An Automated algorithm for the identification of artifacts in mottled and noisy images

Rochester Institute of Technology

RIT Scholar Works

Theses

Thesis/Dissertation Collections

7-12-2006

An Automated algorithm for the identification of

artifacts in mottled and noisy images

Onome Augustine Ugbeme

Follow this and additional works at:

http://scholarworks.rit.edu/theses

This Thesis is brought to you for free and open access by the Thesis/Dissertation Collections at RIT Scholar Works. It has been accepted for inclusion in Theses by an authorized administrator of RIT Scholar Works. For more information, please [email protected].

Recommended Citation

AN AUTOMATED ALGORITHM FOR THE IDENTIFICATION OF

ARTIFACTS IN MOTTLED AND NOISY IMAGES

Approved by:

By

Onome Augustine U gbeme

A Graduate Thesis Submitted

m

Partial Fulfillment

of the

Requirements for the Degree of

MASTER of SCIENCE

m

Electrical Engineering

Professor

Eli Saber

---~--~---~~---Graduate Thesis Advisor

-

Dr. Eli Saber

Professor

_{---~----~~~~~~}

S. A. Dianat

Graduate Committee member - Dr. Sohail Dianat

Prof6ssor

_D_a_n_i e_I_P_h_i

I_li~p_s

___

Graduate Committee member

-

Dr. Daniel Phillips

Vincent J. Amuso

Professor ______

---,----~---,----=---_=_=_----:-_=__~:-Electrical Engineering Dept. Head - Dr

.

Vincent Amuso

Department of Electrical Engineering

COLLEGE OF ENGINEERING

ROCHESTER INSTITUTE OF TECHNOLOGY

ROCHESTER

,

NY

THESIS RELEASE PERMISSION FORM

ROCHESTER INSTITUTE OF TECHNOLOGY

COLLEGE OF ENGINEERING

Title: An Automated algorithm for the identification of artifacts in

mottled and noisy images

I, Gnome Augustine U gbeme, hereby grant permission to the

Wallace Memorial Library to reproduce my thesis in whole or part.

Signature:

Date:

Onome Ugbeme

1

/

~o

I

fJ{,

Abstract

Thisthesisproposes an algorithmfor_{automatically classifying} a specific set ofimage

quality

(IQ)

defects on_noisy and mottled printed documents. A roughinitial estimate of

thedefects' locationi.e. defect detection is_manuallyprovidedwith a scanned image.The

approach then proceeds to derive a more accurate segmentation

by

_performing several

image _processing routines on the digital image. This reveals regions of interest from

which

discriminatory

features are extracted. A classification offour defects is achieved

via a customized tree classifier which employs a combination of_size, shape and region

attributes at _{corresponding} nodes to yield appropriate

binary

decisions. Applications of

this process include automated/assisted diagnosis and repair ofprinters or copiers inthe

field in a

timely

fashion. The proposed algorithm is tested onadatabaseof261 scanned

images provided

by

Xerox Corporation and several synthetic defects _yielding a 96.6%

classification rate.

Acknowledgements

I extend _myutmost gratefulness to Dr. Eli

Saber,

_my thesis advisor, for making this

thesis possible, providing valuable advice and _{supporting my}research. I would also like

to thank Dr.

Wencheng

Wu of Xerox Corporation from whom I have gained an appreciableknowledge inimage_qualityprocedures and guidelines. Special thanks toDr. Daniel Phillips and Dr. Sohail Dianat for

being

a part of_my committee and _providing

helpful comments and questions that helped improve this thesis. The Center for

Electronic

Imaging

Systems

(CEIS),

a NYSTAR-designated Center for Advanced

Technology,

Xerox

Corporation,

and the Electrical

Engineering

Department at the

Rochester Institute of

Technology

supported this project and created the _opportunity to

explore this field.

Finally,

I would like to recognize _my

family

for the unconditional

support

they

gave methroughout theyears.

Table

of

Contents

List

of

figures

2

List

of

tables

3

1.

Introduction

4

2.

Background

10

2.1.

Image

Quality

Defects

10

3.

Methodology

15

3.1.

Image

Pre-processing

17

3.2.

Region

of

Interest Identification

and processing....24

3.3.

Outlier

Processing

26

3.4.

Final Classification

via

Morphological

Reconstruction

28

4.

Experimental

results and

discussions

32

5. Conclusion

38

List

of

figures

Figure 1.1 Sample scanned printed document

highlighting

_manually cropped

regions ofinterest 5

Figure 1.2 Typical Content Based Image Retrieval

(CBIR)

_setup 6

Figure2.1 Defectsofinterest 13

Figure 2.2

Varying

sizes ofDCDdefect 14

Figure 3.1 Defectsegmentationandidentificationframework 16

Figure 3.2

De-screening

process 19

Figure3.3 ComparisonofPCswithalternative grayscale method 22

Figure3.4 Local Normalization 23

Figure3.5 Local Normalization block diagram 24

Figure3.6 ROIfeatureextraction 26

Figure 3.7

Study

ofSNR 26

Figure3.8

Missing

debrisidentificationprocedure 28

Figure 3.9 Withinmask segmentation 29

Figure 3.10 Withinmask segmentation 29

Figure 3.11

Obtaining

thedifference image 30

Figure 3.12 Differenceimages 31

Figure4.3 Additionof

increasing

levelsof gaussian noise and_resulting

List

of

tables

Table 4.1 Classificationresults_usingour proposedmethod 35

Chapter 1

Introduction

Image quality stands_amongoneofthemostimportantattributes ofimage acquisition

and _{printing devices. In} the past

decade,

we have seen a tremendous increase in the

quality ofprinted documents due to significant technological advances in non-impact

printing. Present

day

print engines are required to meet consistent and stable image

customers. The current marketplace demands the best image _quality atcompetitive costs

with minimum downtime.

Hence,

the_ability forprint engine vendorsto _reliably achieve

the highest level of image _quality will ensure them a

leadership

role in the _printing

industry.

Figure1.1 Samplescanned printeddocument

highlighting

_manuallycropped regions ofinterest

containingdefects.

However,

even though the _quality of printed documents has

improved

significantly

overthepast

decade,

current printengines still produce a_varietyofimage _qualitydefects [image:10.535.56.485.152.563.2]

and artifacts (see Figure 1.1). These artifacts are manifested in a _variety of ways and

occur in different

locations

on the printed document. Operator and/or engineer

intervention isoften requiredtoperform corrective actionstoeliminate orreducetheirre

occurrences. In some _cases, new defects never encountered before are diagnosed and

resolved in a

timely

manner. This

dependency

on a trained operator that is capable of

identifying

artifacts and _performing the appropriate corrective action as _quickly as

possibletends to hinderthe efficient flow ofbusiness inprint shops for instance.

Hence,

theneedforonline algorithms capable of_{automatically}

identifying

artifacts and_outlining

corresponding corrective measures is _continuously growing. This will provide end-users

with the _ability to

independently

make a diagnosis of printers or copiers in a

timely

fashion.

One approach to

tackling

this problem is

by

_{utilizing Content Based Information}

Retrieval

(CBIR)

techniques. In an effort to subdue manual annotations oflarge image

databases,

research interests in this area have abounded since the _early 1990s.1'2CBIR

employs visual properties of a _query image

-color, shape, texture,

frequency,

and

regions _ofinterest

(ROI)

-to traverse _{image databases according}to user's interests (see

Figure 1.2). This is achieved

by

_utilizing

highly

descriptive multidimensional feature

vectors which can beextracted from global orlocal positionswithinthe image_using one

or more oftheabove visual content properties. Someapplications requirefeaturesthatare

minimally sensitive to changes

-affine, reflections, perspective,

deformations,

or

luminance3

- inthe image. For

instance,

shapefeatures can be represented with moment

invariants,4

Fourierdescriptors,5or

chain-codes.6

Similarity

measuresbetweenthefeature

computed. Often useddistance metrics include Euclidean,6'7 Minkowski,8 Mahalanobis8

and Kullback-Leibler

(KL)

divergence.10

The decisionof a"matched" image is achieved

by

searching through the similarity values. This can be performed _using an exhaustive

approach or more _efficiently via

indexing

schemes. Popular

indexing

schemes include

R* quad-trees,12

or

Self-Organizing

Maps (SOMs).13' 14

Thus,

an automatic

defect identification system canbe devisedtocompareadefective samplewithimagesin

an _{already existing} online database from which

top

matches can be selected.

Additionally,

the system can be designed to

dynamically

perform updates when new

defectsare encountered.

[image:12.535.71.468.339.526.2]

Userinput Image _Features Extraction

Similarity

Matching

K

Indexing

& Retrieval ? Database Images i k Retrieved results Features Extraction Database ? Features

Figure 1.2 Typical Content Based Image Retrieval

(CBIR)

setup.

Some systems counteract the defect recognition problem via a series of image

processing and pattern classification steps. Iivarinen and Visa14

employed unsupervised

SOMs forthe classification ofbase paper surface defects such asholes and spots. Their

classificationisbasedonfeaturesextractedfromtheinternal _{structure, shape,} andtexture

-7-traits ofthedefective areas. Segmentationand morphological operations are amongstthe

pre-processing steps

they

carried out prior to feature extraction and classification.

Additionally,

welddefect detectionhas beenaccomplished

by

Li and Liao15

viaGaussian

distribution fittings ofhorizontal line profilesfrom thedefect images.

They

also _employ

background_removal, dark imageenhancementandimagenormalizationbeforeproperties

oftheprofiles are exploited as features.

Mery

and

Berti16

used texturalfeatures from co

occurrence matrices and 2-D Gaborfunctions. Prior to the feature extraction _stage,

they

explore Laplacian of Gaussian edge detection to segment "key"

defective areas. More

recently, Ng17

proposed a novel histogram-based

thresholding

scheme to assist in the

segmentation of"smallsized"

sheet metaldefects. The successfulness ofthisapproachis

however limited to high contrast images containing a defect. In general, the algorithms

described above are designed to handle defects that possess a high level contrast to the

surrounding background.

Furthermore,

the algorithms cannot _{be successfully} applied to

the recognition of all printed documents defects which sometimes have a subtle

appearance.

This thesis describes a method to

identify

artifacts resident on scanned copies of

printed documents. These artifacts tend to have large intra class _variations, possess low

contrastor exhibitilluminationvariations. It is assumedthat the defect localizationstage

i.e. arough enclosure ofdefectiveareas, has already beenperformed apriori

by

ahuman

operator. Thus the proposed method is responsible for _{accurately segmenting}the defect

and_providingalabel

by

classifying

discriminatory

features extractedfromthe segmented

regions. To achieve this, the algorithm employs various image _processing steps such as

amongst samples of the same class and facilitate subsequent _shape, size and region

featureanalysis. Theeffectiveness ofthealgorithmis demonstratedon a sampledatabase

of scanned electrophotographic and synthetic images. These images possess significant

variations within each classto

thoroughly

test the _sensitivityofthe methodto intra-class

differences.

Additionally,

this approach differentiates between mottle and deletion

imagesandfurther identifiesthesub-classes withinthe deletioncategory.

The rest ofthisreportisorganized asfollows: Chaptertwo provides abriefoverview

and discussion of the nature of the defects at hand. Chapter three describes the

methodology oftheapproach which includesthe _{image pre-processing} steps carried_out,

the procedures usedto extracttheRegions ofInterest

(ROIs),

the_{reasoning behind} class

decisions _making and a validation ofchosen features. In chapter

four,

a comparison

betweentheproposed grayscale formulationvia principal components and an alternative

grayscale approach is given. This section also discusses the final classification results.

Chapter 2

Background

2.1. Image

Quality

Defects

Printer image quality defects are undesirable qualities of an electrophotographic _copy

produced

by

aprinter or copier. Even though these print engines have been

thoroughly

tested

during

the manufacturing process, theoccurrenceofdefectsremainsinevitable due

to the volume and

diversity

of printed material. Thus a number of standards have

emerged to define objective evaluation methods for print quality. The ISO-1366019

standardisa_widelyaccepted standard fortheobjective characterization ofimage _quality

-10-as it pertains to office equipment. It categorizes these qualities as either: character and

line attributes or large area attributes. Character and line defects include raggedness,

blurriness,

character

darkness,

_{line width, contrast,}

fill,

and extraneous marks. While

large areadefects referto graininess, mottle, and voidto name afew. More recently, the

ISO-_{19751 image}

quality standards foroffice equipmenthas established task teams to

perform research workona more expansive _categoryofdefectsnamely:

1)

Textandline (e.g. sharpor smoothedges andproper_{text color),}

2) Macro-Uniformity

(e.g. streaks,

bands,

and mottle _{defects existing} on a scale greater

than25mm2Field Of View

(FOV)),

3)

Micro-Uniformity

(e.g. streaks,

band,

mottle, and graininess _{defects usually}restricted

toa25mm maximum

FOV),

4)

Gloss

Uniformity

(e.g. differentialgloss and gloss_artifacts),

5)

Color Rendition (e.g. color

fidelity

andcontinuity),and

6)

Effective Resolution.

This thesis addresses a _group ofdefects that can be characterized under the

Macro-Uniformity

attributes.The defects ofinterest include

deletions,

Debris Centered Deletion

(DCD),

debris missing

DCD,

and mottle. Deletions (Figures. 2.1a

-2.1c)

are _usually

manifested as elliptical regions _containing a localized group of pixels lighter than the

uniform background and are in general aresult ofafault in the_charging process. DCDs

(Figures. 2.1d

-2.

If),

onthe other

hand,

resemble adeletionwithan added presence of a

centralized collection oflocalized dark pixels (called a debris). The DCD appears as a

smalldarkspoton adeletion background. Aspecial case ofDCDsexists wherethedebris

is missing due to accidental "rub off

by

the electrophotographic process _revealing the

paper color(see Figures.

2.1g

and 2.1h). Thisyields _abnormally brightgray-levelvalues

compared to its immediate _neighboring pixels. Mottledefects (Figures. 2.1i

-2.1j)

refer

to _{non-uniformity in} the perceived print

density

(i.e. reflectance) and can be gauged

by

the _relationship between light and dark regions. The ISO-13660 standard 18' 19 defines

mottle as _{non-uniformity occurring} on a spatial scale between 1.27mm and 12.7mm. It

characterizes it as large area print _quality attributes _possessing aperiodic fluctuations of

density

at a spatial

frequency

less than 0.4 cycles per millimeter in all directions. Even

thoughmottle hasreceived quiteabitof attentioninthe

industry,

aquantifiable universal

measure has yet to be developed. The best achieved benchmark so far has been

introduced in the ISO-_{19751 image quality standards for} systems.20

Other defects e.g.

streaks and bands are _certainly of significant interest but are not handled in this paper

since their shape properties differ greatly from the elliptical shapes identified

by

the

proposed method.

The axial lengths ofthe deletion and (debris_missing) DCD _vary fromabout

1/32"

to

over

1/10"

(see Figure 2.2 for _example) and can occur on different locations of the

printed document (see Figure 1.1).

Thus,

any proposed recognition scheme should be

nominally invariant to the sizes and locations ofthe defects.

However,

it should also be

able to _employthe size ofthedefectas anindicationofthe defect severity.

Furthermore,

the proposed algorithm should have the abilityto process different sized digital images

(i.e. cropped areas greaterthan

25mm2

for Macro-Uniformity).

The aforementioned assumptions about characteristics of the defect i.e. elliptical

distribution,

localized spatial arrangement ofthe

deletions,

_{non-uniformity} of_{the mottle,}

andcentralizeddebris presence ofthe _{DCD (or missing}

DCD)

aretaken advantage ofin

the generation of

descriptive

features used in the classifier. The successfulness ofthis

thesis will

help

_{in quantifying} an objective approachtowards _{evaluating printing} device [image:18.535.56.488.129.499.2]

defects.

Figure 2.1 Defectsofinterest:

(a)-(c)

Deletion,

(d)-(f)

Debris Centered Deletion

(DCD), (g)-(h)

Debris-missing DCD,

and

(i)-(j)

Mottle.

[image:19.535.86.435.43.190.2]

-13-(a)

(b)

Figure 2.2

Varying

sizes ofDCDdefect:

(a)

Majoraxiallength~ 1/10"

and

(b)

Majoraxial

length- _1/12".

14-Chapter 3

Methodology

The proposed method is summarizedin Figure 3.1. It comprises offourmajor steps:

1)

pre-processing,

2)

ROI identificationandprocessing,

3)

outlier

discovery,

and

4)

final

classification. The input to the algorithm is a scanned color sample (i.e. digital

image)

containing the defect. The

flowchart,

shown in Figure

3.1,

represents a classification

approach based on the expected statistics ofthe regions of interest in the images. It

depicts five leafnodes andfour decision nodes thatare used torepresentthe classes and

-15-decisions respectively. Each of the decision nodes employs size or shape features to

classify the defect

by

_{utilizing empirically} selected thresholds. These thresholds are

obtained

by

_{selecting random samples to} constitute the

training

data,

after which the

feature of interest is extracted from the image. A threshold that separates the

binary

classes atthedecisionnodesisthenchosen.Describednext arethe_{pre-processing}steps.

1' De-screen

'

QueryRGB image Image < Pre-processing

PC Transformto

obtain grayscale

Localnormalization

Tu

I

Obtain ROIs

ROIidentification < &_processing

Compute Hausdorff Distance(HD)and Reetangularity (RECT)

Unknown Sample

Outlier

processing

Perform Debris_Missing Compensation(Optional)

Segment Difference image &

extractdebrissizefeature

Final Classification _-<^ DCD Mottle Missing DCD Deletion

|

=_{Leaf Nodes}

(output)

[image:21.535.48.489.186.641.2]

=_{Decision Nodes}

Figure 3.1 Defectsegmentationandidentification framework.

16-3.1.

Image

Pre-processing

3.1.1.

De-screening

Locating

regions ofinterests involves_applying a segmentation routineto separatethe

image into

foreground

andbackground areas.The foregroundregion consists ofpixels in

defect areas which can be identified _using a

binary

threshold value.

However,

direct

thresholding

of the images without _{pre-processing generally leads} to inaccurate ROI

selections. Scanned imagestend to containhalftone screens (or marking screens)usedto

produce the image on the substrate

during

the electrophotographic,

ink-jet,

or

lithographic marking process. : Researchers have developed

de-screening

proceduresto

counteract this problem. Dunn and

Mathew22

treated the halftone screens as texture

capable of

being

extracted _using a single circularly-symmetric filter. Sharma et

al?1

developed a process responsible for

determining

the

identity

ofthe _{underlying marking}

process

by

_analyzingthepowerspectra of adigital image forthepresenceofhigh energy

at high frequencies.

They

utilize this information for scanner or copier recalibrations in

orderto produce adocument ofhigh

fidelity

withminimal screens. The same

frequency

domain

de-screening

approach is adopted here to minimize the effect ofthe screens on

the classification process. In particular, the power spectrum ofthe image is analyzed to

determine the existence ofhigh energy content (at high

frequencies)

that represents the

signature ofthe underlyinghalftonescreens. The screens arethenreduced via a repetitive

"notch"

frequency-domain

filtering

operation to yield a _sufficiently smooth image for

further analysis. This process is performed

individually

on each channel. A Butterworth

notch

filter5

of ordern= ₁₅

givenby:

17-F = notch

1+

D

Dx(u,v)D2(u,v)

where:

Do

=

Radius of

filter,

Dl(u,v)

=

^(u-M/2-uJ

+(v-N/2-vJ

,

(1)

D2(u,v)

=

^(u-M/2

₊

u0f +(v-N/2 +_v0f ,

M =Number of_{columnsin image} ,

N=Number of_{rowsin image}

.

is utilized. The variables u and v are the

frequency

domain coordinates. The origin of

Fnotchhas beenshiftedto the center

frequency

coordinatesi.e.

Folch(0,

0)

is locatedatu

=

M/2andv=_N/2.

Thus,

thenotchlocationsare_{symmetrically located}at

(uq,

_vo) and

(-uq,

-vo). The order ischosenas n = 15in orderto preserve the contrast ofdebris pixels

intheDCDanddebris missingsamples.

An illustration of this application on a DCD image is shown in Figure 3.2. The

frequency

spectrum(Figure

3.2b)

shows significant power(outlined

by

the

boxes)

athigh

frequencies that correspond to the sinusoidal halftone screens. The black spots (Figures

3.2c

-3.2e)

are indications of successive

frequency

domain

filtering

operations.The final

de-screened image in Figure 3.2f is thus obtained fromthe inverse Fouriertransform of

thefilteredversionsofallthreechannels.

-18-Figure3.2

De-screening

process:

(a)

Originalscanned

document, (b) Frequency

spectrumofred

channel_showingpresence of abnormal_{high energy}athigh

frequencies, (c)-(e)

Successivenotch

filtering, (f)

De-screenedversion ofFigure3.2a.

3.1.2. Grayscale ConversionviaPrincipalComponent Analysis

Special instances arise where the "debris" _{is only} visible in one channel.

Hence,

a

channel

by

_channel,i.e. 3-dimensional processing oftheimage isnotdesirable dueto the

added computational _complexity and potential for inaccurate results.

Therefore,

the

image is transformed to a single grayscale channel. There are numerous RGB to

grayscale methods used in the computer vision literature. Some _{simply employ} the

averageoftheRGB channelsas a_{corresponding}grayscaleimage. Thisapproachdoes not

usuallyretainthesubtledetailssince theresultantgrayscaleisnot_perceptuallyequivalent

to thebrightnessoftheoriginalcolorimage. A betterapproximation ofthebrightnesscan

be derived

by

summingweighted versions ofthe

R,

G,

andB channels. Othermethods

handle the problem via

de-saturation24

i.e. removal ofthe saturation information ofthe

image. In particular, MATLAB's grayscale conversion

(rgb2gray)25

computes a

[image:24.535.63.471.48.311.2]

grayscale image

by

_eliminating the hue and saturation components while _retaining the

luminance. Forour

dataset,

_{these conventional methods often failtopreservethecontrast}

ofthe centered debris with respect to the deletion background.

Rather,

they

tend to be

successful when applied to normal images e.g. portraits, landscape scenes, and road

scenesthathavean appreciable contrast and exhibitamultimodal histogramproperty. To

avoid this

limitation,

a principal component

(PC)

analysis is performed on the RGB

image in orderto convert itto a single channelin an optimumfashion. Figure 3.3 shows

comparisons of the _rgb2gray method with grayscale images obtained from the PC

transformation. This DCDimage (Figure

3.3a)

bears"questionable"

debris which is _only

visible in the blue channel (Figure 3.3d). The blue channel can thus be selected via

contrast selection schemesor

by

_utilizingadditive/subtractivecolortheories5.

Principal component

analysis26

is a linear data reduction approach that _optimally

projects a^-dimensionaldataontoalower-dimensional subspaceina mean-squared error

sense. It does this

by

_performing a coordinate rotation that aligns the transformedaxes

with the directions of maximum variance ofthe data. Let

^F,,^,

and

3

represent the

R,

G and B channels in lexicographic _ordering and =

[*?,

^ 3

]

T be the

corresponding 3 x MNconcatenation. The mean vector m=

[w,

m2

m3]T

and covariance

matrixCarecomputedas follows:

I MN

^Tbl^W' *=l>2,and3.

(2)

C= o-%(jy CT T

arvaVj

_2 o\j,3

ovov ov 0%0">i'

(3)

where:

r -I i MN

. _{c72t=4K-^)2j=^i-j;(^(/)-^)2,and}

avaVt

=E^J-mX^k-mj]^1^^J(l)-mJX^k(l)-mk).

C isa_real,symmetric matrixwhichcanbeexpressed as:

C= UAUT

(4)

where U is a 3 x ₃ orthonormal matrix ofeigenvectors _{corresponding} to the ordered

eigenvalues

Xi

>

fa

>

fa

_{contained in the diagonal matrix A} =

diagfl/,

fa,

fa). The

principalcomponentsofTarethencalculatedby:

Y=_UTx= [YlY2Y3]T

(5)

Hence,

the variance ofthe original information is distributed amongst _{the eigenvalues,}

with the 1st eigenvalue

(fa)

_producing aPC

(Yi)

that accounts fora given percentage of

the totalvariance oftheRGB data.

Therefore,

a75%testisenforcedonthe 1steigenvalue

as ameasure oftheconfidence ofthePCtransformation.The

R, G,

orBchannelthathas

the highest variance is chosen as the grayscale image if this criterion fails. Sample

reordered versions of

Yu Y2,

and

Y3

are shownin Figures3.3f-3.3h.

#

Figure3.3 ComparisonofPCswith alternative grayscale method:

(a)

Original DCDimagewith

"invisible"

debris,

(b)-(d)

corresponding

R, G,

andBchannelsrespectively,

(e)

MATLAB rgb2grayprocedure,

(f)

1st

PC, (g)

2ndPCand

(h)

3rdPC.

3.1.3. LocalNormalization

Once a high contrast grayscaleimage has been obtained, a local normalization

(LN)

procedure is employed to compensate for non-uniform background situations. The LN

process is designed to handle large illumination variations (see Figures 3.4a and

3.4c)

characteristic of anumberofsamplesinthedatabase. This approachis givenby:

f(i,j)-mf(i,j)

g(!,j)=

-vf(Uj)

(6)

where:

f(i,j)

istheselected/transformedimage(grayscale).

m/ij) isanestimationofthelocal meanof

f(i,j)

a/i,j) isanestimationofthelocal standarddeviation

g(i,j)istheoutputimage

The above outlined approach _efficiently removes variations in the image (see Figure

3.4b). The block diagram in Figure 3.5 depictsthe implementation ofthe LNprocedure.

[image:27.535.71.490.49.255.2]

An estimation ofthe image's local mean is obtained

by

filtering

with a 70 x 70 spatial

Gaussian lowpass

filter,

hrfij)

_{of standard}

deviation,

o\= 21. This Gaussian mask also

ensuresthat_anycandidate pixel is assigned ahigherweightwhile_utilizingthecorrection

routine. The []2

and

V[]

symbols (see Figure

3.5)

represents"square"and "square

root"

operations _respectively and are used to complete the computation of the standard

deviation. Thesecond_smoothing

filter,

h2(i,j)

_is equivalenttoh,(i,j).

^g*5BwPfcii'!i 5/

TraSSe

Figure 3.4 LocalNormalization:

(a)

Non-uniformbackgroundsample,

(b)

Localnormalized

version ofFigure

3.4a,

(c)

Binarizedresult ofFigure 3.4aand

(d)

Binarizedresult ofFigure3.4b. [image:28.535.59.483.210.644.2]

JfiJ)

Gaussian Filter,hi(ij)

m/U)

[f

V[]

-.

o/ij)

[image:29.535.100.452.83.166.2]

gfij)

Figure 3.5LocalNormalization block diagram.

3.2. Region

of

Interest

Identification

and

Processing

3.2.1.

Binary

thresholding

The normalized

image,

_{gfij) is further processed}

by

utilizing a median based

thresholding

approach to

identify

the ROIs. Other

thresholding

schemes tend to

immediately

separate the debris without _{providing any deletion boundary. The median}

threshold was thus observed to be more robust to noise and outliers within the image

when compared to a mean-based thresholding. Ifthe median is given

by

TM

_, then the

binary

image

b(ij)

iscomputed as:

b(i,j)

=

I

ifg(i,J)>TM

0,

ifg(i,j)<TM'

(7)

The images inFigures 3.4cand3.4dwere obtained

by

_utilizingEquation9.

3.2.2. Morphological

filtering

(non-Mottle vs.

Mottle-type)

An opening morphology operation27

is then employed to remove _noisy objects from

the

binary

image,

b(ij). Figure 3.6 depicts a comparison of typical

morphologically

openedimages obtained from a

deletion, DCD,

and mottle

images,

respectively. Notice

the different dimensions ofeach image. The_relationship

between

_the

largest

_{region and}

its neighbors is employed as a

discriminatory

feature since ittends to be

invariant

to the

defectsize.

Specifically,

theratio oftherootmean square

(RMS)

ofthearea ofthelargest

object (i.e. the_{largest defect region)} to theRMS ofthe area of other objects is employed

as the feature ofinterest. This is _{equivalenttocomputing the signal-to-noise ratio}

(SNR)

ofthe ROI. This

feature

isthereaftercomparedtoan_empiricallydeterminedthreshold

T0

to yield class

decisions

(atthe first decision _{node) between} mottle and deletion-type (or

non-mottle) images. Figure 3.7 shows typical distributions ofthe number ofpixels in

each regionfora mottle and non-mottletype. Ifthesignal ofinterest is given

by

the(area

of

the)

major ROI then a large SNR is expected for the non-mottle image while a low

SNRistypicalofa mottleimage.

In order to ensure that _{only defects}

bearing

a deletion-type signature are considered

for further _processing, we impose a shape contour test

by

_utilizing the Hausdorff

distance. In particular, the contour ofthe major ROI is compared to a fitted ellipse's

contour which is created _using a similar approach to that found in29.

Additionally,

the

rectangularity ofthemajor ROI measured

by

theratio ofthe area ofthe regionto the

area ofits minimum

bounding

rectangle

(MBR)

is employed as another feature. These

twofeatures i.e. Hausdorff distanceand_{rectangularity,} ensurethat_{only defects exhibiting}

an elliptical shape configuration (or are _{non-rectangular)} are regarded for further

processing.

They

are compared with thresholds

Tt

and

T2

to separate unknown defects

from deletion-typesamples. Thiscompletesthe seconddecisionnode.

350pixels

0>

[image:31.535.64.468.44.301.2]

(e)

(f)

Figure 3.6 ROIfeatureextractionfor:

(a)-(c) Deletion,

DCDandMottle

images,

(d)-(f)

Respective morphologicallyopenedimages.

300 400

Reaions

(a)

Figure 3.7

Study

ofSNRof:

(a)

Typicalmottledimageand

(b)

Typicalnon-mottletype.

3.3. Outlier

Processing

Thefirst_steptowards_{processing only deletion-type}imagesutilizestheconvexhullof

the major ROI as a mask (Figure

3.8a)

for the replacement ofpotential _missing debris

pixels. To obtain these pixels, athreshold is _{automatically} computed and applied to the

masked region. Let p represent a given _{gray level. This} threshold is obtained

by

first [image:31.535.55.488.347.503.2]

generating the histogram

h(p)

forpmi

<p

<pmax

as shown in Figure 3.8b. Thenthe set

containingpossiblebright_{outlier pixels}withinthemaskis definedby:

Kl={p:h(p)>0

and _p> 0.75_-{pmax

-Pmm

)}

(8)

where the constraint /?>0.75

(pmax

-p^) is designed to limit the desired threshold to

only bright pixel values. Define a new set

51

whose elements are the backward

differences betweenadjacent elements of911. Thedesiredthresholdcanbeobtained as:

Outlier Threshold- min(9*

}

where 6 = ₅

for the given dataset. To avoid _conflicts, i.e. two _{(or more) bright} outlier

regions revealed

by

the

threshold,

the largest region is _simply chosen. Then an

empirically determined value

T3

is compared with the difference ofthe outlier and the

neighboring regionsas:

median\0G

}

-median\Np

}

>

T3

(10)

where

Og

=

Potential Outlier group and,

Np

= _Neighborhood

pixels contained in a

bounding

box around

Og

(Figure 3.8d). This helps to quantify the outlier measure ofthe

region in question to determine if it is _{significantly} different from its neighbors. Figure

3.8 shows a successful identification ofthe _missing pixels where the

0G

region (Figure

3.8c &

3.8d)

has beenobtainedvia

thresholding

withthe desired outlierthreshold value.

The

Np

region(Figure

3.8d)

is createdfrom a 1-pixel

boundary

around Oc. Since the

Og

region satisfies the constraint posed in Equation

12,

the bright pixels within that region

canbereplaced withadarkpixel value(i.e. 0).

-27-Create

masked region

histogram

-.**.***,F.<i

^_ii

(a)

64

61

84

79

59

61

Np- -"

"

61

69 110 128

89

64

51 120

143 161

120

74

J4-_-Kt

la

145 102

74

Sameas

69

691

87 115

77

61

84

56

74

71

59

69

Create I-pixel

boundaryaround mask

(d) (c)

Figure3.8

Missing

debris identificationprocedure.

3.3. Final

classification via

Morphological

Reconstruction

To efficiently differentiate between deletionandDCD

defects,

themajordifference is

exploited - the

presence of a _group of dark pixels in an approximate center of the

elliptical region. This is a difficult task as demonstrated in 32 wherein information from

thehistogramwas utilizedto deviseathresholdthatlocalizespossible debrispixels. Due

to the _noisy nature of deletion samples, additional steps

involving

the acceptance or

rejection of segmenteddarkpixels are needed. Forthelow contrastDCD images (Figure

2.1e),

thedebrispixelsare not_successfullyidentifiedafterthresholding. Figure3.9shows

intermediate results obtained

by

_applying different

thresholding

mechanisms to pixels

withinthecreated mask (Figure

3.9b)

foraDCD sample. Thebestresult (Figure

3.9e)

is [image:33.535.52.490.72.385.2]

28-achieved

by

a hole emphasizing _routine which utilizes morphological reconstruction to

fill "holes"

in the image and thereafter compute a difference image. Otsu's

method31

(Figure

3.9c)

doesnot provide adesirableoutcome whilethemodifiedOtsu'sapproach17'

(Figure

3.9d)

is close to the ground truth result but is noisier compared to the

morphological reconstruction approach. Figure 3.10 illustrates a similar situation with a

deletion image.

The fill-hole process is a special morphological

transformation30

(called a geodesic

transformation)

which accepts two images

-a m-arker -and -amask. The mask is used to

restrict the growth (or

decay)

of the marker image

during

regular morphological [image:34.535.55.489.317.590.2]

operations.

Figure 3.9 Withinmask segmentation:

(a)

Original

image, (b)

masked

image, (c)

Otsu's

thresholding,

(d)

Modified Otsu'sthresholding16,24

and

(e)

Morphologicalreconstruction approach.

Figure 3.10Withinmask segmentation:

(a)

Original

image, (b)

masked

image, (c)

Otsu's

thresholding,

(d)

ModifiedOtsu's

thresholding1624

and

(e)

Morphologicalreconstruction approach.

For this process, the marker is set to the maximum value ofthe image except _along its

borderwhere thevalues oftheoriginal image are

kept,

while themaskis represented

by

the image itself. An erosion ofthe marker is then

iteratively

performed until a stable [image:34.535.53.491.463.550.2]

-29-result is achieved. Thefinal eroded marker constitutesthefilledimage andthe holes can

be obtained

by

_subtracting the input image from its filled version (see 27 for a detailed

explanation). A sample ofthisobtaineddifference image is shownin Figure

3.11;

notice

how the _previously unperceivable debris becomes

highly

prominent in the difference

image.

Image fill

C>

Obtainoutlier

group &

measureSNR

?

[image:35.535.62.472.212.327.2]

Difference image

Figure 3.11

Obtaining

thedifferenceimage

Figure 3.12ashows aperspective version ofthedifferenceimage in Figure 3.11. The

grayscale SNR ofthe major peak regionin the difference image is used to _quantify the

disparity

between DCD anddeletionsamples. Ifthepeak regions (bright_outliers) denote

the debris of

interest,

then the DCD are expected to possess a higher SNR level when

compared to the deletion. A low SNR is an indication that no further processing is

requiredandthereforeno prominent outliers exists.

30-Generate 1 difference image

Peakregion

1

or

^

i Ljt.

CO)

906

L

V''

am ^juJU, BSB Bua4 0

.^j

|^^

<l<

*"""*-^_ ^^^^^^ BO SO *""---.._^ *

[image:36.535.62.498.62.341.2]

S H Generate difference image Peakregion

(a)

(b)

Figure 3.12 Difference images:

(a)

DCD (From Figure

3.11),

SNR=_22dB

and

(b)

Deletion, lOdB.

Thenumber ofpixels _representingthe debris isutilized asthe

discriminating

feature

between deletion and DCD samples. As an added vote of_{confidence, the} Mahalanobis

distance8

in a normalized range

[0,

100]

ofthe segmented debris fromthe center ofthe

mask (i.e. convex hull of major ROI as in Figure

3.8a)

is employed as a vote of

classification confidence ofthe DCD. We expect true DCD samples to thus have a low

mahalanobisvalue,while thedeletionsamples possessavote of confidenceequivalentto

a normalized version ofthe SNR ofthe _{difference image. The normalizing} constant in

thiscaseisobtainedfrom a set of

training

data.

Chapter 4

Experimental

Results

and

discussions

Theperformance oftheproposed algorithmistestedon adatabase of273 imagesthat

consists of:

1)

261 scanned images ofRGB format provided

by

Xerox Corporation with

accompanying ground truth

labels,

and

2)

12 Non-defect and synthetic images (Figure

4.1)

comprising of

logos,

and

"rectangular"

shape defects. The scanned images were

comprised of88

DCDs,

80

deletions,

and 93 mottleimages (see Table 4.1). Thesynthetic

images were introduced to test the algorithm's robustness and ensure _only regions

-32-satisfying the "elliptical" ROI _property are processed as deletion-type defects. The

images

are scanned with anEPSONGT 30000 flatbedscanner at a_scanningresolutionof

600 dots per inch

(dpi)

in accordance with the ISO-_{13660 standards. No spatial image}

processing is applied onthe image

(during

_scanning) so asto determine how muchpre

processing the algorithm would be responsible for. MATLAB33 was chosen as the

classification _prototyping environment due its _ability to _easily perform computations

involving

matrices. It however has a speed disadvantage due to its inherent nature to

interpretprograms. Asmentioned

before,

the_{defect is manually localized}and cropped

by

a human operator thus _resulting in digital images of_{varying dimensions} greater than

25mm2. Outofthe88 DCD

images,

_{20 had missing beads}where some could_{possibly be}

classified as deletionsinsteadofDCDs

by

ahumanobserver.

TEXACO

"(a)"

[image:38.535.57.474.348.461.2]

(b)

(c)

(d)

Figure 4.1 Syntheticsamples:

(a)

_DCD,

(b)

Deletionand

(c)-(d)

Logos.

Table 4.1 illustrates a_summary ofthe

algorithms' performance. The classification of

the DCD class yields a 93.2% accuracy. The mis-classified samples are a result oflow

contrast (Figure

4.2a)

images where the bead was not visible. Giventhe lackof contrast

and the absence ofthe bead (Figure

4.2b),

it isentirely possible that the originals could

have beenclassifiedasDCDsordeletions

by

an operator and/orservice engineer.

[image:39.535.72.468.50.438.2]

Figure 4.2 Misclassifiedsamples:

(a)

DCDwithlowcontrast,

(b)

DCDwith a nonvisible

bead,

(c)-(d)

Deletionswith possibledebris missingregions.

Deletion classification yields 96.3% accuracy. Mis-classified samples appear to

possess a potential _missing debris group ofpixels (Figure 4.2c

-4.2d)

and

thereby

are

classified asDCDs

by

our proposed algorithm. Once again, this misclassificationcanbe

attributedtoapotential confusionfromthegroundtruthinformation.

Theresultsforthe remainingcategoriesare also showninTable

4.1,

wherethemottle

and _arbitrary images were classified with a 99% and 100% accuracy respectively. The

-34-total correct classification rate is 96.6% whichis obtained from an average ofthe given

individualaccuracies.

The effectiveness ofthe PC transform forgrayscale conversion is also demonstrated.

Table 4.2 shows the resultant classification obtained _{from using MATLAB's rgb2gray}

method. It can be seen from Table 4.2 that the _{MATLAB rgb2gray functions} results in

lowerclassification_{accuracy especially}forthecaseofDCDs.

CLASSIFICATION RESULTS Number

DCD/D bri

True Class of

* '

Deletion Mottle Unknown

Accuracy

missing J

images

DCD/Debris missing 88 82 5 1 93.2%

Deletion 80 2 78 97.6%

Mottle 93 1 92 99%

Other 12 12 100%

Table 4.1 Classificationresults_usingour proposed method.

CLASSIFICATION RESULTS

Debris missing 88 64 20

Deletion 80 3 77

Mottle 93

Other 12

_, . Numberof DCD/Debris ~ . A.

,, ^, TT . .

True Class . . . Deletion Mottle Unknown

Accuracy

images _missing _ 4 72.7% 96.25% 93 100% 12 100%

Table 4.2Classificationresults_{using MATLAB rgb2gray}approach.

The robustness ofthe algorithm against _varying degrees ofrandom Gaussian noise

was tested.

First,

the noise present in the normal images was quantified

by

_selecting

random samples and extracting arbitrarysized cropped regions fromvarious locations in

the image. Wefoundthe imagestohave anintrinsic noiselevelwith standarddeviationa

~ 0.

04,

thus _anyadditional noisetendsto further depletethecontrast.Repeated testswith

white Gaussiannoise (a=

0.01, 0.02,

_and

0.05)

are summarizedin Figure 4.3 fromwhich

35-it canbe _easilyseenthat the classification_{accuracy is}

inversely

proportional to the level

of noise added. This is dueto a significant

drop

incontrast withadditional noise. Mottle

identification

tends tohold_upwell againstthenoisedueto the_{already noisy}natureofthe

mottle

images;

_{misclassifiedmottle images}

approach adeletion image. Themisclassified

deletion

images

are classifiedas Mottleand tend toworsen with

increasing

noise. As the

levelof noise is

increased

from a =

0.01, 0.02,

and

0.05,

the total classification(average

of all the _{classifications) is} reduced from

91.3%, 86%,

and 73.4%. Since a flatbed

scanner does not

typically

corrupt the digital image with _{noise, this} does not representa

majorlimitationofthealgorithm.

100

90

80

70

60

50

40

30

20

- Mottle

- DCD Deletion

0.01 0.02 0.03 0.04 0.05

[image:41.535.60.482.293.477.2]

GaussianNoiseStandardDeviation, o

Figure 4.3 Additionof

increasing

levelsofgaussiannoiseand_resultingperformance.

The algorithm isresponsible for

locating

and _classifying one majordefectper digital

sample. The total _processingtime takenfora sample of size 118 * _{118 is approximately}

2.5s for a deletion-type image and _{approximately} 2s for a mottle-type image. This is as

expected since the mottle-type images are

immediately

classified at the first decision

node. An increase intheresolutiontoabout 350 x _{350 increasesthecomputationtime to}

approximately 18s for a deletion-type image and approximately20s for a mottle

image.

36-The computation costs involved is not at all

burdening

for the given MATLAB

prototypingenvironment.

Chapter 5

Conclusion

and

Future Work

This thesis proposes an algorithm for automatically

identifying

and _classifying a

specificset ofimage quality defectsinprinteddocuments. The algorithm accepts scanned

versions ofsuspicious defectedprinted media, attempts to _accurately locatemajor defect

regions and indicate the defect type

(Deletion, DCD,

or mottle). Due to large variations

between elements ofthe same class, several _{pre-processing techniques} were carried out

on each image to attain some level of_uniformity amongstthe samples.

Using

a custom

-38-decision_{tree classifier,}

binary

decisions were made

by

_employing simple shape and size

constraints at each decision node. The use of principal component analysis to obtain a

grayscale image helpstopreservethe contrastoftheoriginal RGB sample.

Additionally,

the use oflocalnormalizationprocedures assisted _{in avoiding}misclassification situations

by

making the

background

uniform wherever possible. Also given the _noisy nature and

low contrast of some _{the sample,} an_accuracyof96.6% was still attained.

However,

this

accuracy tends to depreciate with

increasing

noise levels using the pre-computed

thresholds.

Since this procedure has proved to be quite _{successful, the} next _{step involves} an

automation of the defect localization process

by

_possibly

incorporating

the original

electronic documentto

help

localizeareas ofsignificantdefectpresence. Therobustness

ofthe algorithm against

increasing

levels ofnoise will also be improvedupon

by

_making

thecomputation ofthethresholdvaluesdependentonthegiven noiselevel.

39-References

1. Y.

Rui,

T. S.

Huang,

and S.

Chang,

_{"Image Retrieval: Current}

Techniques,

Promising

Directions,

andOpen Issues,"

Journal of Vis. Comm. andImage Rep.

10,

pp. 39-62(1999).

2. A. M. W.

Smeulders,

M.

Worring,

S.

Santini,

A.

Gupta,

and R.

Jain,

"Content-based image retrieval attheend ofthe_early years,"

IEEE Trans, onPatt. Analysis andMachine

Intelligence,

Vol.22,

No.

12,

pp.

1349-1380,

Dec. 2000.

3. H.

Burkhardt,

and S.

Siggelkow,

"Invariant features for

discriminating

between

equivalence classes,"

Nonlinear Model-based Image Vid. Proc. and

Anal,

John

Wiley

and

Sons,

2000.

4. M. K.

Hu,

"Visual pattern recognition

by

moment invariants," in J. K.

Aggarwal,

R. O.

Duda,

and A.

Rosenfeld,

Comp. Meths. in Image

Anal,

IEEE computer

Society,

Los

Angeles, CA,

1977.

5. R. C. Gonzalez and R. E.

Woods,

Digital Image

Processing

2nd

Edition,

Prentice

Hall,

New

Jersey,

2002.

6. B. M.

Mehtre,

M.

Kankanhalli,

andW. F.

Lee,

"Shapemeasuresforcontentbased

image retrieval: A

comparison,"

Information

Processing

& Management

33(3),

1997.

7. Y.

Rui,

T. S.

Huang,

and S.

Mehrotra,

"Content-based image retrieval with

relevance feedback inMARS," _{Proc. ofInt. Conf.} onImage

Proc,

Vol.2,

pp. 815

-818, 1997.

8. W. Y.

Ma,

and B. S.

Manjunath,

"Netra: A toolbox for _navigating large image databases,"

Multimedia

Systems,

Vol.7, No.3,

pp.:184-198, 1999.

9. R. O.

Duda,

P. E.

Hart,

andD. G.

Stork,

Pattern Classification2nd

Edition,

John

Wiley

and

Sons,

2001.

10,

N.

Vasconcelos,

"On _{the complexity} of probabilistic image

retrieval," IEEEInt.

Conf. on Comp. Vis. andPatt.

Recognition,

Vol.

2,

pp.

400-407,

July

2001.

11. N.

Beckmann,

H.

Kriegel,

R.

Schneider,

and B.

Seeger,

"The R*-tree: An

efficient robust access method for points and

rectangles,"

ACM SIGMOS Int.

Conf. on

Mang

Of

Data,

Atlantic

City, May

1990.

12. H.

Samet,

"The quadtree and related hierarchical data

structures,"

ACM

Computing

Surveys, Vol.16, No.2,

pp.

187-260,

1984.

13. H. J.

Zhang,

and D.

Zhong,

"A Schemeforvisual feature-basedimage

indexing,"

Proc. ofSPIEconf. onStorageandRetrieval for Imageand Video Databases

III,

pp.

36-46,

San

Jose,

Feb. 1995.

14. J. Iivarinen and A.

Visa,

"An Adaptive Texture and Shape Based Defect Classification,"

Proc. ofthe 14th Int. Conf. on Patt.

Rec,

Vol.

1,

pp.

117-122,

Brisbane, Australia,

Aug. 1998.

15. Y. Li andT. W.

Liao,

"Weld DefectDetectionBasedonGaussian

Curve,"

IEEE,

1996.

16. D.

Mery

and M. A.

Berti,

"Automatic detection ofwelding defects using texture

features,"

Insight,

Vol.

45(10),

Oct. 2003.

17. H.

Ng,

"Automatic_thresholdingfor Defect Detection,"Proc. ofthe 3rd Int. Conf.

onImageand

Graphics,

2004.

18. J. C.

Briggs,

D. J.

Forrest,

A. H.

Klein,

and M.

Tse,

"Living

with ISO-13660:

Pleasures and Perils," _{Proc. of}the IS&T's

NIP15,

Int. Conf. on Digital

Printing

Tech., Orlando, Florida,

Oct. 1999.

19. ISO/IEC DIS 13660 Draft International

Standard,

"Office Equipment

-Measurement of image _{quality attributes for}

hardcopy

output

-Binary

monochrome text and graphic images," Int. Org. for

Standardization,

ISO/IEC JTC1

SC28,

1996.

20. D. Rene

Rasmussen,

K. D.

Donohue,

Y. S.

Ng,

W. C.

Kress,

F.

Gaykema,

andS.

Zoltner,

"ISO 19751 macro-uniformity."

21. G.

Sharma,

et _al, "Methods and apparatus for

identifying

_marking process and

modifying image data based on image spatial

characteristics,"

US Patent

#6353675B1,

Mar. 2002.

22. D. F. Dunn and N. E.

Mathew,

"Extracting

Color Halftones from Printed

Documents

Using

TextureAnalysis,"

Proc. oftheInt. Conf. onImage Proc., Vol.

1,

pp. 787_-790, Oct. 1997.

23. William K.

Pratt,

Digital Image

Processing,

John

Wiley

&

Sons,

New

York,

1991.

24.

http://gug.sunsite.dk/docs/Grokking-the-GIMP-vL0/node54.html

25. http://www.mathworks.com/access/helpdesk/help/toolbox/images/rgb2gray.html

26. J. E.

Jackson,

A User's Guide to Principal Components.

Wiley

Series in

Probability

andMathematical Statistics. John

Wiley

&

Sons,

New

York,

London,

Sydney,

1991.

27. P.

Soille,

MorphologicalImage Analysis: PrinciplesandApplications.

Springer-Verlag,

1999.

28. D.

Huttenlocker,

G.

Klanderman,

and W.

Rucklidge,

"Comparing

images using

the Hausdorff distance,"

IEEE Trans. Patt.. Anal. Mach. Intell. 15

(9) (1993)

850-863.

29. E. Saber and A. Murat

Tekalp,

"Frontal-view face detection and facial feature

extraction _using color, shape and _{symmetry based} cost

functions,"

Patt. Rec.

Letters,

Vol.

19,

pp.

669-680,

1998.

29. P. L.

Rosin,

"Measuring

shape:_{ellipticity,rectangularity,}and

triangularity,"

Proc.

15th Intern. Conf. onPatt.

Recognition,

2000.

31 N.

Otsu,

"Athresholdselection methodfrom gray-levelhistogram,"IEEE

Transactionson

Systems,

Manand

Cybernetics,

SMC-8,

pp.

62-66,

1978.

32. O. A.

Ugbeme,

E.

Saber,

and W.

Wu,

"An automated defect classification

algorithmforprinted

documents,"

Proc. ofInt. Conf. in

Imaging

Science,

2006.

33. The Math

Works,

Inc.,

MATLAB Reference

Guide,

1994.