h004-01-halcon-matching-introduction(基礎(chǔ)學(xué)習(xí)資料)

MVTec

Software

GmbH

is

a

leadinginternational

manufacturer

of

software

for

machine

visionused

in

all

demanding

areas

of

imaging:

semi-conductor

industry,

web

inspection,

quality

control

and

inspection

applications

ingeneral,

medicine,

surveillance

etc.MVTec"s

innovative

work

is

driven

by

a

commitment

to

be

the

number

one

supplier

for

sophisticated

technologies

in

machine

vision.

MVTec

is

engaged

in

sponsoring

various

activities

in

universities,

thus

participating

in

the

challenging

process

of

understanding

how

machines

can

be

taught

to

see.HALCON

is

the

comprehensive

standard

software

for

machine

vision

withan

integrated

development

environment

(IDE)

that

is

usedworldwide.It

leads

to

cost

savings

and

improved

time

to

market:

HALCON"s

flexible

architecture

facilitates

rapid

development

of

machine

vision,

medicalimaging,

and

image

analysis

applications.

HALCON

provides

outstandingperformance

and

a

comprehensive

support

of

multi-core

platforms,MMX

and

SSE2,

as

well

as

GPU

acceleration.

Itserves

all

industries

with

alibrary

of

more

than

1600

operators

for

blob

analysis,

morphology,matching,

measuring,

identification,

and

3D

vision,

to

name

just

afew.

HALCON

secures

your

investment

by

supporting

a

wide

range

of

operating

systems

and

providing

interfaces

to

hundreds

of

industrial

cameras

and

frame

grabbers,

including

GenICam,

GigE

Vision,

and

IIDC1394.VTec

Software

Gmb

HIntroduction

to

Template

MatchingTask

DescriptionFinding

an

object

in

an

imageGivenImage

with

a

prototype

object(template)

=

reference

imageImage

with

an

object

=

searchimageTransformation

classTo

be

determinedTransformation

parametersdescribing

the

relation

between

theobject

in

the

reference

image

to

theobject

in

the

search

imageReference

imageSearch

imageTypical

ApplicationsPick-and-PlaceMicro

electronicsAlignmentPrint

inspectionRoboticsBesides

other

technologies

HALCON

offers

a

wide

range

of

highly

sophisticated

matching

technologies:HALCON

allows

to

locate

objects

with

arbitrary

orientation

in

3D

(3D

alignment),

it

provides

the

well

established

shape-based

matching

–workingevenwithcolor

images,

the

unique

component

based

matching

and

the

well

proven

normalized

cross

correlation.New

in

HALCON

10

are

two

more

matching

technologies:-

Surface-based

3D

matching,

which

finds

objects

in

arbitrary

orientation

in

3D

data.-

Local

deformable

matching,

which

finds

objects,

even

if

they

are

locally

deformed,

and

even

determines

their

deformation.HALCON

offers

matching

toolsfor

any

requirementThis

figure

helps

to

decide,

which

matching

approach

is

the

best

for

your

application.The

first

question

is

about

searching

objects

which

have

to

be

found

in

2D

space

only

(movement

in

x/y

or

achange

in

distance).

If

only

rigidtransformations

are

present,

you

can

choose

between

shape-based

matching

and

NCC

matching.

Shape-based

matching

is

more

accurate

androbust

against

randomclutter,

occlusion,

inhomogeneous

illumination,

and

varying

edge

polarity.

Besides

this,

shape-based

matching

can

be

usedwith

images

having

an

arbitrary

number

of

channels,

e.g.,

color

images.

NCC

matching

is

also

very

robust

against

shape

variations

and

defocus.It

is

especiallyuseful

to

find

objects

that

must

not

contain

specific

structures.

Note

that

NCC

matching

does

not

support

scaling.

Clutter

andocclusion

can

be

tolerated

if

the

(mean)

gray

values

of

clutter

and

occlusion

are

close

to

those

of

the

occluded

parts

of

the

object.

If

the

objectsare

scaled

in

X-

and/(or)

Y-direction,

(anisotropic)

scaled

matching

can

be

used.

If

compound

objects

shall

be

found,

component-basedmatching

does

the

job

–

with

the

restriction

that

scaling

in

size

is

not

possible.If

theobjects

move

in

3

dimensions,

the

choice

goes

for

the

second

group.

If

the

object

or

relevant

parts

are

planar

buttilted,

the

perspective,

deformable

matching

or

the

descriptor-based

matching

can

be

used.

The

firstis

better

for

objects

with

clear

edges,

while

the

latteris

better

for

textured

structures.

If

a

full

3D

model

is

needed,

3D

matching

can

be

used.Which

matching

method

suits

your

application?StartRigid

transformationComponentsLocaldeformationsColor(Aniso-)ScalingPerspectivedeformations3D3D

contours3D

surfaceAdvantages

of

Template

MatchingOne

method

for

many

applicationsNot

too

many

parametersNo

segmentation

is

necessaryRobustNo

special

knowledge

in

machine

vision

necessaryNecessary

InvariancesThe

object

may

be

transformed

by

a

certain

class

of

transformationsTranslationsRigid

transformationsSimilarity

transformsChanges

in

the

object’s

appearanceNoiseSubpixel

accuracyRealtime

computationOcclusionsClutterIllumination

changesFurther

RequirementsDefocused

imagesFurther

RequirementsMultiple

instancesMultiple

modelsFurther

RequirementsPerspective

distortionsReal-time

computationFurther

RequirementsSearch

complex

pattern

in

front

of

complex

backgroundGray

Value

MatchingObjectSignificant

gray

valuesHomogenous

or

fixed

structureNeighborhoodHomogenous

or

fixed

structureTheory:

Similarity

MeasuresThe

sum

of

absolute

differences·

Let

be

the

template

with

the

domain

andan

imageThe

sum

is

not

invariant

against

changes

in

illumination·

Calculate

the

difference

relative

to

the

mean

valueandInvariant

against

additive

changes

in

illuminationSimilarity

Measures:

DifferenceReference

image

with

templateDissimilarityMatches

with

threshold

20Matches

with

threshold

30Similarity

Measures:

Darker

ObjectsReference

image

with

templateDissimilarityMatches

with

threshold

30Matches

with

threshold

35Similarity

Measures:

Brighter

ObjectsReference

image

with

templateDissimilarityMatches

with

threshold

30Matches

with

threshold

35Similarity

Measures:

Normalized

Diff.Dissimilarity

in

dark

imageMatches

with

threshold

37Matches

with

threshold

37Dissimilarity

in

bright

imageTheory:

Similarity

MeasuresNormalized

cross

correlation:

Both

additive

and

multiplicative

variatioillumination

can

be

compensatedValue

range

of

the

normalized

cross

correlation:Similarity

in

dark

imageSimilarity

Measures:Normalized

Cross

CorrelationMatches

with

threshold

0.75Matches

with

threshold

0.75Matches

with

threshold

0.75Calculation

of

the

similaritymeasure

takes

too

much

time

O(whn).Typically

the

template

is

also

visible

using

a

lower

camera

resolution.This

is

equivalentto

zooming

the

image

down.Typically

the

images

are

zoomed

down

by

steps

of

factor

2.This

sequence

of

smaller

and

smaller

images

is

called

image

pyramid.The

gray

values

of

four

neighboring

pixels

are

summed

up

and

divided

by

four.Image

PyramidsLevel

1Level

4Level

3Level

2Image

PyramidsBehavior

of

the

pattern

in

the

pyramidLevel■4■3■2■1The

highest

possible

pyramid

level

is

selected

where

the

template

still

can

be

detected,

e.g.,

level

4.Reduction

in

the

number

of

pixels

by

a

factorof

4096:

The

image

and

the

template

are

made

smaller

by

a

factorof

64.On

the

highest

level

of

the

pyramid

each

possible

location

is

tested.On

the

higher

levels

typically

a

higher

threshold

must

be

used

because

of

the

zooming

effects.The

candidates

found

on

the

highest

level

are

tracked

(followed)

throughthe

pyramid

down

to

the

lowest

level.Goingtoalower

level

a

broader

search

area

must

be

selected

in

order

to

not

miss

the

correct

match.The

search

ends

atthe

lowest

level.Note:The

search

process

checks

for

matching

candidates

on

the

highest

level

only.Objects

not

found

on

the

highest

level

will

not

be

found

atall.Template

Matching

with

PyramidsLevel

2Level

4Level

3Level

5Template

Matching

with

PyramidsLevel

4

(ROI,

Matches)Level

2Level

3Level

1So

far

the

template

always

appeared

in

the

same

orientation.Similarity

features

do

not

accept

more

than

slight

variations

in

the

orientation.Fora

matching

witharbitrary

orientation

the

template

can

be

rotated

in

advance.To

do

so,

an

appropriate

angle

step

must

be

selected,

e.g.,

1°.With

eachpyramid

level

the

angle

step

can

be

reduced

by

a

factor

of

two.On

the

highest

level

all

positions

and

all

rotations

are

compared.When

going

down

in

the

pyramid

the

neighboring

angle

ranges

also

must

be

checked.Template

Matching

with

RotationBehavior

of

pattern

with

rotation

(gray

values

and

edges)0°18°342°Template

Matching

with

RotationReference

image

with

templateRotation

288oRotation

145oRotation

33oLimits

of

Gray

Value

MatchingLarge

variations

of

the

gray

values

ofthe

objectLarge

variations

of

the

gray

values

inthe

neighborhood

of

the

objectQuestion:

Which

feature

remainsstable?Idea

of

Shape-Based

MatchingSignificant

features