Handwritten character recognition Handwritten character recognition - - PowerPoint PPT Presentation

Handwritten character recognition Handwritten character recognition using elastic matching based on using elastic matching based on a class a class-

• dependent deformation model

dependent deformation model

• S. Uchida and H. Sakoe

Kyushu University, Japan

Introduction Introduction

Human Interface Lab. Kyushu Univ.

3

input pattern

Elastic matching Elastic matching

displacement field v

input fitted to reference reference

Human Interface Lab. Kyushu Univ.

4

Conventional EM techniques Conventional EM techniques

= category-independent = it is assumed that all classes have the same deformation tendency

however

frequent rare may cause

• verfitting to “A” !

Human Interface Lab. Kyushu Univ.

5

Our purpose Our purpose

develop new EM where class-dependent deformation tendency(= eigen-deformation) is considered

Human Interface Lab. Kyushu Univ.

6

Two tasks Two tasks

How to get the eigen-deformations? → Reported already

[Uchida and Sakoe, Pattern Recognition, 2003]

How to utilize the eigen-deformations in EM?

Human Interface Lab. Kyushu Univ.

7

Estimation of Eigen-Deformations

training patterns reference collection of displacement fields using (conventional) EM

…

Principal Component Analysis

… …

eigen-deformations

Human Interface Lab. Kyushu Univ.

8

Estimated eigen Estimated eigen-

• deformations

deformations

u1 u２ u３

apply positively apply negatively

Human Interface Lab. Kyushu Univ.

9

Two tasks Two tasks

How to estimate the eigen-deformations? → Reported already

[Uchida and Sakoe, Pattern Recognition, 2003]

How to utilize the eigen-deformations in EM? → Topic of this presentation

Outline of Outline of proposed technique proposed technique

Human Interface Lab. Kyushu Univ.

11

Class Class-

• dependent deformation model

dependent deformation model

( ) ( )

     + ∑

= M m m c m c

y x u y x P

1

, ,

,

α

uc,1

Deformation model of reference pattern Pc(x,y) (class c):

+α2× …+αM×

=α1×

uc,2 uc,M

Human Interface Lab. Kyushu Univ.

12

) ( =

m

α

<

m

α >

m

α

Geometric interpretation Geometric interpretation

( ) ( )

     + ∑

= M m m c m c

y x u y x P

1

, ,

,

α

...

c

P

pattern space

M-dimensional manifold

Human Interface Lab. Kyushu Univ.

13

( ) ( ) ( )

dy dx y x u y x P y x E

M m m c m c

∫ ∑

              + −

= 2 1

, , ,

,

α

EM based on class EM based on class-

• depend. deform. model
• depend. deform. model

...

m

α

c

P E

m

α

w.r.t. minimize

→

nonlinear function

• f α

reference input pattern space

Human Interface Lab. Kyushu Univ.

14

Tangent plane (M-dimensional)

Solution via linear approximation Solution via linear approximation

...

c

P E

reference

nonlinear linear !!

Easily solvable as a least square problem input Taylor expansion & approx.

Experimental Results Experimental Results

Human Interface Lab. Kyushu Univ.

16

Database Database

English capitals from ETL6

1100 samples×26classes

preprocessing

linear scaling, etc. # 1-100 # 101-600

average reference training patterns

to estimate eigen-d # 601-1100

test patterns

in recog. exp.

Human Interface Lab. Kyushu Univ.

17

Recognition rate Recognition rate

98.8% 98.9% 99.0% 99.1% 99.2% 99.3% 10 20 30 40 50

# eigen-deformations, M

3 eigen-deforms were enough!

Human Interface Lab. Kyushu Univ.

18

Comparison to class Comparison to class-

• in

independent EM (1) dependent EM (1)

98.8% 98.9% 99.0% 99.1% 99.2% 99.3% 10 20 30 40 50

# eigen-deformations, M recognition rate

affine EM

(M=6)

0.23%

（ ３

/４misrecognitions）

Human Interface Lab. Kyushu Univ.

19

Conclusion & Future work Conclusion & Future work

Conclusion

EM based on a class-dependent deformation

model was developed

High accuracy and efficiency were shown

through character recognition experiment

Future work

Solution strategies other than linear approx.

Relation to the sub-space methods