fractional correction term
obviously smaller than PRE (δ) for small |δ|. This results from δ as the Rife estimator. The CQRE integrates the merits of the
the fact that both the phase and modulus information are used Quinn and Rife estimator. From the procedure of CQRE as
to determine the direction of interpolation by Quinn estimator shown in Table 1, the probability of the incorrect interpolation
while only the modulus information is used by Rife estimator. sign of CQRE can be given as
CQRE Q
PE (δ) = PE (δ) PRE (δ) (35)
3) COMBINED QUINN-RIFE ESTIMATOR (CQRE)
By comparing Eq. (17) and Eq. (21), it’s clear that only the Substitution Eq. (35) into Eq. (18), we acquire the MSE of
real parts of two successive ratios of these coefficients are CQRE:
used to construct the Quinn estimator while the modulus are  
used by the Rife estimator. The noise resistance capacity of MSE δ̂CQRE
Quinn estimator is worse than the Rife estimator due to the
(1−|δ|)2 (1 − |δ|)2 + |δ|2
 
neglect of imaginary part when the interpolation direction is =
CQRE
+ (2δ)2 PE (δ) (36)
correct. N sin c2 (|δ|) SNR

117068 VOLUME 7, 2019

S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

where ωα =
W kp − 1
− W kp + 1 and ωβ = 2W kp −




B. PROPOSED COMBINED CANDAN-QUINN-RIFE
ESTIMATOR W kp − 1 − W kp + 1 .
As described in section III (A), although the frequency- Using the independence of W (k), it can be easily veri-
domain peak sample and two adjacent samples are used to fied that ωα and ωβ are independent zero-mean Gaussian
determine the interpolation direction, only the peak sample distributed random variables with the variances 2N σ 2 and
and the adjacent submaximal sample are used to construct 6N σ 2 . Therefore, the estimated δ̂C can be expressed as fol-
δ by the CQRE. The natural question to ask at this stage lows:
is whether all the three samples can be used to construct δ. α
  
B + ωα

Jacobsen and Kootsookos [15] proposed a new three-point δ̂C ≈ Re ≈ Re
β C + ωβ
interpolated Fourier estimator relying on all the three com-  !
B + ωα B C + ωα C
  
plex DFT samples. In [8], Candan presented a derivation for
≈ Re ≈ Re (41)
the Jacobsen formula and present a bias correction. For the C + ωβ 1 + ωβ C

complex single-tone records with a small number of samples,
the bias and MSE analysis of the Candan frequency estimator Expand Eq. (41) by the Taylor series:
is given in [7].  !
B C + ωα C

In this section, for the real-valued single-tone records with
δ̂C ≈ Re
1 + ωβ C

a large number of samples, the statistical performance of the
Candan estimator proposed in [8] is introduced, and then
≈ Re δ + ωα C 1 − ωβ C + h.o.t
 


(42)
by comparing the statistical performance of the proposed
CQRE and Candan estimator, a novel combined estimator is Here the term h.o.t. denotes the higher order terms of ωβ C.

proposed and its MSE is also derived. For large SNR, the Eq. (42) can be approximated with
δωβ ωα δωβ,r ωα,r
 
1) CANDAN ESTIMATOR
δ̂C ≈ Re δ − + =δ− + (43)
The Candan estimator proposed in [8] is given as C C C C
tan π N X (kP − 1) − X (kP + 1)


where ωα,r = Re (ωα ) and ωβ,r = Re ωβ . Using the inde-


δ̂C = Re
π N 2X (kP ) − X (kP − 1) − X (kP + 1) pendence of ωα and ωβ , it can be easily verified that ωα,r and


(37) ωβ,r are independent zero-mean Gaussian distributed random

variables with the variances N σ 2 and 3N σ 2 . Therefore, it’s
In the absence of noise, the estimator given by Eq. (37) can easy to get that
be expressed as  
  E δ̂C = δ (44)
B
δ̂C = cN Re (38)   3N σ 2 2 N σ 2
C var δ̂C − δ = δ + 2
C2 C
where cN = tan(π /N ) , B = S kp − 1
− S kp + 1
, C =
2

π /N





2
 π 2δ2 1 − δ2
2S kp −S kp
− 1 −S kp + 1 . By taking S kp , S kp − 1 = 3δ + 1
2N sin2 (π δ) A2 2σ 2


and S kp + 1 into B and C, they can be written as follows:
  π 2 δ 2 1 − δ 2
2
NA sin (π δ) jφ (kp ) NA sin (π δ) jφ (kp ) 2
= 3δ + 1 (45)
B= e + e 2N sin2 (π δ) SNR
2π (1 + δ) 2π (1 − δ)
2NA sin (π δ) jφ (kp ) From Eq. (44), it can be seen that the Candan estimator is
=
e (39)
2π 1 − δ 2 unbiased. Therefore, the MSE of the Candan estimator can
NA sin (π δ) jφ (kp ) NA sin (π δ) jφ (kp ) be written as
C =2 e − e
2π δ 2π (1 + δ)
   
MSE δ̂C = var δ̂C − δ
NA sin (π δ) jφ (kp ) 2NA sin (π δ) jφ (kp )
+ e =
e (40)  π 2 δ 2 1 − δ 2
2
2π (1 − δ) 2π δ 1 − δ 2 
2
= 3δ + 1 (46)
2N sin2 (π δ) SNR

(39) and (40), it’s easy to know that B C = δ.

From Eq.
Re B C is actually the output of the Jacobsen estimator C. COMBINED CANDAN-QUINN-RIFE
[14] in the absence of noise. We denote Re B C as δ̂J ,


ESTIMATOR (CCQRE)
it’s easy to get that δ̂C = cN δ̂J . For N  π , cN ≈ 1,
 
From Eq. (36) and Eq. (46), it’s clear that both MSE δ̂CQRE
we can get that δ̂C ≈ δ̂J . It means that the Candan estimator  
is equivalent to Jacobsen estimator for N  π . We denote α and MSE δ̂C are non-linear functions of δ and SNR, it’s dif-
and β respectively
as α = X (kP
− 1) − X (kP +
1) = B + ωα ficult to compare the proposed CQRE and Candan estimator
and β = 2X kp − X kp − 1 − X kp + 1 = C + ωβ , directly. Therefore, the numerical analysis is used to compare

VOLUME 7, 2019 117069

 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

It has been proven that the Candan and the CQRE is

unbiased, so from Eq. (48), it’s clear that the CCQRE is
also unbiased. Therefore, the MSE of the CCQRE can be
written as
 
MSE δ̂CCQRE
     h  i


P δ̂ R ≤ η MSE δ̂ C + 1 − P δ̂ R ≤ η
  
 ×MSE δ̂CQRE |δ| ≤ η


=     h  i


P δ̂ R > η MSE δ̂ CQRE + 1 − P δ̂ R > η
  
 ×MSE δ̂C |δ| > η



FIGURE 3. RMSE versus δ under different SNRs. (51)

For |δ| ≤ η,
 
the statistical performance of the two methods. We denote P δ̂R ≤ η = P (|δ| + χ ≤ η)
 
MSE δ̂CQRE χ η − |δ|
 
= P (χ ≤ η − |δ|) = P ≤
RMSE =   (47) σχ σχ
MSE δ̂C 1

1 η − |δ|

= 1 − erfc √ (52)
RMSE versus δ under different SNRs are presented in Fig. 3. 2 2 σχ
From Fig. 3, it’s easy to see that RMSE decreases with the For |δ| > η,
increase of |δ|. When |δ| is smaller than 0.15, RMSE is larger
1 η − |δ|
    1  
than 1, while |δ| is larger than 0.15, RMSE is smaller than 1. P δ̂R > η = 1 − P δ̂R ≤ η = erfc √
It means that the MSE of the CQRE is larger than the Candan 2 2 σχ
estimator for |δ| is smaller than 0.15, while the MSE of (53)
the CQRE is smaller than the Candan estimator for |δ| is
Substitution Eq. (36), Eq. (46), Eq. (52) and Eq. (53) into
larger than 0.15. That is to say, there exists a threshold of
Eq. (51) yields the MSE of the CCQRE.
δ between the CQRE and the Candan estimator. We denote
the threshold as η. For |δ| is smaller than η, the Candan
IV. SIMULATIONS AND EXPERIMENTS
estimator outperforms the CQRE while it performs worse
In this section, the performance comparison of the proposed
than the CQRE for |δ| is larger than η.
method is carried out with the existed state-of-the-art methods
Based on the above analysis, the proposed method named
including Rife, Quinn and Candan estimator in both simu-
as combined Candan-Quinn-Rife estimator (CCQRE) is
lated and real data.
given,
A. SIMULATIONS

δ̂C δ̂R ≤ η
δ̂CCQRE = (48) The signal is of length N = 1024 samples, a sampling
δ̂ δ̂R > η frequency fs = 4000Hz and amplitude A0 = 1. The index of
CQRE
the bin kp corresponding the peak frequency is set to be 79.
Now we analyze the statistical performance of the pro- The snr is defined in decibels as snr = 10 log10 [SNR]. The
posed CCQRE. In [21], it is proven that accuracy of the estimated parameters is measured by MSE
A1

A2
 which is defined as
δ̂R = |δ| + λ2 − λ1 = |δ| + χ (49) Nr 
(A1 + A2 )2 A1   1 X 2
MSE δ̂ = δ̂r − δ (54)
where A1 = S kp , A2 = S kp+κ ,
λ1 = WR k
p ,



Nr
r=1
1 S kp+1 > S kp−1
λ2 = WR kp+κ , κ =


−1 else
, where δ̂r denotes an estimated parameter, δ denotes the actual
  value and Nr is the number of Monte Carlo runs. In this paper,
χ = A1
λ − AA12 λ1 . For λ1 ∼ N 0, N σ 2 ,


(A1 +A2 )2
2 unless otherwise stated, 1000 times Monte Carlo simulations
λ2 ∼ N 0, N σ 2 and λ1 and λ2 are independent with each are run for each SNR and δ, and φ0 ∼ uniform [0,2π ).
other, χ follows Gaussian distribution with 0 mean and σχ2 Case 1 (MSE of δ̂ for CCQRE Versus the Threshold η): As
variance. And σχ2 is [21] analyzed in Section III, the MSE of the CCQRE is dependent
on the threshold η. Therefore, we firstly assess the perfor-
(1 − |δ|)2 (1 − |δ|)2 + |δ|2
 
mance of the CCQRE by plotting the MSE of CCQRE versus
σχ2 = (50)
N sin c2 (|δ|) SNR η. η is set to start from 0 to 0.5 in a step of 0.05 for a fixed snr.

117070 VOLUME 7, 2019

S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

 
FIGURE 5. PE δ̂ achieved by simulation and by theoretical calculation
versus δ.

 
FIGURE 4. MSE δ̂ for CCQRE versus the threshold η. (a) the mean of
MSEs for the varying δ (b) the std of MSEs for the varying δ.

For a fixed η and snr, the fractional correction term δ is set to

 
FIGURE 6. MSE δ̂ achieved by simulation and by theoretical calculation
vary from 0.025 to 0.5 in a step of 0.025. Then the mean and versus δ.
standard deviation (std) of the MSEs for the varying δ with
a fixed η and snr are calculated to show the dependences of
MSE regard to the threshold η.
Fig. 4 shows the mean and std of MSEs for the varying
δ versus the threshold η under different snrs. From Fig. 4,
it’s easy to see that both the mean and std of MSEs for the
varying δ are smallest when η is 0.15 under all the set snrs.
It shows a quite good agreement with the results concluded
from Fig. 3 in Section III. Therefore, in all the following
experiments, we have fixed the threshold η to 0.15.
Case 2 (MSE of δ̂ Versus δ): As analyzed in Section III,
the statistical performance of the direct estimation methods
are dependent on the |δ|, so we just analyze for the fractional  
FIGURE 7. MSE δ̂ achieved by simulation versus δ.
correction term δ > 0. For δ < 0, it is symmetrical with
δ < 0. The fractional correction term δ is set to start from
0.025 to 0.5 in a step of 0.025 and the snr is set to be −3 dB QR
for each δ.  δ are presented
and PE (δ) versus  in Fig.
 5. The accuracy
of MSE δ̂CQRE , MSE δ̂C and MSE δ̂CCQRE versus δ
The accuracy of the probability of the incorrect interpo-
Q
lation sign of Quinn estimator PE (δ) given by Eq. (34) and are presented in Fig. 6. The simulation and theoretical results
CQRE are represented by continuous lines and circles, respectively.
the proposed CQRE PE (δ) given by Eq. (35) versus δ
Both Fig. 5 and Fig. 6 show a quite good agreement between
have been verified in simulations.
 Theaccuracy of the MSE simulation and theoretical results versus δ.
of the proposed CQRE MSE δ̂CQRE given by Eq. (36), Fig. 7 shows the MSEs versus δ between the proposed esti-
 
CCQRE MSE δ̂CCQRE given by Eq. (51) and Candan esti- mator and the state-of-the-art direct estimators. From Fig. 7,
  it can be found that the MSE of the CQRE shows a mono-
mator MSE δ̂C given by Eq. (46) have also been verified tone decrease with δ increasing while the MSE of the
Q
by means of computer simulations. The accuracy of PE (δ) Quinn and Candan estimator shows a monotone increase with

VOLUME 7, 2019 117071

 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

δ increasing. The MSE of Rife estimator and CCQRE

does not change monotonically with δ increasing. The MSE
of Rife shows a very large fluctuation with δ increasing while
the CCQRE shows a slight fluctuation.
From Fig. 7, it can also be found that the CQRE outper-
forms the Rife and Quinn estimator for δ > 0.05. The CQRE
performs better than the Rife estimator but worse than Quinn
estimator for δ < 0.05. However comparing with the Rife and
Quinn estimator, the CQRE performs more robust against for
all δ. The Candan estimator outperforms the CQRE for δ <
0.15 while it performs worse than the CQRE for δ > 0.15.
It is consistent with the theoretical analysis in Section III (B).
 
Comparing all the estimators, it’s easy to see although the FIGURE 9. MSE δ̂ achieved by simulation versus snr with δ = 0.1.
MSE of CCQRE is slightly larger than the Candan estimator
for δ < 0.15, it is most robust against for all δ.
Case 3 (MSE of δ̂ Versus SNR): We now assess the perfor-
mance of the estimators by plotting the MSE versus different
snrs starting from −6 to 6 dB in a step of 1 dB for a fixed
δ. δ is respectively set to be 0.1 and 0.4 which respectively
represent a typical small and large
 fractional
 correction
  term.
The accuracy of MSE δ̂CQRE , MSE δ̂C and
 
MSE δ̂CCQRE versus snr has also been verified by means
of computer simulations. The accuracy of MSE versus
snr with δ = 0.1 and 0.4 are respectively presented
in Fig. 8 (a) and (b). The simulation and theoretical results
are represented by continuous lines and circles, respectively.  
FIGURE 10. MSE δ̂ achieved by simulation versus snr with δ = 0.4.
Fig. 8 shows a quite good agreement between simulation and
theoretical results versus snr.
The MSEs of the estimated fractional correction term ver-
sus snr with δ = 0.1 and δ = 0.4 are respectively presented
in Fig. 9 and Fig. 10. From both Fig. 9 and Fig. 10, it can be
seen that the MSEs of all estimators decrease and approach
the CRLB with the snr increasing.
From Fig. 9, it can be seen that, for δ = 0.1, the Rife
estimator performs worst while the Candan estimator per-
forms best. The CQRE performs better than the Quinn and
Rife estimator while it performs worse than the CCQRE and
Candan estimator. For snr>0 dB, the CQRE and CCQRE
performs closely to the Candan estimator. For snr<0 dB,
the CCQRE performs slightly worse than the Candan esti-
mator. It is caused by that, for δ = 0.1 which is close to the
set threshold 0.15, the probability of δ̂R ≤ η is low as the
snr is low.
According to Fig. 10, we can see that the Quinn estimator
performs worst while the estimator CCQRE performs best.
The MSE of CQRE is very closely to CCQRE. The Candan
estimator performs better than the Quinn estimator but worse
than the other estimators even for high snr.
The means and stds of MSEs for the varying δs or snrs
are shown in Table 2. The means and stds of MSEs for δ
varying from 0.025 to 0.5 in a step of 0.025 with snr=−3dB
as illustrated by Fig. 7 are given by the second line of Table 2.
As shown by the second line, for varying δs with a fixed
snr, both the mean and std of the MSEs for the CCQRE are
 
FIGURE 8. MSE δ̂ achieved by simulation and by theoretical calculation
versus snr (a) δ = 0.1 (b) δ = 0.4. smallest among all the analyzed methods.

117072 VOLUME 7, 2019

S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone

TABLE 2. The means and stds of MSEs for the varying δs or snrs (10−4 ).

FIGURE 12. Amplitude normalization time-domain waveform, amplitude

spectrum and time–frequency distribution of the received 10 single-tones
in sea trial.

FIGURE 11. Schematic diagram of the sea trial.

The means and stds of MSEs for snr varying from −6 to

6 dB in a step of 1 dB with δ = 0.1 and 0.4 as illustrated
by Fig. 9 and Fig. 10 are respectively given by the third
and fourth line of Table 2. As shown by the third and fourth
line, for varying snrs, the mean and std of the MSEs for the
CCQRE is slightly larger than the Candan method but much
smaller than other methods when δ = 0.1. When δ = 0.4,
both the mean and std of the MSEs for the CCQRE are
smallest among all the analyzed methods.
FIGURE 13. Absolute error of the estimated δ for the received
Based on the above analysis, we can draw a conclusion of 10 single-tones in sea trial.
that the CCQRE performs best and most robust among all the
presented estimators.
be seen that the proposed CCQRE performs the same as the
Rife estimator while they outperforms the Quinn and Candan
B. EXPERIMENTS OF SEA TRIAL DATA
estimator. This proves that, even for the single-tones polluted
Schematic diagram of the sea trial is shown in Fig. 11. by the real ocean ambient noise, the CCQRE is also effective
As shown in Fig. 11, the acoustic source is 10 m to the ocean and more robust.
surface and the receiver is 30 m. The distance between the
acoustic source and receiver is about 45 km. The sea trial is
V. CONCLUSION
carried out in the East China Sea and the depth of the sea is
For fine-resolution frequency estimation of a real-valued
100 m. The sea condition is three-level. A single-tone with
single-tone, a direct frequency estimation method named
f0 = 310Hz has been sent from the acoustic source and the
CCQRE is proposed by only utilizing three DFT samples
transmitting source level is 200dB. The sampling frequency
around the peak. The CCQRE method is proposed by inherit-
fs and DFT length are the same as the computer simulations.
ing the advantages of the Rife, Quinn and Candan estimators.
So δ is equal to 0.36 and kp is also 79.
Both theoretical analysis and simulation results show that the
Fig. 12 shows the amplitude normalization time-domain
CCQRE is more robust against the interbin frequency, and
waveform, amplitude spectrum and time–frequency distribu-
has better estimation performance, compared to the original
tion of the received 10 single-tones with f0 = 310Hz in the
three estimators.
sea trial. The received 10 single-tones are recorded by a single
Experiments of sea trail data prove that, even for the single-
hydrophone. It can be found that the received single-tones
tones polluted by the real ocean ambient noise, the proposed
are almost masked by the ocean ambient noise in the time-
CCQRE is also effective and more robust. The proposed
domain, as shown in Fig.12.
CCQRE only utilizes three DFT samples for the frequency
The accuracy of the estimated δ for the sea trial data is
estimation, and thus it is suitable for real-time applications
measured by absolute error (AE) which is defined as
such as sonar and radar signal processing.
 
AE δ̂ = δ̂ − δ (55)
ACKNOWLEDGMENT
Absolute error of the estimated δ for the received 10 single- The authors would like to thank the reviewers for their useful
tones in sea trial is presented in Fig.13. From Fig. 13, it can comments.

VOLUME 7, 2019 117073

 S. Yao et al.: Improved Fine-Resolution Method for Frequency Estimation of Real-Valued Single-Tone
REFERENCES
[1] Y. Ding, C. Lei, X. Xu, K. Sun, and L. Wang, ‘‘Human micro-Doppler
frequency estimation approach for Doppler radar,’’ IEEE Access, vol. 6,
pp. 6149–6159, 2018.
[2] S. Djukanović, T. Popović, and A. Mitrović, ‘‘Precise sinusoid frequency
estimation based on parabolic interpolation,’’ in Proc. 24th Telecommun.
Forum (TELFOR), Belgrade, Serbia, Nov. 2016, pp. 1–4.
[3] D. Slepian, ‘‘Estimation of signal parameters in the presence of noise,’’
Trans. IRE Prof. Group Inf. Theory, vol. 3, no. 3, pp. 68–89, Mar. 1954.
[4] H.-B. Lin and K. Ding, ‘‘Energy based signal parameter estimation method
and a comparative study of different frequency estimators,’’ Mech. Syst.
Signal Process., vol. 25, pp. 452–464, Jan. 2011.
[5] D. C. Rife and R. R. Boorstyn, ‘‘Single tone parameter estimation from
discrete-time observations,’’ IEEE Trans. Inf. Theory, vol. IT-20, no. 5,
pp. 591–598, Sep. 1976.
[6] D. Belega and D. Petri, ‘‘Frequency estimation by two-or three-point inter-
polated Fourier algorithms based on cosine windows,’’ Signal Process.,
vol. 117, pp. 115–125, Dec. 2015.
[7] C. Candan, ‘‘Analysis and further improvement of fine resolution fre-
quency estimation method from three DFT samples,’’ IEEE Signal Process.
Lett., vol. 20, no. 9, pp. 913–916, Sep. 2013.
[8] C. Candan, ‘‘A method for fine resolution frequency estimation from
three DFT samples,’’ IEEE Signal Process. Lett., vol. 18, no. 6,
pp. 351–354, Jun. 2011.
[9] J.-R. Liao and C.-M. Chen, ‘‘Phase correction of discrete Fourier transform
coefficients to reduce frequency estimation bias of single tone complex
sinusoid,’’ Signal Process., vol. 94, pp. 108–117, Jul. 2014.
[10] D. C. Rife and G. A. Vincent, ‘‘Use of the discrete Fourier transform in
the measurement of frequencies and levels of tones,’’ Bell Syst. Tech. J.,
vol. 49, no. 2, pp. 197–228, 1970.
[11] J. Z. Xiang, S. Qing, and C. Wei, ‘‘A novel single tone frequency estimation
by interpolation using DFT samples with zero-padding,’’ in Proc. IEEE
ICSP, Chengdu, China, Mar. 2017, pp. 277–281.
[12] B. G. Quinn, ‘‘Estimating frequency by interpolation using Fourier coef-
ficients,’’ IEEE Trans. Signal Process., vol. 42, no. 5, pp. 1264–1268,
May 1994.
[13] B. G. Quinn, ‘‘Estimation of frequency, amplitude, and phase from the DFT
of a time series,’’ IEEE Trans. Signal Process., vol. 45, no. 3, pp. 814–817,
Mar. 1997.
[14] E. Aboutanios and B. Mulgrew, ‘‘Iterative frequency estimation by interpo-
lation on Fourier coefficients,’’ IEEE Trans. Signal Process., vol. 53, no. 4,
pp. 1237–1242, Apr. 2005.
[15] E. Jacobsen and P. Kootsookos, ‘‘Fast, accurate frequency estimators,’’
IEEE Signal Process. Mag., vol. 24, no. 3, pp. 123–125, May 2007.
[16] U. Orguner and C. Candan, ‘‘A fine-resolution frequency estimator using
an arbitrary number of DFT coefficients,’’ Signal Process., vol. 105, no. 10,
pp. 17–21, Dec. 2014.
[17] Y. Liu, Z. Nie, Z. Zhao, and Q. H. Liu, ‘‘Generalization of iterative Fourier
interpolation algorithms for single frequency estimation,’’ Digit. Signal
Process., vol. 21, no. 1, pp. 141–149, Jan. 2011.
[18] D. Mikluc, D. Bujaković, M. Andrić, and S. Simić, ‘‘Estimation and extrac-
tion of radar signal features using modified B distribution and particle
filters,’’ Frequenz, Jul. 2016, pp. 1–11.
117074
[19] J. Schoukens, R. Pintelon, and H. Van Hamme, ‘‘The interpolated fast
Fourier transform: A comparative study,’’ IEEE Trans. Instrum. Meas.,
vol. 41, no. 2, pp. 226–232, Apr. 1992.
[20] G. Q. Qi and X.-L. Jia, ‘‘Accuracy analysis of frequency estimation of
sinusoid based on interpolated FFT,’’ (in Chinese), Acta Electron. Sinica,
vol. 32, no. 4, pp. 625–629, Apr. 2004.
[21] G. Q. Qi, ‘‘Accuracy analysis and comparison of some FFT-based fre-
quency estimators,’’ (in Chinese), J. Vib. Eng., vol. 19, no. 1, pp. 86–92,
Mar. 2006.
SHUAI YAO received the B.S. degree from
the Ocean University of China, Qingdao, China,
in 2005, and the Ph.D. degree from Southeast
University, Nanjing, China, in 2016, where he is
currently a Lecturer with the Key Laboratory of
Underwater Acoustic Signal Processing of Min-
istry of Education. His research interests include
underwater acoustic signal processing, array sig-
nal processing, parameter estimation, and signal
detection.
QISONG WU received the B.S. and Ph.D. degrees
from Xidian University, Xi’an, China, in 2005 and
2010, respectively. He was a Postdoctoral Asso-
ciate with Duke University, Durham, NC, USA,
from 2010 to 2013, and Villanova University,
Villanova, PA, USA, from 2013 to 2015. He is cur-
rently an Associate Professor with the Key Labora-
tory of Underwater Acoustic Signal Processing of
Ministry of Education, Southeast University. His
research interests include sonar signal processing,
radar signal processing, and the sparse Bayesian learning.
SHILIANG FANG received the M.S. and Ph.D.
degrees from Southeast University, Nanjing,
China, in 1986 and 2009, respectively, where he is
currently a Professor with the Key Laboratory of
Underwater Acoustic Signal Processing of Min-
istry of Education. His research interests include
signal processing, target detection, parameter esti-
mation, and underwater target classification.
VOLUME 7, 2019