1 IO IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-25, NO.

2, APRIL 1911

A New Phase Unwrapping Algorithm

JOSE M. TRIBOLET, STUDENT MEMBER, IEEE

Abstract-A newphaseunwrappingalgorithm is proposed that
combines the information contained in both the phase derivative and
the principal value of the phase into an adaptive numerical integration
scheme. This new algorithm has proven itself to be very reliable and it
c p be easily incorporated in standard homomorphicsignal processors,
I. INTRODUCTION
arg [ X ( * ) ]:i its phase.Let ?(n) be thecomplexcepstrum
of x(n) and X(ei’) its Fourier transform. Then
2 ( e i W >= log x ( e i w >= log IX(eia>I t j arg [ ~ ( e i w ) ] (2)
where we have assumed that the complex logarithm is defined
so that (2) representsAaproperFouriertransform.In that
case, the derivative of X ( e i w ) is well defined:

H OMOMORPHIC signal processing, a nonlinear technique

based on linearfilteringoperations onthecomplex
cepstrum [ 1 1 , was introduced in 1965 by Oppenheim and has
~_d 2 ( e i w ) -- d log X ( e i w )d X-- ( e f w ) / d w
dw dw X(eia) ’ (3)
beenquite successfully applied to avarietyofproblems, and it follows that the derivative of arg [ X ( e i w ) ] is given by
namely, speech, seismic, and EEG data processing. the imaginary part of (3):
Thecomputationofthecomplexcepstrum requires the
definition of a continuous phase curve for the input data [ 11. d arg [ x ( e i W ) ]-
-
XR(eiw)X;(eiw) - X z ( e i w ) x i (e’”)
This phase curve is commonly called the unwrapped phase and dw IX(eiw)12
it can be rigorouslydefinedinterms of the integralof its
derivative [ 11. (4)
Conventionalphaseunwrappinghasbeendoneeither by where the prime denotes d / d w and
appropriate processing ofthe principal value ofthe phase
X’(eiw) = X i ( e i w ) t j X j ( e i w ) = - j FT {nx(n)}. (5)
(Schafer’s algorithm) orby numericalintegration of the
phase derivative. The phasearg [ X ( e i w ) ] can thus beunambiguouslydefined
The algorithm we proposehere is anadaptivenumerical in terms of its derivative arg‘ [ X ( e i w ) ] as
integration scheme that combines the information contained
W
in both the phase derivative andthe principal value of the
arg [ X ( e i w ) ] = arg’ [ X ( e j s ) ]d v (6)
phase, as follows: at each frequency, asetof permissible
phasevaluesis defined by adding integer multiples of 2~ to
the principal value of the phase. The selection of one of these with initial conditions given by
values as the value of the unwrapped phase at this frequency arg [ X ( e i o ) ]= 0. (7)
is done with the help of a phase estimate formed by numerical
integration of the phase derivative with a given step interval. Such phase functions are commonly called the unwrapped
This step interval is adapted until the phase estimate becomes phase of x(n).
arbitrarily close t o one of the permissiblephasevalues. This In order for the complex logarithm in (2) to beproperly
newalgorithmhasprovenitself to be very reliable, at least defined, arg [ X ( e j w ) ] mustbeacontinuousandoddfunc-
as far as our experience with speech, seismic, and lunar data tion.Fromitsdefinitionin (6) and (7), we concludethat
is concerned. the unwrapped phase is indeed a continuous function; it will
alsobe anoddfunction whenever themeanofthe phase
11. THEUNWRAPPED
PHASE:THEORY derivative (arg’ [ X ( e j w ) ]), defined as
Let x(n) be a sequence and X ( e i w ) its Fourier transform.
Let (arg’ [ X ( e i W ) ] =
> & 2n
arg’ [ X ( e j w ) ]dw, (8)
X ( e i w ) = X,(eiw) t X z ( e i a )
equalszero. When such is notthe case, phase unwrapping
= IX(eiw)l exp [ j arg [X(e’”)] ] (1) mustbefollowed by the removal of the linear phase com-
where IX(.)lis the magnitude of the Fourier transform and ponent induced by (arg‘ [ X ( e i w ) ]).

Manuscript received March 5, 1976; revised October 12, 1976. This 111. CONVENTIONALPHASEUNWRAPPING
: BASIC ISSUES
work was supported in part by theAdvanced Research Projects Agency
monitored by ONRunderContractN00014-75-C-0951-NR049-308 In spite of the fact that the unwrapped phase is very pre-
and in part by the National Science Foundation under Grant ENG71- cisely defined by ( 6 ) , one soon realizes that such an equation
02319-A02. can never beexactlyimplementedina digital computer.It
TheauthoriswiththeDepartment of ElectricalEngineering and
Computer Science, Research Laboratory of Electronics, Massachusetts is therefore of interest to investigate how to approximate it
Institute of Technology, Cambridge, MA 02139. or how to alternatively define the unwrapped phase. Perhaps
TRIBOLET: NEW PHASE UNWRAPPING ALGORITHM 171

themoststraightforwardapproximation to (6) is to do nu-

merical integration of the phase derivativeon a set of uni-
formly spaced frequencies
2n
8={(Ok=-k, k = 0 , 1 , " * , N - 1) (9)
N
since, through the use of the fast Fourier transform (FFT),
it is possible to evaluate very efficiently arg' [ X ( e i w k ) ] ,
\dwk E 8. This method depends very critically on the size of
theintegrationstep 2n/N. Unfortunately,it is not possible
I 131

to estimate a priori how large the FFT size N should be in

order to accurately unwrap the phase. As shall be illustrated
in Section V, numerical integration may lead to very signifi-
cant errors, even for reasonablylarge FFT's.
Anotherapproachthathasbeentakenquiteoften is to
compute the principalvalue of the phase
ARG [X(eiwk)]= { arg [X(eiwk)]}mod2 n , Vcdk E 8 I
Fig. 1 . Ambiguities in detecting discontinuities of the principal value
(10) of the phase.

using inverse tangent routines on X(ejw) and then unwrap by

appropriatelyaddingmultiplesof 2n to theprincipal value The phase unwrapping problem amounts to determining the
until the discontinuities induced by the modulo 2n operation correct integer value ZC(Ol)such that
are removed. This algorithm relies on the detection of such arg [X(ein I )] = ARG [X(eis21 )] t 2n Zc(nl). (1 2)
discontinuities,which is donebycomputingthedifference
betweentheprincipal values ofthe phase attwoadjacent This is done through the use of numerical integration of the
frequencies W k - 1 and W k . Whenever this difference is greater phase derivative. We shall adopt here the trapezoidal integra-
than a given threshold, we say that a discontinuity is present. tion rule. Assuming the unwrapped phase to be known at a
This procedure will yield the unwrapped phase whenever the frequency Q0 < Q 1 , we shall definea phase estimateat
frequency sampling is fineenough so thatthedifference a l , a ? g [X(eis21)180]by
betweenanytwoadjacent samples oftheunwrappedphase
is always less than the prespecified threshold. In such a case,
theunequivocalsortingbetweennaturalvariationsofthe t 8 1 - 8 0 [ arg' [X(ein 0 )] t arg' [X(ein 1 )] ] . (1 3)
phase andthediscontinuitiesinducedbythemodulo 277 2
operation is possible. An examplewheresuchsorting will Clearly,thisestimateimproves as thestep interval A 8 =
be equivocal is the following. Consider a sequence x(n) that 8 , - Q 0 becomes smaller. We shall define the phase estimate
hasonezero very close to theunit circle at 0, = ( W k t at C l 1 to be consistent if it lies within a predefined distance
0k-1)/2. Thephase will then change byapproximately n of one of thepermissible phase values at Cl1, that is, if
between O k and W k - 1 , increasing or decreasing depending on
alC(8,)3la?'g [X(ei"1)18~]- ARG [X(ei"~)]
the position of the zero relative to the unit circle. This situa-
tion is illustrated in Fig. 1. At this sampling rate,the al- + 2nZc(81)I < THLD < T . (14)
gorithmshallnotbe able to distinguishbetweena phase
The basic idea of this algorithm is thus to adapt the step
increaseof n, whichcorresponds to the presence ofadis-
size A 8 untilaconsistent phase estimate is found. The
continuityinducedbythemodulo 277 operationbetween
resultant 1,(8,) in (14) is used in (12) to form the unwrapped
W k - 1 and w e , and a decrease of n, which corresponds to the
phase at 8 , . Thisunwrapped phase is then used to form
absence of such discontinuity.
This situation often arises in practice, as shall be illustrated
a?'g [X(ei"~)l8~], a2> and so on.
inSection V. Thisisoneofthemanyexamplesonecould B. Implementation
think of where the information contained in the samples of
theprincipal value ofthe phase is notenoughto allow a The basic idea described previously requires the computa-
tion of the. phase derivative and of the principal value of the
reliable operation of this algorithm.
phase ona set ofnonequallyspacedfrequencies.Forthis
IV. PHASEUNWRAPPING BY ADAPTIVE INTEGRATION algorithm to be practical, one must take full advantage of the
A . Basic Idea FFTalgorithmandreducethenumber of extra discrete
Fourier transform (DFT) computations to a reasonably small
Let nl be an arbitrary frequency value and ARG [X(ein 1 ) ]
number. Let us denote by
be the principal value of the phase at 8 ,.
The set of permissible phase values at Q1 is then given by {Wk = (2n/N)k, k = 0,1, * . , N - 1) (1 5)
{ ARG [X(@ 1 )] t 2n1, V I , integer }. (1 1) the set of uniformly spaced frequencieswithinterval 2n/N
 172 IEEE TRANSACTIONS
ACOUSTICS,
SPEECH,
ON AND SIGNAL PROCESSING,
APRIL 1917

where N = 2M (or, in general, any highly composite number). and principal values are computed, they are stored in a stack
The phase derivative and the principal value of the phase at fashion. As soon as a consistent estimate is found, the corre-
thesefrequenciesmay then be computed usingFFT’s to spondent data are moved out of the stack t o a register that
evaluate the DFT’s of x(n) and nx(n). At each W k , a phase holdsthemost recentconsistentestimateof the phase at
estimate is initiallyformed by one-step trapezoidal integra- somefrequency
within [ ~ k -o k~] ,. New estimates are
tion,starting at W k . 1 . If theresultantestimate is not con- always formed by integrating from the most recent estimate
sistent, the adaptive integration scheme is applied within the to the frequency correspondent to the top of the stack. We
interval [ o ~ . o~k ,] . The step size adaptation was carefully present next a Fortran program along these lines.
designed to minimize thenumberofextra DFT’s required.
The search for consistency is done by consecutively splitting C. A Fortran Program
thestep intervalinhalf. As therequired phase derivatives See program shown below.

CCEPS TITLE: C 09/01/76 JMT

C
PURPOSE C :
C To compute the complex cepstrum cx(n)
C of a sequence x(n).
C
C AUTHOR: Jose M. Tribolet
C room 36-387 M.I.T.
C Cambridge, MA 02139
C
C USAGE :
C Call CCEPS(X,NX,CXE,CXO,NCX,ISNX,ISFX,YR,YI)
C
C DESCRIPTION OF PARAMETERS:
C X:Real array containing the sequence x(n).
C NX:Length of x(n).
C CXE:Real array containing the even-indexed
C samples of cx(n).
C This array must be of dimension NCX/Z+l.
C CX0:Real array containing the odd-indexed
C samples of cx(n).
C This array must be of dimension NCX/2+1.
C NCX:Number of complex cepstrum samples to be
C computed.Program will evaluate cx(n)for
C -NCX/Z-l<n<NCX/2.Since radix2 FFTs are
C used,NCX must be a power2 . of
C NCX must be greater or equal to NX
C 1SNX:Integer value that is either-1 or +1
C depending on whether the data x(n) had
C to be sign reversed or not.
C 1SFX:Integer value that indicates the amount
C of shifting that had to be performed on
C the data x(n).
C YR,YI:Two real auxiliary arrays,each of
C dimension NCX/2+1.
C
C SUBROUTINES CALLED:R2FFT(XR,XI,NPTS,ITYPE)
C
C RZFFT is a radix 2 FFT,with the
C following characteristics:
C XR,XI:Real and imaginary arrays to be transformed.
C NPTS:Number of complex points to be transformed,
C i.e.,the number of elements in XR or X1:must
C be an integer power 2of .
C 1TYPE:Indicates type of FFT and direction:
C ITYPE=+l FFT for real 2*NPTS-point sequence:
C Calling program supplies NPTS-point
C complex sequence formed by loading odd-
C -indexed samples of real sequence into
C XR and even-indexed samples of real se-
C quence into XI(index=1,2, ...,2*
NPTS).
C Program computes the first NPTS of the
C 2*NPTS-point DFT of the 2*NPTS-point re-
C al sequence.Since the DFT of a real N-
C -point sequence is Hermitian symmetric
C mod N,the first transform point and the
C (NPTS+l)th transform point are both pure-
C ly real.By convention,the real part of the
C (NPTS+l)th transform point will be retur-
C ned in the imaginary partof the first
C transform point.TheDFT is thus uniquely
C determined.
C ITYPE=-1 Inverse FFT for Z*NPTS-point Hermitian
TRIBOLET: NEWPHASEUNWRAPPINGALGORITHM

C symmetric complex DFT sequence:

C Calling program supplies the firstNPTS of
C the input sequence with the convention that
C the real part of the (NPTS+l)th pointbe
C stored in the imaginary partof the first
C point.The desired inverseDFT is purely
C rea1:its odd-indexed points are givenby
C XR and its even-indexed points by XI
C ...,
(indexed=1,2, 2*NPTS).
C
C

SUBROUTINE CCEPS(X,NX,CXE,CXO,NCX,ISNX,ISFX,YR,YI)
DIMENSION X(1) ,CXE(l) ,CXO(1) ,YR(l) ,YI(1)
DIMENSION ISK(13) ,SK1(13) ,SK2(13)
INTEGER SP
COMPLEX CO,Cl,CB
DOUBLE PRECISION A,B,C,D,E
DATA TWOPI/6.283185307/

C
C Initialization
C

N=12
L=2**N
NPTS=NCX/2
PI=TWOPI/2.
H=TWOPI/FLOAT(2*NCX)
Hl=H/L
THLD1=.8*PI
THLD2=.5*PI
ISNX=+I

C
C Transform x (n)
C
C Load CXE and CXO arrays.
C
NHALF=NX/2
DO 10 I=l,NHALF
J=2*1-1
CXE(I)=X(J)
10 CXO(I)=X(J+l)
GO TO 20
IF( (NX-2*NHALF) .EQ.O)
NHALF=NHALF+l
CXE(NHALF)=X(NX)
cxo (NHALF)=o.
20 IF(NHALF.EQ.NPTS)GO TO 35
NHALF=NHALF+l
C
DO 30 I=NHALF,NPTS
CXE(I)=O.
30 cxo(I)=o.
C
C FFT
C
35 CALL R2FFT(CXE,CXO,NPTS,+l)
CXE (NPTS+l) (1)
=CXO
cxo(l)=o.
CXO =O.
(NPTS+l)

Transform n*x (n)

Load Y R and YI arrays.

NHALF=NX/2
DO 40 I=l,NHALF
J=2*1-1
YR(I)=(J-~)*x(J)
40 YI(I)=J*X(J+l)
IF( (NX-2”NHALF) .EQ.O)
GO TO 50
NHALF=NHALF+l
YR(NHALF)=(NX-~)*X(NX)
YI (NHALF)
=O.
50 IF(NHALF.EQ.NPTS)GO TO 6 5
NHALF=NHALF+l
DO 60 I=NHALF,NPTS
YR(I)=O.
 IEEE TRANSACTIONS ON ACOUSTICS,SPEECH, AND SIGNAL PROCESSING, APRIL 1917

60 YI(I)=O.
C
C FFT
C
65 CALL RZFFT(YR.YI,NPTS,+l)
YR(NPTS+l)=YI (1)
YI =O.
(NPTS+l)
YI(l)=O.
C
C Check if sign is required.
reversal
C
If (CXE(1) .LT.O.) ISNX=-I

Compute 1ogmagnitude:store in CXE

Compute phase derivative:store in YR.
Compute phase principal va1ue:store YI.
in
For w=(TWOPI/NCX)*I,I=O,l,-.-NPTS.

DVTMN=O.
DO 70 I=l,NPTS+l
A=CXE ) (I
B=CXO ) (I
C=YR (I)
D=YI (I)
E=A*A+B*B
YR(I)=- ( (A*c+B*D)/E)
DVTMN=DVTMN+YR( I)
C
CXE (I) =DLOG (E) /2
IF(ISNX.EQ.l)YI(I)=DATANZ(B,A)
IF(ISNX.EQ.-l)YI (I)=DATAN2 (-B,-A)
70 CONTINUE
DVTMN=(2*DVTMN-YR(l)-YR(NPTS+l) )/NCX
ELPINC=2*DVTMN*H

Phase unwrappingby adaptive integration.

PH=O.
DO 150 1=2,NPTS+1
C
C Store unwrapped phase in CXO.
C

CXO (1-1) =PH

C
C Form phase estimate at w=(TWOPI/NCX)* (1-1) by
C trapezoidal integration with step size TWOPI/NCX.
C

PHAINC=H*(YR(I)+YR(I-1))
IF (ABS (PHAINC-ELPINC) .GT.THLDl)GO
90 TO
PH=CXO (1-1) +PHAINC
C
C Check consistency of estimate.
C

AO=(PH-YI(J.1 )/TWOPI
Al=IFIX(AO) *TWOPI+YI (1)
A2=A1+SIGN(TW0PItAO)
A3=ABS (AI-PH)
A4=ABS (A2-PH)
IF(A3.GT.THLD2.AND.A4.GT.THLDI) GO TO 90

C
C Phase estimate was consistent.
C

PH=Al
IF(A3.GT.A4)PH=A2
IF(ABS(PH-CXO(1-1)) .GT.PI)GO TO
90
TRIBOLET: NEWPHASEUNWRAPPINGALGORITHM 175

GO TO 150

C
C Phase estimate was not consistent:
C Adapt step size.
C
C Initiate software stack.
C
90 SP=1
ISK( 1) =L+1
SKl(l)=YI (I)
SK2(1)=YR(I)
C
C Initiate registers
C
IB=l
Bl=CXO(I-l)
C
BZ=YR(I-l)
C
C If software stack dimension does not
C allow further step reduction,STOP.
C
100 IF((ISK(SP)-IB) .GT.1) GO 110
TO
STOP 'FAILED'

C
C Define intermediate frequencies(1.F.):
C w= (TWOPI/NCX)* (I-2+ (K-1) /L)
C
110 K=(ISK(SP)+IB)/2
C
C Compute DFTs of x(n) and n*x(n)
AT I.F..
C

AO=TWOPI*(1-2.+(K-l)/FLOAT(L))/FLOAT(NCX)
Cl=(O.,O.)
C2=(0.,0.)
PO 120 J=l,NX
ARG=AO* (J-1)
CO=CMF'LX(COS(ARG) ,-SIN(ARG))*X(J)
c1=c1+co
120 C2=C2+CO* (J-1)
C
C Compute phase derivative
and the
C principal valueof the phaseAT I.F..
C

SP=SP+l
ISK(SP)=K
A=REAL (C1)
B=AIMAG (C1)
C=REAL (C2)
D=AIMAG (C2)
IF(ISNX.EQ.+l) SKl(SP)=DATANZ(B,A)
IF(ISNX.EQ.-1) SKl(SP)=DATANZ(-B,-A)
SK2 (SP)=- (A*C+B*D)/(A*A+B*B)

C
C Evaluate estimate at I.F.
C
130 DELTA=Hl*(ISK(SP)-IB)
PHAINC=DELTA* (B2fSK2 ) (SP)
IF (ABS(PHAINC-DELTA*2*DVTMN) .GT.THLDl)GC To 100
PH=Bl+PHAINC
C
C Check consistencyof estimate at I.F.
C
AO= (PH-SK1 (SP) )/TWOPI
Al=IFIX (AO) *TWOPI+SKl (SP)
A2=Al+SIGN(TWOPI,AO)
A3=ABS (Al-PH)
A4=ABS (A2-PH)
IF(A3.LT.THLD2.OR.A4.LT.THLD2)m TO 140
176 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,APRIL 1977

C
C Estimate was not consistent:Reduce step size.
C
C
Go To 100
C
C Estimate was consistent:Update registers.
C
140 PH=Al
IF GT .A4 ) PH=A2
(A3.
IF (ABS (PH-B1) .GT.PI)
100 GO TO
IB=ISK(SP)
Bl=PH
B2=SK2 (SP)
SP=SP-l

C
C When software stack is empty,the unwrapped
C phase atw= TWOPI* (I-l)/NCX is held in the
C B1 register.
C

IF(SP.NE.0) GO TO 130
PH=Bl

C
C End of step size adaptation.
C
150 CONTINUE
C
C
C End of Phase Unwrapping
C
C
C Remove Linear Phase Component
C

ISFX= (ABS f.1)

(PH*2/TWOPI)
IF(PH.LT.O.0) ISFX=-ISFX
H=PH/NPTS
M) 160 I=2,NPTS
160 CXO[I)=CXO(I)-H*(I-l)
C
C Compute cx (nl :IFFT.
C

CXO (1) =CXE (NPTS+l)

CALL RZFFT(CXE,CXO,NPTS,-l)
RETURN
END

V. AN EXAMPLE frequencies,leading to anunacceptableresultfortheun-

Consider the signal shownin Fig. 2. In Fig. 3 itsphase
derivative is depicted,corresponding t o a FFT size of N =
2048, which amounts to four times the data length. Let us,
for example, focus on the low frequencies where large spikes
of the phase derivative occur.Table I containsthefirst25
samples of the phase derivative (scaled by n / N ) , the principal
value of the phase,and the unwrappingresults asgiven by
numericalintegration, Schafer’s algorithm(withathreshold
set for 4.5), and adaptive integration.
Note that both the numerical integration and the adaptive
integration results progress with a similar trend, although the
accumulation of error in the numerical integration estimate,
noticeable after the occurrence of the large derivative spike at
sample k = 21, causes a significant deviation of these estimates
fromthesetof permissible estimates as previouslydefined.
Thisdeviation will bepropagatedfromthispoint t o higher
wrapped phase.
Schafer’s algorithmhas also failed t o successfully unwrap
the phase. For example, at sample k = 10, it actually predicts
an increasein the phase from - 13.86 to -9.71 against the
evidence offered by the phasederivative at thatpoint.The
same happens at samples k = 15 and k = 21. Theresultsof
adaptive integration are shown in Fig. 4 after the removal of
the linearphase component.Theunwrapped phaseagrees
everywhere with the information available. The phasevalues
are selectedfrom the permissiblephasevalues at eachfre-
quency and they do follow the trend predicted by the phase
derivative.
VI. FINALCOMMENTS
Phase unwrapping is, as stated in this paper, a well-defined
operation. Anystable signal x(n) without zeros on the unit
TRIBOLET: NEW PHASE UNWRAPPING ALGORITHM 177

TABLE 1
OF DIFFERENT
COMPARISON PHASE UNWRAPPING TECHNIQUES

0 - .86 .0 0 0.0 .0 0 .00-

1 - .76 -1.61 - 1.62
__ - 1.64 - 1.64
2 - .67 -3.06 - 3.06 - 3.06 - 3.06

512
3 - .70 - 1.86 - 4.43 - 4.42 - 4.42
4 .31 - .86 - 6.00 - 5.97 - 5 "-
5 - .87 __-- - 7.13 - 7.77 - 7.77 __ 1.49
6 -3.03 - .69 - 9.28 - 9.31 -9.31
- 7 - .64

-~8 - .74
- 9 -1.24 -13.96 _________
-1.30 -13.86 -13.86
_ 10~2.85 ~ _- .BO_ _ _ _ _ _ -16.00 _ _ - 9.71 . -16.00
- 11 1.48 - .62 -17.42-11.31 -11.08

- 12 - .60 -18.58

Fig. 2. A s i g d x [ n ] .
____

-
13

14

15
-
-
.66
.94
-1.14
-
-2.50

- .33
:*T&Fo
1.45
, -21.51
-23.59 I -11.11
-19.83
-21.35

-23.68

+I 23
16 - .74 - -25.46 -12.90 -25.41

__ 17 - .64 -1.69
18 - .61 __
-. -2.93
19 - .59
-16.70 ____
2.15 -
- 20 - .64 _____
-17.88-30.50 -30.45

21 -2.73 -1:;: - -20.39 - 3 2 . 5

0 - 1025
22 - .58 2 L - -35.18
-16.32
-37.19 -

Ill II It - 23 - .49 1.47 - __-38.38

-36.23-17.38
I

circle will have a unique unwrapped phase arg [X(ejw)]. The

phaseunwrappingmethodproposedhereperforms signifi-
cantly better than conventional methods.
This method has been used in the homomorphic analysis
of speechandseismicdatawithencouragingresults.Itis
hoped that it will contribute to further research on the use
of homomorphic signal processing.

REFERENCES -39 54

[ l ] A. V. Oppenheim and R. W. Schafer, Digital Sigml Processing. Fig. 4. Unwrappedphase arg [X(eiO)] aftertheremoval of the linear
Cliffs,Englewood NJ: Prentice-Hall, 1975. phase component.