24 IEEETRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO.

1, JANUARY 1981

A .Nonuniform SamplingApproach to Data Compression

Abstracr-A nonuniform sampling approach to digital encoding of values of N and f,, however, will inevitably result in a large
analog sources is proposed. The nonuniform sampler is basically a signal distortion, which can be grossly grouped into: 1) slope
level crossing detector (LCD) which produces a sample wheneverthe overload noise and2) granular noise. Slopeoverload occurs
input to theLCD crosses a threshold level.The information aboutthe
source signal is contained in the time intervals between level crossings
when the input signal changes faster than the DPCM coder can
and in the directions of level crossings. By assigning strings of the follow, while granular noise is proportional to the quantizer
2-tuple “00” to represent the time between level crossings and “01” step size d. Thus, for sources which have intermittent occur:
and ‘‘10’’ to denote the directions level
of crossings, theoutput binary rences of fast changing and relatively quiescent intervals,
sequence of the nonuniform sampling encoder (NSE) contains a high slope overload noise and granular noise will occur in distinct
probability of the0 symbol, which makes it suitable for further simple
run-length encoding (RLE) to attain a “good” overall compression
regions. To minimize the effect of slope overload and granular
ratio. lntroduction ,of prediction converts the NSE to a nonuniform noise, it seems reasonable to sample the fast changing regions
sampling,predictive coding (NSPC) scheme, which, depending on the at a higher rate than the ‘slowly changing regions, as well as
source, can potentially improve the compression ratio. Results ob- assigning different quantizer stepsizes. Adaptive delta modula-
tained in the encoding ofa band-limied Gaussian source and a raster- tion (ADM) samples the source at a futed frequency and as-
scanned black and white still image reveal that an NSE/RLE or NSPCI
RLE system exhibits performance superiorthat to of an adaptive delta
signs varying step sizes to attain the best compromise insignal
modulation
system.. ’ ’
distortion. It is apparent, however, that nonuniform sampling
. .
can enhance thesignal following capability of an encoder.
An asynchronous delta modulation scheme [8] sends a
I. INTRODUCTION
pulse whenever the source signal crosses a threshold level: In
IFFERENTIAL pulse codemodulation (DPCM)’is a thisway more pulses are sent during fast changingintervals
D r e l a t i v e l y simple and practical data compression scheme. and fewer pulses during relatively quiescent intervals, i.e., the
Data compression takes place by virtue of the fact that the signal is sampled nonuniformly. If the quiescent intervals are
source signal is sampled at a uniformrate f, and that the long and the number of these long intervals are large, then the
samples are quantized .to a finite’ number of, say N , bits per average numberof pulses transmitted wouldbe relatively
source sampie. For a given value of N , the quantizer step size small. Yet, the transmitted pulses contain sufficient informa-
d is chosen so that the 2-tuple (N, d) yields a. “good” approx- tion to permit a fairly accurate reconstruction of the source
imation to the signal dynamics. The sampling and quantiza- signal.
tion processes incur information loss, the amount of which is While thenonuniform sampling feature of asynchronous
directly related to thevalues of N and f,. delta modulationpointstowardthe possibility of efficient
Since the source samples are separated by a fured time in- encoding, the continuous-valued time intervals between trans-
terval, theinformationabout a source is contained in the mitted.pulses do not offer compatibility withdigital transmis-
quantized levels, whichare determinedbythe’valuesof N sion or. digital information storage.
and d. The value of N determines the rate of data transmis- In.this paper we employthreshold sensing and digital
sion while that of d influences the signal distortion. Max [12] coding of the time intervals between level crossings to mech-
has described optimum’strategies for theselection of N and d. anize a nonuniform sampling approach to source coding.
The ability of a DPCM system to follow a source signal de- The nonuniform sampling encoded signal carries the informa-
pends o n ’ t h e quantizer size N (and hence d) and the sampling tionaboutthe source signal in the intervals between level
frequency:f,. Since the primary reason behind data compres- crossings and in the directions of level crossings. With a suit-
sion is to reduce the. transmission rateand/ordata storage able.symbol assignment, thenonuniform samplingencoder
requirement, one wouldlike ‘to make N and f, as small as (NSE) output symbols assume a ‘highly ,asymmetricproba-
possible, subject to an acceptable distortion tolerence. Small bility distribution. That is, the NSE output sequence is con-
ditioned to possess a specific form of redundancy which can
Paper approved by the Editor for Data Communication Systems of be removed byfurther distortionless encoding (DLE). This
the IEEE Communications Society for publication after presentation
at the 17th Allerton Conference03 Communication, Control, and Com- feature of the NSE is in direct contrast to the equally likely
puting,Urbana,IL,October1979.ManuscriptreceivedFebruary 4, probability distribution of the output symbols of an optimum
1980; revised June 16, 1980. This work was supported by the Natural delta ,modulator.(DM). Thus, while notmuchadditional
Sciences’and‘ Engineering Councilof Canada under Grant A7779 and a
Postgraduate Scholarship. compaction gain may be obtained by fuhherencodingthe
The authors are with the Department of Electrical Engineering, Un- output of a delta modulator, additional distortionless encoding
iversity of Waterloo, Waterloo, Ont., Canada. of the NSE output sequence is necessary in order to attain a
1 A delta modulation (DM) system can be viewed as a special case
of DPCM with N = 1 bit. In this paper reference made‘to DPCM will high overall compaction gain. Therefore,fortheencoding
also include DM as a special case. scheme described in this paper, it is meaningful only to talk

0090-6778/81/0100-0024$00.75 0 1981 IEEE

MARK AND TODD: NONUNIFORM
SAMPLING
APPROACH TO DATA COMPRESSION 25

of the compression ratio of a concatenated NSE/DLE system,

but not that of the NSE alone.
The paper is organized as follows. Section I1 describes the
nonuniform sampling encoding together with the chosen sym- I
I NON-UNIFORM
SAMPLING
ENCODER
(NSEII
bol assignment t o make the NSE a nonuniform sampling with L ______-________
synchronous transmission device. Section I11 discusses the in- Fig. 1. Nonuniform sampling encoder (NSE).
corporation of prediction t o formulate a nonuniform sampling
predictive coding (NSPC) scheme. A salient feature associated ward crossing. Thus, two symbols are needed t o represent the
with the NSPC scheme lies in that prediction is accomplished level crossings and a third symbol to represent the quantized
by a minimization of a certain function of the slope of the intervals ij. In the present study we make thefollowing binary
error signal rather than the error itself. Section IV proposes 2-tuple assignments:
a simply implementable run-length encoding (RLE) scheme to
operate in tandem with either the NSE or NSPC to attain a 00 - coding of ij

01 - upward crossing
large overall compression ratio. Analytic and simulation
results in the encoding of a band-limited Gaussian source and
a still black and white image source are given in Section V. 10 - downward crossing.

11. NONUNIFORM SAMPLING ENCODING For example, if the signal crosses a level at time ti in the up-
The motivation for the present proposed nonuniform sam- ward direction and ?j = mitq, then (mi - 1) pairs of 0’s are
pling encoding scheme comes from the consideration that, if used t o represent the length ofij and the 2-tuple01 is used to
informationaboutthe source signal is carriedin the times represent the upward crossing. Thebinaryrepresentation is
between level crossings. rather than in the quantized ampli- then
tude as in DPCM systems, the encoder possesses an ability to
follow the signal waveform even on a nonadaptive basis. The
nonuniform sampling problem is depicted in Fig. 1 where the
level crossing detector (LCD), a device consisting of a number
of uniformly spaced levels, is used to monitor the level cros-
It is seen thattwo binary digits are used to represent the
sings. A signal sample is taken whenever a level crossing takes
quantum step t q . Thus, the frequency with which two bits
place. If a pulse is generated and transmitted everytime a level
of information are used t o encode the time intervals is fq =
crossing occurs, the encoding procedureis called asynchronous
l/tq. If is large, the binary sequencegenerated by the non-
delta modulation [8].However, the time b,etween level cros-
uniform sampling encoder (NSE) has a high probability of03,
sings is a continuous variable, which is not suitable for digital
which is particularly suitable for further distortionless, such as
transmission or digitalstorage. Thepresentstudy considers
run-length, encoding.
quantizing the time between level crossings and encoding the
For the purpose of illustrating the signal encoding and re-
quantized time intervals in uniform steps. As a result, while
production, suppose the signal crosses level xi at time ti-
the sampling is nonuniform,theinformation transmission
and level xi+ = xi + d , where d is the amplitude step size of
is synchronous.
Let ti be the time instant at which the source signal crosses
the LCD, at time tj = ti- +
ri, with an upward motion. Then
rii = mjtq. Assuming mi = 4, the binary representation is then
a level and T~ = ti - ti- be the interval between the last level
00 IO0 IO0 101, where the three pairs of binary 0’s’ carry the
crossing at time ti- and the present crossing at time ti. Let
information about the length of i j and the 2-tuple01 indicates
tq be the time quantum so that the quantized time interval
the upward motion. Since ;ii is only an approxhation to rj,
is prescribed by
+
the reproduced sample has value equal xi d , but occurs at
time (ti f tq/2), as illustrated in Fig. 2.
+.I = ( Let f,be the cutoff frequency of the source signal and f, =
k(2fC), k 2 1, be the sampling frequencyunder a uniform
sampling scheme such as DPCM. With an N bit quantizer the
where transmission rate of a DPCM coder is N bitslsource sample,
where the samples are taken at the frequency f,. Let f q =
A
mi = of,, where 0 < 0. The transmission rate of the proposed non-
uniform sampling encoder, referred tothefrequency f,, is
20 bits/source sample.
and the symbol 1.1 denotes the integer smaller than or equal
In order to assign a common measure for both the uniform
to the argument. The error resulting from this time quantiza-
and nonuniform sampling encoding schemes we define a nor-
tion is bounded by &tq/2. The signal distortion due t o time
malized rate F, wherenormalization is with respect to the
quantizationfor aband-limited Gaussian signal is given in
signal cutoff frequency f ,and the subscript x denotes the kind
Section IV.
of encodingscheme. We thus have the definitions
The information about the source signal is contained in ri
and the direction of level crossing, i.e., an upward or a down- FDPCM & Nf,/f, bits/signal bandwidth (bpsb) (3)
26 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 1 , JANUARY 1981

I Jdt I -&I

Fig. 2. Illustration of sample reconstruction (x = reconstructed

point).
Fig. 3. Nonuniform sampling predictive coding (NSPC).
and
where

= 2Dfs/fc bits/signal bandwidth

(bpsb) (4)
wherethesubscript NS denotesthenonuniformsampling
scheme under consideration, in which two bits are sent at the
frequency f q . Since delta modulation can be viewed as DPCM
y(t) = i ’ ( t ) + Z(t)
with N = 1, it follows that
F D M ’ g fs/fc bits/signal bandwidth (bpsb). (5) and { T ~ are
} the timeintervals between adjacent level crossings,
counting backwards from time t. The predictor takes the form
of a tapped-delay line filter with nonuniform tap separations.
111. NONUNIFORM SAMPLING PREDICTIVE CODING Since rj is a random variable, the tap separations are notfxed
The encoding procedure depicted in Fig. 1 is analogous to and the mechanization of a kth-order predictor is nontrivial.
PCM coding in the sense that no prior correlationremoval has Forour NSPC system weuse asimpleone-steppredictor.
beencarriedout.Unlike PCM inwhichtheinformation is Since theslope e’(tj) of the signal e(t) when it crossesthe
carried in the amplitude, .the NSE carries the information in threshold level at time f j is-of interest, the one-step predictor
the times betweenlevel crossings. If the mean interval between at time instantt j is prescribed by
level crossings is long the probability distribution of the NSE
output .binarysequence will
be highlyasymmetricwith
P(0) 9 P(l).. The distortionless encoding stage of an overall
data compression system can offer a high compression ratio.
Therefore,forsources,such as video images, whichexhibit
short fast changing intervalsseparatedbylongrelatively
quiescent intervals, the tandem connection of the NSE and an
entropy coder .can offer veTy good rate-distortion character-
istics. Thepredictorcoefficient g(tj) can beoptimizedby solving
The average length of time between level crossings can be minimizationproblem (6) using a successive pointsteepest
increased by a decrease in the slope of the source signal. This descent method [ 1 5 ] . Thisresults in the following iterative
observation led us to propose the nonuniform sampling pre- equation for computingg(tj):
dictive coding (NSPC) scheme shown in Fig. 3. The NSE per-
formsananalog-to-digitalcoding.The device ( N E - ’ per-
forms a pseudodigital-to-analog conversion. With a knowledge
ofhowasource signalis nonuniformlysampled itmaybe where G is the gradient given by
possible toreconstructthesourcefromthenonuniform
samples [13] . Sincethenonuniformsamplingundercon-
sideration is driven by the source waveform, a technique for
exactreproductionofthe originalsource signal -fromthe
nonuniform samples has hitherto eluded us. For the present
we propose a simple technique of joining the reconstructed and is obtained by solving the equation
samples by straightlines and low-pass filtering the resultant
piecewise linear approximationtoobtainthesmootherror a~{Te’(t)~ Ig(rj+ 1’)
- = 0,
approximation Z(t) (seeFig. 3). . Thepredictorwithcoef- W j
ficients ( g i } i , canbeoptimizedby solving aminimization
problem of the form [ 6 ] i.e., is selected to cause E{ [e’(t)] I g ( r . + l ) } to be amini-
mum. Execution of iterative algorithm (9j begins by assuming
an initial value for &to) and then compute G and /3j by (10)
and (1 1).
MARK AND TODD: NONUNIFORM SAMPLING APPROACH TO DATA COMPRESSION 27

IV. RUN-LENGTH ENCODING

Compared tomore efficientschemes such as Huffman
coding, run-length encoding is suboptimum. Yet, the symbol
representation at the output of either the NSE or the NSPC
lends itself to simple run-length encoding. Moreover, since run-
length is a variable length-block length encoding scheme, the
encodedsymbolsaregeneratedsynchronously.Thus,unlike
the variable-length-to-variable-lengthencoding schemes in
which buffering is required at both the input and the output
of the encoder,run-length needs buffering at the input only.
Because of the manner in which the information is repre-
b-' fj _I
Ti c
sented at the NSE or NSPC output, the following simple run-
length encodingscheme, illustrated for the case n = 3, is a Fig. 4. Time quantization error model.
natural choice:
It is noted that the NSE digitization and run-length encoding
run block length n = 3 can be implemented quiteeasily using a digital counter.
01 000 V. ANALYSIS AND SIMULATION RESULTS
10 001 In this section we present results obtained in the encoding
000 1 010 of a band-limited Gaussian source and a black and white still
image.
0010 01 1
00000 1 100 A . Encoding of a Band-limited Gaussian Source
0000 10 101 I ) Approximate Analysis of Time Quantization Noise:
Consider the level crossing problem depictedin Fig. 4. The
000000 110 timebetween level crossings at times and ti is rj = ti -
unused 111. t i - 1 . Let Fi be- thequantizedapproximationcomputedby
(1). The error is then AT^ = rj - i j with 1 AT, I < t q / 2 . The
The following assumptions are needed for theensuing analysis. reconstructed sample will be the point D instead of B. The
Assumption 1: The upward crossing frequency equals the amplitude error is ei = D - C. If we approximate the segment
downward crossing frequency. between C and B by astraightline, thenfromgeometry,
Assumption 2: The NSE or NSPC output sequence is in- we have
dependentlyand identically distributed (i.i.d.).By Assump-
tion 1,we have d
ej = - AT;.
P(0 1) = P( 10) = f (1 -P(00)).

Let p = P(O0). By Assumption 2 , the mean run-length is given

by Assumption 3: x'(t) and Ari are statistically independent.
1- p 2 " - 1 - 1 Assumption 4: {AT/} is zero mean and uniformly distrib-
I= symbols. uted in the interval ( - t q / 2 , t q / 2 ) .
1- P With the above assumptions, we deduce that
Since there are two bits per symbol, then
1 -p 2 " - 1- 1
1=2 bits. By Assumption 4 and lettingP(A7i) be the probability density
1- P of Ari, we have
The compaction ratio(CR) offered by run-length encoding is
tqP
E { [ AT^] 2 } = (ATi)2 P(ATj)d(ATj) = tq2/12 (15)
CR = i/n - tqI2

(12) On theassumptionthat x'(t) haszero mean,then u,r2 =

E{ [x'(t)]'}. It follows that

which has a limitingvalue

oe2 = E [ e i 2 ] = u x r 2 - tq2/12.
(13) It is recognized thatspectraof image signals have' gradual
rolloffs at the cutoff frequency and that a Gaussian model is
28 IEEE TRANSACTIONS
COMMUNICATIONS,
ON VOL.NO.
COM-29, 1, JANUARY 1981

not so realisticmonochrome
for imagery. To facilitate
a level f, 1 fc
crossing analysis, weband-limited
consider
a Gaussian source 0 20 60 80 100
with spectral density I
-m I I I
40
I I I I I I I
'0

A IflGfc
otherwise.

The analysis in the sequel based on this Gaussian model,

while not too realistic, serves as a useful aid in the selection
of a time quantum for time interval quantization. Since ux2 = I-z
Jf-cf,.Sy> df, we 'have A = ux2/(2fc). The variance ofthe 3
0
-30
slope ii then w

I- fc -401
(JxJ = Fig. 5 . Time quantization noise characteristics for the encoding of a
band-limited Gaussian source.

= (27rfc)2 0 3 3 . \

Substituting (1 7) in (16 ) we have t"

where fq = 1/tq.The normalized distortion is thus given by

D oe2/ux2 = [nfc/(3fq)I
= 0.40 - 20 log1 o(fq/fd) dB
= 6.40 - 20 log1 0 F N S dB (18) Fig. 6 .
-
Geometric illustration of level crossing model at level xo.

where F N S is as defined by (4). A plot of D as a function

sings, denoted byr, is
of FNs is shown in Fig. 5.
2)MeanTimeBetween Level Crossings: Following an m

argument similar to Blake and Lindsey [14] , we use a first- r = V(0) + 2 2

n=l
V ( 4
orderapproximationdepicted in Fig. 6 , where we examine
the process in the proximity of x = x o . The expected number
of level crossings per unit time in thenegative direction is

Pr [xo < x < x o +dx,--<x' <-I

4x0) = where we have used theproperty V(-Xo) = V(Xo) for a
-dt
Gaussian process. For d i n the range

d
0.01 < -< 0.50
(JX
The total frequency of
crossings of this level is
the frequency oflevel crossings can be approximated by

For a Gaussian source the slope and the amplitude at a given

instant arestatistically independent.It can thenbe easily Hence, the mean time betweencrossings is
shown that

3) Selection of Time Quantum tq: The.maximum slope

With x = n - d ; n = 0,+1, +2, e-, the total frequency of cros- ;,,s that can be accommodated bythe NSE without ex-
MARK AND TODD: NONUNIFORM SAMPLING
APPROACH TO DATA COMPRESSION 29

periencing overload is 0 ERROR INNSEENCODER

o ERROR IN NSPC ENCODER
- SAMPLE FUNCTION OF BANDLIMITED
GAUSSIAN SOURCE

For the encoder to have an overload probability of less than

0
4X we must have [9]

40,t <Smax. (23) -I

Forthe band-limited Gaussian source, wehave previously -

shown that Fig. 7. A sample of reconstruction errorsvia simulation.

0,J = 4 n 2fc 2 0, 2 . (24)

NSPC
NSE
Using (22) and (24) in (23), we obtain the inequality
a OPTIMIZED

40 60 80 100 120 140 160 180

from which we have
Fx
Fig. 8. Signal-to-noiseratio performance in the encoding of aband-
limited Gaussian source (without run-length encoding).

The simulation results forthe NSE and NSPC encoders

4 ) Numerical Results: We have carried out asimulation are plotted in Fig. 8 as a functionof FNS= 2fq/fc. Also
studyofthe encodingofaband-limited Gaussian source, plotted is the result of an optimized linear delta modulator
(LDM), i.e., thestep sizeis optimized foreach value of F D M =
which was generated using a 64th-order transversal filter to
fs/fc. It is observed that the nonuniform sampling approach
give the resultant band-limited Gaussian source the following
spectral characteristics: a) passband ripple = k 0.136 percent; offers superior performanceto the optimizedLDM.
b) minimum stopband attenuation = -57.4 dB; c) fc = 0.12 B. Image Coding
at the 3 dB point; and d) a, = 1. With d = 0.25, we require
One of the salientfeatures of the proposed nonuniform
t, < 0.1436 or fq/fc > 58.04. With tq = 0.125 we encoded
the band-limited Gaussian source using NSE and NSPC.A
samplingscheme was the contention that an NSE or NSPC
system could capitalize onthe regions oflowdetail in an
segment ofthe band-limited Gaussian process andthe as-
image source to produceanoutput sequence with a large
sociated errors are shown in Fig. 7. It is observed that the
errors resulting from the NSE and NSPC are comparable; the value of P(OO), so that significant compaction gain canbe
difference lies in that -?;rlSpc > ‘ ~ N S E so that the coded binary obtainedfromadditional run-length orentropy encoding.
On the other hand, the preferred state of an adaptive delta
sequence at the output of the NSPC has a higher p = P(00).
modulation (ADM) system tends t o produce an output se-
Forthe band-limited Gaussian source, we previously
obtained quencewith P(1)P(-l), so thatonecannotexpectany
significant gain fromfurtherentropy coding.Encoding of
“IMAGE” shown in Fig. 10(a) using an ADM with an adapta-
?NSE = - (2 1) tion constant y = 2 [ l l ] reveals that P(1) el‘-1). Also, the
frequency of occurrence of runs of lengths “1” dominates,
as showninTable I. Because of theseproperties, we con-
With the above values of d, a,, and f,, we have ~ N S E= jecture that not much can be gained from entropy encoding
0.72. It follows that % = ? N s ~ / t , = 5.8 andP(O0) = 0.85. the ADM output. In the case of encoding with either NSE or
With n = 5 and p = 0.85 the compression ratio calculated NSPC, the quantum size d , the sampling rate FNS,and the
by (12) is then CR e 2.4. run-length code block size n are preselected. Results obtained
In the simulation we computethe signal-to-noise ratio in the encoding of “IMAGE” are tabulated in Table 11. In the
by the following formula: NSPC case, we let &ti) = 1 [see (7)]. The overall transmission
rate R is thatoftheconcatenated NSE/RLE(NSPC/RLE)
encoder given by

1
R =F N S* - bits/signal bandwidth.
2
j= 1
(Xj -Xj*)2 CR

Since FNS= 2f,/fc, an increase in F N S implies an increase

where xi* denotes the jth reconstructedsample. in f, or adecreasein t,. A smaller value of t, wilI be ac-
 30 IEEETRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 1, JANUARY 1981

TABLE I
DISTRIBUTION OF THE RUN-LENGTHS OF AN ADM (y = 2)
OUTPUT SEQUENCE IN IMAGE CODING
UJN-LENGTH FREQUENCY OF OCCURRENCE
(+1 and -1)
FDM = 8 bpsb FDM = 16 bpsb

1 0,529 0.531

2 0.187 0.187

3 0.163 0.164

‘ 4 0.070 0.065

5 0.024 0.026

6 0.014 0.010

7 0.004 0.007

8 0.001 0.003

9 0.001 0.002

10 0.001 0.001

11 0.001 < 0.001

12 c 0.001

TABLE I1
CONCATENATED NONUNIFORM SAMPLINGIRUN-LENGTH
ENCODING OF “IMAGE” (PEAK-PEAK SIGNAL HEIGHT
NORMALIZED TO 1.0)
OVERALL RATE (bpsb)
. . 1

i:
SIM CAL
4.1 4.00 4 . 2 4

9.5 6.686.32

5.5 7 . 5 6 7.28

companied by a larger number of “00” and, therefore, a larger

value of P(O0). This.wil1 result in a larger,compaction gain in
run-length encoding. The selection of I, and n should be such
as to make both R andthetimequantization noise small.
As this selection of t, and n for a given d is rather ad hoc, the
36 i .i
NSPC

results shown in Table I1 are not necessarily the best. The sig- 32.
nal-to-noise ratio.performance,as a functionofthe overall
transmission rate for the encoding of “IMAGE” is plotted in
Fig. 9. It is noted that the NSE/RLE and NSPC/RLE systems 28.
outperform theADM system.’
Fig. lO(b)-(f) shows thereconstructed “IMAGE” at two 24.
different, rates. To compare performance betyeen the ADM
and NSE/RLE or NSPC/RLE systems,a number of salient
20.
characteristics are of specific interest: 1) for a fwed value of d,
the quantization noise corresponding to a given code rate is
bounded within the quantum interval (-d/2, d / 2 ) ;2) during a
quiescent interval when the signal fluctuates within aninterval
t .
( - d / 2 , d/2), the signal distortionexhibitedbythe NSE or
NSPC would be at its worst condition; and 3) an ADM system 1 I I I I I I I
is capable of adapting itself to yield a very small granular noise 8 16 24 32 40 48 56
RATE (bits/source bondwidth 1, R
Fig. 9. Signal-to-noise ratio performance in the encoding o f “IMAGE.”
2 Huffman encoding the ADM output sequence offersnocom-
paction gain. The rate associated with the ADM curve in Fig. 9 is just
FADM,
MARK
TODD:
AND NONUNIFORM
SAMPLING
APPROACH
TO DATA COMPRESSION 31

following attributes: 1) a graceful reproduction of thelow

activity regions, such as the subject’s forehead and face, and
2 ) the degree with which the sharp edges (e&, sharp black-to-
white transitions) of the high activity regions, such as the sub-
ject’s eyeball, eyebrow,transitionsfrom forehead and back-
ground to thesubject’s hair, can be reproduced.
A comparison of the original image [Fig. lO(a)] with the 8
bpsb ADM encoded image [Fig. 10(b)] reveals that the fore-
head exhibits a relatively graceful reproduction. However, the
8 bpsb ADM is unable to preserve the sharp edges as char-
acterized by the fuzzy transitions from the subject’s 1) fore-
head to theeyebrow, 2) white part ofthe eyeball tothe
black part of the eyeball, and 3) forehead and background t o
the subject’s hair.On theotherhand,the 8 bpsb NSE or
NSPC reproduces the above mentionedsharp edges with a
high degree of fidelity. At the same time the 8 bpsb NSE or
NSPC yields agraceful reproductionofthe lowactivity
regions. It is noted that 4 bpsb NSE or NSPC also preserves
sharp contrasts, but is unable t o maintain a graceful degrada-
tion in the low activity regions. Comparing at the same rate,
NSE or NSPC exhibits superior performance t o ADM, both in
terms of signal-to-noise ratio (Fig. 9) and subjective compar-
ison (Fig. 10).

VI. CONCLUDING REMARKS

Fig. 10. Subjective comparisons in the encoding of “IMAGE.” (a)
Original “IMAGE.” (b) 8 bitslsignal bandwidth ADM. (c) 4 bits/sig-
nal bandwidth NSE. (d) 4 bitslsignal bandwidth NSPC. (e) 8 bits/ We have describeda nonuniform sampling approach to
signal bandwidth NSE. (f) 8 bits/signal bandwidth NSPC. source coding, either with or withoutprediction. The informa-
tion about the source signal is contained in the times between
level crossings, which are codedinto a highly asymmetric
when the quiescent interval is long. This condition corresponds ( 0 , 1) binary sequence withP(0) SP(1). The output from the
to a very small stepsize. When a large change in slope suddenly nonuniform sampling encoder (NSE) is inasuitable format
occurs, however, the ADM tends t o experience a long recovery to permit further simple run-lengthencoding (or entropy
period, during which slope overloadnoise can be ~ignificant.~ coding) toattain a“good” signal-to-noise ratio versus rate
Oneof
the main
drawbacks withuniform sampling characteristic.
schemes, such as ADM, lies in that an equal number of samples Results obtained in the encoding aofband-limited Gaussian
is taken from the low activity regions (relatively quiescent re- source reveal thatthe NSE or NSPC exhibitsSNR vs rate
gions where the slope is small) and high activity regions (sharp characteristicssuperior tothatof an optimized LDM. Of
changes from a relatively light gray level t o a relatively d x k particular significance are the results obtained in image coding,
level, i.e., large slopeconditions). To minimize the effect of which reveal that concatenated NSE/RLE or NSPClRLE en-
slopeoverload, an ADM has t o sample the source at a rate coding is significantly better than adaptive delta modulation
much higher than the Nyquist rate. The proposed nonuniform (ADM) encoding both in terms of SNR versus rate charac-
sampling scheme selectively samples the source, so that more teristics and subjective comparisons.
samples are taken in the high activity regions. Consider the The present studyhas not considered the possibility of non-
image shown in Fig. 10. The subject’s forehead is a low activ- uniformthreshold levels in the generation ofnonuniform
ity region. The transition from the forehead to the eyebrow is samples. Also, it is possible to combine the digital representa-
a high activity region. Also, the transitions between the white tion of the NSE output with the entropy coding stage, e.g., it
part and the blackpartof the subject’seyeballrepresentsa may be’ possible to Huffman encode the {mi} with a prefm
high activity region, while within the black or white region, we code, where the prefut is used toindicatethedirectionof
have a low activity scenery. threshold crossing.
It is in the high activity regions that ADM experiences re- Although the NSE encoder generates an output sequence
construction problems, unless the sampling rate is unduly drawn from an alphabet of three symbols, i.e., 00, 01, 10; it
high. For the purpose of subjective comparison between the is not the same as a three-level delta modulation encoder. As
performanceof ADM and NSE or NSPC,we lookforthe the sampling rate of the NSE is a function of theslope of the
input waveform, it is a waveform following encoder without
feedback.
3 In the image coding with ADM we set a lower bound on the
step size. This lower bound was chosen by trial and error to yield a best The numerical results of Table I1 (last two columns) reveal
compromise between slope overload and granular noise. that the simpler NSE encoder performs as well as, or better
32 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 1 , JANUARY 1981

than, the NSPC encoder when the source is relatively quies- Jon W. Mark (S’6&M’62-SM’80) received the
cent, as is the case of “IMAGE.” B.A.Sc. degree in electrical engineering fromthe
University of Toronto, Toronto, Ont., Canada, in
REFERENCES 1962andtheM.Eng.andPh.D.degrees in
electrical
engineering
from
McMaster Uni-
1 J.B.O’Neal,Jr.,“Predictivequantizingsystems(differential -versity,Hamilton,Ont.,Canada, in 1968and
pulse-code modulation) for the transmission of television signals,” 1970, respectively.
Bell Syst. Tech. J . , vol. 45, pp. 689-721, May-June 1966. From 1962to1968 he waswithCanadian
N. S . Jayant, “Digital coding of speech waveforms: PCM, DPCM, WestinghouseCompany,Ltd.,Hamilton,Ont.,
andDMquantizers,” Proc.IEEE, vol.62,pp.611-632, May Canada,where he wasengagedinthearea of
1974. underwateracousticcommunicationandsonar
J. E. Abate, “Linear and adaptive delta modulation, Proc. IEEE, signalprocessing. He is continuingasaConsultanttoWestinghouse
vol. 5 5 , pp. 298-308, Mar. 1967. Canada, Ltd. From September 1968 to August 1970he was an NRC PIER
N. S . Jayant, “Adaptive delta modulation with a one-bit memory, Fellow at McMaster University, on leave from Canadian Westinghouse
Bell Syst. Tech. J . , vol. 49, pp. 321-342, Mar. 1970. Company, Ltd., studying toward the Ph.D. degree. In September 1970 he
L.H.ZetterbergandJ.Uddenfeldt,“Adaptivedeltamodulation joined the Department of Electrical Engineering, Universityof Waterloo,
with delayed decision,” IEEE Trans. Commun., vol. COM-22, pp. Waterloo, Ont., Canada, where he is currently a Professor and Associate
1195-1 198, Sept. 1974. ChairmanforGraduateStudies. He wasDirectoroftheComputer
W.
J.Mark,“Adaptivepredictive
run-lengthencoding
for CommunicationsNetworksGroupoftheUniversity of Waterloofrom
analoguesources,” Proc. Inst. E l m .E n g . , vol.123,pp. 1189-’ January 1976 to August 1976. From September 1976 to August 1977, he
1196,Nov.1976. was on sabbatical at the IBM Thomas J. Watson Research Center. He had
T.A.HawkesandP. A. Simonpieri, “Signalcoding using previouslyworked in theareas of radar/sonarsignalprocessingand
asynchronousdeltamodulation,” IEEE Trans. Comrnun., vol. adaptive equalization. His current research interests include data com-
COM-22, pp. 346-348, Mar.1974. pression, image and facsimile coding, and computer communication using
H. h o s e T, . Aoki,and K. Watanabe, “Asynchronous delta satellites and terrestrial store-and-forward networks.
modulation systems,’’ Electron. Commun. Japan, pp. 3 4 4 2 , Mar. Dr. Mark is a member of the IEEE Communications Society Computer
1966. Communication Technical Committee and was Secretary of that Com-
R.Steele, DeltaModulationSystems. Bristol,England:Arrow- mittee from June 1976 to June 1979.
smith,1975
E.N.Protonotario,“Slopeoverloadnoise in differentialpulse
code modulation,” IEEE Trans. Inform. Theory, vol.IT-20,.pp.
58-64,Jan.1974.
*
L.H.Goldsteinand B . Liu,“Quantizationnoise in ADPCM
systems,” IEEE Trans. Common., vol.COM-25,pp. 227-238, Terence D. Todd (S’78) was born in Galt, Ont.,
Feb.1977. Canada. He receivedtheB.A.Sc.andM.A.Sc.
J . Max, “Quantizing for minimum distortion,” IRE Trans. Inform. degrees in electricalengineeringfromtheUni-
Theory, vol. IT-6, pp. 7-12, Mar.1960. versity of Waterloo, Waterloo, Ont., Canada, in
L.
J.Yen,
“Onnonuniformsamplingof
bandwidth-limited 1978 and 1979, respectively.
signals,” IRE Trans. Circuit Theory, pp. 251-257, Dec. 1956. Currently; he isemployed by theComputer
I . F.Blakeand W.C. Lindsey“Level-crossingproblemsfor CommunicationsNetworksGroupatthe Uni-
random processes,” IEEE Trans. Inform. Theory, vol. IT-19, pp. versity of Waterloo where he is working on local
295-3 15. May 1973. areacommunicationnetworks. He is presently
C, McMillan, Jr., Mathematical Programming. New York: studying towards the Ph.D. degree in electrical
Wiley, 1970, ch. 5 . engineering on a part-time basis.