Www.monash.edu.au An Improved Shape Descriptor Using Bezier Curves Authors: Ferdous Ahmed Sohel Dr. Gour Chandra Karmakar Professor Laurence Sean Dooley.

www.monash.edu.au

An Improved Shape Descriptor Using Bezier Curves

Authors:

Ferdous Ahmed Sohel

Dr. Gour Chandra Karmakar

Professor Laurence Sean Dooley

Gippsland School of Information Technology

www.monash.edu.au

2

ISDBC: Presentation Outline

• Introduction. • Existing Bezier curve based shape

descriptors.• Proposed ISDBC

– Control point determination. – Control point coding.

• Results.• Conclusions.

www.monash.edu.au

3

Introduction

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

…0110…010100… ShapeDescriptor

Shape Coding:

Applications:

Communication: Mobile multimedia communication, low bit rate coding.

Storage and retrieval: Digital library, indexing, digital archiving.

Interactive editing: Cartoons, digital films.

www.monash.edu.au

4

Existing shape descriptor using BC

• Object shape description by Cinque et al (1998).

Divides the shape into equal segments

in terms of number of boundary points.Easy and simple.Considers encoding of the control

points.

Does not consider the shape

information.The coding is not efficient. 0 20 40 60

0

20

40

60

R1

R2Segment ends

www.monash.edu.au

5

Existing shape descriptor using BC

• Chinese calligraphic character descriptor by Yang et al (2001).

Uses trial-and-error methods.The control points are highly depended on direction of

the shape at the segment-ends.Does not consider the encoding.Allows the control point be out side the shape.

Considers the shape information.

www.monash.edu.au

6

Existing shape descriptor using BC

• Arabic character descriptor by Sarfraz and Khan (2002, 2003).

Does not consider the encoding.Allows the control point be out side the

shape.

Considers the shape information.

Optimally selects the control points in the distortion sense.

20 40 60

10

20

30

40

www.monash.edu.au

7

Existing shape descriptor using BC

• Active shape lip model by Shdaifa et al (2003).

Easy and simple.

The control points are highly depended on

direction of the shape at the segment-ends.Does not consider the encoding.Allows the control point be out side the

shape.30 40 50 60

20

30

40

50

60

c1c2

c3

c4

c5

www.monash.edu.au

8

Existing shape descriptor using BC

• Generic shape descriptor by Sohel et al (2005).

Weak for the shapes having loops

or

repeated or branches.

Considers the shape information.

Uses efficient dynamic fixed length coding (DFLC) method.

Control points are with in the range of shape points

Also the algorithm by Cinque et al is weak for loops and branches!

20 30 40 50 60

0

5

10

15

20

25

30

www.monash.edu.au

9

Proposed ISDBC

• Control point calculation– Divide the shape into segments.

– Find the significant points.

– Insert supplementary points.

– Form the potential control point set.

– Determine the control points.

• Control points encoding– Encoded differentially by modified DFLC.

www.monash.edu.au

10

ISDBC

For the ith segment

|Bi|: number of potential control points.

bI: set of potential control points.

1,3,,

2,,

1,0,0, ;;;4

*3

4

iiBiB Bii

ii

iiii bvbvbvbv

www.monash.edu.au

11

ISDBC

• Control point coding

1st point

l1 2*l1 l1 li 2*li li … lj 2*lj

First Segment Next segment … Last Segment

Distance between consecutive control points is periodic.

1st point

l 2*l l l 2*l l … l 2*l

z (=4l): maximum number of potential control points of a segment.

www.monash.edu.au

12

ISDBC

• Dynamic fixed length coding (DFLC)

— A combination of run-length code and chain code.

— Encodes the control points differentially.

— Differential Direction is encoded by 8-bits.

— The distance is the run-length.P

Q

www.monash.edu.au

13

ISDBC

• The complete shape descriptor

4- bit for

length of L1

1st point

Dir

+

L1

Dir

+

L2

Dir

+

L1

Dir

+

L1

Dir

+

L2

Dir

+

L1

… Dir

+

L1

Dir

+

L2

First Segment Next segment … Last Segment

The starting 4 bits are reserved for the length of L1.

ISDBC can encode a segment of up to 4*216 shape points.

www.monash.edu.au

14

Results

Distortion measure [1]

Class 1: Peak distortion (Max)

Class 2: Mean Square distortion (MS)

[1] F.A. Sohel, L.S. Dooley, and G.C. Karmakar, “Accurate distortion measurement for generic shape coding,” Pattern Recognition Letters, Elsevier Science Inc., in press, available on-line.

www.monash.edu.au

15

Results

20 25 30 35 40 45 50 55 60 65

0

5

10

15

20

25

30

ShapeCinque et alSDBC

20 25 30 35 40 45 50 55 60 65

0

5

10

15

20

25

30

ShapeISDBC

ISDBC

www.monash.edu.au

16

Results

0 10 20 30 40 50 60 70

0

10

20

30

40

50

60

70

Shape SarfarzISDBC

0 10 20 30 40 50 60 70

0

10

20

30

40

50

60

70

Shape Yang ShdaifaISDBC

0 10 20 30 40 50 60 70

0

10

20

30

40

50

60

70

ShapeSDBC ISDBC

www.monash.edu.au

17

Results

Shape Bream Arabic LipTechniques Dmax DMS DMax DMS DMax DMS

ISDBC 3.85 2.76 1.2 0.83 1.5 0.9Sohel et al 5.4 3.6 1.9 1.25 12.1 7.5

Cinque et al 6.0 4.3 2.1 1.35 10.2 5.4Sarfraz 4.22 3.34 1.3 0.95 1.65 1.05

Yang et al 3.9 3.46 1.35 1.2 1.7 1.2Shdaifat et al 6.05 6.55 1.4 1.4 1.8 1.5

Distortion in different shapes by the techniques.

www.monash.edu.au

18

Results

Shape Techniques SR = 5 SR = 6 SR = 7 SR =8

BreamISDBC 193 230 267 304

SDBC 193 230 267 304

Cinque et al 244 292 340 388

Lip ISDBC 193 230 267 304

SDBC 193 230 267 304

Cinque et al 244 292 340 388

Bit requirements

20% improvement in the descriptor size. 23% reduction in the descriptor length for each additional segment.

www.monash.edu.au

19

Conclusion

ISDBC utilises the domain specific shape information both in Dividing the shape into segments Determining the control points.

Keeps the distortion lower.Efficiently encode the control points.

www.monash.edu.au

20

Thank you