Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
STOCHASTIC INFERENCE FOR SPATIAL STATISTICS
by
Lars Smedegaard Andersen
TECHNICAL REPORT No. 120
February 1988
Department of Statistics, GN-22
University of WaShington
Seattle, Washington· 98195 USA
Stochastic inference for spatial statistics.
Lars Smedegaard AndersenAfdelingen for Teoretisk Statistik
Aarhus Universitet, Denmark
present address:Department of Statistics ON - 22
Padelford HallUniversity of WashingtonSeattle, WA 98195. USA
Abstract
Statistical image analysis and analysis of lattice data on lattices of dimension two or higherhave assumed increasing interest over the past few years. For image analysis especially, theemphasis has so far been focused on producing fast and reliable algorithms for restorationthat gradually improve the visual quality of the resulting restored image. In this paper wedevelop and study algorithms with precise mathematical optimality properties that facilitatethe use of classical statistical methods for hypothesis testing, estimation, classification, andassessment of uncertainty.
A foundation for statistical inference for image analysis is developed and designed to fit themost commonly used image models: Markov random fields corrupted by local spatiallyindependent noise process. This foundation is found to be general enough to accommodateinference requiring only that decisions could be made had the true image been observed, andas such is suitable also for highly specialized decision problems. Applications beyond thebasic assumptions are explored.
American Mathematical 1980 sut"ec::t cJtas~rifjc::afj.on.
'''/ltJ·Il.~· Bootstrapestiimaltioltl, optimality,
contract mlIntJ(~r
-2-
1. Introduction
In recent the field of image analysis has to some degree focused attention on
investigation of iterative procedures that gradually improve the visual quality of a given
corrupted imag~: Besag (1986), Geman and Geman(1984), Geman and Graffigne (1986),
and others. These procedures unfortunately lack the possibility of assessing the variability of
the restored image, let alone draw inferencefor the underlying signal. We may draw
attention to a method defined by Haslett and Horgan (1986), and later commented upon by
Derin et al (1984) which produces an approximately optimal image restoration for Markov
mesh random fields corrupted additively with Gaussian errors by applying results from the
one dimensional special case. In this paper we design a method which is entirely spatial and
which avoids limiting assumption of Gaussianity of errors - and for which the approximation
is stochastic, introduced by a Monte Carlo approximation, and can thus be adjusted to any
is the measures restoration algorithms. We,
however, believe there is a definite need for these measures to aid the statistician in choosing
candidates.for image restoration procedures. It is important though to k~pin
the aid
nnages. The rele'vant SUCI:::ess me,asu:re
restoration procedure meets this
extent to which the
USl131Jyis one
is no otnrl0rlS' a
we shall mtt:odllce it po:ssilble to overcome
se(;tt(J~n we
is to
- 3 -
development of resampling techniques for image models and to the deduction from these
optimal restoration schemes under different objectives. The more inference
problems are illustrated in section 3, largely through examples. Section 4 includes
concluding remarks and ideas on how the methodology developed in this paper can be
generalized to more complex image models, such as the hidden Markov models and pattern
synthesis models (Grenander 1981, 1984 ).
1.1 The basic image models.
We shall throughout this paper work in the fonowing set up. Let T Zd be a finite
rectangular lattice, and let xT be an unobserved image on T we assume that xT is a
realization of a stationary discrete Markov random field XT , with respect to a translational
invariant neighborhood system; we avoid edge effects by thinking of T as wrapped around a
toms in Rd +1. The observation available is YT' and we assume that
p(YT I XT ) = n p(Yt I X t )
t E T
and furthermore we assume that the set ofpossible vaIuesof X t and Y t are A and B
respectively, where A and B are finite sets, the matrix Ut =p ( Y t I Xt ) is assumed to be
square and invertible.
(1.1 )
Contrary tothe usually imposed assumption of independent identically distributed errors,
independent of XT , (1.1) requires only spatial independence of the error mechanism. For
clarity of exposition we shall consider only translational invariant error mechanisms,
= U , for all t. The assumption that
tranSl.atl,oml1 invariant are UIlJleC,essary restm::;tl()1lS as is SP~~Cilll al~SUltnp'tlorls
is error me~chall1sm
simlplil:::ity we
are knl[)Wlt1: course UStlalJly not case, we to
estiimate p:ar81:net:ers on
Adopting an empirjical is
-4-
random field, inference for xT should thus be carried out using the posterior distribution
, and estimation of parameters for the prior ( often called hyper parameters ) should
be based on the marginal distribution p (YT ) of the observed image. Knowing the matrix U
means that the device through which we receive the signal is known. Non - parametric
models for the image process can also be estimated using similar methods for direct
estimation of the local conditional distributions. When resampling from the non parametric
models we shall use the term bootstrap resampling, since the grey scales A and B are finite
choosing a non - parametric model is, in reality the same as expanding the number of
parameters in a parametric model.
The image models described above are usually chosen for their simplicity rather than
through realistic considerations of mechanisms generating the images. For more structured
models an alternative was proposed by Grenander (1981) that allows for locally highly
specialized models - these models even though appealing have been studied only
sporadically in the literature, mainly because of the even greater analytical problems than for
the Markov random fields models. There is, however, no doubt that applications of far more
advanced and specialized ima.ge modelsthantlIe ones we study here will be introduced in the
future.
2. Resampling methods for Markov random fields.
Given a Markov random field the local conditional distributions are on the form
is the set of clilClm~s
( vc(Xt;XH,» ,c E C,
respelct to
(2.1)
--t R are
are as
nOtm~llizing constant mak:l11lg
) ( )
-5-
or at least a Markov random field with the same local conditional distributions as
by (2.1) can be simulated Metropolis algorithm:
1 Choose a starting value xfO), from the set of configurations Ar .
2 Iteratively with n, pick a pixel t at random ( such that all pixels are visited infinitely
often) and let x.}nJ{t} =xj':..{g and generate x?) using (2.1).
it can be proved, by a Markov chain argument, that xf71) is a recurrent Markov chain with
state space AT. The stable distribution ofxfn ), (as n ~ (0), exists and is a Markov random
field with local conditional distributions (2.1). Usually the<full simultaneous distribution of
Xr is given by (2.2).
2.1 local conditional distributions.
As we argued in section 1 the relevant distribution for restoration of the observed image Yr
is that of Xr I Xr , denoted here for convenience by PYr (xr ); formally this may be written as
P(Yr I xr )P (xr )py/xr)= ()
P Yr
where by the local conditional distributions of Xr conditional on Yr are given by
(2.3)
J7 (x I x ) =Yr t HI P(Yt I X't)p(x't I XH)IE A
(2.4)
is COlldi1tiOllallty on true is
we see no is
we can geI1lera.te as repilicate s~tml)le8 as we
wish from I YT =
-6
The optimality results we are to state
Local optimality ofresamp,ling restoration schemes.
Contrary to the restoration algorithms of simulated annealing and iterative conditional modes
the resampling restoration schemes typically will not produce a global or local maximum
posterior likelihood estimate. For ordinary forecasting problems these approaches are
justified by reference to the Gaussian time series; for image analysis, however, the posterior
distriJ,ution is remarkably flat close to its optimum. J:<urtllerrnoJr~
eventhe most likely configuration in AT is
and prove here will be of definite subjective nature, such as minimum mean squared error,
minimum misclassification restoration etc. This, however, is also the key to the pixel to pixel
variation measure since we acquire more information about the distribution of XT I YT than
just a point estimate of the mode. By introduction of objective functions and coupling of
pixels we are in the position where we can make minimum cost decisions, and for reasonably
small blocks of pixels, B , say the full distribution of XB I YT can be well approximated by
simple~{)nte Carlo.
Definition 2.1
A structure identifier, F, say is a function from AT (= x At)~M, where M is atE T
metric space.
structure ide~ntj,fiers
as salllple ave~raJ~es, mean sQ1Jlare~d s::!!mple error. paraml~ter estumllOI'S,
cl1:1faeterisltics etc. a structure ide~ntj,fie,r,
a po~~tetlOr
t:orec~lStrng. Bellow we
structure identjJl1ers: term
- 7 -
broad sense - coupling of pixels is allowed.
Definition 2.2
A local objective function is a real valued multi linear structure identifier, F, depending
on xT' such that
Fxr(X'T)= 1: It,x, (x't)tE T
and
It.x (x) = O,lt,x (y);;::: O,lor all x,y e A, and t e T.
With this definition in hand we are able to state and prove theorems of optimality for the
resampling estimation schemes.
Theorem 2.1
Let x; ,... ,x: be resampling replicates from (2.3) using Metropolis algorithm as
described above given a local objective function, defined by I : A2 ~ R there exist
o
E{ as n ~ co
where the mittimum is taken over all ttmlctl~(mS 'If
over ( ).
eXJ,ectatllon is
E{ f' • } :: E {E { 1 ' .
- 8 -
= E{ E{ft,x,(+(n 1"'" *» I YT }}teT
wbere by it suffices to prove tbat
The right band side is calculated by
1: f u; (-1 )p (x ;Xt I YT ) ,xeA
(2.5)
where y minimizes (2.5). Since Pt, the bootstrap empirical distribution function for Xt
converges to p (.,Xt I YT ) with probability one y minimizing
1: ft.x(Y)Pt(x)xeA
converges with probability one to y, where by the theorem follows, since A and Bare
finite.
To appreciate the above•theorem fully we tum to a few examples; let f t ,.x (y ) = 1{x :;t:y }, then
the majority vote estimate (among the resampling replicates) converges witb probability one
to the minimum mean misc1assification estimate, as the number of iterations <per resampling
replicate, and the number of bootstrap samples tend to infinity. In the case where A and B
fwlb~~r asswnptions
for f and tbe local conditional distributions.
o
bas a unique
it
E{ } -7 '" E{ ) } , as n
9-
where the minimum is taken over functions 'If: AT -t AT, and the mean is over
( , Y T )·
Proof
As above, only minimization of (2.5) is performed through differentiation. As to
conditions sufficient for the existence and uniqueness of the stable distribution for the
simulation Markov chain see Andersen (1988,b) and Revuz (1975).
o
As a special case of corollary 2.1 we get a resampling approximation to the classical two
filter and Kalman filter for time series models that are special cases of Markov random fields
on one dimensional lattices (Solo (1982) and Meinhold and Singpurvalla (1983) ), along
with a natural generalization of these methods to non Gaussian cases and spatial Gaussian
and non Gaussian Markov random fields, see Andersen (1988,b), Ripley(1981), Besag
(1974,1975) and Green, Jennisonand/Seheult (1985). Gaussian Markov random field
models the mean squared error optimal restoration produces an approximation to the
maximum posterior probability estimate, ie., the same as simulated annealing.
that C011pIJInjl;
in special cases; assume for instance that after obtaining the restored image special interest is
focused on a region, B, say by resampling again an optimal estimate can found for or
strllctll1re identifier thereof, as one
can m<lI'gin<lLlly or cOlldrtiOlla!J[y on
corlSUllllrLg for some ch()ic(~s structure l(lentlJtl.er
1
- 10-
The local asymptotic measure of perfonnance of the restoration:
lim E {Itn~oo
) I YT }
is consistently estimated, but in general not unbiasedly, by
L It ,x (xt)Pt (x) ,xeA
where Pt (x ) is the observed frequency of the event Xt = X , among the resampling
replicates.
In practise we need to choose the number of samples from XT I YT = YT, and in doing this
we also set the internal precision of the stochastic approximation by Monte Carlo. In general
it is our belief that fewer samples of higher quality are preferable to more samples of lesser
quality, which in practice amounts to letting the local updating procedure from Metropolis
algorithm run longer and taking fewer samples. Using these procedures, however, for simple
structure identifiers does not require excessive computer time.
2.3 Exal1tples.
Using Metropolis algorithm we shall present two examples, both employing artificially
generated data, from a Markov random field, . In order to study more structured images,
similar procedures can be developed for the pattern synthesis models see Grenander (1981)
a problem we plan to look into in near future.
1
o
Let ={O,l assume is a MarKC'V r~lJldom
assume is lsOJmeJtric
nearest next nearest nei.ghl,ors,
{---} + ,t +
- 11-
) is the number of ones among the first four nearest neighbors of twhere n l(t) and
and the second nearest neighbors, respec;tively The corresponding conditional
local conditional distributions are easily calculated
We choose to set the parameters at a=(-5.2,1.2,1.0), and p=0.2. Figure 2.1 a) shows
the true signal, b) shows the corrupted image, corrupted through a symmetric binary
noisy channel, in c) we show the minimum misclassification restoration estimate of the
ttuesignaland in d) we show the minimum mean squared error restoration, both using
200 bootstrap samples and 50,OOOlocal updates per replicate. Finally e) and f) show the
estimated local misclassification rates, and the highest 5 percent estimated local
misclassification rates overlaid on the estimated image. Not surprisingly figure 2.1 e)
picks out the boundary with high success.
Insert figure no. 2.1 here
Example 2.2
This example differs from the previous one by having three unordered grey levels. For
restoration purposes local srnoothingalgoritmnsaremown to Work very Wellfor A
categorical although local smoothing becomes less useful. We have chosen the local
conditional distributions such that
o
T11(j»+ (n
cOJ:Te~,pondJtn.glyIS
sec:ond nearest neiighboJrs error prc.ce:ss is de:;;cribed
pI-p
= { p=
- 12-
Figure a), the origiual image was generated from (2.7) usiug 5 x 106 local upclates,
and = 1.05,~= Siuce T is wrapped around a torus, we iu only
major connected patches. b) is the corrupted image, p= 0.35, the minimum
misclassification restoration usiug 200 resampliug replicates and 5 x 104 local up(1at(~s
per replicate shows 150 misclassmcations, d). Figure e) shows this a local variability
measure, again picking out the iuterior boundaries very successfully - f) shows this
"estimated" iuterior boundary process overlaid the restored image.
Insert figure no. 2.2 here
3. Stochastic inference.
The practical inference problems of spatial statistics are many-sided and iu image analysis
often highly context dependent - there are, however, basic problems iu fonnalized hypothesis
testiugsfu~e thehypothesisiare noteaselydescribediutenns oithe Markov random fields
models. A counter example is that of parameter reduciug hypothesis for the Gibbsian
parameters; the estimation problem was studied iu Andersen (l988,a), and we shall assume
that the distortion mechanism pennits consistent estimation of the Gibbsian parameters under
both the full model and the null hypothesis, usiug the maximnm quasi likelihood estimate.
3.1 Examples.
1
..-...t R a structure ide~ntj.tier,
res:anlplilng we two mCtdets
a test can two dis1tributicJns. test is so
so is
one realiz:lticln no
13
practical irnportance.
As a natural choice of structure identifier we suggest the total quasi likelihood,
(3.1)
where the last tenn on the right hand side is the pseudo likelihood function, Besag
(1974,1975). this choice of structure identifier produces a test that will be sensitive
towards hypotheses corresponding to models changing the posterior probabilities for the
restored image significantly. As to applications for image analysis it will often be more
relevant though to use a structure identifier which is chosen specificly to the context.
o
The classification problem and pattern recognition problems are solved more easily in
principal. They require the ability to classify any given directly observed image, as well as a
cost function for misclassification.
Example 3.2
Consider figure 3.1 below, a) showing the original image, generated by the same
Markov l'andomfield as mexatllple 2.2, super imposed on a), in the smaller frame is a
black object drawn by hand. Figure 3.1 b) shows the corrupted image: The original
image received through a symmetric noisy channel with
In c) is the restored using the miumnmn nlis(;la~jsif'ica,tion e:stitl11ation
the postertor dJlstrilbut:ion
T, reslJecttively F 1 is area
lar~test area
T inte.rsectin~ B a COl1tnec:ted
is area lar~;est cOrltnec:ted set mtersectin~B,
all
ext~rimeJllt from
is the estlm~lte, counting the number of
and f) it that 1
biased toward volumes then the true of 163, by repetition
the same original image it seems that 1 and F 2 are biased toward smaller Vo]lulliles
while F 3 seems biased toward larger volumes; the biases of course delpeflld actual
realization of the error process. The true volume of B is 163, and the volume of the
estimated object B is 174, different from EF 1 I YT' EF2 I YT and E F 3 I YT'
o
Insert figure 3.1 here
3.2 Optimality ofdecision rules.
Having chosen the Markov random field image models as the general frame work we must
accept that all configurations over T are possible, the model parameters can not be used to
confine the true inlage to ha.ve a specific structure. For pattern recognition and classification
purposes this characteristic makes•the introduction ofstruCtUre identifiersaJrnost impossible
to avoid; in order to make decisions on the basis of images, using the Markov random field
models, we must be able to classify any given configuration on AT. As usual we also require
the knowledge of a cost function.
It is a serious disadvantage of Markov random field models that a given
feature can not be directly as a askingif :::::
and shape of a
can not
trallslalted into a sta1tistjlcaI hVl)otlnesiS
classific~lti(Jtn pU!1JtOS~~S we can resample COIlCfI.ltlOt1311!y on
is classi:fyable
a structure IdentIfier
tenus
Theorem
Let F : AT -t (I, ... , k } be a structure identifier, and let W be a cost function such
that W is invertible, Wi,i = 0 and Wi,j ~ 0, then the classification rule determined by the
minimum cost identifier of F (xi) ), i = I , ... , 11., with respect to I YT ) ,
i = 1 , ... , k is the minimum cost classifier of xT into
Proof
The proof is trivial since we can approximate the distribution function to F (XT ) I YT
arbitrarily well, see theorem 2.1.
o
Insert figure 3.2·here
Example 33
Consider figure 3.2, in a) is a realization of the Markov random field from example 2.2 -
black patches are separated by only two pixels; b) shows the corrupted image and b) the
minimum misclassification restor:ati(m of a), misclassification rate: 1.96 peI'Cel1tt; we are
iIlt€~rested in finding a minimum m1:scl.:tSsjlficau<m to answer
mrpo1;edon
mlsclasS:ltt<:atllon costs
Fll,2,3 -tIO,l},
- 16-
F = 1 if the largest set of connected black pixels intersecting the top frame and the
largest set of connected black pixels intersecting the lower frame of figure a) is the
same.
F = 0 otherwise.
In 135 of the 200 replicates F = 0 was observed. the procedure, thus, correctly classified
the observed image in figure 3.2 b).
Figure 3.3 shows the minimum misclassification restoration given that the black "blobs"
are figure 3.3 a) is really two objects, misclassification rate: 1.95 percent. Note again.
that since the Markov random field models can not force global features for the
restoration there is no guarantee that F (XT ) = O. In figure 3.3 b) we have displayed the
pixels of disagreement between figures 3.2 c) and 3.3 a) as black pixels overlaid on
figure 3.3 a), there is a total of 13 black pixels in figure 3.3 b).
o
[nsert figure 3.3 here
4. Concluding remarks
The methodology suggested in this paper has been developed to suit the tasks faced in
statistical image analysis under special model assumptions. However, by straight forward
generalizations the results can be retailored for more diverse areas of statistical analysis
lattice data such as Kriging, field experiments in agricultural and statistic,il] ITlecllalrucs.
The idea has the of a
lie,c:ib:ility to the
Mar!kc,v nmdom field 'uuu'"', ass'mnptl<ms.
is seen most
an
to- one
structure Ide~ntllh.er.
toin nocontext delJeDrdellt
- 17-
fonnulate, in mathematically tractable how the eye and the extract
the "important" information from an no both
weakness and the strength of the method in that the formulation a tealtm:e
characteristic is next to impossible in some cases based on all the information one can extract
visually an example consider the problem of classifying of cats and
and brain do this in split seconds, but describing this process mathematically is not possible
outside simplified examples far too simple to be of any practical value). In the dependency of
these methods on chosen feature identifier lies also a great advantage in that the analysis in a
well defined way can be made dependent on the particular context in which the data were
coUectedandthe purpose for which the results
One of the serious disadvantages of using the Markov random field models is that all
configurations of grey levels over T occur with positive probability. The pattern synthesis
models ( Grenander (1981) ) allow for restriction in the space of configurations, AT; two
alternative ways of obtaining this are:
1), Assume we are interestedoruy in realizations for whichxa isfixed,then replicates from
XT I YT =YT , Xa =xT can be generated by leaving xa fixed in all of the local updates
when resampling.
is a set of configurations that contain some known feature, or is allowable, then a
sm:npJle of replicates
XI,T"'" ,T
can pnme<l to
I,T ... , ,T
a sarrlple = E
to:rlr.to~~ral)hy a is needed
can to some extent
asSUlTlptllon that A and B from 1.1 the same and reallZJlI1,:?; that
the that does not depend on t - which was for reasons
exposition only - is in fact immaterial, as is the stationarity of , the MarKcrv random field
part of the model.
Acknowledgements
The author is grateful to The Danish Natural Science Research Council for financial support
during his stay at Department of Statistics, UniversitYQf Washington. Inspir~tion and helpful
discussions with students, faculty and guests at University of Wasbington are also gratefully
appreciated.
A11IdeJrSeltl, L. S. (1988,a) : Consistent parameter estimation for corrup'tedMatrkc.v nmdom
with applications to report no. 117, University
Washington, (submittedfor publication)
Andersen, L. S. (1988,b): Spatial generalizations of the two - filter. (under preparation).
Stat.
: Spatial interaction and the statistical !'I1'\llthr<!l<!
pp. 75 - 83.
~v~fe1l11~_ J. Roy.
Besag, J. (1975): Spatial analysis of non lattice data. The Statistician, 24, pp. 179 195
Besag, J. (1986) : Restoration of dirty pictures. J. Roy. Stat. Soc., B, 47, (with discussion).
Derin, H., Elliott, H., Cristi, R. and Geman, D. (1984): Bayes smoothing algorithms for
segmentation of binary images modeled by Markov random fields. IEEE. Trans. Patt. Anal.
Mash. Intel., 6, pp. 707 - 720.
Geman, V~IU.laJL1, D. ~tO(;has1~ic re~laxlation, Gibbs distrib1JltiOIlS,
Trans. Patt. Anal. Mash. Intel., 6, pp. -741.
Geman,
20-
Haslett, J. and Horgan, G. (1986) : Linear models in spatial discriminant (to apl)ear)
Haslett, J. and Horgan, G. (1987) : Spatial discriminant analysis - A linear discriminant
function for the black I white case. Theory and P.
Devijver and J. Kittler, NATO-ASI F, Springer-Verlag pp. 47-
Lehmann, E. L. (1959) : Testing statistical hypotheses. Wiley, New York.
Meinhold, R. J. and Singpurwalla, N. D. (1983): Understanding the Kalman filter. ArneI'.
Statistician 43, pp. 123 - 127.
Revuz, D. (1975) : Markov chains, volume II North-Holland, Amsterdam.
Ripley, RD. (1981) : Spatial Statistics, Wiley and Sons, New York.
Ripley, RD. (1987): Stochastic simulations. Wiley and Sons, New York.
Solo, V. (1982) : Smoothing estimation of stochastic processes: two filter formulas. IEEE
trans. aut. contr., AC-27, pp. 473 - 476.
Figure 2.1 a}. Origin;!.1 image
realization of Markov l'lIndom
Figure 2.2 e),
from 200 independent replica1e8. pt =p {Xt =xi I Yr >,image shows Pt (1-p, ).
...
Figure 2.2 f), The highest .5 pemmt estimated
misclas.sificatlon m1e8 overlayed on the restored image.
II
,
II
I
I
- hI
- II
- II
- II
- -J:-
- --
-
-Figure 3.1 d), empericaldistribution functioo of F to the
mr"m value is
Figure 3.1 b}, Corrupted image 35 per cent
misdassitied pixels.
~: Jo r
l\)......o
......<0o
......<0 o
......01o
......'!o
Figure 3.1 c}, the minimum misclll88ificatiOn resoration,
using the algorithm, and the parameters from example
2.2. 200 bootstrap samples with ooסס5 local updates per
sample. 'The super imposed object shows up clear in the
nailer frame The number of misclassified pixels is
IIs 1
187.
- Ii
II
iI
II
1o
......<0o
......01o
Figure emperiell.l dlstributloo
mean value is 175.8.
Figure 3.1 f). empericai diairibution function of F $, the
mean value is 177.3.
Figure 3.3 b), disagreements betweenfiguf<IIl3.2 e) and
3.3 a) overillyed3ll black pixels on figure 3.3 a).
6. PERFORMING ORG. REPORT NUMBER
iI.
10. PROGRAM ELEMENT, PROJECT, TASKAREA a WOI'lK UNIT NUMBERS
98195
from ControWn'·Offlce)
ANO AOORESS
NAME ANO AOORESS
St.:ltist:i.cs, GN-22
Lars
IS., OECLASSIFICATION/OOWNGRAOINGSCHEOULE
16, OISTRIBUTION STATEMENT (of this Report)
APPROVED FOR PUBLIC RELEASE: DISTRIBUTION UNLIMITED.
17. OISTRIBUTION STATEMENT (of the abe tract entered In Block :ZO, If different from Report)
SUPPI..EM EN T ARY NOTES
Bootstrap CQ11lP .......,u5' conditional inference, conditioning, decision theory,resam~.lj.ng, variability.
EOITION OF I NOV 6S IS OBSOLETE
",
ing
isMarkov random
process. Thisinference requir
made had the true image been observed,also for highly specialized decision problems.basic assumptions are explored.
Figure 2.1 b), Observed image 20 percent
miclassification rate.
Figure 2.1 a). Original image 100 x 100 pixels,
realization of Markov random field after IOO local
updates.
,.
Figure 2.1 c), Minimum misclMsification restoration,
on the basis of 200 replicate samples and ooסס1 local
updates per sample. Misclassification rate 2.07 percent,
estimated misclassification rate 2.7 percent.
Figure 2.1 d}, Minimum mean squared error estimate.
Estimated mean squared error 0.022, actual mean
squared error 0.017.
"iLl'liI
Figure 2.1 The highest .5 percent estiima,ted
miscla:ssification rates overlayed on the restored image.
ratesEstimated local miscl,as£;ifi.:atioo
Figure 2.2 a), Orginal image 100 x 100 pixels,
realization of Markov random field after 5 x 106 local
updates. parameters are "'I = 1.05, "'2 = 0.9
Figure 2.2 c), Minimum misclassification restoration,
on the basis of 200 replicate samples and 50000 local
updates per sample. Misclassification rate \.50 percent,
estimated misclassilication rate 2.92 percent.
Figure image percent
miclassilication rate, symmetric noisy channel. Actually
observed misclassified pixels 36.5 percent.
Figure 2.2 d), The 150 errors overlayed on the
errors are black.
Figure 2.2 f), The highest 5 percent estimated local
misclassification rates overlayed on the restored image.
......01o
Figure 3.1 a), Original image on 100 x 100 square
lattice, the image consists of a realization from the
Markov random field in example 2.2, and a black object
super imposed in the smaller frame.
Figure 3.1 c), the minimum misclassification resoration,
using the algorithm, and the parameters from example
2.2. 200 boot'ltrap samples with ooסס5 local updates per
sample. TIle super imposed object shows up clear in the
The number of misclassified pixels is
Figure 3.1 e), emperical distribution function of F z, tbe
mean value is 175.8.
......01o
.....'!o
.....(0o
rv.....o
.....01o
.....'!o
......(0o
o
Figure 3.1 b), Corrupted image 35 per cent
misclassified pixels.
"I
II
II
JI"~ ~-_._-~---~~~ --_.,
- II
- -J-
-J
-Figure 3.1 d), emperical distribution function of F t, tbe
mean valueis 175.4.
-
~L_-J
II
"IiI
- II
- -J-
- -
-
-Figure 3.1 fl, emperical distribution function of F 3, the
mean value is 177.3.
Figure 3.2 a), Realization of Markov random field, as in
example 2.2; the black ball in the upper half of the
image was super imposed artificially.
Figure 3.2 c), Restoration using 200 resampled
replicates from Markov random fields model; example
2.2.
Figure 3.2 b), 35 percent corruption of a), symmetric
noisy channel.
Figure 3.3 a). minimum misclassilication restoration
conditional on F (XT ) =0, restoration on the basis of
135 replicates.
Figure 3.3 b), disagreements between figures 3.2 c) and
3.3 a) overlayed as black pixels on figure 3.3 a).