MINISTRY OF EDUATION OF MOLDOVA

Free International University of Moldova

Faulty of Informatics, Engineering and Design
Department of Informational Technologies and Engineering

Accepted for defense

Accepted for defense

Dean of the Faculty

Head of the Department

Iuri Dubovehi, Dr, conf. univ.

___________________________

___________________________

____________________2016

____ _________________2016

LICENSE PROJECT
Image Complexity Determination System

Author
Rodideal Gheorghe, student gr. C-41
Project Supervisor
Veaceslav L. Perju, Dr. Hab., conf. univ.

hiinu 2016
MINISTERUL EDUAIEI AL REPUBLIII MOLDOVA

Universitatea Liber Internaional din Moldova

Faultatea Informati Inginerie i Dizain
atedra Tehnologii Informaionale i Inginerie

Admis la susinere
Deanul Faultii
Iuri Dubovehi, Dr, conf. univ.

Admis la susinere
ef atedr

___________________________
_____________________2016

___________________________
____ _________________ 2016

TEZA DE LIEN
Sistemul de Determinare a Complexitii Imaginilor

Exeutant
Rodideal Gheorghe, Studentul grupei C-42

ondutorul tezei
Perju Veaeslav, Dr. hab., conf. univ.

hiinu 2016

 Rodideal Gheorghe, 2016

ef atedr Tehnologii Informaionale i Inginerie

Iuri Dubovehi, Dr, onf. Univ.
_________________________
___ ________________ 20...

S AR I N A
pentru teza de lien a studentului grupei C-42
Rodideal Gheorghe
Tema: Sistemul de Determinare a Complexitii Imaginilor
aprobat prin ordinul nr. ________ din ___ _____________ 20...
oninutul notei expliative: 1. Analiza algoritmilor, metodelor i sistemelor existente de
determinare a complexitatii imaginelor; 2. Elaborarea, realizarea i eretarea unui algoritm
noude determinare a pixelilor din imagini; 3. Elaborarea i eretarea sistemului de gestioinare a
complexitatii imaginelor.
Lista materialului graphi: 1. lasifiarea algoritmilor, metodelor i sistemelor existente de
determinare; 2. Strutura algoritmului nou de determinare a complexitatii; 3. Shema-blo a
softului elaborat de determinarea complexitatii; 4. Rezultatele eretrilor algoritmului de
determinare a complexitatii imaginelor; 5. Strutura sistemului de determinare; 6. Rezultatele
eretrilor sistemului de determinare a complexitatii imaginelor.
Data nmnrii sarinii: ___ _____________ 20...
Exeutant
Rodideal Gheorghe,
Studentul grupei C-42
ondutorul tezei
Perju Veaeslav, Dr. hab., prof. univ.

ADNOTARE

Rodideal Gheorghe, Sistemul de Determinare a Complexitii Imaginelor, tez de lien

la speialitatea Calculatoare, hiinu, 2016.

Aest proietul uprinde introduerea, trei apitole, onluzii u reomandri,

bibliografia din 18 titluri. Ea este perfetat pe XX pagini, onine XX figuri i X tabele si
formule XX
uvinte-heie: algoritmi, detetarea, reunoaterea , sistem, multidimensional,marcarea
Domeniul de studiu al tezei este prelurarea semnalelor.
Sopul i obietivele lurrii onstau n eretarea, sistem de extragerea conturilor
multidimensionale.
Noutatea i originalitatea lurrii onst n determinarea complexitatii imaginelor in
pofida marii vitezei de procesare a datelor.
Semnifiaia teoreti a lurrii onst n dezvoltarea unui software are implementeaz
mai muli algoritmi pentru a gestiona complexitatea.
Valoarea apliativ a lurrii onst faptul aest sistem poate fi implementat n
diverse domenii preum ar fi n militrie, n sisteme de gestionbare a navelor spatiale i oriunde
unde avem nevoie de marcarea conturilor prin intermediului imagini.
Implementarea rezultatelor obinute. Sistemul realizat a fost onfigurat i testat pe mai multe
alulatoare i pe mai multe imagini.

ABSTRAT

Rodideal Gheorghe, Image Complexity Determination System thesis for speialty

Computers, hisinau, 2016.

The thesis ontains the introdution, three hapters, onlusions and reommendations,
bibliography of XX titles. It onsists of XX pages, inluding XX figures and X tables and XX
formule.
Keywords:algoritmi,detetarea,reunoaterea,sistem,multidimensional,marcareadetetion.
Field of study of the thesis is information proessing.
Goals and objetives inlude researhing, multidimensional extraction system accounts..
Novelty and originality of this work is use of drivers to onnet with images.
The theoretial signifiane developing a software that implements several algorithms for
Image Complexity Detection.
Appliative value of the work is that this system an be edge detection in various
ompanies and areas.
Implementation results. The system developed has been onfigured and tested on
multiple omputers and multiple images.

Contents
Introduction.....................................................................................................................................9
1.

THE ANALYSIS OF THE SYSTEMS AND METHODS.....................................................11

1.1.

Th nd for dfinition of img comlity[7]..........................................................11

1.2.

Comlity[8].................................................................................................................11

1.3.

Entropy and mutual Information[8]................................................................................15

1.4.

Th ntroic modl of visul comlity[6]...................................................................18

1.5.

Th rimnt on rcivd tim[6].............................................................................19

1.6.

Comrison of msurs nd dt vlidtion[6].............................................................21

1.7.

Image partitioning[8]......................................................................................................22

1.8.

Complexity measures[8].................................................................................................23

1.9.

Comlity as m rorty[13]..................................................................................26

1.10.

Visul m comlity tsting mthods[13]...............................................................27

1.11.

Dfinitions suggstd in visul scinc[7]..................................................................28

1.12.

Comrisons with rimntl stimts of comlity [7].....................................29

1.13.

Estimting img comlity[5].................................................................................30

1.14.

The method of complexity: 2x2 matrices with binary entries[4]................................33

1.15.

The analysis of existing methods of the image complexity determination.................41

1.16.

Conclusion...................................................................................................................44

Bibliografie:...................................................................................................................................45

Introduction
In this paper, i introduce a new framework based on information theory and image segmentation
to study the complexity of an image. Different authors have established a relationship between
aesthetics and complexity. In 1928, G.D. Birkhoff introduced the concept of the aesthetic
measure, dened as the ratio between order and complexity The complexity is roughly the
number of elements that the image consists of and the order is a measure for the number of
regularities found in the image Using information theory, M. Bense transformed Birkhoffs
measure into an informational measure: redundance divided by statistical information. To
compute the complexity, he introduced the assumption that an input pattern can always be
described as a two dimensional grid of discrete symbols from a pre-dened repertoire. On the
other hand, he observed that order corresponds to the possibility of perceiving large structures A.
Moles held that an aesthetic measure is closely related to image complexity, and based his
measure of image complexity on information theory P.MachadoandA. Cardosoestablished that
anaesthetic visual measure depends on two factors: processing complexity and image complexity
. They consider that images that are simultaneously visually complex and easy to process are the
images that have a higher aesthetic value. From the above discussed works, it appears that
complexity is at the core of aesthetics. With the guideline that under standing complexity can
shedlight on a esthetics, we will explore image complexity from an information theoretic
perspective. Image complexity has also been related to entropy of the image intensity histogram.
However, this measure does not take into account the spatial distribution of pixels, neither the
fact that a complexity measure must take into account at what level one wants to describe an
object. For instance, a random sequence requires a long description if all details are to be
described, but a very short one if a rough picture is required . In image processing, an image is
segmented by grouping theimagespixelsintounitsthatarehomogeneousinrespect to one or more
characteristics, or features. Segmentation of nontrivial images is one of the most difcult tasks in
image processing. Image segmentation algorithms are generally based on one of two basic
properties of intensity values: discontinuity and similarity. In the rst category, the approach is to
partition the image based on abrupt changes in intensity, such as edges in an image. The principal
approaches in the second category are based in partitioning an image into regions that are similar
according to a set of predened criteria. Thresholding, region growing, and regionsplitting and
merging are examples of methods in this category . This paper is organized as follows. In this
chapter, we present an algorithm which splits an image in relatively homogeneous regions using
abinary space partition (BSP)ora quad-tree. In nex chappter , complexity is dened by using two

measures which take into account the level at which the image is considered. Finally, in next
chapter , we present our conclusions and future research.

1. THE ANALYSIS OF THE SYSTEMS AND METHODS

1.1. Th nd for dfinition of img comlity[7]
Th conct of img comlity is widly usd by comutr scintists nd by nginrs who
dsign nd construct informtion ntworks nd systms for th nlysis, rcognition,
rconstruction, nd visuliztion of imgs. Th conct is lso usd by nuroscintists, not only
thos intrstd in th mchnisms of objct rcognition but lso thos concrnd with lrning
nd mmory. It is mor difficult, for ml, to mmoriz coml img thn siml on.
Yt thr is no grd dfinition for th comlity of n img. Diffrnt dfinitions hv bn
offrd nd diffrnt lgorithms imlmntd for stimting comlity. rticulrly influntil
hs bn Kolmogorovs (1965) dfinition: th comlity of n objct s th lngth of th
shortst rogrm tht cn construct th objct from bsic lmnts, or dscrition lngug. In
othr words, rltiv comlity my b stimtd by th otiml mthod to obtin objcts from
qul comonnts. Thus, riodic squncs (lik sinusoidl grtings) hv low Kolmogorov
comlity, s thy cn b constructd from th sm rting tmlt squnc.
Kolmogorovs notion of comlity is msur of rndomnss, on tht hs somthing in
common with Clud Shnnons ntroy of n informtion sourc. If mny objcts undrgo th
sm rocdur, it is ossibl to dscrib objcts nd th connctions btwn thm with th
sm lngug. It should b notd tht Kolmogorov himslf nvr rltd his comlity
msur to visul objcts.
Kolmogorov comlity is criticizd for not bing sily comutbl (rkio nd Hyvrinn
2009), nd it cnnot immditly b rltd to th mchnisms of visul rocssing. But for
stimtion of img qulity nd img similrity siml msur of img comlity is
ndd.
Img rocssing lgorithms rsult in JG, GIF, nd othr formts to stor nd rocss imgs.
But ths lgorithms do not lwys us known mchnisms of informtion rocssing in humn
vision, though visul scintists hv trditionlly ttmtd to rlt img comlity to th
mchnisms of visul rocssing.[7]

Comlity[8]
In the last two decades, the study of complexity has become a very active research area in many
different fields (automata, information theory, computer science, physics, biology, neuro-science,
etc.) . But, what is complexity? Websters dictionary (1986) defines a complex object to be an
arrangement of parts, so intricate as to be hard to understand or deal with. According to W. Lis,

the meaning of this quantity should be very close to certain measures of difficulty concerning the
object or the system in question: the difficulty in constructing an object, the difficulty in describ ing a system, the difficulty in reaching a goal, the difficulty in performing a task. There are many
definitions of complexity corresponding to the different ways of quantifying these difficulties.
A list of complexity measures provided by Seth Lloyd is grouped under three questions: how
hard is it to describe, how hard is it to create, and what is its degree of organization? In the first
group, entropy is widely applicable for indicating randomness. It also measures uncertainty,
ignorance, surprise, or information. In the second group, the computational complexity quantifies
the amount of computational resources (usually time or space) needed to solve a problem.
Finally, in the third group, mutual information expresses the concept of complexity that
quantifies the degree of structure or correlation of a system or the amount of information shared
between the parts of a system as a result of this organizational structure.
To our knowledge, the only framework existing until now dealing with image complexity is
defined in, which deals with comparing the performance of ATR applications.
In this context, image complexity is defined as a measure of the inherent difficulty of finding a
true target in a given image. Such a metric should predict the performance of a large class of
ATRs on diverse imagery, without advanced knowledge of the targets. A split and merge
segmentation algorithm is first applied that partitions an image into compact regions of uniform
gray-level, no larger than the expected target size. Recursive thresholding determines the splits.
After the segmentation procedure is applied, the target similarity of each region is estimated. The
distribution of this similarity is taken as a basis for complexity measurement. For instance, if
there are many regions with target similarity near the maximum the image is relatively complex.
Three complexity measures are then given. The first is the number of regions whose targetsimilarity exceeds a given threshold, the second measures the distance from the body of the distribution to the most significant outlier, and the third is the weighted average of the distance to
all outliers.[8]

Pic 1.1.1: Input and output distributions for the partitioning of channel.

Img comlity[26]
s mntiond bov, in ordr to trct nd trck th trgt utomticlly it is ncssry for us to
msur th img comlity. Considring th liction nd rsrchs in ctul liction
nd rsrch, w dfin th img comlity s follows: th img comlity is msur of
th inhrnt difficulty of trcting nd trcking trgt.
Bsd on scl, th img comlity dscrition cn b clssifid into two ctgoris: on is
bsd on th globl chrctristics nd th othr on is bsd on rgionl chrctristics. Bsd
on for rticulr trgt or not, th img comlity dscrition cn b clssifid into two
ctgoris: on is in connction with rticulr trgt nd th othr on is nothing to do with th
trgt. Bsd on diffrnt dscritors, th img comlity dscrition cn b clssifid into
thr ctgoris: Thr r gry lvl bsd, dg bsd nd sh bsd.
bout th clssifiction of th dscrition mthods of img comlity, shown in Fig. 1.2.1:
Fig. 1.2.1 Tonomy of img comlity mtrics

Trgt
Indndnt

Trgt
Dndnt

Gry-lvl Globl Rgionl Rgionl

dg

Globl Rgionl Rgionl

sh

Rgionl

Bcus of diffrnt rsrch uros, domstic nd forign scholrs hv diffrnt focuss on

img comlity. In this r, considring th vribility of trgt in rl-tim nd utomtic
trgt trction w dscrib th img comlity ccording to globl chrctristics nd tk
no ccount of th scil trgt.[26]

Img comlity mtrics[26]

trs nd Richrd hv dscribd th img comlity using th globl fturs. Thy
usd gry lvl nd dg chrctristic to dscrib th img comlity, which lost th sc
distribution of gry lvl. Thus this mthod cnt msur th img comlity ccurtly.
In this r, considring th dt chrctristics of th img itslf nd th dmnds of
rcticl liction, w nlyz imgs from th rnc of gry lvl, th rnc of
trgt nd th rndomnss of img ttur. Bsd on th globl ftur bout th rnc of
gry lvl, th rnc of trgt nd th rndomnss of img ttur, this mthod dscrit

th img comlity using th informtion ntroy, th dg ntroy nd ttur ntroy. Th

rimnts show tht this mthod could msur img comlity qulittivly <> Th
comlity givn by this mthod is ccording with th difficulty of th work to trct nd trck
th trgt in tht img.
rnc of gry lvl
Th mor informtion in th img, th mor coml in th img will . Th rnc
of gry lvl cn rflct th gry lvl rich or not. Th informtion ntroy is usd to dscrib
th informtion contind in gry lvl. Th formul of informtion ntroy clcultion is s
follows:
(1.4.1)

In formul 1, N is th numbr of gry lvls nd n is numbr of th ils in ch gry

lvl. Th lrgr th H, th mor coml th img will b.
rnc of dgs
Th quntity nd th comlity of th trgt cn b rssd with dgs. Thus, w cn
tk dvntg of dg ntroy to chrctriz th rnc of th trgt. Th rnc of
th trgt cn rflct th imgs comlity. For ml, trcting nd trcking th intrsting
trgt is quit difficult whn th numbr of th trgts is lrg. Th formul of dg ntroy
clcultion is s follows:
R =-(g / )log2(g / )

(1.4.2)

In formul 2, dg is th numbr of th dg oints in th img. is th numbr of img

ils. In this r, w us th Cnny ortor to trct dgs nd clcult th numbr of th
dg oints.
Rndomnss of img tturs
Informtion ntroy nd dg ntroy cnt rss th sc distribution of gry lvl.
For ml, som imgs hv th sm informtion ntroy nd dg ntroy but thy hv
diffrnt img comlity. Thr for w should hv som othr chrctristic to rss th
sc distribution of gry lvl.
Ttur ftur is kind of non-sctrl fturs, which is wy to msur th stil
distribution of gry lvl. Ttur nlysis of imgs ws dvlod in th 1970s. Thr r
mny dscrition mthods bout img ttur. trcting ttur fturs bsd on gry lvl
co-occurrnc mtri is clssic sttisticl nlysis mthod. Rsrch on gry lvl cooccurrnc mtri hs long history. Most of scholrs gr tht this is vry rlibl mthod 6-7.

Hr w minly us th gry lvl co-occurrnc mtri to trct th ttur informtion 8. Th

formul of gry lvl co-occurrnc mtri comuttion is s follows:
In formul 3, on th right sid of th qution th molculr is th numbr of il
couls. Thos il couls hv som kind of stil rltions whos gry lvls r g1 nd g2
rsctivly. On th right sid of th qution th dnomintor is th totl numbr of th il
couls (# indict th numbr of th following fctor). In tht wy, w gt th normlizd .

(1.4.3)
From th gry lvl co-occurrnc mtri, w cn gt th scond-ordr momnts,
contrst, corrltion, ntroy nd sris of ttur dscritions. Considring th rl-tim nd
dimnsionl consistncy, w just us gry lvl co-occurrnc mtri to clcult th ttur
ntroy to dscrib th rndomnss of th img ttur. Th formul of ttur ntroy
clcultion is s follows:

(1.4.4)

Th grtr th vlu S is, th mor rndomnss th ttur distribution, trcting nd

trcking th intrsting trgt is mor difficult. In this cs, w considr tht th img is mor
coml.

Fig. 1.4.1 Tow imgs with th sm informtion ntroy nd dgs, but diffrnt in ntroy

From wht discussd bov, w us th informtion ntroy, th ttur ntroy nd th

dg ntroy to msur th img comlity.[26]

1.2. Entropy and mutual Information[8]

The Shannon entropy H (X) of a discrete random variable X with values in the set X =
{xi, x2xn} is defined as

(1.5.1)
where n = |X|, pt = Pr[X = xt] for i {1,..n}. As logpt represents the information
associated with the result xt, the entropy gives us the average information or uncertainty of a
random variable. The logarithms are taken in entropy is expressed in bits. We use the convention
that 0log0 = 0. We can use interchangeably the notation H (X) or H(p) for the entropy, where p =
{p\, p

p n } is the corresponding probability distribution.

If we consider another random variable Y with marginal probability distribution q,

corresponding to values in the set Y = {yi, y2, ym}, the conditional entropy is defined as
(1.5.2)
where m = |Y| and pt|j = Pr[X = at|Y = bj] is the conditional probability. H (X |Y)
corresponds to the uncertainty in the channel input from the receivers point of view, and vice
versa for H(Y|X). Note that in general H(X|Y) = H(Y|X) and H (X) > H (X Y) > 0.
The mutual information (MI) between two random variables X and Y is defined as
(1.5.3)

Pij = Pr[X = xi, Y = yj] is the joint probability. Mutual information represents the amount
of information that one random variable, the output of the channel, contains about a second
random variable, the input of the channel, and vice versa. I(X, Y) is a measure of the shared
information or dependence between X and Y.[8]
Method based on Entropy
Given an image with N pixels and an intensity histogram with n( pixels in bin i, we
define a discrete information channel where input X represents the bins of the histogram, with
probability distribution {pt} = { N }, output Y the pixel-to- pixel image partition, with uniform
distribution {q j} = { N }, and conditional probability {pj|t} of the channel is the transition
probability from bin i of the histogram to pixel j of the image. This information channel can be
represented by

(1.5.4)

As we have seen in content 1.4, mutual information I(X, Y) is a measure of the

dependence or shared information between X and Y. It can be seen that in the previous channel
(1.5.4), given a pixel, there is no uncertainty about the corresponding bin of the histogram, or
equivalently I(X, Y) = H(X). From the data processing inequality , any clustering over X or Y
will reduce I (X, Y).
An information channel can be defined for each color component of an image. All the
algorithms presented in this paper can be applied to each channel component of a color system.
[8]

Pic 1.5.1: Lena image with luminance Y709 for differents values of Pe (i) and MIR (ii). The (RMSE, PSNR) values for (i)
are (45.47, 14.66), (30.02, 18.27), (14.03, 24.88), and (8.21, 29.54), respectively. For (ii) we have (32.98, 17.45), (16.23,
23.61), (9.71, 28.07), and (6.25, 31.90), respectively.

Pic 1.5.2: Lena image in RGB system with Pe = 0.4. We obtain (a) 1835 (r = 0.70), (b) 3692 (r = 1.41), and(c) 4179 (r =
1.57) regions. The merging image (d) has RMSE=13.20 and PSNR=25.32.

ic 1.5.3 (a) BSP (r = 0.010, MIR = 42.34) (b) Quad-tree (r = 0.010, MIR = 39.10) (c) Quad-tree (r = 0.014, MIR = 42.34)

1.3. Th ntroic modl of visul comlity[6]

To crt mthmticl modl of th visul comlity bsd on stil rmtrs w
hv rviwd mny of th locl nd globl fturs from litrtur. Globl fturs r suitd to
driv singl vlus from th gnrl rortis of n img. Locl fturs r ndd to tk
into ccount clssicl vrbl lntions for th mning of comlity: mny vrsus fw,
curvd nd/or dtild vrsus linr nd lnr, coml tturs vrsus flt rs. Using locl
fturs lso hls rducing mbiguitis in rsults.
Local Features Extraction
oints of intrst my b idntifid by using locl ortors. W chos two wll- known
locl fturs: th img dgs nd th locl symmtris comutd by th Discrt Symmtry
Trnsform (DST). dg dtctors highlight img zons with brut chngs in luminosity lvl,
ssocitd with surfc discontinuity. Th rtionl is strightforwrd: th mor dgs, th mor
objcts (or th mor surfcs), nd grtr rcivd comlity. DST trcts zons of th
img in which th locl gry lvls show high dgr of rdil symmtry (whr th dgr of
loclity dnds on th rdius of th locl window usd). It is intrsting to not tht oints of
intrst dtctd by DST r to b rltd with oints to which shifts of gz r dirctd
rformd by humns wtching th sm img. rt from th nturl ttrction of symmtry,
this lso mns tht th mor th oints of intrst in n img, th mor coml th img is
rcivd s. Mor scificlly, DST comuts locl symmtris of n img bsd on msur
of il momnts of body round its cntr of grvity. In th img cs, th ils insid
circulr window r considrd s oint msss, with thir mss rssd by thir gry vlu g.
n ntroic Msur of Comlity
W r now intrstd in globl lgorithm tht cn outut singl vlu for ch
filtrd imgs, whil rsrving its clss of comlity. W dcidd to invstigt th
usfulnss for this tsk of th fuzzy ntroic distnc functions dtild in . Thr r lnty of
rsons for considring ths functions mong mny othrs usully mloyd in this kind of

tsk: first, soft comuting roch using fuzzy vlus sms rorit whn w r trying
to dscrib sitution whr binry logic is too strict. s for th ntroic distnc function, w
cn rformult our min qustion from How coml is this img? to Wht is th distnc of this
img from th simlst ossibl img in th dfind ftur sc? .

This roch lds to th us of

stndrd distnc functions, which rsct th usul rortis of idntity, symmtry nd

tringulr inqulity, ugmntd by ntroic functions. W chos th following functions:
G

o(n) = Tog(27 (nlog(n) + (1 - n)log(1 - n). [7] z Gi(n) = i (1- - nn1), G2(n) = 4n(i - n)

whr, = --=i \hiI, nd hi r th gry lvls of th img ils norml

izd in th rng [0,1]. It cn b sily shown tht Gj (j = 0,1, 2) stisfis th rortis of
distnc function, nd it tks vlus in th intrvl [0,1]. For ch Gj function, th first 15% of
inut vlus is md to mor thn hlf th rng of outut vlus. Most of th img would
hv comlity ind in th first qurtil, so w ctd to obtin bttr clssifiction
through th nsion of ctly this rng of inut vlus. W considrd thr comlity
clsss of imgs, thr diffrnt filtrs, nd thr ntroy distnc functions. In Tbl 1 w
show th mn vlu for ch ctgory. It is sy to s tht ll choics of functions nd filtrs
giv rsults in lin with our cttions, with slightly bttr rformnc obtind by th us
of function G2 nd th symmtry filtr without thrsholding. Not tht th rortd vlus of ntroy r normlizd in th intrvl [0,1] with th sm rocdur dscribd in Sction 3.

Fig. 1.4.1. Rsults from filtring nd lying ntroy functions to ll imgs in our ool, nd thn tking mn vlus
for ch ctgory

1.4. Th rimnt on rcivd tim[6]

W dvisd n rimnt to dmonstrt tht subjctiv msur of th rcivd tim
cn b usd s n indirct msur of th comlity of n img. W skd numbr of
voluntrs to obsrv som imgs on comutr scrn, nd rcordd thir rcivd durtion
of th obsrvtion.
W run two sts of rimnts, on in Univrsity of Surrys CVSS grou (50
individuls) nd th othr in th Dirtimnto di sicologi dllUnivrsit di lrmo (15
individuls). In ordr to minimiz th culturl bis, ll rimntl subjcts hd univrsity
bckgrounds. rticints wr rt of th stff nd undrgrdut studnts, on voluntring

bsis, without ny knowldg of our rsrchs ims. rivcy of th subjcts ws tkn cr of

ccording to th Itlin lw on rsonl dt; only initils, g nd gndr wr rcordd for
ch subjct.

Fig. 1.5.1 mls of tst imgs, clssifid by intuitiv comlity: high comlity (to); mdium comlity (middl);
low comlity (bottom)

Th rimnts wr hld in dim light room to rduc visul distrction, giving tim to
th rticint for drknss dttion. ll th usul rgonomic rcutions, such s using
qusi-soundroof room, wr tkn, nd th subjct ws llowd to choos thir own rfrrd
osition nd visul ngl. Th imgs wr rsntd full scrn. Th softwr usd ws hommd using th multimdi rogrmming nvironmnt Mcromdi Drmwvr 2004 M on
n l Mcintosh comutr with TFT LCD monitor. Th chosn imgs wr comutr
scns of intings, dividd in thr ctgoris rrsnting diffrnt lvls of visul comlity,
bsd on th rsnc or bsnc of crtin clsss of fturs nd cu oints. Figur 1 shows
mls of inting usd in this study.
ch img wr rsntd for fid riod of tim (90 scs.), with no tmorl clus;
th rimnt lso hd controlld dsign in ordr to minimiz sid ffcts: lights dimmd nd
uniform, subjct lon in soundroof room. Th subjct ws lrtd to focus thir ttntion on
th contnts of th dislyd imgs. Th imgs usd for th rimnts wr chosn
ccording to th intuitiv hyothsis tht th comlity of scn incrss with th numbr of
objcts nd thir rltiv osition, nd with its ovrll structur [8]. Th chosn imgs wr
intings, dividd in thr ctgoris rrsnting diffrnt lvls of visul comlity. Hr th
stimt tim rcivd by ch subjct is rortd. W considr it s subjctiv msurs of
comlity for th thr ctgoris of imgs introducd bov. In th following it will b
dnotd s
Th sml mn vlu nd th vrinc of th rcivd tim (jT, uT )
r rortd in Tbl 1.5.2. It is vidnt tht th mn rcivd tim dcrss with th
comlity of th img. In ordr to comr th rcivd tim with th objctiv msur of
comlity, nd to highlight our intrst in rltiv diffrncs btwn stimtions md by th
sm subjct whn wtching diffrnt imgs, th following normlizd tim msurs r
introducd:

Th roosd normliztion llows us bttr comrison with th rsults obtind from

th mthmticl modl, crrid out in th nt sction. In this contt, 0 nd 1 hv no strict
numricl significnc, but should b intrrtd mor lik subjctiv dgrs of comlity,
which suits bst with our fuzzy modl.
Rsults r in grmnt with our modl of tim rction: coml imgs (ctgory I)
roduc shortr tim stimtions thn imgs in ctgory II nd th sm is tru for ctgoris II
nd III.[6]

Fig. 1.5.2 Mn nd Normlizd Tim stimtion

1.5. Comrison of msurs nd dt vlidtion[6]

s shown by comring th ntris of Tbls 1 nd 2, our rimntl dt mtch thos
of th mthmticl modl. In fct, imgs with high ntroic comlity ind gnrt, on
vrg, shortr stimtion of th rcivd tim. Thrfor, ctgory I hs th shortst
vlutions nd ctgory III th longst. Th strong nti-corrltion btwn th ntroic
msur of comlity nd th mntl clock is shown in Fig. 1.15.1. W crrid out strict
vlidtion of th rsults using rovn dt nlysis mthods in ordr to scrtin th rltion
btwn dt nd modl, minimizing th ffcts drivd from th us of mn vlus nd th
crdinlity of th dtst. To vrify tht corrltion btwn th rimntl dt nd th
mthmticl rsults ists, w clcultd th cofficint of corrltion btwn th rsults

Fig. 1.6.1 nti-corrltion btwn th stimt of th mntl clock nd th msur of comlity vi ntroy functions

from th rimntl dt nd th comlmnts of th ntroic msurs of comlity using

Srmns . vn in th worst cs, th robbility of dt nd modl squncs bing
corrltd is mor thn 0.98. To confirm tht th corrltion is not du to th siz of th dt-st,
w crrid out mny nonrmtric bootstr tsts, using 10,000 virtul sts. In ch tst th
diffrnc btwn th mn obtind from th dt nd by th bootstr mthod ws undr
10~4. s w workd mostly with mn vlus, w lso usd th jckknif tchniqu, rclculting th rsults s mny tims s th numbr of imgs in our st, lving out on img
ch tim; ll jckknif sts hd th sm distribution of vlus, with smll numric diffrncs.
[6]

1.6. Image partitioning[8]

In this section, we present a greedy algorithm which partitions an image in quasihomogeneous regions. The optimal partitioning algorithm is NP-complete. To do this partition, a
natural approach could consider the above channel (1.5.4) as the starting point for the image
partitioning, designing a pixel clustering algorithm which minimizes the loss of MI. This process
can be described by a Markov chain, X ^ Y ^ Y, where Y = f (Y) represents a clustering of Y.
However, due to the computational cost of this algorithm, a completely opposite strategy
has been adopted: a top-down splitting algorithm takes the full image as the unique initial
partition and progressively subdivides it with vertical or horizontal lines (BSP) chosen according
to the maximum MI gain for each partitioning step. Note that other types of lines could be used,
obtaining a varied polygonal subdivision. Our splitting process is represented over the channel
(see Fig. 1.1.1)
X Y.

(1.7.1)

The channel varies at each partition step because the number of regions is increased and,
consequently, the marginal probabilities of YY and the conditional probabilities of YY over X also
change. This process can be interpreted in the following way: the choice of the partition which
maximizes the MI increases the chances of guessing the intensity of a pixel chosen randomly
from the knowledge of the region it pertains to.
The algorithm proposed here generates a partitioning tree for a given probability of error Pe by
maximizing the mutual information gain at each partitioning step. This algorithm is based on
Fano's inequality and was introduced by Sethi and Sarvarayudu in the context of pattern
recognition. Similar algorithms with different split criteria have been used in learning and DNA
segmentation.

Given the error probability Pe allowed in partitioning, Fano's inequality provides us with
a lower bound for the gain of mutual information. Taking the equality, we obtain the minimum
value of MI needed in the partitioning algorithm for a given probability of error:
Imin(X, Y) = H(X) - H(Pe) - Pe log(B - 1),

(1.7.2)

where B is the number of bins of the histogram. Note that Imin(X, Y) is calculated from the initial
channel.
The partitioning process can then be seen as follows. At each partitioning step, the tree
acquires information from the original image. The total I ( X , Y ) captured by the tree can be
obtained adding up the mutual information available at the non-terminal nodes of the tree
weighted by the relative area of the region, i.e., the relative number of pixels, corresponding to
each node. The mutual information Ii of an interior node i is only the information gained with its
corresponding splitting. Thus, the total mutual information acquired in the process is given by
(1.7.3)
where T is the number of non-terminal nodes and n is the number of pixels corresponding to
node i . It is important to stress that this process of extracting information enables us to decide
locally which is the best partition. Partitioning stops when I ( X , Y ) > I min( X , Y ) . Alternatively,
a predefined ratio of mutual information ( M I R ) can be given as a stopping criterion. Note that
I ( X , Y ) is the MI of the channel obtained at the end of the process.
This process can also be visualized from equation
H ( X ) = I ( X , Y) + H ( X | Y) ,

(1.7.4)

where the acquisition of information increases I ( X ,Y) and decreases H ( X | Y) , producing a

reduction of uncertainty due to the equalization of the regions. Observe that the maximum
mutual information that can be achieved is H ( X ) .[8]

1.7. Complexity measures[8]

A measure of complexity of an object is a measure of complexity of a task performed on
that object. The concept of complexity is closely related to the difficulty of understanding an
object, which, at the same time, is related to the accuracy of the description of it . On the other
hand, the measure of complexity must take into account at what level one wants to describe the
object. Thus, we can describe every detail of an object or only its non-random regularities. According to this, an important group of complexity measures tries to capture the organizational
structure or the degree of regularity versus randomness. In this section, we are going to present
two complexity measures rooted in these criteria and based on image partitioning.

To introduce our complexity framework, we will reinterpret the previous partitioning

approach from the point of view of the maximization of the Jensen-Shannon divergence. This
perspective, although equivalent to the maximization of mutual information, is more appropriate
to deal with image complexity and has been introduced in the study of the DNA complexity .
First, we define a complexity measure, the Jensen- Shannon divergence, which expresses
the image compositional complexity (ICC) of an image. This measure can be interpreted as the
spatial heterogeneity of an image from a given partition. From, the Jensen-Shannon divergence
applied to an image is given by
(1.8.1)
where R is the number of regions of the image, X, is the random variable associated with region
i, representing the intensity histogram of this region, m is the number of pixels of region i, and N
is the total number of pixels of the image. Observe that for the information channel (1.6.4), the
Jensen- Shannon divergence coincides with the MI. The compositional complexity (1.7.1) fulfils
the following properties:

It increases with a finer partition.

It is null for a single partition.

For a random image and a coarse resolution it would be close to 0.

For a random image and the finest resolution it would be maximum and equal to H (X).

Thus, given an image partition, we can express the heterogeneity of an image using the JSdivergence applied to the probability distribution of each region.
We can also ask which partition maximizes the compositional complexity, for a given number of
regions. This partition should extract the maximum information of the image and create the
maximum heterogeneity between the generated parts. Finding this partition is an NP-complete
problem. We have approached the solution to this problem in using a greedy algorithm.
Our second measure is the number of needed regions in the partitioned image to extract a given
ratio of information. It is related to the complexity in describing an image, and depends on the
accuracy level given by Pe or MIR. The justification for our measure is that the number of
regions is the number of leaves of the tree created in the partitioning process. The coding of this
tree (or equivalently the description of the image) will be clearly dependent on this number. This
is further justified by taking into account that our algorithm tries to create homogeneous regions
with the minimum splitting. In this case, the error probability of the channel is interpreted as the

compression error and thus the number of regions is also related to the difficulty of compression.
[8]
Results
We use a uniform partition to test the compositional complexity on the images in Fig. 1.1.1. The
results obtained are shown in Fig. 1.7.2 for the number of partitions running from 2 x 2 to the
number of pixels in the respective images. We observe that the relative ordering of the
complexities depends on the resolution level (number of partitions). For instance, the earth rise
image appears to be the most complex at resolution 4 x 4 while the wild flowers appears as the
least one. However, this behavior is reversed at high resolution.
In Figure 1.7.1 we can analyze the behavior of the second proposed complexity measure. While
the lines in the graph in Fig 1.7.2 cross themselves, the ones in Figure 1.7.1 keep a regular
ordering. Observe their exponential growing with MIR that is due to the increasing cost of the MI
extraction. It is important to note that for MIR = 0.5 we obtain a good qualitywith a few number
of regions. With respect to the number of regions, the most complex image appears to be the
Baboon and the least one is the Earth rise.
It can also be shown (Figure 1.7.3) that while blurring an image will cause a loss of complexity,
increasing the contrast causes the opposite effect. For instance, for a MIR = 1 and the luminance
channel Y709, the contrasted Lena image of Figure 1.7.3.b (r = 91.7) needs more regions than
the original Lena image (r = 89.4) and the blurred image of Figure 1.7.3.a (r = 48.3) needs less
regions.[8]

Figure 1.7.1 : Ratio of the number of regions r with respect to MIR for the images of Fig. 1 with luminance Y709.

Figure1.7.2: Compositional complexity ICC over the number of regions R of the partitioned images of Fig. 1 with
luminance Y709. The number of partitions goes from 2 x 2 to the number of pixels N in the respective images.

Figure 1.7.3: Lena image: (a) Out of focus and (b) more contrasted than its original.

1.8. Comlity as m rorty[13]

Comlity hs bn th crtogrhrs objct of intrst for mny yrs, s it influncs
rdbility nd ffctivnss of crtogrhic roducts. Comlity rsults from numbr of
symbols on th m, thir divrsity nd th distnc btwn thm (dnsity). Comlity my b
considr s intrction btwn ths lmnts rlting to two fundmntl ms scts syntctic nd smntic, hnc it corrsonds to two comlity scts - visul nd intllctul
comlity (Mcchrn, 1982). Th intllctul comlity is minly dtrmind by th
mount of rsntd informtion, th chrctr of its rsnttion, rocssing lvl nd th
clssifiction mthod s wll s numbr of clsss. vn if th m grhics is roritly
slctd nd objcts rsntd on th m r lgibl nough, th usr my hv difficultis in
undrstnding its contnt if th mount of rsntd informtion is too high (Hung, 2002).
Th visul comlity rsults from stil divrsity of visul m structur nd dnds
on dgr of tnsivnss, gnrliztion nd th dgr of visul vribl ordr. Th visul
comlity cn b rgrdd s th oosit to th rdbility. Wingrt (1974) rovd miriclly
tht th high img dnsity (ovrlodd with dtils) significntly rducs th stil structur

informtion trction ccurcy. Brtin (1967) dscribd rdbility s th bility to distinguish

th vribls from th bckground nd considrd tht it is ffctd by grhicl dnsity,
divrsity nd rsolution connctd with th numbr of symbols, thir siz nd roortions,
whrby grhicl dnsity ws rgrdd s th most imortnt fctor.[13]

1.9. Visul m comlity tsting mthods[13]

Th ms comlity s n objctiv ftur cn b studid clusivly t th visul lvl
sinc only t this lvl it is ossibl to srt subjctiv nd objctiv lyrs, hnc rform
justifid comrison. In th initil stg of rsrch on th visul ms comlity most of th
works wr concrnd with th thmtic ms in rsct to which it ws ossibl to us
mtrics tht llowd quntifying thir comlity in siml wy. ccording to Mcchrn
(1982), numbr of olygons, dgs nd nods on th m lrgly rflcts its visul comlity.
Mullr (1976) lid such comlity dtrminnt in his works on chorolt ms. Th
rsults of his works rtly rflctd th rsult of rvious studis on visul comlity crrid
out by Gttrll (1974), who notd tht th cofficints chrctrizing th visul comlity
should b insrbly rltd to such m fturs s th numbr of oint signturs or th lin
lngth dfining thir boundris. Th mning of msurbl nods, dgs nd links btwn th
lmnts on th m ws dly studid by gnhofr (gnhofr t l, 1994). Th lins nd th
nods wr lso crucil for bi (bi t l, 1992) nd Ilg (1990) in th studis on imgs
comlity nd th ossibility of thir rconstruction vi th utomtic digitiztion rocss.
Mrsy (1990) roosd th clcultion mthod stimting grhicl comlity similr to
Mcchrns (1982) utilizing th thory of grhs nd bsd on th wightd numbr of dgs
on th m. In Ditzls works (1983), th grh thory ws lso lid. Th rimntl
studis of Murry nd Liu (1994) should lso b quotd. Thy took dvntg of gogrhic
informtion systms in which dt is dislyd in th form of grhs, which rsmbl ms. It
turnd out tht th grhicl m comlity should b dfind tking into ccount its stil
vribility, nd not only siml msurs such s numbr of lins or numbr of rticulr ty of
surfc objcts. Bsing on th formntiond works McCrty nd Slisbury (Mcchrn,
1982) dvlod msur, which llows dtrmining th comlity of contour ms. Th
similr indics tking into ccounts th stil distribution of m grhicl dnsity wr workd
u by lying th frctl dimnsion (Burrough, McDonnll, 1998) nd th mthod of stil
utocorrltion (Bonhm-Crtr, 1994).
ntroy is nothr vry romising quntittiv msur, which llows dtrmining th
grhicl lod of th nlyzd m (H t l, 1997). Tht msur hs dirct connction with
th m informtion contnt nd is connctd with th ttmts to chrctriz quntittivly

trnsmission of informtion through th communiction systm. Th works on th mthmticl

bckground of trnsfrring th informtion by th communiction systm nd dtrmining its
informtion contnt with th us of ntroy wr rformd by Shnnon nd Wvr (1949).
srious drwbck of Shnnon nd Wvr mthod, which Li nd Hung (2002) ointd out, is
th lck of ossibility for considrtion of stil distribution of objcts. Thrfor Liu nd
Hung ostultd tht comlity msurs should lso tk this sct into considrtion nd
otd for th cofficints such s Thissn olygon. Th most ctiv rsrchr of ntroy
msurmnt lictions in crtogrhicl rctic ws Bjork (2003). Tking dvntg of
usful informtion conct, h showd how th chngs of symbols usd on th ms, thir
ccurcy nd stimtion of disordr cn ffct th ffctivs of m drfting nd rciving
rocss. Dt comrssion tchniqu (drivd from IT) is nothr vry intrsting roch to
th roblm of controlling m visul comlity (Covny, Highfild, 1995).
Divrsity of th bov dscribd msurs of m visul comlity is consqunc of
divrsifid lictions of individul msurs nd diffrnt undrstnding of wht th
comlity is. Thrfor, in mny css th msurs mk us of vrious, totlly diffrnt
chrctristics of th invstigtd m. Howvr, sinc non of ths msurs my srv for
utomtic dtrmintion of comlity of such grhiclly comlictd objcts s city ms,
novl roch ws ndd for ths lictions. For tht uros digitl img rocssing
tchniqus hv bn roosd nd succssfully lid by th uthors.[13]

1.10. Dfinitions suggstd in visul scinc[7]

In th nintnth cntury visul rsolution ws found to dnd on th sh of tst
objcts. Lttrs r mor coml if thy hv mor lins nd/or othr dtils. t th sm
viwing distnc, lttrs of highr comlity r lss ccurtly rcognizd thn r lttrs of
low comlity (Duk-ldr 1962). Such diffrnc in rcognition of siml nd coml lttrs
ld Lndolt nd Snlln to th id of using th sm chrctr in diffrnt orinttions for
ssssing visul cuity.
Ltr it ws suggstd tht comlity could b ssssd by th numbr of img fturs
such s crossing lins, cornrs, turns, tc (ttnv 1957; Vitz nd Todd 1971). furthr
stimtion of comlity ws suggstd by ttnv (1957) nd ltr dotd by othr uthors
(lli t l 2006): comlity is roortionl to th squrd rimtr of th img dividd by
ink r. By such dfinition, circl or disk r th simlst imgs.
nw hyothsis bout wht undrlis stimts of img comlity cm with th
liction of stil-frquncy nlysis to th dscrition of visul functions. Cmbll,
Robson, nd Blkmor (Blkmor nd Cmbll 1969; Cmbll nd Robson 1968) suggstd

tht th visul systm might us stil-frquncy nlysis to rocss visul imgs. This ld to
th ssumtion tht th mor coml n objct, th mor high frquncis rquird in its
sctrum for rcognition. This roch is widly usd by nginrs who mloy stilfrquncy or wvlt dscrition of imgs. In ths trms comlity cn thn b dfind s th
numbr of ctiv stil frquncis or s th numbr of ctiv wvlts.
similr, but mor formlizd dfinition of img comlity cn b found in Nsnn
t l (1993). Ths rsrchrs suggst stimting th comlity of imgs s th roduct of th
squrd mdin of th distribution of stil frquncis nd th img r. Th incrsing
numbr of rltiv frquncis, or hrmonics, in th sctrum rsults in lrgr mdins. With n
incrsing numbr of non-riodic lins nd stroks in th objct, th numbr of high hrmonics
lso incrss. In this cs th mdin of such n img sctrum lso incrss nd lds to
highr stimt of img comlity, rovidd th r is th sm. ccording to this dfinition,
riodic ttrns with lrgr r r mor coml. For sinusoidl grtings, th img
comlity, thus dfind, is roortionl to th numbr of brs in th grting. Thus Nsnn t
ls (1993) dfinition incororts trditionl concts of img comlity introducd rlir in
th nintnth cntury.[7]

1.11. Comrisons with rimntl stimts of comlity [7]

It is not clr whthr rticulr dfinition of comlity cn b lid only to
scil clss of imgs or cn b gnrlizd to ll tys of visul objcts, s fw studis r
known whr rimntl stimtion of img comlity is comrd to modling. For
ml, in Nsnn t l (1993) th fficincy cofficints of dtction wr comrd with
comlity for only four filtrd imgs on noisy bckground.
To study comlity, ttnv (1957) usd blck-nd-whit synthsizd bstrct imgs
s stimuli. To gnrt th imgs, grid ws filld with dots, thn th dots wr rndomly
connctd, nd th r btwn thm ws filld in with blck ink. Svrl chrctristics of
objcts wr discussd tht cn b usd for stimtion of img comlity, lthough t th sm
tim ttnv notd tht comlity is n ill-dfind vribl (g 221). Substs of imgs
diffrd in such rortis s symmtry, curvdnss (th oints wr connctd ithr by rcs, or
lins, or mid), mtri grin, ngulr vribility, th squr of rimtr dividd by r of ch
sh nd numbr of turns. 168 obsrvrs wr skd to mk rtings of th comlity of th
imgs using svn-ctgory scl. Th rsults showd tht mtri grin nd curvdnss did
not hv ny imct on comlity scling; symmtricl shs wr, in gnrl, stimtd s
mor coml if thy hd th sm numbr of indndnt turns, but th most significnt
vribl ws th numbr of turns, which ccountd for most of th vribility of scling.

Consquntly, th uthor concludd tht img comlity is dtrmind ssntilly by th

numbr of turns in th img.
In numbr of rvious n ttmt hs bn md to corrlt th hysicl chrctristics
of imgs, th rortis of th humn visul systm, nd subjctiv stimts of comlity. It
ws shown tht wll-known stimuli rrsntd s blck nd whit imgs hv thir own
miniml sizs for rcognition of ll dtils. Ths rimntlly obtind miniml sizs r in
grmnt with thir thorticl stimts clcultd s th rquird numbr of smling
lmnts which r hgonl clustrs of svn cons. It ws shown tht th miniml sizs r
lrgr whn th stimulus is subjctivly mor coml.
In summry, svrl studis hv ttmtd to stimt visul comlity in rimnts,
but only limitd comuttionl msurs hv bn lid to ths rsults. stil-frquncy
roch ws suggstd by Nsnn t l but ws not tstd for n tndd st of imgs.
Dsit critics of th liction of stil- frquncy nlysis to vision, mny rsrchrs us
this roch. Modls oftn us filtring of th img with Diffrnc of Gussins (DOG) or
Gbor-lik tchs simulting rctiv filds t th first lvl of rocssing. In our currnt work
w lso ly Fourir nlysis to comut thorticl stimts of comlity.[7]

1.12. Estimting img comlity[5]

Givn gnrl comlity msur C() for n img on cn try to stimt similritis
btwn imgs. niv ssumtion would b tht th diffrnc |C(o) C(i)| tlls th
similrity btwn imgs 0 nd 1. Unfortuntly such gnrl comlity msur dos not
ist. Th closst thing tht ists is th Kolmogorov comlity or lgorithmic ntroy K() of
th img (or ny string) . Kolmogorov comlity is not comutbl, howvr.
vn if th comlity msur C() istd or Kolmogorov comlity wr comutbl, thir
vlu s msurs of similrity would b qustionbl. Intuitivly, th similrity btwn imgs
dos not lwys qul to th diffrnc in comlity. This is bcus th contt lys n
imortnt rol vn t th syntctic lvl, lthough not s much s in th smntic lvl.
n obvious wy of introducing th contt in th ictur is to stimt th joint comlity of
imgs. This is still t vry low lvl but stimting th comlity in th contt of othr
img vrsus th comlity of singl img is mor informtiv thn rbitrry comlity
vlus lon. Hnc w r intrstd in th distnc tht is dfind s

D(o,i) = C(o|i) min{C (o), C(i)},

(1.13.1)

ssuming tht th joint comlity is symmtric, i.. C( 0|1) = C(1|0). lso on wnts to
nsur tht th distnc is normlizd roritly.

s it ws notd bov th idl comlity msur dos not ist nd Kolmogorov comlity
is not comutbl. On cn roimt th idl comlity msur in diffrnt mnnrs,
howvr. Shnnons informtion thory introducd th conct of ntroy, which is sily
stimtd from dt. ntroy cn b sn lso s sttisticl msur of comlity. vn
though Kolmogorov comlity is not comutbl it cn b roimtd using comrssion
bsd mthods. Comlity cn lso b stimtd from modl tht roimts th log-df of
dt s w do in this r.
Rltiv ntroy s distnc msur
Givn discrt robbility distribution Shnnons ntroy H() is dfind s
H() =

() log ().

(1.13.2)

ntroy is nturl msur of comlity, sinc it stimts th dgr of uncrtinty with

rndom vribls. Intuitivly it is ling: Th mor uncrtin w r bout n outcom of n
vnt, th mor coml th hnomnon (dt, img, tc.) is.
Givn nothr distribution Q, th Kullbck-Liblr divrgnc is dfind s
K L ( IIQ) = () log QH.

(1.13.3)

KL-divrgnc is lso clld rltiv ntroy nd it cn b intrrtd s th mount of tr bits

tht is ndd to cod smls from using cod from Q. If th distributions r th sm, th
nd for tr informtion is zro nd th divrgnc is zro s wll. KL-divrgnc is
nonngtiv but not symmtric nd s such it cn not b usd dirctly s msur of distnc or
dissimilrity btwn distributions. Th symmtry is sy to obtin, howvr, just by clculting
nd summing th KL-divrgnc from Q to nd from to Q, hnc th symmtric vrsion is
simly
K L S (,Q)= K L ( | |Q) + KL(Q||).

(1.13.4)

This is not tru mtric but it cn b usd dirctly s msur of distnc or dissimilrity
btwn distributions.
Using th symmtric vrsion of KL-divrgnc (q. 1.13.4) s th ir-wis distnc btwn
two imgs is stright forwrd. It is not quit th idl distnc msur in q. 1.13.1, but it
cturs th id of stimting th comlity in th contt of nothr img.
lgorithmic comlity
Kolmogorov comlity K() of string is th lngth of shortst rogrm using givn
dscrition lngug L on univrsl Turing mchin U tht roducs th string .
K() = min{|| : U() = },

(1.13.5)

whr || dnots th lngth of th rogrm . Kolmogorov comlity is not comutbl.

Conditionl Kolmogorov comlity K(0|1) of string 0 givn string 1 is th lngth of shortst

rogrm tht roducs outut 0 from inut 1
K(o|1) = min{|| : U(|1) = o}.

(1.13.6)

Normlizd informtion distnc [7] is bsd on th Kolmogorov comlity nd is dfind s

N I D ( 0, i)
m{K(0|1), K(1 |0)} m{K(0), K(1)}

(1.13.7)

s Kolmogorov comlity is not comutbl, NID nithr is comutbl. It cn b

roimtd, howvr, using th normlizd comrssion distnc (NCD) [7]. NCD
roimts NID by using rl world comrssor C nd it is dfind s
N C D ( 0, i)
C(0, 1) min{C(0), C(1)} m{C(0), C(1)}

(1.13.8)

To us th NCD for msuring ir-wis distncs btwn imgs on just comrsss imgs
srtly nd conctntd nd obsrvs th diffrnc btwn th comrssion rsults.
Method that Using IC s n roimtion for ntroy
rcticl roimtion of ntroy cn b ttind by fiing som modl which
roimts th log-df. W roos hr to us this roch, in connction with th modl of
indndnt comonnt nlysis (IC), or quivlntly srs coding. Ths modls r widly
usd in sttisticl img modlling. In IC, th df is roimtd s

(1.13.9)
whr n is th dimnsion of th sc, th w r linr fturs, collctd togthr in th mtri
i

W.

Th function G is non-qudrtic function which msurs th srsity of th fturs;

tyiclly G ( u ) = |u| or G ( u ) = logcosh(u) r usd. Th lttr cn b considrd s

smooth roimtion of th formr, which imrovs th convrgnc of th lgorithm.
numbr of lgorithms hv bn dvlod for stimtion of th IC modl, in rticulr th
mtri of fturs W.

ftr th modl hs bn stimtd, w cn thn roimt th comlity of s

(1.13.10)

whr th cttion is tkn, in rctic, ovr th sml.

n intuitiv intrrttion of th nsuing comlity msur is lso ossibl. First, not tht in
IC, th vrinc of th w is fid to on. Th first trm on th right-hnd-sid in (10) cn thus
T

b considrd s msur of srsity. In othr words, it msurs th non-Gussin sct of

th comonnts, comltly nglcting th vrinc-covrinc structur of th dt. In fct, this
trm is minimizd by srs comonnts. Wht is intrsting is tht th scond trm dos
msur th covrinc structur. In fct, w hv in IC th wll-known idntity

1.12.11
whr C() is th covrinc mtri of th dt. This formul shows tht th scond trm in
(1.12.10) is siml function of th dt covrinc mtri. In fct, log | dt W| is mimum if
th dt covrinc hs minimum dtrminnt. minimum dtrminnt for covrinc
mtri is obtind if th vrincs r smll in gnrl, or, wht is mor intrsting for our
uross, if som of th rojctions of th dt hv vry smll vrincs. Sinc in IC, w
constrin th vrincs of th comonnts to b qul to on, only th lttr cs is ossibl.
Thus, our ntroy msur bcoms smll if th dt is concntrtd in subsc of limitd
dimnsion.
Thus, this msur of ntroy (comlity) is smll if th comonnts r vry srs, or if th
dt is concntrtd in subsc of limitd dimnsion, both of which r in lin with our
intuition of structur of multivrit dt.
rcticlitis Rmmbring

th idl comlity distnc in q. 1 w rsnt som rmrks bout

th us of IC modl.
ssuming tht w wnt to stimt th distnc btwn two imgs, w stimt th IC modl
from both imgs srtly nd combind.
Th comlity vlu tht w gt using q. 10 is normlizd in similr mnnr s th NCD in
q. 8.
In rctic th IC modl for imgs is stimtd from dt tht contins lrg numbr of
rndomly smld img tchs.[3]

1.13. The method of complexity: 2x2 matrices with binary entries[4]

We apply the model for complexity to the simplest example: the set of all 2X2 matrices
with entries either 0 or + . Consider the four distinct matrices that represent all possible matrices
of this type. We count arrays that can be transformed into each other by rotations or reflections,
and by changing +'s into 0's, as similar. The four distinct examples are labeled (1) through (4) in
Figure 1.14.1.

Fig. 1.14.1. Four distinct 2X2 arrays with 0 and + entries.

0 0

+ 0

+ +

0 0

0 0

0 0

The design temperature T measures the degree of internal contrast; the density of
differentiations; the smallness of subdivisions. This is the raw amount of information contained
in each design. The least complex pattern is a uniform placement of the same symbol over the
array, so a measure of temperature is based on the number of symbols in a region the basic
criterion is that a uniform surface, which has no differentiations, should have T = 0. We define T
as the number of different elements (smallest units) minus one. The smallest units in this
example are the 0 or + symbols. T for the four 2X2 arrays takes values 0 or 1 .
T(1) = 0, T(2) = T(3) = T(4) = 1

(1.14.2)

The design harmony H is the symmetry content of each square, measured according to
the following criteria. If any individual symmetry is present, it rates a 1, otherwise a 0.

h1 = reflectional symmetry about the x-axis

h2 = reflectional symmetry about the y-axis

h3 = reflectional symmetry about the diagonal y = x

h4 = reflectional symmetry about the diagonal y = -x

h5 = 90 rotational symmetry (either +90 or -90)

h6 = 180 rotational symmetry

H is the sum total of all the individual contributions hi . Each hi takes values of 0 or 1, so
H in this example can take values from 0 to 6. For the four arrays 1 through 4 in Figure 1, we
have:

H(1) = 6, H(2) = 1, H(3) =1, H(4) = 3

(1.14.3)

The life and complexity are computed from equation (1) using T from Equation (2) and H
from Equation (1.16.3).

array

(1)

(2)

(3)

Life L

Complexity C

Fig 1.14.2 Values of L and C for the four 2X2 arrays in Figure 1.

The relative numbers have meaning for comparing between different arrays. Square (4)
has the greatest degree of life L - three times that of array (2) or (3). The reader will agree that
this is our intuitive feeling for the relative degree of interest among all four arrays. Array (4) is
more interesting to most people than either (2) or (3). On the other hand, array (1) generates no
interest at all because of its uniformity, and that is reflected in its null L value.
The complexity C distinguishes only between array (1), which has no complexity, and the
other three, which have about the same degree of disorganized complexity.
These calculations reveal L to be an important quantity in visualization: it measures the
difference between organized and disorganized complexity. To the best of the author's
knowledge, this is the first time that this is done. Our measurements show that the model clearly
links what we perceive as the visual interest or "life" of a design with a numerical value that can
be computed. This is the key to our model, and underlines why we consider it so important. The
complexity C does not distinguish between what is interesting and what merely has substructure.
These points become strikingly evident in more complex examples, which we consider next.

Method that Using hierarchy to generalize to higher dimensions

The rest of this paper considers 6X6 arrays. One basic component of problem reduction is
geometric: in addition to the entire field of 36 cells we consider smaller arrays which are similar
to the original array. Thus we look at the nine 2X2 arrays and four 3X3 arrays. We call these
arrays subblocks. Values for the complexity of these geometric arrays will be computed in the
next section, where we define the complexity of a complete pattern design. A second type of
problem reduction is symbolic: we count how many symbols the area being searched has. The
area may be the entire pattern of 36 cells, a subblock, or a series of cells associated by an
algorithm.

The two types of problem reduction must interact in any computation of complexity,
since some subblocks may have the full alphabet of symbols while near groupings of cells with a
single symbol might be present. These subblocks permit a recursive formulation based on the
symbols: the entire pattern is a 3x3 array whose elements are 2x2 subblocks. A measure of the
total pattern complexity is based on this subblock reduction.
The design temperature T equals the number of different symbols minus one. These
measurements have to be done hierarchically on three different scales. First on the 6X6 level,
then on each of four 3X3 subblocks, then on each of nine 2X2 subblocks. It will be useful to
label these subblocks in terms of letters and numbers. Let the index n take values a, b, c, d,
and m run from 1 to 9. The regular 3X3 and 2X2 subdivisions of a 6X6 matrix will be denoted as
follows:

Fig. 1.14.3. Subdivisions of a 6X6 array into 3X3 and 2X2 subblocks.

For simplicity, we will ignore all other subblocks. There exist 4X4 subblocks also, and
other 3X3 and 2X2 subblocks that do not coincide with the above subdivisions. (A more accurate
count should include all possible subarrays, which are indeed picked up by the human eye when
the mind computes L and C).
For each different size, the T or H value is the sum of all the T or H values for the
different subblocks.
T(3X3) = SIGMA n = 14 Tn(3X3), T(2X2) = SIGMA m = 19 Tm(2X2)

(1.14.4)

The same goes for H. The three contributions are then combined as a weighted sum of each
different scale.
T = T(6X6) + T(3X3)/2 + T(2X2)/3

(1.14.5)

The weights are 1 for the 6X6 array, 1/2 for the 3X3 subblocks, and 1/3 for the 2X2
subblocks. These give the proper equipartition when we count the subdivisions in terms of their

width. (The significance of this weighting for 1/f scaling will be discussed in a separate paper).
Note that we are not finding the average over the number of matrices. Thus, even though there
are four 3X3 matrices, their width is 1/2 of the original 6X6 matrix, so we divide by 2 instead of
4. Similarly, the width of the 2X2 matrices is 1/3 of the original 6X6 matrix. We will use the
above combination for computing both T and H totals.
One could ignore weighting altogether, and simply add all contributions from all sizes of
submatrices. However, that would skew the numbers so that the smaller elements contribute
much more than the larger elements. Elements of different size contribute simultaneously to our
perception of the whole, so it is necessary to count them in the proper balance.
The harmony H is generalized from the previous example by including measures of
similarity at a distance. Different subblocks may interact with each other. This makes it
necessary to count translational symmetry, which did not apply when dealing with isolated 2X2
arrays. In addition to the six symmetry measures h i , i = 1,...,6 given in the previous section, we
introduce three measures of translational symmetry:

h7 = similarity to another element (yes or no gives a 1 or 0)

h8 = relation to another element by a translation, plus a reflection about either the x-axis
or the y-axis (a "glide reflection").

h9 = relation to another element by a translation, plus a rotation by either +90, -90, or

180.
Note that h8 and h9 will sometimes double-count h7 in cases of high symmetry. That is

justified mathematically. Two different subblocks may be similar as they are oriented, and also
be similar after a reflection or a rotation. A subblock may be related by glide reflection to another
subblock, and by glide rotation to yet another, which counts as 2. (We do not consider each glide
rotation by different multiples of 90 separately, because that would lead to more complication
than we want in this model. Also, empirical experiments show that glide reflections about the
two diagonal axes do not provide a strong visual connection, and for that reason they are not
counted here).
The design harmony is defined as the sum of the hi, i = 1,..., 9. Each hi takes values 0 or 1,
so H for a given array (of any size) ranges from 0 to 9. As in the case of T, these computations
have to be done on three different levels, 6X6, 3X3, and 2X2 (Equation (1.16.4)), then combined
with the appropriate weights in Equation (1.16.5).

Six 6x6 arrays with four different entries

The complexity model is applied to the six 6X6 arrays listed in Fig. 1.16.3. Each location
(cell) contains one of four symbols. The goal is to guess the complexity accurately after as short
a visual inspection as possible: the result obtained is a rank-ordering of the arrays in terms of
decreasing complexity. The extreme scores attainable, maximum and minimum, become more
widely separated as more information is put into the model.

++++++

++00**

+#*#*#

++++++

++00**

#0#*#*

++++++

######

*#++*#

++++++

######

#*00#*

++++++

**0#++

*#*#+#

++++++

**+*++

#*#*#0

+0#*0#

++++++

+#**#+

*#0+**

+0000+

#0##0#

#0*#0+

+0*#0+

*#++#*

0#+0+#

+0#*0+

*#++#*

+*+#**

+0000+

#0##0#

#0+*0+

++++++

+#**#+

Figure 1.14.4. Six 6X6 arrays with four different entries.

Example Fig. 1.16.4 is of high complexity, constructed by an algorithm based on the

rings: 12 elements, B: 20 elements, Y. Each ring is a cycle of all four symbols, followed by a
first-to-last shift to begin a new cycle; the middle ring B starts with a different cycle, than: the
outer ring Y. Example 5 is low in complexity: two types of structure present are the rings of +
and 0. Example (6) is prepared from array number (3) by taking its (3X3) subblock (a) and
reflecting it four ways. Array (6) consequently has a very high degree of internal symmetry,
though still lower than the totally symmetric array (1).

The computations for T and H are straightforward, and all details are given in the
Appendix. Unlike the simple 2X2 case treated earlier, this is not an exhaustive classification of
all possible 6X6 arrays with four different entries. We just pick a sample of arrays to show how
the method works in practice. The matrices chosen have very different internal structure that
illustrates various possibilities.
Before reading further, we suggest that the reader study the above matrices and rankorder them in terms of decreasing C and L. Remember that C measures the intensity of design
complications. In art, C measures the level of visual excitement, which often arises from chaotic
aspects of a design. The life L measures the degree of organized complexity in a design; the
visual interest comes from the degree to which elements interact coherently. The name "life" is
chosen because continuing to increase L mathematically brings one closer to the structure of
living organisms. This comparison is useful for the test that we propose. The reader can decide
which of the above six arrays most resembles something that could be organic, then rank-order
them in L based on this impression.
This example requires the following amended definition of C and L, instead of Equation
(1.16.1):
L = T H , C = T ( 50 - H )

(1.14.6)

array

(1)

(2)

(3)

(4)

(5

Temperature T

8.5

11.7

16.3

Harmony H

51

23

15.7

3.3

Life L

196

184

54

Complexity C

230

401

761

Table 1.14.5. Computation of T, H, L, and C for the six arrays in Figure 1.16.3.

These numbers are larger than the earlier 2X2 case, and that reflects the existence of
substructure. The amended definition of C = T(50 - H) instead of Equation (1.16.1) is necessary
so that our numbers are all positive. The number 50 is arbitrary, and is needed only as an upper
scale. (To be more precise, we should have chosen 51, the maximum H from array (1) in Table
1.16.2, but that makes little difference for the comparison).
The six arrays can be rank-ordered in terms of decreasing L and C.
Ranking in decreasing L: (6), (2), (3), (5), (4), (1).
Ranking in decreasing C: (4), (3), (5), (6), (2), (1).

These measures are in accord with our perception of pattern complexity. Most readers
will agree that these rankings correspond to what they have already concluded from direct
observation. The six examples are decreasing in complexity in the same order as our feelings. We
demonstrate here a strong correlation between the subconscious process of perception and a
simple quantitative model. Our model can be refined by incorporating more and more input, but
even at this stage, it is remarkably accurate in predicting our emotional response to a design.
Even though the design temperature of array (6) is high, the number of internal
symmetries organizes the complexity so that C is lowered and L is raised. Contrast this to the
high-T, low-H array (4) - it has very little internal organization, which raises C and lowers L.
Array (6) shows how C essentially differs from T. Could one not skip the additional
complications of measuring symmetries in this model and simply compute T as the complexity
of a design? The answer is no, because the ranking in decreasing T is 4, 6, 3, 5, 2, 1, which is
entirely distinct from the ranking in decreasing C. Without example (6), however, the two
rankings would coincide.
Superficially, arrays (3), (4) and (6) look pretty similar. Their complex substructure
emerges after looking at them analytically for some time. The more we look, the more hidden
substructure is discovered as the mind moves the symmetry links from its subconscious to its
conscious cognitive state. The point is that we perceive all of this information in an instant. By
asking ourselves which array most resembles organic forms we can make an immediate
assessment without having to analyze whether the substructure is organized or disorganized. The
human mind is evidently programmed and equipped with tools to make this judgment.[4]

1.14.

The analysis of existing methods of the image complexity determination

Method
1. Method based on
Entropy

2. rnc of
gry lvl

3. rnc of
dgs

Description

Advantages

This method give us an globl lgorithm

This roch lds to th us of stndrd

tht cn outut singl vlu for ch

distnc functions, which rsct th usul

filtrd imgs, whil rsrving its clss of

rortis of idntity, symmtry nd

comlity.

tringulr inqulity, ugmntd by ntroic

Disadvantages

functions.
Th rnc of gry lvl cn rflct th Th mor informtion in th img, th

Th informtion ntroy is usd to

gry lvl.

mor coml in th img will be will be

dscrib th informtion contind

beneficial on the image complexity

in gry lvl only.

determination.
Th quntity nd th comlity of th With this method w cn tk dvntg of
trgt cn b rssd with dgs.

dg ntroy to chrctriz th rnc

of th trgt. Th rnc of th trgt

4. Rndomnss of
img tturs

cn rflct th imgs comlity.

For ml, som imgs hv th sm Informtion ntroy nd dg ntroy cnt Fot this method w should hv
informtion ntroy nd dg ntroy but rss th sc distribution of gry lvl.

som othr chrctristic to

thy hv diffrnt img comlity.

rss th sc distribution of
gry lvl.

5. 2x2 matrices

We apply the model for complexity to the These calculations give us an important This method is used more on the

with binary

simplest example: the set of all 2X2 quantity in visualization: it measures the temperature estimation measures

entries

matrices with entries either 0 or + . Consider difference

6. Method that
Using IC s n
roimtion
for ntroy

between

organized

and the degree of internal contrast; the

the four distinct matrices that represent all disorganized complexity.

density of differentiations; the

possible matrices of this type.

rcticl roimtion of ntroy cn b

smallness of subdivisions.
Ths modls r usd onli in
sttisticl img modlling.

ttind by fiing som modl which

roimts th log-df.

We can us this roch, in connction

with th modl of indndnt comonnt
nlysis (IC), or quivlntly srs
coding.

7. Method that
Using hierarchy
to generalize to
higher

This method using 6x6 arrays. One basic

component of problem reduction is
geometric: in addition to the entire field of
36 cells we consider smaller arrays which
are similar to the original array.

This model introduced a useful

distinction for discussing
complexity in theoretical terms.

dimensions

8. 6x6 arrays with

four different
entries

9. Image
partitioning

The goal is to guess the complexity

accurately after as short a visual inspection
as possible: the result obtained is a rankordering of the arrays in terms of decreasing
complexity. The extreme scores attainable,
maximum and minimum, become more
widely separated as more information is put
into the model.

6X6 case showed the considerable power of

the model. We are in fact measuring the
organizational entropy (degree of disorder),
which is the negative of the degree of
connections established via visual
symmetries.

This method is also used more on

the temperature estimation
measures the degree of internal
contrast; the density of
differentiations; the smallness of
subdivisions.

In this method, we present a greedy

algorithm which partitions an image in
quasi-homogeneous regions. The optimal
partitioning algorithm is NP-complete. To
do this partition, a natural approach could
consider the above channel as the starting
point for the image partitioning, designing a
pixel clustering algorithm which minimizes
the loss of MI.

This process can be interpreted in the following way: the choice of the partition
which maximizes the MI increases the
chances of guessing the intensity of a pixel
chosen randomly from the knowledge of the
region it pertains to.

The algorithm of this method not

generates a partitioning tree for a
given probability of error Pe.

1.15.

Conclusion

The many advantages of our approach include that: it shows sensitivity to habitat features
at the community level; it is inexpensive, simple and accessible to all; it allows for the
monitoring of forests at multiple scene scales in both space and time; it can provide additional
information of ecological relevance to sensor networks; it can be added to the bag of sampling
devices of most field protocols; and the use of structural complexity as an EO is practically and
theoretically attractive. The list of disadvantages includes that: it is a methodological approach in
its infancy that will require confirmation from other systems; photographic settings will have to
be fully standardized and their calibration addressed (e.g., image resolution versus extent); and
MIG estimates will need to be correlated to other measures of plant architecturesuch as canopy
closure, canopy cover and vertical structureto further interpret themechanisms beyond our
present definition of structuralcomplexity in an image. Monitoring forest dynamics at a high
resolution in spaceand time offers the possibility of discerning the ecologicalsignature of these
systems. Signature variations could provideinformation on the integrity and stability of
ecologicalprocesses, both globally and locally. The detection of localdisturbances assessed by a
change in structural complexitycould help alert ecologists and guide their actions to siteswhere
the integrity is threatened. By revisiting the same sitesweek after week one quickly realizes how
dynamic anecosystem may be. Phototropism, flooding events, springand fall phenology, growth
and senescence, flowering time,grazing and disease perturbations, falling trees, and gapdynamics
are some of the many processes that structure theforest habitat on a relatively short temporal
window. A holisticapproach capable of integrating these processes in time and space would
certainly benefit scientists and decision makers.

Bibliografie:
1. Honghai Yu, Stefan Winkler, IMAGE COMPLEXITY AND SPATIAL
INFORMATION ( ECE Department, University of Illinois at Urbana-Champaign, USA
Advanced Digital Sciences Center (ADSC), University of Illinois at Urbana-Champaign,
Singapore 2011).
http://vintage.winklerbros.net/Publications/qomex2013si.pdf

2. Perreira Da Silva, Vincent Courboulay, Pascal Estraillier IMAGE COMPLEXITY

MEASURE BASED ON VISUAL ATTENTION(Matthieu. IEEE International
Conference on Image Processing - ICIP, Sep 2011, Bruxelles, Belgium. to be published,
2011).
https://hal.archives-ouvertes.fr/hal-00617725/document

3. Jukka Perki , Aapo Hyvarinen Modelling image complexity by independent component

analysis, with application to content-based image retrieval (Helsinki Institute for
Information Technology, Department of Computer Science, Department of Mathematics
and Statistics).
https://www.cs.helsinki.fi/u/ahyvarin/papers/Perkio09ICANN.pdf

4. Allen Klinger, Nikos A. Salingaros Complexity and Visual Images (Department of

Computer Science, University of California at Los Angeles 2013).
http://web.cs.ucla.edu/~klinger/image_complexity_7_1_97.html.bak

5. Richard Alan Peters , Robin N. Strickland Image Complexity Metrics for Automatic
Target Recognizers(Department of Electrical and Computer Engineering Vanderbilt
University Nashville, TN 37235, Department of Electrical and Computer Engineering
Vanderbilt University University of Arizona Tucson, AZ 85721 2008).
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.39.5833&rep=rep1&type=pdf

6. Maurizio Cardaci Vito Di Ges`u, Maria Petrou, and Marco Elio On the Evaluation of
Images Complexity: A Fuzzy Approach(Tabacchi Universit`a di Palermo, C.I.T.C., Italy
2010).
http://www.math.unipa.it/~digesu/Time_work.pdf

7. Chikhman V. , Bondarko V., Danilova M. , Goluzina A. , Shelepin Y. Complexity of

images: experimental and computational estimates compared(Pavlov Institute of
Physiology, Russian Academy of Sciences 2012).
http://www.ncbi.nlm.nih.gov/pubmed/23094454

8. J. Rigau, M.Feixas, M.Sbert Theoretic Framework for Image Complexity (Institut

d'Informtic
ai
Aplicacions,
Universitatde
Girona,
Spain2005).
https://www.researchgate.net/publication/220795259_An_InformationTheoretic_Framework_for_Image_Complexity

9. Mario Ignacio Chacn Murgua, Alma Delia Corral Senz and Rafael Sandoval
Rodrguez Chihuahua A Fuzzy Approach on Image Complexity Measure(Institute of
Technology,
DSP
&
Vision
Laboratory
2009).
http://www.scielo.org.mx/pdf/cys/v10n3/v10n3a6.pdf

10. Veaceslav L. Perju Determination of the image complexity feature pattern

recognition(2003).
http://www.math.md/nrofdownloads.php?file=/files/csjm/v11-n3/v11-n3-(pp231-247).pdf

11. Roque Marin, Eva Onaindia, Alberto Bugarin Current Topics in Artificial Intelligence(
11th Conference of the Spanish Association for Artificial Intelligence 2005).
https://books.google.md/books?
id=F2rLaFtIrd0C&pg=PA347&lpg=PA347&dq=image+complexity+determination&sour
ce=bl&ots=WccGEEW40C&sig=zZVBXQmOhMBc2hNHKffgKVMk02s&hl=ru&sa=X
&ved=0ahUKEwiGqfbt4MfLAhUljXIKHYJBBPIQ6AEISTAJ#v=onepage&q=image
%20complexity%20determination&f=false

12. Marin, Eva Onaindia, Alberto Bugarin Metode de Prelucrare a Semnalelor

(Universitatea Tibiscus Timisoara 2007)
http://anale-informatica.tibiscus.ro/download/lucrari/1-2-18-Mark.pdf

13. Agata Ciokosz-Styk1, Adam Styk Advanced image processing for maps graphical
complexity estimationInstitute of Geodesy and Cartography, Poland
http://icaci.org/files/documents/ICC_proceedings/ICC2013/_extendedAbstract/260_proce
eding.pdf

14. Universitatea Babes-Bolyai, Faculutatea si Informatica Prelucrarea Semanlelor II

http://www.cs.ubbcluj.ro/~per/BN/Prel_Img%20II.pdf

15. Ioan Ispas Modelare i modele matematice n recunoaterea obiectelor i clasificarea

automat a imaginilor (Catedra de Matematic-Informatic, Universitatea Petru Maior,
Trgu
Mure).
http://www.upm.ro/facultati_departamente/stiinte_litere/conferinte/situl_integrare_europ
eana/Lucrari/Ispas.pdf

16. Constantin VERTAN PRELUCRAREA SI ANALIZA IMAGINILOR 2001.

http://imag.pub.ro/ro/cursuri/archive/carte_pai.pdf

17. K C
https://ru.wikipedia.org/wiki/%D0%9A%D0%BE
%D0%BB%D0%BC%D0%BE%D0%B3%D0%BE%D1%80%D0%BE
%D0%B2%D1%81%D0%BA%D0%B0%D1%8F_%D1%81%D0%BB%D0%BE
%D0%B6%D0%BD%D0%BE%D1%81%D1%82%D1%8C

18. .. , ..

.
http://izvestia.asu.ru/2013/1-1/info-comp/TheNewsOfASU-2013-1-1-info-comp-09.pdf

19. . . , . .
( . . .
, , 15.11.2012).
http://www.jip.ru/2012/362-371-2012.pdf

20. . . , . .
( , .
2013).
http://jre.cplire.ru/iso/feb13/9/text.html

21. .., .., ..

http://www.fesmu.ru/elib/Article.aspx?id=204188

22.

https://habrahabr.ru/post/229757/
23. ., . . CGM
Journal, 2003.

24. Gupta D. Computer Gesture Recognition: Using the Constellation method.// Caltech
undergraduate Research Journal, 2001, vol.1, 1. pp. 26-31.

25. Vezhnevets V. Face and facial feature tracking for natural Human-Computer Interface.//
GraphiCon 2002.

26. .., .., ..

web-.// GraphiCon 2002.

27. Song Qinghuan, Chen Zhongbia Sun Shaohuac, Hu Jinlonga, Cheng Huaa, Institute
of Optics and Electronics, Chinese Academy of Sciences, Chengdu, Sichuan A scene
recognition method based on image complexity China University of Chinese Academy
of Sciences, Beijing , China
http://spie.org/Publications/Proceedings/Paper/10.1117/12.2069491