0% found this document useful (0 votes)
57 views20 pagesPredictive Coding of Images
This paper presents a predictive coding method for images that utilizes both adaptive intra-frame and motion compensation predictors. The adaptive predictors are designed to minimize prediction error by continuously adjusting weight coefficients based on the performance of each mode, without the need for overhead information transmission. Additionally, the paper addresses the sensitivity of predictive coding to transmission errors and proposes a self-adjustment method for the decoder to maintain synchronization.
Uploaded by
Staffan NyströmCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
57 views20 pagesPredictive Coding of Images
This paper presents a predictive coding method for images that utilizes both adaptive intra-frame and motion compensation predictors. The adaptive predictors are designed to minimize prediction error by continuously adjusting weight coefficients based on the performance of each mode, without the need for overhead information transmission. Additionally, the paper addresses the sensitivity of predictive coding to transmission errors and proposes a self-adjustment method for the decoder to maintain synchronization.
Uploaded by
Staffan NyströmCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 20
Signal Processing: Image Communication 4 (1992) 457-476 437
Elsevier
Predictive coding of images using an adaptive intra-frame
predictor and motion compensation
G. Tziritas and J.C. Pesquet
Laboratoire des Signaux et Systemes (C.N.R.S.), Ecole Supérieure d'Electricité, Plateww de Moulon,
91192 Gif-sur- Yuerte Cedex, France
Received 7 May 1990
Revised 3 July 1991
Abstract. In this article we present an image predictive coding method using both intra- and inter-frame predictors. The intra-
frame predictor is an adaptive FIR filter using the well-known LMS algorithm to track continuously spatial local characteristies
of the intensity. The inter-frame predictor is motion-adaprtive using a pelrecursive method estimating the displacement vector.
‘Weight coefficients are continuously adapted in order to favor the prediction mode wich performs better between intra-frame
‘and motion compensation mode, I is « backwards adaptation which does not necessitate the transmission of an overhead
information. Neither the weight coefficients nor displacement vectors are transmitted. Apart from the quantized predietion
error, it may be necessary to transmit the detection of a discontinuity of the displacement vector. For the examined imaze
sequence significant improvement is obtained in comparison with only adaptive intra-frame or only motion compensation
mode, We give and discuss the extension of a known adaptive quantize for 2D signals, A crucial problem in predictive coding,
particularly with adaptive techniques, is the sensitivity to transmission errors. A method ensuring the self-adjustment of the
decoder in the presence of transmission errors, which do not affect the pixel synchronization, is proposed for the intraframe
mode, Neither overhead information nor error-correcting codes are needed.
Keywords. Predictive image coding: adaptive prediction; motion compensation; adaptive quantization; channel errors,
1, Introduction
Among different image coding methodologies, predictive coding can be simply implemented and produces
good results at higher rates. In this paper, we are interested in adaptive methods of predictive coding. The
block diagram of the predictive coder /decoder discussed in this paper is shown in Fig. 1. It contains an
adaptive predictor and an adaptive quantizer, the adaptation being realized using the quantized prediction
error 2, which is the transmitted information only. In Fig. 1, x is the original signal, X is the signal
reconstructed by the decoder, ¢ is the predicted signal and e is the prediction error. As a result of eventual
transmission errors the received prediction error may be different, ¢’, and similarly will be concerning the
predicted, $', and the reconstructed, 3", signals.
Knowing that there exist simultaneously spatial and temporal redundancies, we consider a hybrid struc-
ture of intra- and inter-frame prediction (Fig. 2). The spatial part of the predictor is a linear filter, Given
that the statistical characteristics of the luminance vary spatially and that this signal is characterized by
many nonstationarities, the predictive filter must be space-varying. An approach which permits tracking of
the variations is an adaptive filter. A method of adapting the 2D filter consists of switching between different
filters, switching based on a classification rule. Sometimes this technique requires the transmission of
(0923-5965 /92/S05.00 ©; 1992 Elsevier §
fence Publishers B.V. All rights reserved 458 G. Triritas, IC. Pesquet { Predictive coding of images
+ n rs
a5 Quantizer_ |S} Encoder
XK
Paco |
¥ = 7
Channel
+.
ef ale
[aca]
Net
Fig. 1. The structure of the adaptive predictive coder/decoder,
Fig. 2. An adaptive intra-int
frame predictor.
overhead information, i.e. the class of each pixel. Another adaptive method is to identify the optimum
linear filter for a block of the frame using a least squares criterion [11]. The coefficients of the filter must
also be transmitted in this case. To avoid the transmission of additional information we here use a continu-
ously adapted filter, the adaptation heing based on previously reconstructed pels. The adaptive algorithm
used is that of Least Mean Squares (LMS) first introduced by Widrow. The temporal part of the predictor
is adapted using an estimator of the displacement vector. A pel-recursive estimator is used, similar to the
one of Walker and Rao [17] and of Cafforio and Rocca [3]. We discuss the displacement estimator in more
detail in Section 2.
Thus we have an adaptive filter for spatial prediction and a motion adaptive inter-frame prediction. It
is obvious that in some regions one type of the described predictions performs better than another, in the
‘Spl Pressing; are Comma G. Teinitas, J.C. Pesquet | Predictive coding of images 459
sense that the prediction error is minimized. We must then adapt between these two types of prediction.
We here propose to continuously adapt a weight coefficient for the inter-frame predictor and the coefficients
of a traversal intra-frame filter. The predictor can then be written as
Tonns)= atk, )Im—k,n=1; 1) + bln —6, n= a; 1-1), a)
ides
where 7 is the luminance or the intensity, J is the reconstructed intensity, m (respectively n) is the row
(respectively column) index, r is the temporal coordinate, S'is a quarter-plane or a nonsymmetric half-
plane domain, # (respectively 6) is the horizontal (respectively vertical) component of the estimated displace-
ment vector. The coefficients of the spatial filter {a(k, /)} and the weight coefficient 6 are continuously
adapted using the LMS algorithm.
The prediction error which is quantized and transmitted is not stationary and its probability distribution
is unknown. A reduction of the distortion is obtained, if the quantizer is not fixed, but adapted to the
statistics of the prediction error. An extension of known techniques in 1D predictive coding is given in this
paper for a scalar three-level quantizer,
It is known that predictive coding is sensitive to transmission errors, even with constant length codes.
The sensitivity to transmission errors is greater if the predictor is adaptive and this adaptation is based on
the prediction error. A topic also studied in this paper is the adjustment of the decoder in the presence of
transmission errors. We only consider errors which do not affect the pixel synchronization and we demon-
strate that the self-adjustment of the decoder can be obtained using some regularization constraints.
The organization of the paper is as follows. In Section 2 the algorithm of motion estimation is briefly
presented and discussed. In Section 3 the algorithm of adaptive updating of the coefficients is given. In
Section 4 the adaptive quantizer used in this paper is presented and discussed. Section 5 presents constraints
and modifications used to obtain the self-adjustment of the decoder in the case of transmission errors.
Section 6 gives some results of the algorithms proposed in this paper to illustrate and evaluate their
performances.
2. Motion compensation
The displacement vector is estimated for each pixel. The algorithm used here to estimate the displacement
vector is similar to that of Walker and Rao [17] and Cafforio and Rocca [3]. It is a pel-recursive intensity-
based algorithm which is presented below. The estimator is composed of three parts. The first part is an
a priori estimator using the estimated values of the displacement vector in the causal neighborhood of the
current pixel. The second part is a detector of discontinuities on the velocity ficld. The third part is an a
posteriori estimator based on the minimization of the Displaced Frame Difference (DFD)
2.1. A priori estimation of the displacement vector
The a priori estimator used here is based on the hypothesis that the displacement vector is slowly varying.
Therefore it is assumed that the motion is locally translatory. The simplest a priori estimation (u’, 0°) at
pixel (nt, n) under this assumption, knowing the already estimated values (ii, 6) of the displacement vector,
is obtained from the previous pixel in the raster scan order:
fy foe
[fom ]_ semen ?| for n>0, 2
o'(m.n) | L6(m,n—1)
Vol. 4 N9, 6, November 1992 400 G. Tairitas, J.C. Pesquet / Predictive coding of images
and for the first column:
[pono | fern a cy
v%(m, 0) Len 1, 0)
2.2. Detection of a discontinuity
As the estimation method is based on the intensity using local measurements, the algorithm may estimate
false displacements. Discontinuities on the real velocity vector field also exist, because in natural scenes
independent 3D rigid movements, or edges between diflerent surfaces which are subject to the same 3D
rigid motion may cxist. An efficient estimate of the velocity vector necessitates the joint detection of all
these discontinuities. This detection must be based on the correctness of the a priori estimation of the
displacement vector and consequently on the prediction error of the intensity. The prediction error based
on the displacement estimation is the displaced frame difference given in
em, n) = Hanns = Kans 0 n= t= 1). QB)
Two approaches, which are described below, can be considered.
(A) The test of discontinuity is realized at the current pixel using the current a priori DFD. If a disconti-
nuity is detected, a reset code is transmitted to set the a priori estimation to zero. To determine the test we
assume that (m,n; £), knowing the displacement vector, is a Laplacian (or Gaussian) distributed variable
with mean Im~ v,n—u; ¢~ 1) and standard deviation & (=hypothesis 0: no discontinuity). Under hypo-
thesis of discontinuity /(m, 1; 1) is assumed Laplacian-distributed (or respectively Gaussian) with mean
Hm, n;1~ 1) and the same standard deviation. The test is determined by maximizing the probability «
posteriori, We suppose known probabilities of discontinuity (p;) und no discontinuity (p)> pi). Some
calculations give the following test: in the Laplacian case a discontinuity is detected if
D |= [lon ~ tom, n; 1 Y)>F in”, (4)
v2 Pi
Mon, n; 1)—I(m—v", n= a":
and in the Gaussian case if
Po
pL
(lon, m1) Tan= v", n=? t= 1) (em, m2) — Hom, ni t= 2)? > 20? In (5)
(2) The test of discontinuity is realized in a causal neighborhood Sy of the current pixel. If'a discontinuity
is detected, the a priori estimation is set to zero, but there is no need to transmit a reset code. Using similar
assumptions as in case (1), we obtain the following test: a discontinuity is detected in the Laplacian
case if
DY Von—k,n-t ) Im kv, n—1-w; t-1)|— an —k, nF, Mm k, nF t- 1)
hess
card[Sa} tn (6)
v2 A
Signal Processing: age Cummmuniation G Triritas, J.C. Pesquet / Predictive coding of tmages 461
card[ Ss] being the cardinal of set of pixels Sy, and in the Gaussian case if
LY don-kyn- lh )-Ton-k 08 nw; 1-1)
tires
= (Km-k, n~1; t)— (mk, n= 1; t-1)) > 20? card[Sa] in. (7)
2.3. A posteriori estimation of the displacement vector
Independently of the detection of a discontinuity the displacement vector must be updated at each point.
The criterion which is optimized is the square of the a posteriori displaced frame difference:
era, n) = Hm, 3 )— Hon v nuit VD) (8)
under some regularization constraints, Finally the following quadratic form must be minimized:
lu, v) = 2 m, n) + Af (uw + (v— 9)
where A is a regularization constant measuring the confidence on the a priori estimation in connection with
the estimation error. In reality this criterion is not directly quadratic on the unknown parameters w and o.
A first order development is admitted in order to obtain the linearization of e(m, 1),
em, n) = e%(m, n) + gu) + ¢,(0- 0"), (0)
where g,=J,(m—v°, n—u?; t— 1) and g,=1,(m—v°, n—u"; 1-1) are the horizontal and the vertical gradi-
ents of the intensity. Using this approximation we obtain the following solution:
|e 2» ee n) + (m,n) a ay
dian) Lean) aAteitesle,
In practice the reconstructed image intensity is used to determine g, and g,, in order to have exactly the
same estimator at the decoder. For the same reason the DFD is calculated using the reconstructed intensity.
24. Convergence properties of the algorithm
Another way to interpret (11) is that of recursive stochastic adaptive algorithms, which have interesting
performance in tracking the eventually changing displacement vector. The algorithm based on (2) and (11),
without the discontinuity detection, is known as the Normalized Least Mean Square algorithm (NLMS).
In [2] it is proven that thc NLMS algorithm according to (2) and (11) is almost surely and mean square
exponentially convergent, if {g., gy} constitutes
{a) a stationary ergodic process having positive-definite covariance, or
(b) a nonstationary g-mixing process with mixing constants {9,,.) satisfying
LL VOnn< 462 G Toirtas, J.C. Posquet { Predictive coding of images
and
fe “ fef Ronn) Kann) n)
2 VLi.Gn, i,m») Pn, n)
Mace MN nino nwo
being positive-definite (E being the notation for the average). Let us remind that a g-mixing process is one
for which the distance future is weakly dependent upon the present and that the mixing constants measure
this dependence.
This convergence property means that, if the real 2D motion is translatory, and the linear model according
to (10) is valid, then the algorithm according to (2) and (11) converges to the real displacement vector. It
also means that, if the real displacement vector is varying, the algorithm according to (2) and (11) converges
toa vector which ensures the minimization of the DED.
3. Adaptative intra—inter-frame predic
Two methods can be used in order to adapt the predictor to the spatiotemporal characteristics of the
signal. The first uses a forward segmentation where the best predictor is selected from a set of predictors,
For example, an inter-frame predictor may be better for parts of the frame without motion or with slow
motion, and an intra-frame predictor may be better for parts with fast motion. The second method uses a
backward scheme of adaptation of the predictor without segmentation. The predictor is a weighted sum of
predictors and the predictor coefficients are adapted continuously, pel by pel, to the local characteristics of
the intensity. In this paper we adopt the second method. We use four directive predictors: a horizontal,
vertical, a diagonal and a temporal one. The last one is based on displacement estimation. In (1), @(0, 1)
is the horizontal coefficient, a(1, 0) the vertical, a(1, 1) the diagonal and 6 the temporal, A similar scheme
is used by Pirsch. The work of Pirsch [16] is limited to the case of only two predictors: an intra-frame
using three fixed coefficients and an inter-frame without motion compensation.
To adapt the predictor coefficients we use the Least Mean Square (LMS) algorithm for its simplicity and
good performances in nonstationary situations. The LMS adaptive filter has been used in one-dimensional
predictive coding, and in a wide range of one-dimensional signal processing applications. A thorough study
of adaptive FIR filters is given by Macchi and Belanger [10]. An extension of the LMS algorithm for two-
dimensional signals is discussed by Hadhoud and Thomas [8]. Alexander and Rajala [1] used the LMS
algorithm in image coding and obtained a reduction of the distortion of about 6 dB compared with a fixed
coefficient DPCM. We also have given results of applying the LMS algorithm in intra- and inter-frame
image coding, with or without adaptive quantization [14].
The stochastic gradient algorithm minimizes the mean square prediction error
E{U(m, n; t)— Tm, n; 0°}. 2)
The coefficients (here three intra-frame and one inter-frame coefficients) are updated by the following
expression:
(0, 1) (0, 1) Tm, n= 131)
a(1, 0) a(1, 0) Tm-1,n3 1)
alt) | 7) a1) FHM! -tn-iit) | 3)
% Some B Nonny Ton—6,n—it; 1-1)
Sigal Froceing: Inge Commanicaton G. Teiritas, J.C. Pesquet / Predictive coding of images 463
where Zi the quantized prediction error and 1 is known as the adaptation step-size. The updated coefficients
at pixel (im, n) are used to predict the intensity at the next pixel (m, +1). Tt has been proved that without
quantization the inequality
2
O0, (is)
em, 2) = Qn eC, n)), (16)
sonny =F sem, ”, an
‘s(n, 0)
with 2(-1, 2) =2(m, 1) =I) for all m0 and n>0, and F(-) a positive valued function which
satisfies
FOQ)=FA(-1I)>1, F(0)<1, (18)
F(1) is called the step-size adaptation parameter. A possible choice for these parameters is F(0)F(1)
The parameter z(m, n) may be interpreted as the state variable of the quantizer.
The parameter 2(m, ) is bounded in order to avoid that s(i, k) converges to zero or diverges:
0<6K and />L), a bound for X(m, n) is given by
> KL =
im, a) < int fiat ke Drvtrgle Me 4009)
Fem SS i paar aR 2, Ee Orie eae
wiulze
where (r; e!, re!) belongs to the convergence region in the z-transform domain with 00 and 30, (24)
s=0 imo
‘melee
with
KL
(m,n) =w(m,n) = SY alk, Da(m—k, n=),
Kao 10
Kirulee
The solution of this 2D linear difference equation on e(m,n) depends on the input (m,n) and on the
initial conditions. Let us consider the following initial conditions for the predictor: $(m,)=0, for m<0
or n<0. Then, e(m,n)=x(m, n), for m<0 or n<0. Let us write
e(m, n)=e™(m, ny +e"(m, n) (25)
where the part e(m,n) depends on the input (m,n) with zero initial conditions, and e(m, n) is the
solution of the homogeneous 2D linear difference equation resulting from (24), with (m,n) =0, depending
on the initial conditions only. Concerning the first term e'(m, n), we can write
Koo kt
em n= YY alk, Dem—k.n-D + wlmny— YY alk, Denk, n=), (26)
0 imo ko imo
weviao iewigo
Let us consider the impulse response h(k, !) of the above recursive filter. We can write
mon kok
(n,m) = LE blk, Df wom—k,n-D= YY alky, Wem—k- ky, 2-1-1) ). 7)
io Fo kaa yao
dyrnire
Knowing that the quantized prediction error is bounded by 42/5, we obtain the following bound:
(em, I} <[ Ef boom mI +4 FY lath, HI) EE We 28)
wo Fo
:
Wereelao
The second term e“(im, ) due to the initial conditions is studied in Appendix B, where it is proved that
its space-averaged mean absolute value converges to 7ero, i.e.
lim ayel's! *5'1eeionmi}= 29)
Mew MN yco nce
Vol 4 No.6, November 1992 466, G. Tiritas, J.C. Pesquet | Predictive coding of images
Using (25), (28) and (29), we finally find that
lim su Nye S, Seem, mit
‘ Pune ba ‘i
Moved
<(e.+ (ato. I+ (a, Ol +a. DD ) ELM DL (30)
Ko
5
where ¢, =E{\w(mn, 2)]}
The above results, according to the study of [5], permit to establish the existence of a unique stationary
joint distribution of the input and decoded 2D processes.
5. Decoding
the presence of transmission errors
‘The main problem with predictive coding, and particularly with aduptive methods, was found to be its
instability in the presence of transmission errors. It is important to ensure the robustness of the decoder
without rate augmentation. The criterion of adaptation of the predictor must then be modified. We study
this problem in the case of intraframe coding without motion compensation (=0).
‘The algorithm presented in Section 3 may be divergent in the presence of transmission errors. A mod-
ification of the initial algorithm, constraining the output of the predictor, leads to the stabilization of the
filter in the decoder. The criterion minimized for the constrained LMS algorithm is
I= E{e°(m, n)} + nE{ ln, n) — ion, ny}, Gl)
where im, n) is the output of a fixed coefficient filter, whose inverse is a stable filter, and 7 is a positive
constant which may be interpreted as a regularization factor.
For stationary signals, the minimization of the above criterion gives the following solution.
1
A.
H+—_ a, (32)
I+ +n
where H (vector containing the filter coefficients) is the optimal solution resulting from the LMS algorithm
Thus H, is the barycenter of (H, 1) and (Hr, 1). Therefore Hy may be interpreted as an a priori conservative
estimation of the predictor coefficients with confidence degree 7. A similar approach is introduced in [12]
for another type of adaptive predictor. The updated equation of the constraind stochastic gradient algorithm
is
(0, 1) a(0, 1) Tin,n-1id)
(1,0) | =] a(1,0) +4(&(m, n)— n(Fim, n)— idm, n)))| Tom, ns) (3)
11) timmy LAC, WD) Jima Tm=-1,n-1;0)
A study of this algorithm is given in [13], where its performance is evaluated, and in [15], where its
robustness is proved for some specific signals.
Ifan adaptive quantizer is used, it is also vulnerable to transmission errors. It must be modified, to avoid
or to dissipate the effects of transmission errors. In [7] a technique is introduced for this purpose in the
Signal Procesing: Image Conniction G. Toirkas, IC. Pesquet { Predictive coding of images 467
case of a 1D adaptive quantizer. We usc the same technique here by modifying (15) to diminish the effect
of a transmission error
Hm, n—1)'s(m—1, 0)
SCM, 1) Paes 34)
( zm InP eo
where 0<7<. In fact, the above equation gives a linear recursive relation on In a(n, 2)
2 -{ 2m, 2) \
In s(n) = 7 In z(m— Ln) +7 In 20m, n= 1) 7? In can) (35)
smn
The 2D linear recursive filter on In z(m, n) is separable and obviously stable for 0< <1. Thus effects of
errors on Fle(m, n)/s(m, n)) will be dissipated,
The theoretical results of [13] and [15], and simulation results of the following section demonstrate that
the decoder is readjusted if the transmission errors do not affect the pixel synchronization. This supposes
that fixed length codes are used and only bit inversions are considered. The case of variable length codes
is more difficult; it is not considered here,
6. Simulations and results
The coder described in this paper has been simulated for the ‘CAR® sequence of images provided by
CCETT (COST 211 bis European normalization). Figure 4 gives the first image of the sequence. The car
is in movement and the camera pans the scene. A strong additional noise disturbs the intensity und dillerent
types of spatial details are present in the scene. To appreciate the importance of the motion, the difference
between the first two frames is given in Fig, 5. The empirical standard deviation of the difference is 30.1
The quality criterion is given by the mean square distortion
= a = = Gm, n) ~ Tom, ny, (36)
Fig. 4. The first frame of the sequence ‘CAR’ (CCETT), 468, G, Tsiritas, IC. Pesquet / Predictive coding of images
Fig. 5. The difference between the first two frames.
or equivalently by /D. The power of the prediction error is given in order to appreciate the performance
of the predictor. Finally, the entropy of the quantized prediction error is given, this quantity being closely
linked to the compression rate.
Table 1 gives numerical results for 60 fields of the sequence using the adaptive quantizer of Section 4.
We distinguish two cases of predictors using a motion estimator: in case A the coder transmits the addresses
of points of discontinuity and in case B it does not transmit them. The detection of discontinuities is
described in Section 2, and the rate of discontinuities detected for the frame of Table 1 was about 1%.
is the power of the prediction error using the adaptive quantizer and H is the entropy of the 3-level
quantized prediction error (H also includes the discontinuity side information in the case of coders A).
SNR is the signal-to-noise ratio given in dB and defined by
SNR =10 tog 22. @)
In our simulations each field is composed of 264 lines and 674 points per line. For the obtained value of
the entropy and for fields of 288 by 720 points, the bit-rate is about 15 Mbits/sec for the luminance
component. The pictures given in the paper are frames composed of 512 by 512 points. Figure 6 gives the
distortion for the 60 fields of the sequence using the hybrid predictor of case A, except the first two fields
which are coded using an adaptive intra-field predictor,
Table 1
‘Numerical results for the ‘CAR’ sequence
Hybrid Hybrid Motion Motion Adaptive
prediction prediction compensation compensation —_intraframe
4 B 4 B
133 139 160 162 164
XB 75 18 6 87 90
SNR 30.7 303 295 293 291
" 13 133 136 1.36 Lal G. Teiritas, J.C. Pesquet / Predictive coding of images 469)
80
0
50 —
ot % 3 4 30
Fig, 6. Distortion of the ‘CAR’ sequence using the hybrid coder A
Figures 7, 8 and 9 give the prediction error before quantization for the hybrid adaptive, the motion
compensated and the adaptive intraframe predictor, respectively. The prediction error is dilated by a factor
2 in these three figures.
If the rate of transmission errors is about 107%, the square root of the distortion is 9.1 at the coder, using
the stabilization method of Section 5, and 9.3 at the decoder for frame ‘CAR_00’. This distortion is
practically visually imperceptible. In Fig. 10 reconstructed at the decoder frame ‘CAR 00” is given, when
the pure LMS algorithm is used for adapting an intra-frame predictor and when only one transmission
error occurs. In Fig. 11 is given the reconstructed at the decoder frame “CAR_00’, when the modified LMS
algorithm is used for adapting an intra-frame predictor and when the rate of transmission errors is about
10°, We give in Fig. 12 frame ‘CAR 00’ as it is reconstructed at the decoder, when a block of 256
transmission errors occurred at the 255th line of the frame. The square root of the distortion is 9.5. All
these results illustrate the small sensitivity of the proposed adaptive method to transmission errors.
7. Conclusion
An adaptive intra-inter-frame predictive coding method is presented. The simulation of the proposed
method for a very critical image sequence illustrated a certain improvement in comparison with
Fig, 7, The prediction error before quantization for the hybrid adaptive predictor.
Vol. 4. No, 6. November 1992 G. Teiritas, LC. Pesquet } Predietive coding of images
Fig. 10. The decoded frame “CAR_00” with the LMS intra-frame predictor after one transmission error. G Teiritas, J.C. Pesquet / Predictive coding of images a
Fig. 11. ‘The decnded Frame ‘CAR_00" with the moditied LMS intra-trame predictor with a rate of transmission errors of 10 4
Fig. 12. The decoded frame *CAR_00° with the modified LMS intra-frame predictor after a block of 256 transi
1 ett088.
non-adaptive methods or only motion compensation techniques. Only small complexity increase from
motion compensation pel-recursive methods is needed. The crucial problem of transmission crrors is consid-
ered in the case of adaptive intraframe prediction. Only some algorithmic complexity is added in order to
make the decoder in practice non-sensitive to channel errors. Knowing that pel predictive methods produce
good results at higher rates, which are improved using adaptive techniques, we think that replacing the pel
prediction by block prediction, and scalar quantization by vector quantization, good results could be
obtained at lower bit-rates.
Appendix A. Bounding the output of 2D quarter-plane linear filter
Firstly we prove the following proposition.
Vol. 4, No.6, November 192 an G. Toiritas, J.C. Posquet / Predictive coding of images
PROPOSITION. Let h(m, n) be the impulsional response of a first quarter-plane 2D linear filter and H(z, w)
its transformation in the (z, w) domain. If the filter is stable, then
«2 » 1
YY lcm, | < inf ———_§ sup |H(r, e!, ry e!)], (Al)
mao 9-0 rue (Ln) = ra) 4.82
where (rye, r2!*) belongs to the region of convergence of H(z, w), with O —1. The above equation leads to
| y y leon, ny]
mao nao
_. fd=a@)0-p%)
Seer Garp) Me)
w
[1.09 N+
es ee
N-
1-p
Jat
dla+ lal, 1)
(a- -a
[is. I)|M+ (B.7) G. Teiritas, J.C. Pesquet { Predictive coding of imaves 475
where ¢,=Ef|x(m, »)]}. Finally, we find
{Merwe
EY Sy ieP%m.mih
Ue
extn a
< ae 1,0)|+la(l, DDN
You might also like
